11 |
|
import java.util.Collection; |
12 |
|
import java.util.Collections; |
13 |
|
import java.util.List; |
14 |
– |
import java.util.Random; |
14 |
|
import java.util.concurrent.AbstractExecutorService; |
15 |
|
import java.util.concurrent.Callable; |
16 |
|
import java.util.concurrent.ExecutorService; |
18 |
|
import java.util.concurrent.RejectedExecutionException; |
19 |
|
import java.util.concurrent.RunnableFuture; |
20 |
|
import java.util.concurrent.TimeUnit; |
22 |
– |
import java.util.concurrent.atomic.AtomicInteger; |
23 |
– |
import java.util.concurrent.atomic.AtomicLong; |
24 |
– |
import java.util.concurrent.locks.ReentrantLock; |
25 |
– |
import java.util.concurrent.locks.Condition; |
21 |
|
|
22 |
|
/** |
23 |
|
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
37 |
|
* ForkJoinPool}s may also be appropriate for use with event-style |
38 |
|
* tasks that are never joined. |
39 |
|
* |
40 |
< |
* <p>A {@code ForkJoinPool} is constructed with a given target |
41 |
< |
* parallelism level; by default, equal to the number of available |
42 |
< |
* processors. The pool attempts to maintain enough active (or |
43 |
< |
* available) threads by dynamically adding, suspending, or resuming |
44 |
< |
* internal worker threads, even if some tasks are stalled waiting to |
45 |
< |
* join others. However, no such adjustments are guaranteed in the |
46 |
< |
* face of blocked IO or other unmanaged synchronization. The nested |
47 |
< |
* {@link ManagedBlocker} interface enables extension of the kinds of |
40 |
> |
* <p>A static {@link #commonPool()} is available and appropriate for |
41 |
> |
* 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 |
> |
* use). |
46 |
> |
* |
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 |
> |
* blocked IO or other unmanaged synchronization. The nested {@link |
55 |
> |
* ManagedBlocker} interface enables extension of the kinds of |
56 |
|
* synchronization accommodated. |
57 |
|
* |
58 |
|
* <p>In addition to execution and lifecycle control methods, this |
62 |
|
* {@link #toString} returns indications of pool state in a |
63 |
|
* convenient form for informal monitoring. |
64 |
|
* |
65 |
< |
* <p> As is the case with other ExecutorServices, there are three |
66 |
< |
* main task execution methods summarized in the following |
67 |
< |
* table. These are designed to be used primarily by clients not |
68 |
< |
* already engaged in fork/join computations in the current pool. The |
69 |
< |
* main forms of these methods accept instances of {@code |
70 |
< |
* ForkJoinTask}, but overloaded forms also allow mixed execution of |
71 |
< |
* plain {@code Runnable}- or {@code Callable}- based activities as |
72 |
< |
* well. However, tasks that are already executing in a pool should |
73 |
< |
* normally instead use the within-computation forms listed in the |
74 |
< |
* table unless using async event-style tasks that are not usually |
75 |
< |
* joined, in which case there is little difference among choice of |
73 |
< |
* methods. |
65 |
> |
* <p>As is the case with other ExecutorServices, there are three |
66 |
> |
* 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> |
97 |
|
* </tr> |
98 |
|
* </table> |
99 |
|
* |
100 |
< |
* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is |
101 |
< |
* used for all parallel task execution in a program or subsystem. |
102 |
< |
* Otherwise, use would not usually outweigh the construction and |
103 |
< |
* bookkeeping overhead of creating a large set of threads. For |
104 |
< |
* example, a common pool could be used for the {@code SortTasks} |
105 |
< |
* illustrated in {@link RecursiveAction}. Because {@code |
106 |
< |
* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon |
107 |
< |
* daemon} mode, there is typically no need to explicitly {@link |
108 |
< |
* #shutdown} such a pool upon program exit. |
109 |
< |
* |
108 |
< |
* <pre> {@code |
109 |
< |
* static final ForkJoinPool mainPool = new ForkJoinPool(); |
110 |
< |
* ... |
111 |
< |
* public void sort(long[] array) { |
112 |
< |
* mainPool.invoke(new SortTask(array, 0, array.length)); |
113 |
< |
* }}</pre> |
100 |
> |
* <p>The common pool is by default constructed with default |
101 |
> |
* parameters, but these may be controlled by setting three {@link |
102 |
> |
* System#getProperty properties} with prefix {@code |
103 |
> |
* 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 |
> |
* java.lang.Thread.UncaughtExceptionHandler |
108 |
> |
* Thread.UncaughtExceptionHandler}. Upon any error in establishing |
109 |
> |
* these settings, default parameters are used. |
110 |
|
* |
111 |
|
* <p><b>Implementation notes</b>: This implementation restricts the |
112 |
|
* maximum number of running threads to 32767. Attempts to create |
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 |
131 |
< |
* queues. Workers take these tasks and typically split them into |
132 |
< |
* subtasks that may be stolen by other workers. Preference rules |
133 |
< |
* give first priority to processing tasks from their own queues |
134 |
< |
* (LIFO or FIFO, depending on mode), then to randomized FIFO |
135 |
< |
* steals of tasks in other queues. |
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. |
137 |
> |
* WorkQueues |
138 |
|
* ========== |
139 |
|
* |
140 |
|
* Most operations occur within work-stealing queues (in nested |
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 |
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 |
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. |
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 |
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 |
191 |
< |
* it. For example, some sets of tasks profit from cache |
192 |
< |
* affinities, but others are harmed by cache pollution effects.) |
190 |
> |
* information, we do not usually have a basis for exploiting it. |
191 |
> |
* For example, some sets of tasks profit from cache affinities, |
192 |
> |
* but others are harmed by cache pollution effects.) |
193 |
|
* |
194 |
|
* WorkQueues are also used in a similar way for tasks submitted |
195 |
|
* to the pool. We cannot mix these tasks in the same queues used |
196 |
|
* for work-stealing (this would contaminate lifo/fifo |
197 |
< |
* processing). Instead, we loosely associate submission queues |
197 |
> |
* processing). Instead, we randomly associate submission queues |
198 |
|
* with submitting threads, using a form of hashing. The |
199 |
|
* ThreadLocal Submitter class contains a value initially used as |
200 |
|
* a hash code for choosing existing queues, but may be randomly |
201 |
|
* repositioned upon contention with other submitters. In |
202 |
< |
* essence, submitters act like workers except that they never |
203 |
< |
* take tasks, and they are multiplexed on to a finite number of |
204 |
< |
* shared work queues. However, classes are set up so that future |
205 |
< |
* extensions could allow submitters to optionally help perform |
206 |
< |
* tasks as well. Pool submissions from internal workers are also |
207 |
< |
* allowed, but use randomized rather than thread-hashed queue |
208 |
< |
* indices to avoid imbalance. Insertion of tasks in shared mode |
209 |
< |
* requires a lock (mainly to protect in the case of resizing) but |
210 |
< |
* we use only a simple spinlock (using bits in field runState), |
211 |
< |
* because submitters encountering a busy queue try or create |
212 |
< |
* others so never block. |
202 |
> |
* 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 |
|
* |
216 |
< |
* Management. |
216 |
> |
* Management |
217 |
|
* ========== |
218 |
|
* |
219 |
|
* The main throughput advantages of work-stealing stem from |
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 |
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 |
233 |
|
* and their negations (used for thresholding) to fit into 16bit |
234 |
|
* fields. |
235 |
|
* |
236 |
< |
* Field "runState" contains 32 bits needed to register and |
237 |
< |
* deregister WorkQueues, as well as to enable shutdown. It is |
238 |
< |
* only modified under a lock (normally briefly held, but |
239 |
< |
* occasionally protecting allocations and resizings) but even |
240 |
< |
* when locked remains available to check consistency. |
236 |
> |
* 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 |
> |
* block when needed. |
243 |
|
* |
244 |
|
* Recording WorkQueues. WorkQueues are recorded in the |
245 |
< |
* "workQueues" array that is created upon pool construction and |
246 |
< |
* expanded if necessary. Updates to the array while recording |
247 |
< |
* new workers and unrecording terminated ones are protected from |
248 |
< |
* each other by a lock but the array is otherwise concurrently |
249 |
< |
* readable, and accessed directly. To simplify index-based |
250 |
< |
* operations, the array size is always a power of two, and all |
251 |
< |
* readers must tolerate null slots. Shared (submission) queues |
252 |
< |
* are at even indices, worker queues at odd indices. Grouping |
253 |
< |
* them together in this way simplifies and speeds up task |
254 |
< |
* scanning. To avoid flailing during start-up, the array is |
255 |
< |
* presized to hold twice #parallelism workers (which is unlikely |
256 |
< |
* to need further resizing during execution). But to avoid |
257 |
< |
* dealing with so many null slots, variable runState includes a |
258 |
< |
* mask for the nearest power of two that contains all current |
256 |
< |
* workers. All worker thread creation is on-demand, triggered by |
257 |
< |
* task submissions, replacement of terminated workers, and/or |
245 |
> |
* "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 |
> |
* tolerate null slots. Worker queues are at odd indices. Shared |
252 |
> |
* (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 |
> |
* |
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 |
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. |
272 |
< |
* |
273 |
< |
* All uses of the workQueues array check that it is non-null |
274 |
< |
* (even if previously non-null). This allows nulling during |
275 |
< |
* termination, which is currently not necessary, but remains an |
276 |
< |
* option for resource-revocation-based shutdown schemes. It also |
276 |
< |
* helps reduce JIT issuance of uncommon-trap code, which tends to |
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 |
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 |
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 |
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 |
< |
* execute. When a submission is added or another worker adds a |
317 |
< |
* task to a queue that previously had fewer than two tasks, they |
318 |
< |
* signal waiting workers (or trigger creation of new ones if |
319 |
< |
* fewer than the given parallelism level -- see signalWork). |
320 |
< |
* These primary signals are buttressed by signals during rescans; |
321 |
< |
* together these cover the signals needed in cases when more |
322 |
< |
* tasks are pushed but untaken, and improve performance compared |
323 |
< |
* to having one thread wake up all workers. |
316 |
> |
* 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 |
> |
* 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 |
|
* |
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 |
< |
* time out and terminate if the pool has remained quiescent for |
335 |
< |
* SHRINK_RATE nanosecs. This will slowly propagate, eventually |
336 |
< |
* terminating all workers after long periods of non-use. |
334 |
> |
* 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 |
|
* |
340 |
|
* Shutdown and Termination. A call to shutdownNow atomically sets |
341 |
< |
* a runState bit and then (non-atomically) sets each workers |
342 |
< |
* runState status, cancels all unprocessed tasks, and wakes up |
341 |
> |
* a plock bit and then (non-atomically) sets each worker's |
342 |
> |
* qlock status, cancels all unprocessed tasks, and wakes up |
343 |
|
* 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 |
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 |
< |
* ============== |
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 |
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 |
361 |
< |
* progress. Instead we combine two tactics: |
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. |
367 |
|
* method tryCompensate() may create or re-activate a spare |
368 |
|
* thread to compensate for blocked joiners until they unblock. |
369 |
|
* |
370 |
< |
* A third form (implemented in tryRemoveAndExec and |
371 |
< |
* tryPollForAndExec) amounts to helping a hypothetical |
372 |
< |
* compensator: If we can readily tell that a possible action of a |
373 |
< |
* compensator is to steal and execute the task being joined, the |
374 |
< |
* joining thread can do so directly, without the need for a |
375 |
< |
* compensation thread (although at the expense of larger run-time |
376 |
< |
* stacks, but the tradeoff is typically worthwhile). |
370 |
> |
* 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 |
|
* |
378 |
|
* The ManagedBlocker extension API can't use helping so relies |
379 |
|
* only on compensation in method awaitBlocker. |
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 workers array to locate stealers, but |
397 |
< |
* often doesn't because stealers leave hints (that may become |
398 |
< |
* stale/wrong) of where to locate them. A stealHint is only a |
399 |
< |
* hint because a worker might have had multiple steals and the |
400 |
< |
* hint records only one of them (usually the most current). |
401 |
< |
* Hinting isolates cost to when it is needed, rather than adding |
402 |
< |
* to per-task overhead. (2) It is "shallow", ignoring nesting |
403 |
< |
* and potentially cyclic mutual steals. (3) It is intentionally |
396 |
> |
* requires a linear scan of workQueues array to locate stealers, |
397 |
> |
* but often doesn't because stealers leave hints (that may become |
398 |
> |
* 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 |
|
* 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_DEPTH) and fall back to suspending |
409 |
< |
* the worker and if necessary replacing it with another. |
408 |
> |
* to find work (see MAX_HELP) and fall back to suspending the |
409 |
> |
* worker and if necessary replacing it with another. |
410 |
> |
* |
411 |
> |
* 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 |
|
* 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. Currently, in keeping with on-demand |
423 |
< |
* signalling policy, we compensate only if blocking would leave |
424 |
< |
* less than one active (non-waiting, non-blocked) worker. |
425 |
< |
* Additionally, to avoid some false alarms due to GC, lagging |
426 |
< |
* counters, system activity, etc, compensated blocking for joins |
427 |
< |
* is only attempted after rechecks stabilize in |
428 |
< |
* ForkJoinTask.awaitJoin. (Retries are interspersed with |
429 |
< |
* Thread.yield, for good citizenship.) |
430 |
< |
* |
431 |
< |
* Style notes: There is a lot of representation-level coupling |
432 |
< |
* among classes ForkJoinPool, ForkJoinWorkerThread, and |
433 |
< |
* ForkJoinTask. The fields of WorkQueue maintain data structures |
434 |
< |
* managed by ForkJoinPool, so are directly accessed. There is |
435 |
< |
* little point trying to reduce this, since any associated future |
436 |
< |
* changes in representations will need to be accompanied by |
437 |
< |
* algorithmic changes anyway. All together, these low-level |
438 |
< |
* implementation choices produce as much as a factor of 4 |
439 |
< |
* performance improvement compared to naive implementations, and |
440 |
< |
* enable the processing of billions of tasks per second, at the |
441 |
< |
* expense of some ugliness. |
442 |
< |
* |
443 |
< |
* Methods signalWork() and scan() are the main bottlenecks so are |
444 |
< |
* especially heavily micro-optimized/mangled. There are lots of |
445 |
< |
* inline assignments (of form "while ((local = field) != 0)") |
446 |
< |
* which are usually the simplest way to ensure the required read |
447 |
< |
* orderings (which are sometimes critical). This leads to a |
448 |
< |
* "C"-like style of listing declarations of these locals at the |
449 |
< |
* heads of methods or blocks. There are several occurrences of |
450 |
< |
* the unusual "do {} while (!cas...)" which is the simplest way |
451 |
< |
* to force an update of a CAS'ed variable. There are also other |
452 |
< |
* coding oddities that help some methods perform reasonably even |
453 |
< |
* when interpreted (not compiled). |
454 |
< |
* |
455 |
< |
* The order of declarations in this file is: (1) declarations of |
456 |
< |
* statics (2) fields (along with constants used when unpacking |
457 |
< |
* some of them), listed in an order that tends to reduce |
458 |
< |
* contention among them a bit under most JVMs; (3) nested |
459 |
< |
* classes; (4) internal control methods; (5) callbacks and other |
460 |
< |
* support for ForkJoinTask methods; (6) exported methods (plus a |
461 |
< |
* few little helpers); (7) static block initializing all statics |
462 |
< |
* in a minimally dependent order. |
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 |
> |
* methods tryCompensate and awaitJoin. |
439 |
> |
* |
440 |
> |
* Common Pool |
441 |
> |
* =========== |
442 |
> |
* |
443 |
> |
* The static commonPool always exists after static |
444 |
> |
* 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 |
> |
* returns quickly (at the most helping to signal some common pool |
455 |
> |
* 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 |
> |
* 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 |
> |
* CAS'ed variable. There are also other coding oddities (including |
481 |
> |
* several unnecessary-looking hoisted null checks) that help |
482 |
> |
* 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. |
527 |
|
* Default ForkJoinWorkerThreadFactory implementation; creates a |
528 |
|
* new ForkJoinWorkerThread. |
529 |
|
*/ |
530 |
< |
static class DefaultForkJoinWorkerThreadFactory |
530 |
> |
static final class DefaultForkJoinWorkerThreadFactory |
531 |
|
implements ForkJoinWorkerThreadFactory { |
532 |
< |
public ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
532 |
> |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
533 |
|
return new ForkJoinWorkerThread(pool); |
534 |
|
} |
535 |
|
} |
536 |
|
|
537 |
|
/** |
538 |
< |
* Creates a new ForkJoinWorkerThread. This factory is used unless |
539 |
< |
* overridden in ForkJoinPool constructors. |
540 |
< |
*/ |
541 |
< |
public static final ForkJoinWorkerThreadFactory |
542 |
< |
defaultForkJoinWorkerThreadFactory; |
543 |
< |
|
544 |
< |
/** |
545 |
< |
* Permission required for callers of methods that may start or |
546 |
< |
* kill threads. |
538 |
> |
* 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 |
< |
private static final RuntimePermission modifyThreadPermission; |
553 |
< |
|
554 |
< |
/** |
491 |
< |
* If there is a security manager, makes sure caller has |
492 |
< |
* permission to modify threads. |
493 |
< |
*/ |
494 |
< |
private static void checkPermission() { |
495 |
< |
SecurityManager security = System.getSecurityManager(); |
496 |
< |
if (security != null) |
497 |
< |
security.checkPermission(modifyThreadPermission); |
552 |
> |
static final class Submitter { |
553 |
> |
int seed; |
554 |
> |
Submitter(int s) { seed = s; } |
555 |
|
} |
556 |
|
|
557 |
|
/** |
558 |
< |
* Generator for assigning sequence numbers as pool names. |
559 |
< |
*/ |
560 |
< |
private static final AtomicInteger poolNumberGenerator; |
561 |
< |
|
505 |
< |
/** |
506 |
< |
* Bits and masks for control variables |
507 |
< |
* |
508 |
< |
* Field ctl is a long packed with: |
509 |
< |
* AC: Number of active running workers minus target parallelism (16 bits) |
510 |
< |
* TC: Number of total workers minus target parallelism (16 bits) |
511 |
< |
* ST: true if pool is terminating (1 bit) |
512 |
< |
* EC: the wait count of top waiting thread (15 bits) |
513 |
< |
* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits) |
514 |
< |
* |
515 |
< |
* When convenient, we can extract the upper 32 bits of counts and |
516 |
< |
* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e = |
517 |
< |
* (int)ctl. The ec field is never accessed alone, but always |
518 |
< |
* together with id and st. The offsets of counts by the target |
519 |
< |
* parallelism and the positionings of fields makes it possible to |
520 |
< |
* perform the most common checks via sign tests of fields: When |
521 |
< |
* ac is negative, there are not enough active workers, when tc is |
522 |
< |
* negative, there are not enough total workers, when id is |
523 |
< |
* negative, there is at least one waiting worker, and when e is |
524 |
< |
* negative, the pool is terminating. To deal with these possibly |
525 |
< |
* negative fields, we use casts in and out of "short" and/or |
526 |
< |
* signed shifts to maintain signedness. |
527 |
< |
* |
528 |
< |
* When a thread is queued (inactivated), its eventCount field is |
529 |
< |
* negative, which is the only way to tell if a worker is |
530 |
< |
* prevented from executing tasks, even though it must continue to |
531 |
< |
* scan for them to avoid queuing races. |
532 |
< |
* |
533 |
< |
* Field runState is an int packed with: |
534 |
< |
* SHUTDOWN: true if shutdown is enabled (1 bit) |
535 |
< |
* SEQ: a sequence number updated upon (de)registering workers (15 bits) |
536 |
< |
* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits) |
537 |
< |
* |
538 |
< |
* The combination of mask and sequence number enables simple |
539 |
< |
* consistency checks: Staleness of read-only operations on the |
540 |
< |
* workers and queues arrays can be checked by comparing runState |
541 |
< |
* before vs after the reads. The low 16 bits (i.e, anding with |
542 |
< |
* SMASK) hold (the smallest power of two covering all worker |
543 |
< |
* indices, minus one. The mask for queues (vs workers) is twice |
544 |
< |
* this value plus 1. |
545 |
< |
*/ |
546 |
< |
|
547 |
< |
// bit positions/shifts for fields |
548 |
< |
private static final int AC_SHIFT = 48; |
549 |
< |
private static final int TC_SHIFT = 32; |
550 |
< |
private static final int ST_SHIFT = 31; |
551 |
< |
private static final int EC_SHIFT = 16; |
552 |
< |
|
553 |
< |
// bounds |
554 |
< |
private static final int MAX_ID = 0x7fff; // max poolIndex |
555 |
< |
private static final int SMASK = 0xffff; // mask short bits |
556 |
< |
private static final int SHORT_SIGN = 1 << 15; |
557 |
< |
private static final int INT_SIGN = 1 << 31; |
558 |
< |
|
559 |
< |
// masks |
560 |
< |
private static final long STOP_BIT = 0x0001L << ST_SHIFT; |
561 |
< |
private static final long AC_MASK = ((long)SMASK) << AC_SHIFT; |
562 |
< |
private static final long TC_MASK = ((long)SMASK) << TC_SHIFT; |
563 |
< |
|
564 |
< |
// units for incrementing and decrementing |
565 |
< |
private static final long TC_UNIT = 1L << TC_SHIFT; |
566 |
< |
private static final long AC_UNIT = 1L << AC_SHIFT; |
567 |
< |
|
568 |
< |
// masks and units for dealing with u = (int)(ctl >>> 32) |
569 |
< |
private static final int UAC_SHIFT = AC_SHIFT - 32; |
570 |
< |
private static final int UTC_SHIFT = TC_SHIFT - 32; |
571 |
< |
private static final int UAC_MASK = SMASK << UAC_SHIFT; |
572 |
< |
private static final int UTC_MASK = SMASK << UTC_SHIFT; |
573 |
< |
private static final int UAC_UNIT = 1 << UAC_SHIFT; |
574 |
< |
private static final int UTC_UNIT = 1 << UTC_SHIFT; |
575 |
< |
|
576 |
< |
// masks and units for dealing with e = (int)ctl |
577 |
< |
private static final int E_MASK = 0x7fffffff; // no STOP_BIT |
578 |
< |
private static final int E_SEQ = 1 << EC_SHIFT; |
579 |
< |
|
580 |
< |
// runState bits |
581 |
< |
private static final int SHUTDOWN = 1 << 31; |
582 |
< |
private static final int RS_SEQ = 1 << 16; |
583 |
< |
private static final int RS_SEQ_MASK = 0x7fff0000; |
584 |
< |
|
585 |
< |
// access mode for WorkQueue |
586 |
< |
static final int LIFO_QUEUE = 0; |
587 |
< |
static final int FIFO_QUEUE = 1; |
588 |
< |
static final int SHARED_QUEUE = -1; |
589 |
< |
|
590 |
< |
/** |
591 |
< |
* The wakeup interval (in nanoseconds) for a worker waiting for a |
592 |
< |
* task when the pool is quiescent to instead try to shrink the |
593 |
< |
* number of workers. The exact value does not matter too |
594 |
< |
* much. It must be short enough to release resources during |
595 |
< |
* sustained periods of idleness, but not so short that threads |
596 |
< |
* are continually re-created. |
597 |
< |
*/ |
598 |
< |
private static final long SHRINK_RATE = |
599 |
< |
4L * 1000L * 1000L * 1000L; // 4 seconds |
600 |
< |
|
601 |
< |
/** |
602 |
< |
* The timeout value for attempted shrinkage, includes |
603 |
< |
* some slop to cope with system timer imprecision. |
604 |
< |
*/ |
605 |
< |
private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10); |
606 |
< |
|
607 |
< |
/** |
608 |
< |
* The maximum stolen->joining link depth allowed in tryHelpStealer. |
609 |
< |
* Depths for legitimate chains are unbounded, but we use a fixed |
610 |
< |
* constant to avoid (otherwise unchecked) cycles and to bound |
611 |
< |
* staleness of traversal parameters at the expense of sometimes |
612 |
< |
* blocking when we could be helping. |
613 |
< |
*/ |
614 |
< |
private static final int MAX_HELP_DEPTH = 16; |
615 |
< |
|
616 |
< |
/* |
617 |
< |
* Field layout order in this class tends to matter more than one |
618 |
< |
* would like. Runtime layout order is only loosely related to |
619 |
< |
* declaration order and may differ across JVMs, but the following |
620 |
< |
* empirically works OK on current JVMs. |
558 |
> |
* 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 |
< |
|
564 |
< |
volatile long ctl; // main pool control |
565 |
< |
final int parallelism; // parallelism level |
566 |
< |
final int localMode; // per-worker scheduling mode |
567 |
< |
int nextPoolIndex; // hint used in registerWorker |
568 |
< |
volatile int runState; // shutdown status, seq, and mask |
569 |
< |
WorkQueue[] workQueues; // main registry |
629 |
< |
final ReentrantLock lock; // for registration |
630 |
< |
final Condition termination; // for awaitTermination |
631 |
< |
final ForkJoinWorkerThreadFactory factory; // factory for new workers |
632 |
< |
final Thread.UncaughtExceptionHandler ueh; // per-worker UEH |
633 |
< |
final AtomicLong stealCount; // collect counts when terminated |
634 |
< |
final AtomicInteger nextWorkerNumber; // to create worker name string |
635 |
< |
final String workerNamePrefix; // Prefix for assigning worker names |
563 |
> |
static final class EmptyTask extends ForkJoinTask<Void> { |
564 |
> |
private static final long serialVersionUID = -7721805057305804111L; |
565 |
> |
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 |
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 |
< |
* for push, or under lock for trySharedPush, and accessed by |
586 |
< |
* other threads only after reading (volatile) base. Both top and |
587 |
< |
* base are allowed to wrap around on overflow, but (top - base) |
588 |
< |
* (or more commonly -(base - top) to force volatile read of base |
589 |
< |
* before top) still estimates size. |
585 |
> |
* 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 |
|
* |
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 |
< |
* its publication in the queue. (Although we can avoid one case |
602 |
< |
* of this when locked in trySharedPush.) All removals entail a |
603 |
< |
* CAS to null. The array is always a power of two. To ensure |
604 |
< |
* safety of Unsafe array operations, all accesses perform |
665 |
< |
* explicit null checks and implicit bounds checks via |
666 |
< |
* power-of-two masking. |
601 |
> |
* 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 |
|
* |
606 |
|
* In addition to basic queuing support, this class contains |
607 |
|
* fields described elsewhere to control execution. It turns out |
608 |
< |
* to work better memory-layout-wise to include them in this |
609 |
< |
* class rather than a separate class. |
608 |
> |
* to work better memory-layout-wise to include them in this class |
609 |
> |
* rather than a separate class. |
610 |
|
* |
611 |
|
* Performance on most platforms is very sensitive to placement of |
612 |
|
* instances of both WorkQueues and their arrays -- we absolutely |
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 |
< |
* properties of JVM field layout rules.) We also take care in |
624 |
< |
* allocating and sizing and resizing the array. Non-shared queue |
625 |
< |
* arrays are initialized (via method growArray) by workers before |
626 |
< |
* use. Others are allocated on first use. |
623 |
> |
* 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 |
|
*/ |
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 set larger to |
632 |
< |
* reduce cacheline sharing among queues. |
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 << 8; |
638 |
> |
static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
639 |
|
|
640 |
|
/** |
641 |
|
* Maximum size for queue arrays. Must be a power of two less |
646 |
|
*/ |
647 |
|
static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M |
648 |
|
|
649 |
< |
volatile long totalSteals; // cumulative number of steals |
649 |
> |
// Heuristic padding to ameliorate unfortunate memory placements |
650 |
> |
volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; |
651 |
> |
|
652 |
|
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 |
< |
int rescans; // remaining scans until block |
712 |
< |
int nsteals; // top-level task executions since last idle |
713 |
< |
final int mode; // lifo, fifo, or shared |
655 |
> |
int hint; // steal or signal hint (index) |
656 |
|
int poolIndex; // index of this queue in pool (or 0) |
657 |
< |
int stealHint; // index of most recent known stealer |
658 |
< |
volatile int runState; // 1: locked, -1: terminate; else 0 |
657 |
> |
final int mode; // 0: lifo, > 0: fifo, < 0: shared |
658 |
> |
int nsteals; // number of steals |
659 |
> |
volatile int qlock; // 1: locked, -1: terminate; else 0 |
660 |
|
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 |
< |
ForkJoinTask<?> currentJoin; // task being joined in awaitJoin |
666 |
> |
volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin |
667 |
|
ForkJoinTask<?> currentSteal; // current non-local task being executed |
724 |
– |
// Heuristic padding to ameliorate unfortunate memory placements |
725 |
– |
Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a; |
668 |
|
|
669 |
< |
WorkQueue(ForkJoinWorkerThread owner, int mode) { |
669 |
> |
volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; |
670 |
> |
volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d; |
671 |
> |
|
672 |
> |
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode, |
673 |
> |
int seed) { |
674 |
> |
this.pool = pool; |
675 |
|
this.owner = owner; |
676 |
|
this.mode = mode; |
677 |
+ |
this.seed = seed; |
678 |
|
// Place indices in the center of array (that is not yet allocated) |
679 |
|
base = top = INITIAL_QUEUE_CAPACITY >>> 1; |
680 |
|
} |
681 |
|
|
682 |
|
/** |
683 |
< |
* Returns number of tasks in the queue |
683 |
> |
* 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; |
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. |
707 |
> |
* Pushes a task. Call only by owner in unshared queues. (The |
708 |
> |
* shared-queue version is embedded in method externalPush.) |
709 |
|
* |
710 |
|
* @param task the task. Caller must ensure non-null. |
711 |
< |
* @param p, if non-null, pool to signal if necessary |
747 |
< |
* @throw RejectedExecutionException if array cannot |
748 |
< |
* be resized |
711 |
> |
* @throw RejectedExecutionException if array cannot be resized |
712 |
|
*/ |
713 |
< |
final void push(ForkJoinTask<?> task, ForkJoinPool p) { |
714 |
< |
ForkJoinTask<?>[] a; |
713 |
> |
final void push(ForkJoinTask<?> task) { |
714 |
> |
ForkJoinTask<?>[] a; ForkJoinPool p; |
715 |
|
int s = top, m, n; |
716 |
|
if ((a = array) != null) { // ignore if queue removed |
717 |
< |
U.putOrderedObject |
718 |
< |
(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task); |
717 |
> |
int j = (((m = a.length - 1) & s) << ASHIFT) + ABASE; |
718 |
> |
U.putOrderedObject(a, j, task); |
719 |
|
if ((n = (top = s + 1) - base) <= 2) { |
720 |
< |
if (p != null) |
721 |
< |
p.signalWork(); |
720 |
> |
if ((p = pool) != null) |
721 |
> |
p.signalWork(this); |
722 |
|
} |
723 |
|
else if (n >= m) |
724 |
< |
growArray(true); |
724 |
> |
growArray(); |
725 |
|
} |
726 |
|
} |
727 |
|
|
728 |
+ |
/** |
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 |
+ |
} |
753 |
+ |
return a; |
754 |
+ |
} |
755 |
+ |
|
756 |
|
/** |
757 |
< |
* Pushes a task if lock is free and array is either big |
758 |
< |
* enough or can be resized to be big enough. |
768 |
< |
* |
769 |
< |
* @param task the task. Caller must ensure non-null. |
770 |
< |
* @return true if submitted |
757 |
> |
* Takes next task, if one exists, in LIFO order. Call only |
758 |
> |
* by owner in unshared queues. |
759 |
|
*/ |
760 |
< |
final boolean trySharedPush(ForkJoinTask<?> task) { |
761 |
< |
boolean submitted = false; |
762 |
< |
if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { |
763 |
< |
ForkJoinTask<?>[] a = array; |
764 |
< |
int s = top, n = s - base; |
765 |
< |
try { |
766 |
< |
if ((a != null && n < a.length - 1) || |
767 |
< |
(a = growArray(false)) != null) { // must presize |
768 |
< |
int j = (((a.length - 1) & s) << ASHIFT) + ABASE; |
769 |
< |
U.putObject(a, (long)j, task); // don't need "ordered" |
782 |
< |
top = s + 1; |
783 |
< |
submitted = true; |
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 |
|
} |
785 |
– |
} finally { |
786 |
– |
runState = 0; // unlock |
771 |
|
} |
772 |
|
} |
773 |
< |
return submitted; |
773 |
> |
return null; |
774 |
|
} |
775 |
|
|
776 |
|
/** |
777 |
< |
* Takes next task, if one exists, in FIFO order. |
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<?> poll() { |
782 |
< |
ForkJoinTask<?>[] a; int b, i; |
783 |
< |
while ((b = base) - top < 0 && (a = array) != null && |
784 |
< |
(i = (a.length - 1) & b) >= 0) { |
785 |
< |
int j = (i << ASHIFT) + ABASE; |
786 |
< |
ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
801 |
< |
if (t != null && base == b && |
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; |
793 |
|
} |
794 |
|
|
795 |
|
/** |
796 |
< |
* Takes next task, if one exists, in LIFO order. |
812 |
< |
* Call only by owner in unshared queues. |
796 |
> |
* Takes next task, if one exists, in FIFO order. |
797 |
|
*/ |
798 |
< |
final ForkJoinTask<?> pop() { |
799 |
< |
ForkJoinTask<?> t; int m; |
800 |
< |
ForkJoinTask<?>[] a = array; |
801 |
< |
if (a != null && (m = a.length - 1) >= 0) { |
802 |
< |
for (int s; (s = top - 1) - base >= 0;) { |
803 |
< |
int j = ((m & s) << ASHIFT) + ABASE; |
804 |
< |
if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null) |
805 |
< |
break; |
806 |
< |
if (U.compareAndSwapObject(a, j, t, null)) { |
823 |
< |
top = s; |
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 |
+ |
Thread.yield(); // wait for lagging update (very rare) |
814 |
+ |
} |
815 |
|
} |
816 |
|
return null; |
817 |
|
} |
836 |
|
} |
837 |
|
|
838 |
|
/** |
851 |
– |
* Returns task at index b if b is current base of queue. |
852 |
– |
*/ |
853 |
– |
final ForkJoinTask<?> pollAt(int b) { |
854 |
– |
ForkJoinTask<?>[] a; int i; |
855 |
– |
ForkJoinTask<?> task = null; |
856 |
– |
if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) { |
857 |
– |
int j = (i << ASHIFT) + ABASE; |
858 |
– |
ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
859 |
– |
if (t != null && base == b && |
860 |
– |
U.compareAndSwapObject(a, j, t, null)) { |
861 |
– |
base = b + 1; |
862 |
– |
task = t; |
863 |
– |
} |
864 |
– |
} |
865 |
– |
return task; |
866 |
– |
} |
867 |
– |
|
868 |
– |
/** |
839 |
|
* Pops the given task only if it is at the current top. |
840 |
+ |
* (A shared version is available only via FJP.tryExternalUnpush) |
841 |
|
*/ |
842 |
|
final boolean tryUnpush(ForkJoinTask<?> t) { |
843 |
|
ForkJoinTask<?>[] a; int s; |
851 |
|
} |
852 |
|
|
853 |
|
/** |
854 |
< |
* Polls the given task only if it is at the current base. |
854 |
> |
* Removes and cancels all known tasks, ignoring any exceptions. |
855 |
|
*/ |
856 |
< |
final boolean pollFor(ForkJoinTask<?> task) { |
857 |
< |
ForkJoinTask<?>[] a; int b, i; |
858 |
< |
if ((b = base) - top < 0 && (a = array) != null && |
859 |
< |
(i = (a.length - 1) & b) >= 0) { |
860 |
< |
int j = (i << ASHIFT) + ABASE; |
861 |
< |
if (U.getObjectVolatile(a, j) == task && base == b && |
862 |
< |
U.compareAndSwapObject(a, j, task, null)) { |
863 |
< |
base = b + 1; |
864 |
< |
return true; |
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 |
> |
// Specialized execution methods |
878 |
> |
|
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 |
|
} |
896 |
– |
return false; |
895 |
|
} |
896 |
|
|
897 |
|
/** |
898 |
< |
* If present, removes from queue and executes the given task, or |
899 |
< |
* any other cancelled task. Returns (true) immediately on any CAS |
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 |
> |
* If present, removes from queue and executes the given task, |
907 |
> |
* or any other cancelled task. Returns (true) on any CAS |
908 |
|
* or consistency check failure so caller can retry. |
909 |
|
* |
910 |
< |
* @return false if no progress can be made |
910 |
> |
* @return false if no progress can be made, else true; |
911 |
|
*/ |
912 |
|
final boolean tryRemoveAndExec(ForkJoinTask<?> task) { |
913 |
< |
boolean removed = false, empty = true, progress = true; |
913 |
> |
boolean stat = true, removed = false, empty = true; |
914 |
|
ForkJoinTask<?>[] a; int m, s, b, n; |
915 |
|
if ((a = array) != null && (m = a.length - 1) >= 0 && |
916 |
|
(n = (s = top) - (b = base)) > 0) { |
940 |
|
} |
941 |
|
if (--n == 0) { |
942 |
|
if (!empty && base == b) |
943 |
< |
progress = false; |
943 |
> |
stat = false; |
944 |
|
break; |
945 |
|
} |
946 |
|
} |
947 |
|
} |
948 |
|
if (removed) |
949 |
|
task.doExec(); |
950 |
< |
return progress; |
950 |
> |
return stat; |
951 |
|
} |
952 |
|
|
953 |
|
/** |
954 |
< |
* Initializes or doubles the capacity of array. Call either |
955 |
< |
* by owner or with lock held -- it is OK for base, but not |
950 |
< |
* top, to move while resizings are in progress. |
951 |
< |
* |
952 |
< |
* @param rejectOnFailure if true, throw exception if capacity |
953 |
< |
* exceeded (relayed ultimately to user); else return null. |
954 |
> |
* Polls for and executes the given task or any other task in |
955 |
> |
* its CountedCompleter computation |
956 |
|
*/ |
957 |
< |
final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) { |
958 |
< |
ForkJoinTask<?>[] oldA = array; |
959 |
< |
int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; |
960 |
< |
if (size <= MAXIMUM_QUEUE_CAPACITY) { |
961 |
< |
int oldMask, t, b; |
962 |
< |
ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size]; |
963 |
< |
if (oldA != null && (oldMask = oldA.length - 1) >= 0 && |
964 |
< |
(t = top) - (b = base) > 0) { |
965 |
< |
int mask = size - 1; |
966 |
< |
do { |
967 |
< |
ForkJoinTask<?> x; |
968 |
< |
int oldj = ((b & oldMask) << ASHIFT) + ABASE; |
969 |
< |
int j = ((b & mask) << ASHIFT) + ABASE; |
970 |
< |
x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj); |
971 |
< |
if (x != null && |
972 |
< |
U.compareAndSwapObject(oldA, oldj, x, null)) |
973 |
< |
U.putObjectVolatile(a, j, x); |
974 |
< |
} while (++b != t); |
957 |
> |
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 |
> |
} |
972 |
> |
else |
973 |
> |
break; // restart |
974 |
> |
} |
975 |
> |
if ((r = r.completer) == null) |
976 |
> |
break outer; // not part of root computation |
977 |
|
} |
974 |
– |
return a; |
975 |
– |
} |
976 |
– |
else if (!rejectOnFailure) |
977 |
– |
return null; |
978 |
– |
else |
979 |
– |
throw new RejectedExecutionException("Queue capacity exceeded"); |
980 |
– |
} |
981 |
– |
|
982 |
– |
/** |
983 |
– |
* Removes and cancels all known tasks, ignoring any exceptions |
984 |
– |
*/ |
985 |
– |
final void cancelAll() { |
986 |
– |
ForkJoinTask.cancelIgnoringExceptions(currentJoin); |
987 |
– |
ForkJoinTask.cancelIgnoringExceptions(currentSteal); |
988 |
– |
for (ForkJoinTask<?> t; (t = poll()) != null; ) |
989 |
– |
ForkJoinTask.cancelIgnoringExceptions(t); |
990 |
– |
} |
991 |
– |
|
992 |
– |
// Execution methods |
993 |
– |
|
994 |
– |
/** |
995 |
– |
* Removes and runs tasks until empty, using local mode |
996 |
– |
* ordering. |
997 |
– |
*/ |
998 |
– |
final void runLocalTasks() { |
999 |
– |
if (base - top < 0) { |
1000 |
– |
for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; ) |
1001 |
– |
t.doExec(); |
978 |
|
} |
979 |
+ |
return false; |
980 |
|
} |
981 |
|
|
982 |
|
/** |
983 |
|
* Executes a top-level task and any local tasks remaining |
984 |
|
* after execution. |
1008 |
– |
* |
1009 |
– |
* @return true unless terminating |
985 |
|
*/ |
986 |
< |
final boolean runTask(ForkJoinTask<?> t) { |
1012 |
< |
boolean alive = true; |
986 |
> |
final void runTask(ForkJoinTask<?> t) { |
987 |
|
if (t != null) { |
988 |
< |
currentSteal = t; |
1015 |
< |
t.doExec(); |
1016 |
< |
runLocalTasks(); |
1017 |
< |
++nsteals; |
988 |
> |
(currentSteal = t).doExec(); |
989 |
|
currentSteal = null; |
990 |
+ |
if (base - top < 0) { // process remaining local tasks |
991 |
+ |
if (mode == 0) |
992 |
+ |
popAndExecAll(); |
993 |
+ |
else |
994 |
+ |
pollAndExecAll(); |
995 |
+ |
} |
996 |
+ |
++nsteals; |
997 |
+ |
hint = -1; |
998 |
|
} |
1020 |
– |
else if (runState < 0) // terminating |
1021 |
– |
alive = false; |
1022 |
– |
return alive; |
999 |
|
} |
1000 |
|
|
1001 |
|
/** |
1002 |
< |
* Executes a non-top-level (stolen) task |
1002 |
> |
* Executes a non-top-level (stolen) task. |
1003 |
|
*/ |
1004 |
|
final void runSubtask(ForkJoinTask<?> t) { |
1005 |
|
if (t != null) { |
1006 |
|
ForkJoinTask<?> ps = currentSteal; |
1007 |
< |
currentSteal = t; |
1032 |
< |
t.doExec(); |
1007 |
> |
(currentSteal = t).doExec(); |
1008 |
|
currentSteal = ps; |
1009 |
|
} |
1010 |
|
} |
1011 |
|
|
1012 |
|
/** |
1013 |
< |
* Computes next value for random probes. Scans don't require |
1039 |
< |
* a very high quality generator, but also not a crummy one. |
1040 |
< |
* Marsaglia xor-shift is cheap and works well enough. Note: |
1041 |
< |
* This is manually inlined in several usages in ForkJoinPool |
1042 |
< |
* to avoid writes inside busy scan loops. |
1013 |
> |
* Returns true if owned and not known to be blocked. |
1014 |
|
*/ |
1015 |
< |
final int nextSeed() { |
1016 |
< |
int r = seed; |
1017 |
< |
r ^= r << 13; |
1018 |
< |
r ^= r >>> 17; |
1019 |
< |
r ^= r << 5; |
1020 |
< |
return seed = r; |
1015 |
> |
final boolean isApparentlyUnblocked() { |
1016 |
> |
Thread wt; Thread.State s; |
1017 |
> |
return (eventCount >= 0 && |
1018 |
> |
(wt = owner) != null && |
1019 |
> |
(s = wt.getState()) != Thread.State.BLOCKED && |
1020 |
> |
s != Thread.State.WAITING && |
1021 |
> |
s != Thread.State.TIMED_WAITING); |
1022 |
> |
} |
1023 |
> |
|
1024 |
> |
/** |
1025 |
> |
* If this owned and is not already interrupted, try to |
1026 |
> |
* interrupt and/or unpark, ignoring exceptions. |
1027 |
> |
*/ |
1028 |
> |
final void interruptOwner() { |
1029 |
> |
Thread wt, p; |
1030 |
> |
if ((wt = owner) != null && !wt.isInterrupted()) { |
1031 |
> |
try { |
1032 |
> |
wt.interrupt(); |
1033 |
> |
} catch (SecurityException ignore) { |
1034 |
> |
} |
1035 |
> |
} |
1036 |
> |
if ((p = parker) != null) |
1037 |
> |
U.unpark(p); |
1038 |
|
} |
1039 |
|
|
1040 |
|
// Unsafe mechanics |
1041 |
|
private static final sun.misc.Unsafe U; |
1042 |
< |
private static final long RUNSTATE; |
1042 |
> |
private static final long QLOCK; |
1043 |
|
private static final int ABASE; |
1044 |
|
private static final int ASHIFT; |
1045 |
|
static { |
1048 |
|
U = getUnsafe(); |
1049 |
|
Class<?> k = WorkQueue.class; |
1050 |
|
Class<?> ak = ForkJoinTask[].class; |
1051 |
< |
RUNSTATE = U.objectFieldOffset |
1052 |
< |
(k.getDeclaredField("runState")); |
1051 |
> |
QLOCK = U.objectFieldOffset |
1052 |
> |
(k.getDeclaredField("qlock")); |
1053 |
|
ABASE = U.arrayBaseOffset(ak); |
1054 |
|
s = U.arrayIndexScale(ak); |
1055 |
|
} catch (Exception e) { |
1061 |
|
} |
1062 |
|
} |
1063 |
|
|
1064 |
+ |
// static fields (initialized in static initializer below) |
1065 |
+ |
|
1066 |
|
/** |
1067 |
< |
* Class for artificial tasks that are used to replace the target |
1068 |
< |
* of local joins if they are removed from an interior queue slot |
1079 |
< |
* in WorkQueue.tryRemoveAndExec. We don't need the proxy to |
1080 |
< |
* actually do anything beyond having a unique identity. |
1067 |
> |
* Creates a new ForkJoinWorkerThread. This factory is used unless |
1068 |
> |
* overridden in ForkJoinPool constructors. |
1069 |
|
*/ |
1070 |
< |
static final class EmptyTask extends ForkJoinTask<Void> { |
1071 |
< |
EmptyTask() { status = ForkJoinTask.NORMAL; } // force done |
1084 |
< |
public Void getRawResult() { return null; } |
1085 |
< |
public void setRawResult(Void x) {} |
1086 |
< |
public boolean exec() { return true; } |
1087 |
< |
} |
1070 |
> |
public static final ForkJoinWorkerThreadFactory |
1071 |
> |
defaultForkJoinWorkerThreadFactory; |
1072 |
|
|
1073 |
|
/** |
1074 |
< |
<<<<<<< ForkJoinPool.java |
1075 |
< |
* Per-thread records for (typically non-FJ) threads that submit |
1076 |
< |
* to pools. Cureently holds only psuedo-random seed / index that |
1077 |
< |
* is used to chose submission queues in method doSubmit. In the |
1078 |
< |
* future, this may incorporate a means to implement different |
1095 |
< |
* task rejection and resubmission policies. |
1074 |
> |
* Per-thread submission bookkeeping. Shared across all pools |
1075 |
> |
* to reduce ThreadLocal pollution and because random motion |
1076 |
> |
* to avoid contention in one pool is likely to hold for others. |
1077 |
> |
* Lazily initialized on first submission (but null-checked |
1078 |
> |
* in other contexts to avoid unnecessary initialization). |
1079 |
|
*/ |
1080 |
< |
static final class Submitter { |
1098 |
< |
int seed; // seed for random submission queue selection |
1080 |
> |
static final ThreadLocal<Submitter> submitters; |
1081 |
|
|
1082 |
< |
// Heuristic padding to ameliorate unfortunate memory placements |
1083 |
< |
int p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, pa, pb, pc, pd, pe; |
1082 |
> |
/** |
1083 |
> |
* Permission required for callers of methods that may start or |
1084 |
> |
* kill threads. |
1085 |
> |
*/ |
1086 |
> |
private static final RuntimePermission modifyThreadPermission; |
1087 |
|
|
1088 |
< |
Submitter() { |
1089 |
< |
// Use identityHashCode, forced negative, for seed |
1090 |
< |
seed = System.identityHashCode(Thread.currentThread()) | (1 << 31); |
1091 |
< |
} |
1088 |
> |
/** |
1089 |
> |
* Common (static) pool. Non-null for public use unless a static |
1090 |
> |
* construction exception, but internal usages null-check on use |
1091 |
> |
* to paranoically avoid potential initialization circularities |
1092 |
> |
* as well as to simplify generated code. |
1093 |
> |
*/ |
1094 |
> |
static final ForkJoinPool commonPool; |
1095 |
|
|
1096 |
< |
/** |
1097 |
< |
* Computes next value for random probes. Like method |
1098 |
< |
* WorkQueue.nextSeed, this is manually inlined in several |
1099 |
< |
* usages to avoid writes inside busy loops. |
1112 |
< |
*/ |
1113 |
< |
final int nextSeed() { |
1114 |
< |
int r = seed; |
1115 |
< |
r ^= r << 13; |
1116 |
< |
r ^= r >>> 17; |
1117 |
< |
return seed = r ^= r << 5; |
1118 |
< |
} |
1119 |
< |
} |
1096 |
> |
/** |
1097 |
> |
* Common pool parallelism. Must equal commonPool.parallelism. |
1098 |
> |
*/ |
1099 |
> |
static final int commonPoolParallelism; |
1100 |
|
|
1101 |
< |
/** ThreadLocal class for Submitters */ |
1102 |
< |
static final class ThreadSubmitter extends ThreadLocal<Submitter> { |
1103 |
< |
public Submitter initialValue() { return new Submitter(); } |
1101 |
> |
/** |
1102 |
> |
* Sequence number for creating workerNamePrefix. |
1103 |
> |
*/ |
1104 |
> |
private static int poolNumberSequence; |
1105 |
> |
|
1106 |
> |
/** |
1107 |
> |
* Return the next sequence number. We don't expect this to |
1108 |
> |
* ever contend so use simple builtin sync. |
1109 |
> |
*/ |
1110 |
> |
private static final synchronized int nextPoolId() { |
1111 |
> |
return ++poolNumberSequence; |
1112 |
|
} |
1113 |
|
|
1114 |
+ |
// static constants |
1115 |
+ |
|
1116 |
|
/** |
1117 |
< |
* Per-thread submission bookeeping. Shared across all pools |
1118 |
< |
* to reduce ThreadLocal pollution and because random motion |
1119 |
< |
* to avoid contention in one pool is likely to hold for others. |
1117 |
> |
* Initial timeout value (in nanoseconds) for the thread |
1118 |
> |
* triggering quiescence to park waiting for new work. On timeout, |
1119 |
> |
* the thread will instead try to shrink the number of |
1120 |
> |
* workers. The value should be large enough to avoid overly |
1121 |
> |
* aggressive shrinkage during most transient stalls (long GCs |
1122 |
> |
* etc). |
1123 |
|
*/ |
1124 |
< |
static final ThreadSubmitter submitters = new ThreadSubmitter(); |
1124 |
> |
private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec |
1125 |
|
|
1126 |
|
/** |
1127 |
< |
* Top-level runloop for workers |
1127 |
> |
* Timeout value when there are more threads than parallelism level |
1128 |
|
*/ |
1129 |
< |
final void runWorker(ForkJoinWorkerThread wt) { |
1137 |
< |
// Initialize queue array and seed in this thread |
1138 |
< |
WorkQueue w = wt.workQueue; |
1139 |
< |
w.growArray(false); |
1140 |
< |
// Same initial hash as Submitters |
1141 |
< |
w.seed = System.identityHashCode(Thread.currentThread()) | (1 << 31); |
1129 |
> |
private static final long FAST_IDLE_TIMEOUT = 200L * 1000L * 1000L; |
1130 |
|
|
1131 |
< |
do {} while (w.runTask(scan(w))); |
1132 |
< |
} |
1131 |
> |
/** |
1132 |
> |
* The maximum stolen->joining link depth allowed in method |
1133 |
> |
* tryHelpStealer. Must be a power of two. Depths for legitimate |
1134 |
> |
* chains are unbounded, but we use a fixed constant to avoid |
1135 |
> |
* (otherwise unchecked) cycles and to bound staleness of |
1136 |
> |
* traversal parameters at the expense of sometimes blocking when |
1137 |
> |
* we could be helping. |
1138 |
> |
*/ |
1139 |
> |
private static final int MAX_HELP = 64; |
1140 |
|
|
1141 |
< |
// Creating, registering and deregistering workers |
1141 |
> |
/** |
1142 |
> |
* Increment for seed generators. See class ThreadLocal for |
1143 |
> |
* explanation. |
1144 |
> |
*/ |
1145 |
> |
private static final int SEED_INCREMENT = 0x61c88647; |
1146 |
|
|
1147 |
|
/** |
1148 |
< |
* Tries to create and start a worker |
1148 |
> |
* Bits and masks for control variables |
1149 |
> |
* |
1150 |
> |
* Field ctl is a long packed with: |
1151 |
> |
* AC: Number of active running workers minus target parallelism (16 bits) |
1152 |
> |
* TC: Number of total workers minus target parallelism (16 bits) |
1153 |
> |
* ST: true if pool is terminating (1 bit) |
1154 |
> |
* EC: the wait count of top waiting thread (15 bits) |
1155 |
> |
* ID: poolIndex of top of Treiber stack of waiters (16 bits) |
1156 |
> |
* |
1157 |
> |
* When convenient, we can extract the upper 32 bits of counts and |
1158 |
> |
* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e = |
1159 |
> |
* (int)ctl. The ec field is never accessed alone, but always |
1160 |
> |
* together with id and st. The offsets of counts by the target |
1161 |
> |
* parallelism and the positionings of fields makes it possible to |
1162 |
> |
* perform the most common checks via sign tests of fields: When |
1163 |
> |
* ac is negative, there are not enough active workers, when tc is |
1164 |
> |
* negative, there are not enough total workers, and when e is |
1165 |
> |
* negative, the pool is terminating. To deal with these possibly |
1166 |
> |
* negative fields, we use casts in and out of "short" and/or |
1167 |
> |
* signed shifts to maintain signedness. |
1168 |
> |
* |
1169 |
> |
* When a thread is queued (inactivated), its eventCount field is |
1170 |
> |
* set negative, which is the only way to tell if a worker is |
1171 |
> |
* prevented from executing tasks, even though it must continue to |
1172 |
> |
* scan for them to avoid queuing races. Note however that |
1173 |
> |
* eventCount updates lag releases so usage requires care. |
1174 |
> |
* |
1175 |
> |
* Field plock is an int packed with: |
1176 |
> |
* SHUTDOWN: true if shutdown is enabled (1 bit) |
1177 |
> |
* SEQ: a sequence lock, with PL_LOCK bit set if locked (30 bits) |
1178 |
> |
* SIGNAL: set when threads may be waiting on the lock (1 bit) |
1179 |
> |
* |
1180 |
> |
* The sequence number enables simple consistency checks: |
1181 |
> |
* Staleness of read-only operations on the workQueues array can |
1182 |
> |
* be checked by comparing plock before vs after the reads. |
1183 |
|
*/ |
1184 |
< |
private void addWorker() { |
1185 |
< |
Throwable ex = null; |
1186 |
< |
ForkJoinWorkerThread w = null; |
1187 |
< |
try { |
1188 |
< |
if ((w = factory.newThread(this)) != null) { |
1189 |
< |
w.start(); |
1190 |
< |
return; |
1184 |
> |
|
1185 |
> |
// bit positions/shifts for fields |
1186 |
> |
private static final int AC_SHIFT = 48; |
1187 |
> |
private static final int TC_SHIFT = 32; |
1188 |
> |
private static final int ST_SHIFT = 31; |
1189 |
> |
private static final int EC_SHIFT = 16; |
1190 |
> |
|
1191 |
> |
// bounds |
1192 |
> |
private static final int SMASK = 0xffff; // short bits |
1193 |
> |
private static final int MAX_CAP = 0x7fff; // max #workers - 1 |
1194 |
> |
private static final int EVENMASK = 0xfffe; // even short bits |
1195 |
> |
private static final int SQMASK = 0x007e; // max 64 (even) slots |
1196 |
> |
private static final int SHORT_SIGN = 1 << 15; |
1197 |
> |
private static final int INT_SIGN = 1 << 31; |
1198 |
> |
|
1199 |
> |
// masks |
1200 |
> |
private static final long STOP_BIT = 0x0001L << ST_SHIFT; |
1201 |
> |
private static final long AC_MASK = ((long)SMASK) << AC_SHIFT; |
1202 |
> |
private static final long TC_MASK = ((long)SMASK) << TC_SHIFT; |
1203 |
> |
|
1204 |
> |
// units for incrementing and decrementing |
1205 |
> |
private static final long TC_UNIT = 1L << TC_SHIFT; |
1206 |
> |
private static final long AC_UNIT = 1L << AC_SHIFT; |
1207 |
> |
|
1208 |
> |
// masks and units for dealing with u = (int)(ctl >>> 32) |
1209 |
> |
private static final int UAC_SHIFT = AC_SHIFT - 32; |
1210 |
> |
private static final int UTC_SHIFT = TC_SHIFT - 32; |
1211 |
> |
private static final int UAC_MASK = SMASK << UAC_SHIFT; |
1212 |
> |
private static final int UTC_MASK = SMASK << UTC_SHIFT; |
1213 |
> |
private static final int UAC_UNIT = 1 << UAC_SHIFT; |
1214 |
> |
private static final int UTC_UNIT = 1 << UTC_SHIFT; |
1215 |
> |
|
1216 |
> |
// masks and units for dealing with e = (int)ctl |
1217 |
> |
private static final int E_MASK = 0x7fffffff; // no STOP_BIT |
1218 |
> |
private static final int E_SEQ = 1 << EC_SHIFT; |
1219 |
> |
|
1220 |
> |
// plock bits |
1221 |
> |
private static final int SHUTDOWN = 1 << 31; |
1222 |
> |
private static final int PL_LOCK = 2; |
1223 |
> |
private static final int PL_SIGNAL = 1; |
1224 |
> |
private static final int PL_SPINS = 1 << 8; |
1225 |
> |
|
1226 |
> |
// access mode for WorkQueue |
1227 |
> |
static final int LIFO_QUEUE = 0; |
1228 |
> |
static final int FIFO_QUEUE = 1; |
1229 |
> |
static final int SHARED_QUEUE = -1; |
1230 |
> |
|
1231 |
> |
// bounds for #steps in scan loop -- must be power 2 minus 1 |
1232 |
> |
private static final int MIN_SCAN = 0x1ff; // cover estimation slop |
1233 |
> |
private static final int MAX_SCAN = 0x1ffff; // 4 * max workers |
1234 |
> |
|
1235 |
> |
// Instance fields |
1236 |
> |
|
1237 |
> |
/* |
1238 |
> |
* Field layout of this class tends to matter more than one would |
1239 |
> |
* like. Runtime layout order is only loosely related to |
1240 |
> |
* declaration order and may differ across JVMs, but the following |
1241 |
> |
* empirically works OK on current JVMs. |
1242 |
> |
*/ |
1243 |
> |
|
1244 |
> |
// Heuristic padding to ameliorate unfortunate memory placements |
1245 |
> |
volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; |
1246 |
> |
|
1247 |
> |
volatile long stealCount; // collects worker counts |
1248 |
> |
volatile long ctl; // main pool control |
1249 |
> |
volatile int plock; // shutdown status and seqLock |
1250 |
> |
volatile int indexSeed; // worker/submitter index seed |
1251 |
> |
final int config; // mode and parallelism level |
1252 |
> |
WorkQueue[] workQueues; // main registry |
1253 |
> |
final ForkJoinWorkerThreadFactory factory; |
1254 |
> |
final Thread.UncaughtExceptionHandler ueh; // per-worker UEH |
1255 |
> |
final String workerNamePrefix; // to create worker name string |
1256 |
> |
|
1257 |
> |
volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; |
1258 |
> |
volatile Object pad18, pad19, pad1a, pad1b; |
1259 |
> |
|
1260 |
> |
/* |
1261 |
> |
* Acquires the plock lock to protect worker array and related |
1262 |
> |
* updates. This method is called only if an initial CAS on plock |
1263 |
> |
* fails. This acts as a spinLock for normal cases, but falls back |
1264 |
> |
* to builtin monitor to block when (rarely) needed. This would be |
1265 |
> |
* a terrible idea for a highly contended lock, but works fine as |
1266 |
> |
* a more conservative alternative to a pure spinlock. See |
1267 |
> |
* internal ConcurrentHashMap documentation for further |
1268 |
> |
* explanation of nearly the same construction. |
1269 |
> |
*/ |
1270 |
> |
private int acquirePlock() { |
1271 |
> |
int spins = PL_SPINS, r = 0, ps, nps; |
1272 |
> |
for (;;) { |
1273 |
> |
if (((ps = plock) & PL_LOCK) == 0 && |
1274 |
> |
U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK)) |
1275 |
> |
return nps; |
1276 |
> |
else if (r == 0) { // randomize spins if possible |
1277 |
> |
Thread t = Thread.currentThread(); WorkQueue w; Submitter z; |
1278 |
> |
if ((t instanceof ForkJoinWorkerThread) && |
1279 |
> |
(w = ((ForkJoinWorkerThread)t).workQueue) != null) |
1280 |
> |
r = w.seed; |
1281 |
> |
else if ((z = submitters.get()) != null) |
1282 |
> |
r = z.seed; |
1283 |
> |
else |
1284 |
> |
r = 1; |
1285 |
> |
} |
1286 |
> |
else if (spins >= 0) { |
1287 |
> |
r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift |
1288 |
> |
if (r >= 0) |
1289 |
> |
--spins; |
1290 |
> |
} |
1291 |
> |
else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) { |
1292 |
> |
synchronized (this) { |
1293 |
> |
if ((plock & PL_SIGNAL) != 0) { |
1294 |
> |
try { |
1295 |
> |
wait(); |
1296 |
> |
} catch (InterruptedException ie) { |
1297 |
> |
try { |
1298 |
> |
Thread.currentThread().interrupt(); |
1299 |
> |
} catch (SecurityException ignore) { |
1300 |
> |
} |
1301 |
> |
} |
1302 |
> |
} |
1303 |
> |
else |
1304 |
> |
notifyAll(); |
1305 |
> |
} |
1306 |
|
} |
1159 |
– |
} catch (Throwable e) { |
1160 |
– |
ex = e; |
1307 |
|
} |
1162 |
– |
deregisterWorker(w, ex); |
1308 |
|
} |
1309 |
|
|
1310 |
|
/** |
1311 |
< |
* Callback from ForkJoinWorkerThread constructor to assign a |
1312 |
< |
* public name. This must be separate from registerWorker because |
1168 |
< |
* it is called during the "super" constructor call in |
1169 |
< |
* ForkJoinWorkerThread. |
1311 |
> |
* Unlocks and signals any thread waiting for plock. Called only |
1312 |
> |
* when CAS of seq value for unlock fails. |
1313 |
|
*/ |
1314 |
< |
final String nextWorkerName() { |
1315 |
< |
return workerNamePrefix.concat |
1316 |
< |
(Integer.toString(nextWorkerNumber.addAndGet(1))); |
1314 |
> |
private void releasePlock(int ps) { |
1315 |
> |
plock = ps; |
1316 |
> |
synchronized (this) { notifyAll(); } |
1317 |
> |
} |
1318 |
> |
|
1319 |
> |
/** |
1320 |
> |
* Performs secondary initialization, called when plock is zero. |
1321 |
> |
* Creates workQueue array and sets plock to a valid value. The |
1322 |
> |
* lock body must be exception-free (so no try/finally) so we |
1323 |
> |
* optimistically allocate new array outside the lock and throw |
1324 |
> |
* away if (very rarely) not needed. (A similar tactic is used in |
1325 |
> |
* fullExternalPush.) Because the plock seq value can eventually |
1326 |
> |
* wrap around zero, this method harmlessly fails to reinitialize |
1327 |
> |
* if workQueues exists, while still advancing plock. |
1328 |
> |
* |
1329 |
> |
* Additonally tries to create the first worker. |
1330 |
> |
*/ |
1331 |
> |
private void initWorkers() { |
1332 |
> |
WorkQueue[] ws, nws; int ps; |
1333 |
> |
int p = config & SMASK; // find power of two table size |
1334 |
> |
int n = (p > 1) ? p - 1 : 1; // ensure at least 2 slots |
1335 |
> |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16; |
1336 |
> |
n = (n + 1) << 1; |
1337 |
> |
if ((ws = workQueues) == null || ws.length == 0) |
1338 |
> |
nws = new WorkQueue[n]; |
1339 |
> |
else |
1340 |
> |
nws = null; |
1341 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
1342 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1343 |
> |
ps = acquirePlock(); |
1344 |
> |
if (((ws = workQueues) == null || ws.length == 0) && nws != null) |
1345 |
> |
workQueues = nws; |
1346 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1347 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1348 |
> |
releasePlock(nps); |
1349 |
> |
tryAddWorker(); |
1350 |
> |
} |
1351 |
> |
|
1352 |
> |
/** |
1353 |
> |
* Tries to create and start one worker. Adjusts counts etc on |
1354 |
> |
* failure. |
1355 |
> |
*/ |
1356 |
> |
private void tryAddWorker() { |
1357 |
> |
long c; int u; |
1358 |
> |
while ((u = (int)((c = ctl) >>> 32)) < 0 && |
1359 |
> |
(u & SHORT_SIGN) != 0 && (int)c == 0) { |
1360 |
> |
long nc = (long)(((u + UTC_UNIT) & UTC_MASK) | |
1361 |
> |
((u + UAC_UNIT) & UAC_MASK)) << 32; |
1362 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1363 |
> |
ForkJoinWorkerThreadFactory fac; |
1364 |
> |
Throwable ex = null; |
1365 |
> |
ForkJoinWorkerThread wt = null; |
1366 |
> |
try { |
1367 |
> |
if ((fac = factory) != null && |
1368 |
> |
(wt = fac.newThread(this)) != null) { |
1369 |
> |
wt.start(); |
1370 |
> |
break; |
1371 |
> |
} |
1372 |
> |
} catch (Throwable e) { |
1373 |
> |
ex = e; |
1374 |
> |
} |
1375 |
> |
deregisterWorker(wt, ex); |
1376 |
> |
break; |
1377 |
> |
} |
1378 |
> |
} |
1379 |
|
} |
1380 |
|
|
1381 |
+ |
// Registering and deregistering workers |
1382 |
+ |
|
1383 |
|
/** |
1384 |
< |
* Callback from ForkJoinWorkerThread constructor to establish and |
1385 |
< |
* record its WorkQueue |
1384 |
> |
* Callback from ForkJoinWorkerThread to establish and record its |
1385 |
> |
* WorkQueue. To avoid scanning bias due to packing entries in |
1386 |
> |
* front of the workQueues array, we treat the array as a simple |
1387 |
> |
* power-of-two hash table using per-thread seed as hash, |
1388 |
> |
* expanding as needed. |
1389 |
|
* |
1390 |
|
* @param wt the worker thread |
1391 |
+ |
* @return the worker's queue |
1392 |
|
*/ |
1393 |
< |
final void registerWorker(ForkJoinWorkerThread wt) { |
1394 |
< |
WorkQueue w = wt.workQueue; |
1395 |
< |
ReentrantLock lock = this.lock; |
1396 |
< |
lock.lock(); |
1393 |
> |
final WorkQueue registerWorker(ForkJoinWorkerThread wt) { |
1394 |
> |
Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps; |
1395 |
> |
wt.setDaemon(true); |
1396 |
> |
if ((handler = ueh) != null) |
1397 |
> |
wt.setUncaughtExceptionHandler(handler); |
1398 |
> |
do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed, |
1399 |
> |
s += SEED_INCREMENT) || |
1400 |
> |
s == 0); // skip 0 |
1401 |
> |
WorkQueue w = new WorkQueue(this, wt, config >>> 16, s); |
1402 |
> |
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 |
|
try { |
1407 |
< |
int k = nextPoolIndex; |
1408 |
< |
WorkQueue[] ws = workQueues; |
1409 |
< |
if (ws != null) { // ignore on shutdown |
1410 |
< |
int n = ws.length; |
1411 |
< |
if (k < 0 || (k & 1) == 0 || k >= n || ws[k] != null) { |
1412 |
< |
for (k = 1; k < n && ws[k] != null; k += 2) |
1413 |
< |
; // workers are at odd indices |
1414 |
< |
if (k >= n) // resize |
1415 |
< |
workQueues = ws = Arrays.copyOf(ws, n << 1); |
1416 |
< |
} |
1417 |
< |
w.poolIndex = k; |
1418 |
< |
w.eventCount = ~(k >>> 1) & SMASK; // Set up wait count |
1419 |
< |
ws[k] = w; // record worker |
1420 |
< |
nextPoolIndex = k + 2; |
1421 |
< |
int rs = runState; |
1422 |
< |
int m = rs & SMASK; // recalculate runState mask |
1203 |
< |
if (k > m) |
1204 |
< |
m = (m << 1) + 1; |
1205 |
< |
runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m; |
1407 |
> |
if ((ws = workQueues) != null) { // skip if shutting down |
1408 |
> |
int n = ws.length, m = n - 1; |
1409 |
> |
int r = (s << 1) | 1; // use odd-numbered indices |
1410 |
> |
if (ws[r &= m] != null) { // collision |
1411 |
> |
int probes = 0; // step by approx half size |
1412 |
> |
int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2; |
1413 |
> |
while (ws[r = (r + step) & m] != null) { |
1414 |
> |
if (++probes >= n) { |
1415 |
> |
workQueues = ws = Arrays.copyOf(ws, n <<= 1); |
1416 |
> |
m = n - 1; |
1417 |
> |
probes = 0; |
1418 |
> |
} |
1419 |
> |
} |
1420 |
> |
} |
1421 |
> |
w.eventCount = w.poolIndex = r; // volatile write orders |
1422 |
> |
ws[r] = w; |
1423 |
|
} |
1424 |
|
} finally { |
1425 |
< |
lock.unlock(); |
1425 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1426 |
> |
releasePlock(nps); |
1427 |
|
} |
1428 |
+ |
wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex))); |
1429 |
+ |
return w; |
1430 |
|
} |
1431 |
|
|
1432 |
|
/** |
1433 |
< |
* Final callback from terminating worker, as well as failure to |
1434 |
< |
* construct or start a worker in addWorker. Removes record of |
1435 |
< |
* worker from array, and adjusts counts. If pool is shutting |
1436 |
< |
* down, tries to complete termination. |
1433 |
> |
* Final callback from terminating worker, as well as upon failure |
1434 |
> |
* to construct or start a worker. Removes record of worker from |
1435 |
> |
* array, and adjusts counts. If pool is shutting down, tries to |
1436 |
> |
* complete termination. |
1437 |
|
* |
1438 |
< |
* @param wt the worker thread or null if addWorker failed |
1438 |
> |
* @param wt the worker thread or null if construction failed |
1439 |
|
* @param ex the exception causing failure, or null if none |
1440 |
|
*/ |
1441 |
|
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1442 |
|
WorkQueue w = null; |
1443 |
|
if (wt != null && (w = wt.workQueue) != null) { |
1444 |
< |
w.runState = -1; // ensure runState is set |
1445 |
< |
stealCount.getAndAdd(w.totalSteals + w.nsteals); |
1446 |
< |
int idx = w.poolIndex; |
1447 |
< |
ReentrantLock lock = this.lock; |
1448 |
< |
lock.lock(); |
1449 |
< |
try { // remove record from array |
1444 |
> |
int ps; |
1445 |
> |
w.qlock = -1; // ensure set |
1446 |
> |
long ns = w.nsteals, sc; // collect steal count |
1447 |
> |
do {} while (!U.compareAndSwapLong(this, STEALCOUNT, |
1448 |
> |
sc = stealCount, sc + ns)); |
1449 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
1450 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1451 |
> |
ps = acquirePlock(); |
1452 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1453 |
> |
try { |
1454 |
> |
int idx = w.poolIndex; |
1455 |
|
WorkQueue[] ws = workQueues; |
1456 |
|
if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) |
1457 |
< |
ws[nextPoolIndex = idx] = null; |
1457 |
> |
ws[idx] = null; |
1458 |
|
} finally { |
1459 |
< |
lock.unlock(); |
1459 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1460 |
> |
releasePlock(nps); |
1461 |
|
} |
1462 |
|
} |
1463 |
|
|
1467 |
|
((c - TC_UNIT) & TC_MASK) | |
1468 |
|
(c & ~(AC_MASK|TC_MASK))))); |
1469 |
|
|
1470 |
< |
if (!tryTerminate(false) && w != null) { |
1470 |
> |
if (!tryTerminate(false, false) && w != null) { |
1471 |
|
w.cancelAll(); // cancel remaining tasks |
1472 |
|
if (w.array != null) // suppress signal if never ran |
1473 |
< |
signalWork(); // wake up or create replacement |
1473 |
> |
tryAddWorker(); // create replacement |
1474 |
> |
if (ex == null) // help clean refs on way out |
1475 |
> |
ForkJoinTask.helpExpungeStaleExceptions(); |
1476 |
|
} |
1477 |
|
|
1478 |
|
if (ex != null) // rethrow |
1479 |
< |
U.throwException(ex); |
1479 |
> |
ForkJoinTask.rethrow(ex); |
1480 |
|
} |
1481 |
|
|
1482 |
+ |
// Submissions |
1483 |
+ |
|
1484 |
|
/** |
1485 |
< |
* Tries to add and register a new queue at the given index. |
1486 |
< |
* |
1487 |
< |
* @param idx the workQueues array index to register the queue |
1488 |
< |
* @return the queue, or null if could not add because could |
1489 |
< |
* not acquire lock or idx is unusable |
1490 |
< |
*/ |
1491 |
< |
private WorkQueue tryAddSharedQueue(int idx) { |
1492 |
< |
WorkQueue q = null; |
1493 |
< |
ReentrantLock lock = this.lock; |
1494 |
< |
if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) { |
1495 |
< |
// create queue outside of lock but only if apparently free |
1496 |
< |
WorkQueue nq = new WorkQueue(null, SHARED_QUEUE); |
1497 |
< |
if (lock.tryLock()) { |
1498 |
< |
try { |
1499 |
< |
WorkQueue[] ws = workQueues; |
1500 |
< |
if (ws != null && idx < ws.length) { |
1501 |
< |
if ((q = ws[idx]) == null) { |
1502 |
< |
int rs; // update runState seq |
1503 |
< |
ws[idx] = q = nq; |
1504 |
< |
runState = (((rs = runState) & SHUTDOWN) | |
1505 |
< |
((rs + RS_SEQ) & ~SHUTDOWN)); |
1485 |
> |
* Unless shutting down, adds the given task to a submission queue |
1486 |
> |
* at submitter's current queue index (modulo submission |
1487 |
> |
* range). Only the most common path is directly handled in this |
1488 |
> |
* method. All others are relayed to fullExternalPush. |
1489 |
> |
* |
1490 |
> |
* @param task the task. Caller must ensure non-null. |
1491 |
> |
*/ |
1492 |
> |
final void externalPush(ForkJoinTask<?> task) { |
1493 |
> |
WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a; |
1494 |
> |
if ((z = submitters.get()) != null && plock > 0 && |
1495 |
> |
(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 && |
1496 |
> |
(q = ws[m & z.seed & SQMASK]) != null && |
1497 |
> |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock |
1498 |
> |
int b = q.base, s = q.top, n, an; |
1499 |
> |
if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) { |
1500 |
> |
int j = (((an - 1) & s) << ASHIFT) + ABASE; |
1501 |
> |
U.putOrderedObject(a, j, task); |
1502 |
> |
q.top = s + 1; // push on to deque |
1503 |
> |
q.qlock = 0; |
1504 |
> |
if (n <= 2) |
1505 |
> |
signalWork(q); |
1506 |
> |
return; |
1507 |
> |
} |
1508 |
> |
q.qlock = 0; |
1509 |
> |
} |
1510 |
> |
fullExternalPush(task); |
1511 |
> |
} |
1512 |
> |
|
1513 |
> |
/** |
1514 |
> |
* Full version of externalPush. This method is called, among |
1515 |
> |
* other times, upon the first submission of the first task to the |
1516 |
> |
* pool, so must perform secondary initialization (via |
1517 |
> |
* initWorkers). It also detects first submission by an external |
1518 |
> |
* thread by looking up its ThreadLocal, and creates a new shared |
1519 |
> |
* queue if the one at index if empty or contended. The plock lock |
1520 |
> |
* body must be exception-free (so no try/finally) so we |
1521 |
> |
* optimistically allocate new queues outside the lock and throw |
1522 |
> |
* them away if (very rarely) not needed. |
1523 |
> |
*/ |
1524 |
> |
private void fullExternalPush(ForkJoinTask<?> task) { |
1525 |
> |
int r = 0; // random index seed |
1526 |
> |
for (Submitter z = submitters.get();;) { |
1527 |
> |
WorkQueue[] ws; WorkQueue q; int ps, m, k; |
1528 |
> |
if (z == null) { |
1529 |
> |
if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed, |
1530 |
> |
r += SEED_INCREMENT) && r != 0) |
1531 |
> |
submitters.set(z = new Submitter(r)); |
1532 |
> |
} |
1533 |
> |
else if (r == 0) { // move to a different index |
1534 |
> |
r = z.seed; |
1535 |
> |
r ^= r << 13; // same xorshift as WorkQueues |
1536 |
> |
r ^= r >>> 17; |
1537 |
> |
z.seed = r ^ (r << 5); |
1538 |
> |
} |
1539 |
> |
else if ((ps = plock) < 0) |
1540 |
> |
throw new RejectedExecutionException(); |
1541 |
> |
else if (ps == 0 || (ws = workQueues) == null || |
1542 |
> |
(m = ws.length - 1) < 0) |
1543 |
> |
initWorkers(); |
1544 |
> |
else if ((q = ws[k = r & m & SQMASK]) != null) { |
1545 |
> |
if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
1546 |
> |
ForkJoinTask<?>[] a = q.array; |
1547 |
> |
int s = q.top; |
1548 |
> |
boolean submitted = false; |
1549 |
> |
try { // locked version of push |
1550 |
> |
if ((a != null && a.length > s + 1 - q.base) || |
1551 |
> |
(a = q.growArray()) != null) { // must presize |
1552 |
> |
int j = (((a.length - 1) & s) << ASHIFT) + ABASE; |
1553 |
> |
U.putOrderedObject(a, j, task); |
1554 |
> |
q.top = s + 1; |
1555 |
> |
submitted = true; |
1556 |
|
} |
1557 |
+ |
} finally { |
1558 |
+ |
q.qlock = 0; // unlock |
1559 |
+ |
} |
1560 |
+ |
if (submitted) { |
1561 |
+ |
signalWork(q); |
1562 |
+ |
return; |
1563 |
|
} |
1278 |
– |
} finally { |
1279 |
– |
lock.unlock(); |
1564 |
|
} |
1565 |
+ |
r = 0; // move on failure |
1566 |
+ |
} |
1567 |
+ |
else if (((ps = plock) & PL_LOCK) == 0) { // create new queue |
1568 |
+ |
q = new WorkQueue(this, null, SHARED_QUEUE, r); |
1569 |
+ |
if (((ps = plock) & PL_LOCK) != 0 || |
1570 |
+ |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1571 |
+ |
ps = acquirePlock(); |
1572 |
+ |
if ((ws = workQueues) != null && k < ws.length && ws[k] == null) |
1573 |
+ |
ws[k] = q; |
1574 |
+ |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1575 |
+ |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1576 |
+ |
releasePlock(nps); |
1577 |
|
} |
1578 |
+ |
else |
1579 |
+ |
r = 0; // try elsewhere while lock held |
1580 |
|
} |
1283 |
– |
return q; |
1581 |
|
} |
1582 |
|
|
1583 |
|
// Maintaining ctl counts |
1584 |
|
|
1585 |
|
/** |
1586 |
< |
* Increments active count; mainly called upon return from blocking |
1586 |
> |
* Increments active count; mainly called upon return from blocking. |
1587 |
|
*/ |
1588 |
|
final void incrementActiveCount() { |
1589 |
|
long c; |
1591 |
|
} |
1592 |
|
|
1593 |
|
/** |
1594 |
< |
* Activates or creates a worker |
1594 |
> |
* Tries to create or activate a worker if too few are active. |
1595 |
> |
* |
1596 |
> |
* @param q the (non-null) queue holding tasks to be signalled |
1597 |
|
*/ |
1598 |
< |
final void signalWork() { |
1599 |
< |
/* |
1600 |
< |
* The while condition is true if: (there is are too few total |
1601 |
< |
* workers OR there is at least one waiter) AND (there are too |
1602 |
< |
* few active workers OR the pool is terminating). The value |
1603 |
< |
* of e distinguishes the remaining cases: zero (no waiters) |
1604 |
< |
* for create, negative if terminating (in which case do |
1605 |
< |
* nothing), else release a waiter. The secondary checks for |
1606 |
< |
* release (non-null array etc) can fail if the pool begins |
1607 |
< |
* terminating after the test, and don't impose any added cost |
1608 |
< |
* because JVMs must perform null and bounds checks anyway. |
1609 |
< |
*/ |
1610 |
< |
long c; int e, u; |
1611 |
< |
while ((((e = (int)(c = ctl)) | (u = (int)(c >>> 32))) & |
1612 |
< |
(INT_SIGN|SHORT_SIGN)) == (INT_SIGN|SHORT_SIGN)) { |
1613 |
< |
WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p; |
1614 |
< |
if (e == 0) { // add a new worker |
1615 |
< |
if (U.compareAndSwapLong |
1317 |
< |
(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) | |
1318 |
< |
((u + UAC_UNIT) & UAC_MASK)) << 32)) { |
1319 |
< |
addWorker(); |
1320 |
< |
break; |
1598 |
> |
final void signalWork(WorkQueue q) { |
1599 |
> |
int hint = q.poolIndex; |
1600 |
> |
long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p; |
1601 |
> |
while ((u = (int)((c = ctl) >>> 32)) < 0) { |
1602 |
> |
if ((e = (int)c) > 0) { |
1603 |
> |
if ((ws = workQueues) != null && ws.length > (i = e & SMASK) && |
1604 |
> |
(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { |
1605 |
> |
long nc = (((long)(w.nextWait & E_MASK)) | |
1606 |
> |
((long)(u + UAC_UNIT) << 32)); |
1607 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1608 |
> |
w.hint = hint; |
1609 |
> |
w.eventCount = (e + E_SEQ) & E_MASK; |
1610 |
> |
if ((p = w.parker) != null) |
1611 |
> |
U.unpark(p); |
1612 |
> |
break; |
1613 |
> |
} |
1614 |
> |
if (q.top - q.base <= 0) |
1615 |
> |
break; |
1616 |
|
} |
1617 |
< |
} |
1323 |
< |
else if (e > 0 && ws != null && |
1324 |
< |
(i = ((~e << 1) | 1) & SMASK) < ws.length && |
1325 |
< |
(w = ws[i]) != null && |
1326 |
< |
w.eventCount == (e | INT_SIGN)) { |
1327 |
< |
if (U.compareAndSwapLong |
1328 |
< |
(this, CTL, c, (((long)(w.nextWait & E_MASK)) | |
1329 |
< |
((long)(u + UAC_UNIT) << 32)))) { |
1330 |
< |
w.eventCount = (e + E_SEQ) & E_MASK; |
1331 |
< |
if ((p = w.parker) != null) |
1332 |
< |
U.unpark(p); // release a waiting worker |
1617 |
> |
else |
1618 |
|
break; |
1334 |
– |
} |
1619 |
|
} |
1620 |
< |
else |
1620 |
> |
else { |
1621 |
> |
if ((short)u < 0) |
1622 |
> |
tryAddWorker(); |
1623 |
|
break; |
1338 |
– |
} |
1339 |
– |
} |
1340 |
– |
|
1341 |
– |
/** |
1342 |
– |
* Tries to decrement active count (sometimes implicitly) and |
1343 |
– |
* possibly release or create a compensating worker in preparation |
1344 |
– |
* for blocking. Fails on contention or termination. |
1345 |
– |
* |
1346 |
– |
* @return true if the caller can block, else should recheck and retry |
1347 |
– |
*/ |
1348 |
– |
final boolean tryCompensate() { |
1349 |
– |
WorkQueue[] ws; WorkQueue w; Thread p; |
1350 |
– |
int pc = parallelism, e, u, ac, tc, i; |
1351 |
– |
long c = ctl; |
1352 |
– |
|
1353 |
– |
if ((e = (int)c) >= 0) { |
1354 |
– |
if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 && |
1355 |
– |
e != 0 && (ws = workQueues) != null && |
1356 |
– |
(i = ((~e << 1) | 1) & SMASK) < ws.length && |
1357 |
– |
(w = ws[i]) != null) { |
1358 |
– |
if (w.eventCount == (e | INT_SIGN) && |
1359 |
– |
U.compareAndSwapLong |
1360 |
– |
(this, CTL, c, ((long)(w.nextWait & E_MASK) | |
1361 |
– |
(c & (AC_MASK|TC_MASK))))) { |
1362 |
– |
w.eventCount = (e + E_SEQ) & E_MASK; |
1363 |
– |
if ((p = w.parker) != null) |
1364 |
– |
U.unpark(p); |
1365 |
– |
return true; // release an idle worker |
1366 |
– |
} |
1367 |
– |
} |
1368 |
– |
else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) { |
1369 |
– |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1370 |
– |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1371 |
– |
return true; // no compensation needed |
1372 |
– |
} |
1373 |
– |
else if (tc + pc < MAX_ID) { |
1374 |
– |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1375 |
– |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1376 |
– |
addWorker(); |
1377 |
– |
return true; // create replacement |
1378 |
– |
} |
1624 |
|
} |
1625 |
|
} |
1381 |
– |
return false; |
1626 |
|
} |
1627 |
|
|
1628 |
< |
// Submissions |
1628 |
> |
// Scanning for tasks |
1629 |
|
|
1630 |
|
/** |
1631 |
< |
* Unless shutting down, adds the given task to a submission queue |
1388 |
< |
* at submitter's current queue index. If no queue exists at the |
1389 |
< |
* index, one is created unless pool lock is busy. If the queue |
1390 |
< |
* and/or lock are busy, another index is randomly chosen. |
1631 |
> |
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1632 |
|
*/ |
1633 |
< |
private void doSubmit(ForkJoinTask<?> task) { |
1634 |
< |
if (task == null) |
1635 |
< |
throw new NullPointerException(); |
1395 |
< |
Submitter s = submitters.get(); |
1396 |
< |
for (int r = s.seed;;) { |
1397 |
< |
WorkQueue q; int k; |
1398 |
< |
int rs = runState, m = rs & SMASK; |
1399 |
< |
WorkQueue[] ws = workQueues; |
1400 |
< |
if (rs < 0 || ws == null) // shutting down |
1401 |
< |
throw new RejectedExecutionException(); |
1402 |
< |
if (ws.length > m && // k must be at index |
1403 |
< |
((q = ws[k = (r << 1) & m]) != null || |
1404 |
< |
(q = tryAddSharedQueue(k)) != null) && |
1405 |
< |
q.trySharedPush(task)) { |
1406 |
< |
signalWork(); |
1407 |
< |
return; |
1408 |
< |
} |
1409 |
< |
r ^= r << 13; // xorshift seed to new position |
1410 |
< |
r ^= r >>> 17; |
1411 |
< |
if (((s.seed = r ^= r << 5) & m) == 0) |
1412 |
< |
Thread.yield(); // occasionally yield if busy |
1413 |
< |
} |
1633 |
> |
final void runWorker(WorkQueue w) { |
1634 |
> |
w.growArray(); // allocate queue |
1635 |
> |
do { w.runTask(scan(w)); } while (w.qlock >= 0); |
1636 |
|
} |
1637 |
|
|
1416 |
– |
|
1417 |
– |
// Scanning for tasks |
1418 |
– |
|
1638 |
|
/** |
1639 |
|
* Scans for and, if found, returns one task, else possibly |
1640 |
|
* inactivates the worker. This method operates on single reads of |
1641 |
< |
* volatile state and is designed to be re-invoked continuously in |
1642 |
< |
* part because it returns upon detecting inconsistencies, |
1641 |
> |
* volatile state and is designed to be re-invoked continuously, |
1642 |
> |
* in part because it returns upon detecting inconsistencies, |
1643 |
|
* contention, or state changes that indicate possible success on |
1644 |
|
* re-invocation. |
1645 |
|
* |
1646 |
< |
* The scan searches for tasks across queues, randomly selecting |
1647 |
< |
* the first #queues probes, favoring steals 2:1 over submissions |
1648 |
< |
* (by exploiting even/odd indexing), and then performing a |
1649 |
< |
* circular sweep of all queues. The scan terminates upon either |
1650 |
< |
* finding a non-empty queue, or completing a full sweep. If the |
1651 |
< |
* worker is not inactivated, it takes and returns a task from |
1652 |
< |
* this queue. On failure to find a task, we take one of the |
1653 |
< |
* following actions, after which the caller will retry calling |
1654 |
< |
* this method unless terminated. |
1655 |
< |
* |
1656 |
< |
* * If not a complete sweep, try to release a waiting worker. If |
1657 |
< |
* the scan terminated because the worker is inactivated, then the |
1439 |
< |
* released worker will often be the calling worker, and it can |
1440 |
< |
* succeed obtaining a task on the next call. Or maybe it is |
1441 |
< |
* another worker, but with same net effect. Releasing in other |
1442 |
< |
* cases as well ensures that we have enough workers running. |
1443 |
< |
* |
1444 |
< |
* * If the caller has run a task since the the last empty scan, |
1445 |
< |
* return (to allow rescan) if other workers are not also yet |
1446 |
< |
* enqueued. Field WorkQueue.rescans counts down on each scan to |
1447 |
< |
* ensure eventual inactivation, and occasional calls to |
1448 |
< |
* Thread.yield to help avoid interference with more useful |
1449 |
< |
* activities on the system. |
1646 |
> |
* The scan searches for tasks across queues (starting at a random |
1647 |
> |
* index, and relying on registerWorker to irregularly scatter |
1648 |
> |
* them within array to avoid bias), checking each at least twice. |
1649 |
> |
* The scan terminates upon either finding a non-empty queue, or |
1650 |
> |
* completing the sweep. If the worker is not inactivated, it |
1651 |
> |
* takes and returns a task from this queue. Otherwise, if not |
1652 |
> |
* activated, it signals workers (that may include itself) and |
1653 |
> |
* returns so caller can retry. Also returns for true if the |
1654 |
> |
* worker array may have changed during an empty scan. On failure |
1655 |
> |
* to find a task, we take one of the following actions, after |
1656 |
> |
* which the caller will retry calling this method unless |
1657 |
> |
* terminated. |
1658 |
|
* |
1659 |
< |
* * If pool is terminating, terminate the worker |
1659 |
> |
* * If pool is terminating, terminate the worker. |
1660 |
|
* |
1661 |
|
* * If not already enqueued, try to inactivate and enqueue the |
1662 |
< |
* worker on wait queue. |
1663 |
< |
* |
1664 |
< |
* * If already enqueued and none of the above apply, either park |
1665 |
< |
* awaiting signal, or if this is the most recent waiter and pool |
1666 |
< |
* is quiescent, relay to idleAwaitWork to check for termination |
1667 |
< |
* and possibly shrink pool. |
1662 |
> |
* worker on wait queue. Or, if inactivating has caused the pool |
1663 |
> |
* to be quiescent, relay to idleAwaitWork to check for |
1664 |
> |
* termination and possibly shrink pool. |
1665 |
> |
* |
1666 |
> |
* * If already enqueued and none of the above apply, possibly |
1667 |
> |
* (with 1/2 probability) park awaiting signal, else lingering to |
1668 |
> |
* help scan and signal. |
1669 |
|
* |
1670 |
|
* @param w the worker (via its WorkQueue) |
1671 |
< |
* @return a task or null of none found |
1671 |
> |
* @return a task or null if none found |
1672 |
|
*/ |
1673 |
|
private final ForkJoinTask<?> scan(WorkQueue w) { |
1674 |
< |
boolean swept = false; // true after full empty scan |
1675 |
< |
WorkQueue[] ws; // volatile read order matters |
1676 |
< |
int r = w.seed, ec = w.eventCount; // ec is negative if inactive |
1677 |
< |
int rs = runState, m = rs & SMASK; |
1678 |
< |
if ((ws = workQueues) != null && ws.length > m) { |
1679 |
< |
ForkJoinTask<?> task = null; |
1680 |
< |
for (int k = 0, j = -2 - m; ; ++j) { |
1681 |
< |
WorkQueue q; int b; |
1682 |
< |
if (j < 0) { // random probes while j negative |
1683 |
< |
r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1); |
1684 |
< |
} // worker (not submit) for odd j |
1685 |
< |
else // cyclic scan when j >= 0 |
1686 |
< |
k += (m >>> 1) | 1; // step by half to reduce bias |
1687 |
< |
|
1688 |
< |
if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) { |
1689 |
< |
if (ec >= 0) |
1690 |
< |
task = q.pollAt(b); // steal |
1691 |
< |
break; |
1692 |
< |
} |
1693 |
< |
else if (j > m) { |
1694 |
< |
if (rs == runState) // staleness check |
1695 |
< |
swept = true; |
1696 |
< |
break; |
1674 |
> |
WorkQueue[] ws; int m; |
1675 |
> |
int ps = plock; // read plock before ws |
1676 |
> |
if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) { |
1677 |
> |
int ec = w.eventCount; // ec is negative if inactive |
1678 |
> |
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
1679 |
> |
int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN; |
1680 |
> |
do { |
1681 |
> |
WorkQueue q; ForkJoinTask<?>[] a; int b; |
1682 |
> |
if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 && |
1683 |
> |
(a = q.array) != null) { // probably nonempty |
1684 |
> |
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1685 |
> |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
1686 |
> |
U.getObjectVolatile(a, i); |
1687 |
> |
if (q.base == b && ec >= 0 && t != null && |
1688 |
> |
U.compareAndSwapObject(a, i, t, null)) { |
1689 |
> |
if ((q.base = b + 1) - q.top < 0) |
1690 |
> |
signalWork(q); |
1691 |
> |
return t; // taken |
1692 |
> |
} |
1693 |
> |
else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) { |
1694 |
> |
w.hint = (r + j) & m; // help signal below |
1695 |
> |
break; // cannot take |
1696 |
> |
} |
1697 |
|
} |
1698 |
< |
} |
1699 |
< |
w.seed = r; // save seed for next scan |
1700 |
< |
if (task != null) |
1701 |
< |
return task; |
1702 |
< |
} |
1703 |
< |
|
1704 |
< |
// Decode ctl on empty scan |
1705 |
< |
long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns; |
1706 |
< |
if (!swept) { // try to release a waiter |
1707 |
< |
WorkQueue v; Thread p; |
1708 |
< |
if (e > 0 && a < 0 && ws != null && |
1709 |
< |
(v = ws[((~e << 1) | 1) & m]) != null && |
1710 |
< |
v.eventCount == (e | INT_SIGN) && U.compareAndSwapLong |
1711 |
< |
(this, CTL, c, ((long)(v.nextWait & E_MASK) | |
1712 |
< |
((c + AC_UNIT) & (AC_MASK|TC_MASK))))) { |
1713 |
< |
v.eventCount = (e + E_SEQ) & E_MASK; |
1714 |
< |
if ((p = v.parker) != null) |
1715 |
< |
U.unpark(p); |
1716 |
< |
} |
1717 |
< |
} |
1718 |
< |
else if ((nr = w.rescans) > 0) { // continue rescanning |
1719 |
< |
int ac = a + parallelism; |
1720 |
< |
if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 && |
1721 |
< |
w.eventCount == ec) |
1722 |
< |
Thread.yield(); // 1 bit randomness for yield call |
1723 |
< |
} |
1724 |
< |
else if (e < 0) // pool is terminating |
1725 |
< |
w.runState = -1; |
1726 |
< |
else if (ec >= 0) { // try to enqueue |
1727 |
< |
long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); |
1728 |
< |
w.nextWait = e; |
1729 |
< |
w.eventCount = ec | INT_SIGN; // mark as inactive |
1730 |
< |
if (!U.compareAndSwapLong(this, CTL, c, nc)) |
1731 |
< |
w.eventCount = ec; // back out on CAS failure |
1732 |
< |
else if ((ns = w.nsteals) != 0) { // set rescans if ran task |
1733 |
< |
if (a <= 0) // ... unless too many active |
1525 |
< |
w.rescans = a + parallelism; |
1526 |
< |
w.nsteals = 0; |
1527 |
< |
w.totalSteals += ns; |
1528 |
< |
} |
1529 |
< |
} |
1530 |
< |
else{ // already queued |
1531 |
< |
if (parallelism == -a) |
1532 |
< |
idleAwaitWork(w); // quiescent |
1533 |
< |
if (w.eventCount == ec) { |
1534 |
< |
Thread.interrupted(); // clear status |
1535 |
< |
ForkJoinWorkerThread wt = w.owner; |
1536 |
< |
U.putObject(wt, PARKBLOCKER, this); |
1537 |
< |
w.parker = wt; // emulate LockSupport.park |
1538 |
< |
if (w.eventCount == ec) // recheck |
1539 |
< |
U.park(false, 0L); // block |
1540 |
< |
w.parker = null; |
1541 |
< |
U.putObject(wt, PARKBLOCKER, null); |
1698 |
> |
} while (--j >= 0); |
1699 |
> |
|
1700 |
> |
long c, sc; int e, ns, h; |
1701 |
> |
if ((h = w.hint) < 0) { |
1702 |
> |
if ((ns = w.nsteals) != 0) { |
1703 |
> |
if (U.compareAndSwapLong(this, STEALCOUNT, |
1704 |
> |
sc = stealCount, sc + ns)) |
1705 |
> |
w.nsteals = 0; // collect steals |
1706 |
> |
} |
1707 |
> |
else if (plock != ps) // consistency check |
1708 |
> |
; // skip |
1709 |
> |
else if ((e = (int)(c = ctl)) < 0) |
1710 |
> |
w.qlock = -1; // pool is terminating |
1711 |
> |
else if (ec >= 0) { // try to enqueue/inactivate |
1712 |
> |
long nc = ((long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK))); |
1713 |
> |
w.nextWait = e; // link and mark inactive |
1714 |
> |
w.eventCount = ec | INT_SIGN; |
1715 |
> |
if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc)) |
1716 |
> |
w.eventCount = ec; // unmark on CAS failure |
1717 |
> |
else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK)) |
1718 |
> |
idleAwaitWork(w, nc, c); |
1719 |
> |
} |
1720 |
> |
else if (w.eventCount < 0) { // block |
1721 |
> |
Thread wt = Thread.currentThread(); |
1722 |
> |
Thread.interrupted(); // clear status |
1723 |
> |
U.putObject(wt, PARKBLOCKER, this); |
1724 |
> |
w.parker = wt; // emulate LockSupport.park |
1725 |
> |
if (w.eventCount < 0) // recheck |
1726 |
> |
U.park(false, 0L); |
1727 |
> |
w.parker = null; |
1728 |
> |
U.putObject(wt, PARKBLOCKER, null); |
1729 |
> |
} |
1730 |
> |
} |
1731 |
> |
if (h >= 0 || (h = w.hint) >= 0) { // signal others before retry |
1732 |
> |
w.hint = -1; // reset |
1733 |
> |
helpSignal(null, h, true); |
1734 |
|
} |
1735 |
|
} |
1736 |
|
return null; |
1737 |
|
} |
1738 |
|
|
1739 |
|
/** |
1740 |
< |
* If inactivating worker w has caused pool to become quiescent, |
1741 |
< |
* check for pool termination, and, so long as this is not the |
1742 |
< |
* only worker, wait for event for up to SHRINK_RATE nanosecs On |
1743 |
< |
* timeout, if ctl has not changed, terminate the worker, which |
1744 |
< |
* will in turn wake up another worker to possibly repeat this |
1745 |
< |
* process. |
1740 |
> |
* If inactivating worker w has caused the pool to become |
1741 |
> |
* quiescent, checks for pool termination, and, so long as this is |
1742 |
> |
* not the only worker, waits for event for up to a given |
1743 |
> |
* duration. On timeout, if ctl has not changed, terminates the |
1744 |
> |
* worker, which will in turn wake up another worker to possibly |
1745 |
> |
* repeat this process. |
1746 |
|
* |
1747 |
|
* @param w the calling worker |
1748 |
+ |
* @param currentCtl the ctl value triggering possible quiescence |
1749 |
+ |
* @param prevCtl the ctl value to restore if thread is terminated |
1750 |
|
*/ |
1751 |
< |
private void idleAwaitWork(WorkQueue w) { |
1752 |
< |
long c; int nw, ec; |
1753 |
< |
if (!tryTerminate(false) && |
1754 |
< |
(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 && |
1755 |
< |
(ec = w.eventCount) == ((int)c | INT_SIGN) && |
1756 |
< |
(nw = w.nextWait) != 0) { |
1757 |
< |
long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout |
1758 |
< |
((c + AC_UNIT) & AC_MASK) | (c & TC_MASK)); |
1565 |
< |
ForkJoinTask.helpExpungeStaleExceptions(); // help clean |
1566 |
< |
ForkJoinWorkerThread wt = w.owner; |
1567 |
< |
while (ctl == c) { |
1568 |
< |
long startTime = System.nanoTime(); |
1751 |
> |
private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) { |
1752 |
> |
if (w != null && w.eventCount < 0 && |
1753 |
> |
!tryTerminate(false, false) && (int)prevCtl != 0) { |
1754 |
> |
int dc = -(short)(currentCtl >>> TC_SHIFT); |
1755 |
> |
long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT; |
1756 |
> |
long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop |
1757 |
> |
Thread wt = Thread.currentThread(); |
1758 |
> |
while (ctl == currentCtl) { |
1759 |
|
Thread.interrupted(); // timed variant of version in scan() |
1760 |
|
U.putObject(wt, PARKBLOCKER, this); |
1761 |
|
w.parker = wt; |
1762 |
< |
if (ctl == c) |
1763 |
< |
U.park(false, SHRINK_RATE); |
1762 |
> |
if (ctl == currentCtl) |
1763 |
> |
U.park(false, parkTime); |
1764 |
|
w.parker = null; |
1765 |
|
U.putObject(wt, PARKBLOCKER, null); |
1766 |
< |
if (ctl != c) |
1766 |
> |
if (ctl != currentCtl) |
1767 |
|
break; |
1768 |
< |
if (System.nanoTime() - startTime >= SHRINK_TIMEOUT && |
1769 |
< |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1770 |
< |
w.runState = -1; // shrink |
1771 |
< |
w.eventCount = (ec + E_SEQ) | E_MASK; |
1768 |
> |
if (deadline - System.nanoTime() <= 0L && |
1769 |
> |
U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) { |
1770 |
> |
w.eventCount = (w.eventCount + E_SEQ) | E_MASK; |
1771 |
> |
w.qlock = -1; // shrink |
1772 |
|
break; |
1773 |
|
} |
1774 |
|
} |
1776 |
|
} |
1777 |
|
|
1778 |
|
/** |
1779 |
+ |
* Scans through queues looking for work (optionally, while |
1780 |
+ |
* joining a task); if any present, signals. May return early if |
1781 |
+ |
* more signalling is detectably unneeded. |
1782 |
+ |
* |
1783 |
+ |
* @param task if non-null, return early if done |
1784 |
+ |
* @param origin an index to start scan |
1785 |
+ |
* @param once if only the origin should be checked |
1786 |
+ |
*/ |
1787 |
+ |
private void helpSignal(ForkJoinTask<?> task, int origin, boolean once) { |
1788 |
+ |
WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s; |
1789 |
+ |
if ((u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 && |
1790 |
+ |
(ws = workQueues) != null && (m = ws.length - 1) >= 0) { |
1791 |
+ |
outer: for (int k = origin, j = once? 0 : m; j >= 0; --j) { |
1792 |
+ |
WorkQueue q = ws[k++ & m]; |
1793 |
+ |
for (int n = m;;) { // limit to at most m signals |
1794 |
+ |
if (task != null && task.status < 0) |
1795 |
+ |
break outer; |
1796 |
+ |
if (q == null || |
1797 |
+ |
((s = (task == null ? -1 : 0) - q.base + q.top) <= n && |
1798 |
+ |
(n = s) <= 0)) |
1799 |
+ |
break; |
1800 |
+ |
if ((u = (int)((c = ctl) >>> 32)) >= 0 || |
1801 |
+ |
(e = (int)c) <= 0 || m < (i = e & SMASK) || |
1802 |
+ |
(w = ws[i]) == null) |
1803 |
+ |
break outer; |
1804 |
+ |
long nc = (((long)(w.nextWait & E_MASK)) | |
1805 |
+ |
((long)(u + UAC_UNIT) << 32)); |
1806 |
+ |
if (w.eventCount == (e | INT_SIGN) && |
1807 |
+ |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1808 |
+ |
w.eventCount = (e + E_SEQ) & E_MASK; |
1809 |
+ |
if ((p = w.parker) != null) |
1810 |
+ |
U.unpark(p); |
1811 |
+ |
if (--n <= 0) |
1812 |
+ |
break; |
1813 |
+ |
} |
1814 |
+ |
} |
1815 |
+ |
} |
1816 |
+ |
} |
1817 |
+ |
} |
1818 |
+ |
|
1819 |
+ |
/** |
1820 |
|
* Tries to locate and execute tasks for a stealer of the given |
1821 |
|
* task, or in turn one of its stealers, Traces currentSteal -> |
1822 |
|
* currentJoin links looking for a thread working on a descendant |
1827 |
|
* leaves hints in workers to speed up subsequent calls. The |
1828 |
|
* implementation is very branchy to cope with potential |
1829 |
|
* inconsistencies or loops encountering chains that are stale, |
1830 |
< |
* unknown, or of length greater than MAX_HELP_DEPTH links. All |
1600 |
< |
* of these cases are dealt with by just retrying by caller. |
1830 |
> |
* unknown, or so long that they are likely cyclic. |
1831 |
|
* |
1832 |
|
* @param joiner the joining worker |
1833 |
|
* @param task the task to join |
1834 |
< |
* @return true if found or ran a task (and so is immediately retryable) |
1834 |
> |
* @return 0 if no progress can be made, negative if task |
1835 |
> |
* known complete, else positive |
1836 |
|
*/ |
1837 |
< |
final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) { |
1838 |
< |
ForkJoinTask<?> subtask; // current target |
1839 |
< |
boolean progress = false; |
1840 |
< |
int depth = 0; // current chain depth |
1841 |
< |
int m = runState & SMASK; |
1842 |
< |
WorkQueue[] ws = workQueues; |
1843 |
< |
|
1844 |
< |
if (ws != null && ws.length > m && (subtask = task).status >= 0) { |
1845 |
< |
outer:for (WorkQueue j = joiner;;) { |
1846 |
< |
// Try to find the stealer of subtask, by first using hint |
1847 |
< |
WorkQueue stealer = null; |
1848 |
< |
WorkQueue v = ws[j.stealHint & m]; |
1849 |
< |
if (v != null && v.currentSteal == subtask) |
1850 |
< |
stealer = v; |
1851 |
< |
else { |
1852 |
< |
for (int i = 1; i <= m; i += 2) { |
1853 |
< |
if ((v = ws[i]) != null && v.currentSteal == subtask) { |
1854 |
< |
stealer = v; |
1855 |
< |
j.stealHint = i; // save hint |
1856 |
< |
break; |
1837 |
> |
private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) { |
1838 |
> |
int stat = 0, steps = 0; // bound to avoid cycles |
1839 |
> |
if (joiner != null && task != null) { // hoist null checks |
1840 |
> |
restart: for (;;) { |
1841 |
> |
ForkJoinTask<?> subtask = task; // current target |
1842 |
> |
for (WorkQueue j = joiner, v;;) { // v is stealer of subtask |
1843 |
> |
WorkQueue[] ws; int m, s, h; |
1844 |
> |
if ((s = task.status) < 0) { |
1845 |
> |
stat = s; |
1846 |
> |
break restart; |
1847 |
> |
} |
1848 |
> |
if ((ws = workQueues) == null || (m = ws.length - 1) <= 0) |
1849 |
> |
break restart; // shutting down |
1850 |
> |
if ((v = ws[h = (j.hint | 1) & m]) == null || |
1851 |
> |
v.currentSteal != subtask) { |
1852 |
> |
for (int origin = h;;) { // find stealer |
1853 |
> |
if (((h = (h + 2) & m) & 15) == 1 && |
1854 |
> |
(subtask.status < 0 || j.currentJoin != subtask)) |
1855 |
> |
continue restart; // occasional staleness check |
1856 |
> |
if ((v = ws[h]) != null && |
1857 |
> |
v.currentSteal == subtask) { |
1858 |
> |
j.hint = h; // save hint |
1859 |
> |
break; |
1860 |
> |
} |
1861 |
> |
if (h == origin) |
1862 |
> |
break restart; // cannot find stealer |
1863 |
|
} |
1864 |
|
} |
1865 |
< |
if (stealer == null) |
1865 |
> |
for (;;) { // help stealer or descend to its stealer |
1866 |
> |
ForkJoinTask[] a; int b; |
1867 |
> |
if (subtask.status < 0) // surround probes with |
1868 |
> |
continue restart; // consistency checks |
1869 |
> |
if ((b = v.base) - v.top < 0 && (a = v.array) != null) { |
1870 |
> |
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1871 |
> |
ForkJoinTask<?> t = |
1872 |
> |
(ForkJoinTask<?>)U.getObjectVolatile(a, i); |
1873 |
> |
if (subtask.status < 0 || j.currentJoin != subtask || |
1874 |
> |
v.currentSteal != subtask) |
1875 |
> |
continue restart; // stale |
1876 |
> |
stat = 1; // apparent progress |
1877 |
> |
if (t != null && v.base == b && |
1878 |
> |
U.compareAndSwapObject(a, i, t, null)) { |
1879 |
> |
v.base = b + 1; // help stealer |
1880 |
> |
joiner.runSubtask(t); |
1881 |
> |
} |
1882 |
> |
else if (v.base == b && ++steps == MAX_HELP) |
1883 |
> |
break restart; // v apparently stalled |
1884 |
> |
} |
1885 |
> |
else { // empty -- try to descend |
1886 |
> |
ForkJoinTask<?> next = v.currentJoin; |
1887 |
> |
if (subtask.status < 0 || j.currentJoin != subtask || |
1888 |
> |
v.currentSteal != subtask) |
1889 |
> |
continue restart; // stale |
1890 |
> |
else if (next == null || ++steps == MAX_HELP) |
1891 |
> |
break restart; // dead-end or maybe cyclic |
1892 |
> |
else { |
1893 |
> |
subtask = next; |
1894 |
> |
j = v; |
1895 |
> |
break; |
1896 |
> |
} |
1897 |
> |
} |
1898 |
> |
} |
1899 |
> |
} |
1900 |
> |
} |
1901 |
> |
} |
1902 |
> |
return stat; |
1903 |
> |
} |
1904 |
> |
|
1905 |
> |
/** |
1906 |
> |
* Analog of tryHelpStealer for CountedCompleters. Tries to steal |
1907 |
> |
* and run tasks within the target's computation. |
1908 |
> |
* |
1909 |
> |
* @param task the task to join |
1910 |
> |
* @param mode if shared, exit upon completing any task |
1911 |
> |
* if all workers are active |
1912 |
> |
* |
1913 |
> |
*/ |
1914 |
> |
private int helpComplete(ForkJoinTask<?> task, int mode) { |
1915 |
> |
WorkQueue[] ws; WorkQueue q; int m, n, s, u; |
1916 |
> |
if (task != null && (ws = workQueues) != null && |
1917 |
> |
(m = ws.length - 1) >= 0) { |
1918 |
> |
for (int j = 1, origin = j;;) { |
1919 |
> |
if ((s = task.status) < 0) |
1920 |
> |
return s; |
1921 |
> |
if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) { |
1922 |
> |
origin = j; |
1923 |
> |
if (mode == SHARED_QUEUE && |
1924 |
> |
((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)) |
1925 |
|
break; |
1926 |
|
} |
1927 |
+ |
else if ((j = (j + 2) & m) == origin) |
1928 |
+ |
break; |
1929 |
+ |
} |
1930 |
+ |
} |
1931 |
+ |
return 0; |
1932 |
+ |
} |
1933 |
|
|
1934 |
< |
for (WorkQueue q = stealer;;) { // Try to help stealer |
1935 |
< |
ForkJoinTask<?> t; int b; |
1936 |
< |
if (task.status < 0) |
1937 |
< |
break outer; |
1938 |
< |
if ((b = q.base) - q.top < 0) { |
1939 |
< |
progress = true; |
1940 |
< |
if (subtask.status < 0) |
1941 |
< |
break outer; // stale |
1942 |
< |
if ((t = q.pollAt(b)) != null) { |
1943 |
< |
stealer.stealHint = joiner.poolIndex; |
1944 |
< |
joiner.runSubtask(t); |
1934 |
> |
/** |
1935 |
> |
* Tries to decrement active count (sometimes implicitly) and |
1936 |
> |
* possibly release or create a compensating worker in preparation |
1937 |
> |
* for blocking. Fails on contention or termination. Otherwise, |
1938 |
> |
* adds a new thread if no idle workers are available and pool |
1939 |
> |
* may become starved. |
1940 |
> |
*/ |
1941 |
> |
final boolean tryCompensate() { |
1942 |
> |
int pc = config & SMASK, e, i, tc; long c; |
1943 |
> |
WorkQueue[] ws; WorkQueue w; Thread p; |
1944 |
> |
if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) { |
1945 |
> |
if (e != 0 && (i = e & SMASK) < ws.length && |
1946 |
> |
(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { |
1947 |
> |
long nc = ((long)(w.nextWait & E_MASK) | |
1948 |
> |
(c & (AC_MASK|TC_MASK))); |
1949 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1950 |
> |
w.eventCount = (e + E_SEQ) & E_MASK; |
1951 |
> |
if ((p = w.parker) != null) |
1952 |
> |
U.unpark(p); |
1953 |
> |
return true; // replace with idle worker |
1954 |
> |
} |
1955 |
> |
} |
1956 |
> |
else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 && |
1957 |
> |
(int)(c >> AC_SHIFT) + pc > 1) { |
1958 |
> |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1959 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1960 |
> |
return true; // no compensation |
1961 |
> |
} |
1962 |
> |
else if (tc + pc < MAX_CAP) { |
1963 |
> |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1964 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1965 |
> |
ForkJoinWorkerThreadFactory fac; |
1966 |
> |
Throwable ex = null; |
1967 |
> |
ForkJoinWorkerThread wt = null; |
1968 |
> |
try { |
1969 |
> |
if ((fac = factory) != null && |
1970 |
> |
(wt = fac.newThread(this)) != null) { |
1971 |
> |
wt.start(); |
1972 |
> |
return true; |
1973 |
|
} |
1974 |
+ |
} catch (Throwable rex) { |
1975 |
+ |
ex = rex; |
1976 |
|
} |
1977 |
< |
else { // empty - try to descend to find stealer's stealer |
1646 |
< |
ForkJoinTask<?> next = stealer.currentJoin; |
1647 |
< |
if (++depth == MAX_HELP_DEPTH || subtask.status < 0 || |
1648 |
< |
next == null || next == subtask) |
1649 |
< |
break outer; // max depth, stale, dead-end, cyclic |
1650 |
< |
subtask = next; |
1651 |
< |
j = stealer; |
1652 |
< |
break; |
1653 |
< |
} |
1977 |
> |
deregisterWorker(wt, ex); // clean up and return false |
1978 |
|
} |
1979 |
|
} |
1980 |
|
} |
1981 |
< |
return progress; |
1981 |
> |
return false; |
1982 |
|
} |
1983 |
|
|
1984 |
|
/** |
1985 |
< |
* If task is at base of some steal queue, steals and executes it. |
1985 |
> |
* Helps and/or blocks until the given task is done. |
1986 |
|
* |
1987 |
|
* @param joiner the joining worker |
1988 |
|
* @param task the task |
1989 |
+ |
* @return task status on exit |
1990 |
|
*/ |
1991 |
< |
final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) { |
1992 |
< |
WorkQueue[] ws; |
1993 |
< |
int m = runState & SMASK; |
1994 |
< |
if ((ws = workQueues) != null && ws.length > m) { |
1995 |
< |
for (int j = 1; j <= m && task.status >= 0; j += 2) { |
1996 |
< |
WorkQueue q = ws[j]; |
1997 |
< |
if (q != null && q.pollFor(task)) { |
1998 |
< |
joiner.runSubtask(task); |
1999 |
< |
break; |
1991 |
> |
final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) { |
1992 |
> |
int s = 0; |
1993 |
> |
if (joiner != null && task != null && (s = task.status) >= 0) { |
1994 |
> |
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
1995 |
> |
joiner.currentJoin = task; |
1996 |
> |
do {} while ((s = task.status) >= 0 && !joiner.isEmpty() && |
1997 |
> |
joiner.tryRemoveAndExec(task)); // process local tasks |
1998 |
> |
if (s >= 0 && (s = task.status) >= 0) { |
1999 |
> |
helpSignal(task, joiner.poolIndex, false); |
2000 |
> |
if ((s = task.status) >= 0 && |
2001 |
> |
(task instanceof CountedCompleter)) |
2002 |
> |
s = helpComplete(task, LIFO_QUEUE); |
2003 |
> |
} |
2004 |
> |
while (s >= 0 && (s = task.status) >= 0) { |
2005 |
> |
if ((!joiner.isEmpty() || // try helping |
2006 |
> |
(s = tryHelpStealer(joiner, task)) == 0) && |
2007 |
> |
(s = task.status) >= 0) { |
2008 |
> |
helpSignal(task, joiner.poolIndex, false); |
2009 |
> |
if ((s = task.status) >= 0 && tryCompensate()) { |
2010 |
> |
if (task.trySetSignal() && (s = task.status) >= 0) { |
2011 |
> |
synchronized (task) { |
2012 |
> |
if (task.status >= 0) { |
2013 |
> |
try { // see ForkJoinTask |
2014 |
> |
task.wait(); // for explanation |
2015 |
> |
} catch (InterruptedException ie) { |
2016 |
> |
} |
2017 |
> |
} |
2018 |
> |
else |
2019 |
> |
task.notifyAll(); |
2020 |
> |
} |
2021 |
> |
} |
2022 |
> |
long c; // re-activate |
2023 |
> |
do {} while (!U.compareAndSwapLong |
2024 |
> |
(this, CTL, c = ctl, c + AC_UNIT)); |
2025 |
> |
} |
2026 |
|
} |
2027 |
|
} |
2028 |
+ |
joiner.currentJoin = prevJoin; |
2029 |
+ |
} |
2030 |
+ |
return s; |
2031 |
+ |
} |
2032 |
+ |
|
2033 |
+ |
/** |
2034 |
+ |
* Stripped-down variant of awaitJoin used by timed joins. Tries |
2035 |
+ |
* to help join only while there is continuous progress. (Caller |
2036 |
+ |
* will then enter a timed wait.) |
2037 |
+ |
* |
2038 |
+ |
* @param joiner the joining worker |
2039 |
+ |
* @param task the task |
2040 |
+ |
*/ |
2041 |
+ |
final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) { |
2042 |
+ |
int s; |
2043 |
+ |
if (joiner != null && task != null && (s = task.status) >= 0) { |
2044 |
+ |
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
2045 |
+ |
joiner.currentJoin = task; |
2046 |
+ |
do {} while ((s = task.status) >= 0 && !joiner.isEmpty() && |
2047 |
+ |
joiner.tryRemoveAndExec(task)); |
2048 |
+ |
if (s >= 0 && (s = task.status) >= 0) { |
2049 |
+ |
helpSignal(task, joiner.poolIndex, false); |
2050 |
+ |
if ((s = task.status) >= 0 && |
2051 |
+ |
(task instanceof CountedCompleter)) |
2052 |
+ |
s = helpComplete(task, LIFO_QUEUE); |
2053 |
+ |
} |
2054 |
+ |
if (s >= 0 && joiner.isEmpty()) { |
2055 |
+ |
do {} while (task.status >= 0 && |
2056 |
+ |
tryHelpStealer(joiner, task) > 0); |
2057 |
+ |
} |
2058 |
+ |
joiner.currentJoin = prevJoin; |
2059 |
|
} |
2060 |
|
} |
2061 |
|
|
2062 |
|
/** |
2063 |
< |
* Returns a non-empty steal queue, if one is found during a random, |
2064 |
< |
* then cyclic scan, else null. This method must be retried by |
2065 |
< |
* caller if, by the time it tries to use the queue, it is empty. |
2063 |
> |
* Returns a (probably) non-empty steal queue, if one is found |
2064 |
> |
* during a random, then cyclic scan, else null. This method must |
2065 |
> |
* be retried by caller if, by the time it tries to use the queue, |
2066 |
> |
* it is empty. |
2067 |
> |
* @param r a (random) seed for scanning |
2068 |
|
*/ |
2069 |
< |
private WorkQueue findNonEmptyStealQueue(WorkQueue w) { |
1686 |
< |
int r = w.seed; // Same idea as scan(), but ignoring submissions |
2069 |
> |
private WorkQueue findNonEmptyStealQueue(int r) { |
2070 |
|
for (WorkQueue[] ws;;) { |
2071 |
< |
int m = runState & SMASK; |
2072 |
< |
if ((ws = workQueues) == null) |
2071 |
> |
int ps = plock, m, n; |
2072 |
> |
if ((ws = workQueues) == null || (m = ws.length - 1) < 1) |
2073 |
|
return null; |
2074 |
< |
if (ws.length > m) { |
2075 |
< |
WorkQueue q; |
2076 |
< |
for (int n = m << 2, k = r, j = -n;;) { |
2077 |
< |
r ^= r << 13; r ^= r >>> 17; r ^= r << 5; |
2078 |
< |
if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) { |
2079 |
< |
w.seed = r; |
2080 |
< |
return q; |
2081 |
< |
} |
2082 |
< |
else if (j > n) |
2074 |
> |
for (int j = (m + 1) << 2; ;) { |
2075 |
> |
WorkQueue q = ws[(((r + j) << 1) | 1) & m]; |
2076 |
> |
if (q != null && (n = q.base - q.top) < 0) { |
2077 |
> |
if (n < -1) |
2078 |
> |
signalWork(q); |
2079 |
> |
return q; |
2080 |
> |
} |
2081 |
> |
else if (--j < 0) { |
2082 |
> |
if (plock == ps) |
2083 |
|
return null; |
2084 |
< |
else |
1702 |
< |
k = (j++ < 0) ? r : k + ((m >>> 1) | 1); |
1703 |
< |
|
2084 |
> |
break; |
2085 |
|
} |
2086 |
|
} |
2087 |
|
} |
2095 |
|
*/ |
2096 |
|
final void helpQuiescePool(WorkQueue w) { |
2097 |
|
for (boolean active = true;;) { |
2098 |
< |
w.runLocalTasks(); // exhaust local queue |
2099 |
< |
WorkQueue q = findNonEmptyStealQueue(w); |
2098 |
> |
ForkJoinTask<?> localTask; // exhaust local queue |
2099 |
> |
while ((localTask = w.nextLocalTask()) != null) |
2100 |
> |
localTask.doExec(); |
2101 |
> |
// Similar to loop in scan(), but ignoring submissions |
2102 |
> |
WorkQueue q = findNonEmptyStealQueue(w.nextSeed()); |
2103 |
|
if (q != null) { |
2104 |
< |
ForkJoinTask<?> t; |
2104 |
> |
ForkJoinTask<?> t; int b; |
2105 |
|
if (!active) { // re-establish active count |
2106 |
|
long c; |
2107 |
|
active = true; |
2108 |
|
do {} while (!U.compareAndSwapLong |
2109 |
|
(this, CTL, c = ctl, c + AC_UNIT)); |
2110 |
|
} |
2111 |
< |
if ((t = q.poll()) != null) |
2111 |
> |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
2112 |
|
w.runSubtask(t); |
2113 |
|
} |
2114 |
|
else { |
2120 |
|
} |
2121 |
|
else |
2122 |
|
c = ctl; // re-increment on exit |
2123 |
< |
if ((int)(c >> AC_SHIFT) + parallelism == 0) { |
2123 |
> |
if ((int)(c >> AC_SHIFT) + (config & SMASK) == 0) { |
2124 |
|
do {} while (!U.compareAndSwapLong |
2125 |
|
(this, CTL, c = ctl, c + AC_UNIT)); |
2126 |
|
break; |
2130 |
|
} |
2131 |
|
|
2132 |
|
/** |
2133 |
< |
* Gets and removes a local or stolen task for the given worker |
2133 |
> |
* Gets and removes a local or stolen task for the given worker. |
2134 |
|
* |
2135 |
|
* @return a task, if available |
2136 |
|
*/ |
2137 |
|
final ForkJoinTask<?> nextTaskFor(WorkQueue w) { |
2138 |
|
for (ForkJoinTask<?> t;;) { |
2139 |
< |
WorkQueue q; |
2139 |
> |
WorkQueue q; int b; |
2140 |
|
if ((t = w.nextLocalTask()) != null) |
2141 |
|
return t; |
2142 |
< |
if ((q = findNonEmptyStealQueue(w)) == null) |
2142 |
> |
if ((q = findNonEmptyStealQueue(w.nextSeed())) == null) |
2143 |
|
return null; |
2144 |
< |
if ((t = q.poll()) != null) |
2144 |
> |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
2145 |
|
return t; |
2146 |
|
} |
2147 |
|
} |
2148 |
|
|
2149 |
|
/** |
2150 |
< |
* Returns the approximate (non-atomic) number of idle threads per |
2151 |
< |
* active thread to offset steal queue size for method |
2152 |
< |
* ForkJoinTask.getSurplusQueuedTaskCount(). |
2153 |
< |
*/ |
2154 |
< |
final int idlePerActive() { |
2155 |
< |
// Approximate at powers of two for small values, saturate past 4 |
2156 |
< |
int p = parallelism; |
2157 |
< |
int a = p + (int)(ctl >> AC_SHIFT); |
2158 |
< |
return (a > (p >>>= 1) ? 0 : |
2159 |
< |
a > (p >>>= 1) ? 1 : |
2160 |
< |
a > (p >>>= 1) ? 2 : |
2161 |
< |
a > (p >>>= 1) ? 4 : |
2162 |
< |
8); |
2163 |
< |
} |
2164 |
< |
|
2165 |
< |
// Termination |
2166 |
< |
|
2167 |
< |
/** |
2168 |
< |
* Sets SHUTDOWN bit of runState under lock |
2169 |
< |
*/ |
2170 |
< |
private void enableShutdown() { |
2171 |
< |
ReentrantLock lock = this.lock; |
2172 |
< |
if (runState >= 0) { |
2173 |
< |
lock.lock(); // don't need try/finally |
2174 |
< |
runState |= SHUTDOWN; |
2175 |
< |
lock.unlock(); |
2150 |
> |
* Returns a cheap heuristic guide for task partitioning when |
2151 |
> |
* programmers, frameworks, tools, or languages have little or no |
2152 |
> |
* idea about task granularity. In essence by offering this |
2153 |
> |
* method, we ask users only about tradeoffs in overhead vs |
2154 |
> |
* expected throughput and its variance, rather than how finely to |
2155 |
> |
* partition tasks. |
2156 |
> |
* |
2157 |
> |
* In a steady state strict (tree-structured) computation, each |
2158 |
> |
* thread makes available for stealing enough tasks for other |
2159 |
> |
* threads to remain active. Inductively, if all threads play by |
2160 |
> |
* the same rules, each thread should make available only a |
2161 |
> |
* constant number of tasks. |
2162 |
> |
* |
2163 |
> |
* The minimum useful constant is just 1. But using a value of 1 |
2164 |
> |
* would require immediate replenishment upon each steal to |
2165 |
> |
* maintain enough tasks, which is infeasible. Further, |
2166 |
> |
* partitionings/granularities of offered tasks should minimize |
2167 |
> |
* steal rates, which in general means that threads nearer the top |
2168 |
> |
* of computation tree should generate more than those nearer the |
2169 |
> |
* bottom. In perfect steady state, each thread is at |
2170 |
> |
* approximately the same level of computation tree. However, |
2171 |
> |
* producing extra tasks amortizes the uncertainty of progress and |
2172 |
> |
* diffusion assumptions. |
2173 |
> |
* |
2174 |
> |
* So, users will want to use values larger, but not much larger |
2175 |
> |
* than 1 to both smooth over transient shortages and hedge |
2176 |
> |
* against uneven progress; as traded off against the cost of |
2177 |
> |
* extra task overhead. We leave the user to pick a threshold |
2178 |
> |
* value to compare with the results of this call to guide |
2179 |
> |
* decisions, but recommend values such as 3. |
2180 |
> |
* |
2181 |
> |
* When all threads are active, it is on average OK to estimate |
2182 |
> |
* surplus strictly locally. In steady-state, if one thread is |
2183 |
> |
* maintaining say 2 surplus tasks, then so are others. So we can |
2184 |
> |
* just use estimated queue length. However, this strategy alone |
2185 |
> |
* leads to serious mis-estimates in some non-steady-state |
2186 |
> |
* conditions (ramp-up, ramp-down, other stalls). We can detect |
2187 |
> |
* many of these by further considering the number of "idle" |
2188 |
> |
* threads, that are known to have zero queued tasks, so |
2189 |
> |
* compensate by a factor of (#idle/#active) threads. |
2190 |
> |
* |
2191 |
> |
* Note: The approximation of #busy workers as #active workers is |
2192 |
> |
* not very good under current signalling scheme, and should be |
2193 |
> |
* improved. |
2194 |
> |
*/ |
2195 |
> |
static int getSurplusQueuedTaskCount() { |
2196 |
> |
Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; |
2197 |
> |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) { |
2198 |
> |
int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK; |
2199 |
> |
int n = (q = wt.workQueue).top - q.base; |
2200 |
> |
int a = (int)(pool.ctl >> AC_SHIFT) + p; |
2201 |
> |
return n - (a > (p >>>= 1) ? 0 : |
2202 |
> |
a > (p >>>= 1) ? 1 : |
2203 |
> |
a > (p >>>= 1) ? 2 : |
2204 |
> |
a > (p >>>= 1) ? 4 : |
2205 |
> |
8); |
2206 |
|
} |
2207 |
+ |
return 0; |
2208 |
|
} |
2209 |
|
|
2210 |
+ |
// Termination |
2211 |
+ |
|
2212 |
|
/** |
2213 |
< |
* Possibly initiates and/or completes termination. Upon |
2214 |
< |
* termination, cancels all queued tasks and then |
2213 |
> |
* Possibly initiates and/or completes termination. The caller |
2214 |
> |
* triggering termination runs three passes through workQueues: |
2215 |
> |
* (0) Setting termination status, followed by wakeups of queued |
2216 |
> |
* workers; (1) cancelling all tasks; (2) interrupting lagging |
2217 |
> |
* threads (likely in external tasks, but possibly also blocked in |
2218 |
> |
* joins). Each pass repeats previous steps because of potential |
2219 |
> |
* lagging thread creation. |
2220 |
|
* |
2221 |
|
* @param now if true, unconditionally terminate, else only |
2222 |
|
* if no work and no active workers |
2223 |
+ |
* @param enable if true, enable shutdown when next possible |
2224 |
|
* @return true if now terminating or terminated |
2225 |
|
*/ |
2226 |
< |
private boolean tryTerminate(boolean now) { |
2226 |
> |
private boolean tryTerminate(boolean now, boolean enable) { |
2227 |
> |
if (this == commonPool) // cannot shut down |
2228 |
> |
return false; |
2229 |
|
for (long c;;) { |
2230 |
|
if (((c = ctl) & STOP_BIT) != 0) { // already terminating |
2231 |
< |
if ((short)(c >>> TC_SHIFT) == -parallelism) { |
2232 |
< |
ReentrantLock lock = this.lock; // signal when no workers |
2233 |
< |
lock.lock(); // don't need try/finally |
2234 |
< |
termination.signalAll(); // signal when 0 workers |
1810 |
< |
lock.unlock(); |
2231 |
> |
if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) { |
2232 |
> |
synchronized (this) { |
2233 |
> |
notifyAll(); // signal when 0 workers |
2234 |
> |
} |
2235 |
|
} |
2236 |
|
return true; |
2237 |
|
} |
2238 |
< |
if (!now) { |
2239 |
< |
if ((int)(c >> AC_SHIFT) != -parallelism || runState >= 0 || |
2238 |
> |
if (plock >= 0) { // not yet enabled |
2239 |
> |
int ps; |
2240 |
> |
if (!enable) |
2241 |
> |
return false; |
2242 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
2243 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
2244 |
> |
ps = acquirePlock(); |
2245 |
> |
int nps = SHUTDOWN; |
2246 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
2247 |
> |
releasePlock(nps); |
2248 |
> |
} |
2249 |
> |
if (!now) { // check if idle & no tasks |
2250 |
> |
if ((int)(c >> AC_SHIFT) != -(config & SMASK) || |
2251 |
|
hasQueuedSubmissions()) |
2252 |
|
return false; |
2253 |
|
// Check for unqueued inactive workers. One pass suffices. |
2254 |
|
WorkQueue[] ws = workQueues; WorkQueue w; |
2255 |
|
if (ws != null) { |
2256 |
< |
int n = ws.length; |
1822 |
< |
for (int i = 1; i < n; i += 2) { |
2256 |
> |
for (int i = 1; i < ws.length; i += 2) { |
2257 |
|
if ((w = ws[i]) != null && w.eventCount >= 0) |
2258 |
|
return false; |
2259 |
|
} |
2260 |
|
} |
2261 |
|
} |
2262 |
< |
if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) |
2263 |
< |
startTerminating(); |
2262 |
> |
if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) { |
2263 |
> |
for (int pass = 0; pass < 3; ++pass) { |
2264 |
> |
WorkQueue[] ws = workQueues; |
2265 |
> |
if (ws != null) { |
2266 |
> |
WorkQueue w; |
2267 |
> |
int n = ws.length; |
2268 |
> |
for (int i = 0; i < n; ++i) { |
2269 |
> |
if ((w = ws[i]) != null) { |
2270 |
> |
w.qlock = -1; |
2271 |
> |
if (pass > 0) { |
2272 |
> |
w.cancelAll(); |
2273 |
> |
if (pass > 1) |
2274 |
> |
w.interruptOwner(); |
2275 |
> |
} |
2276 |
> |
} |
2277 |
> |
} |
2278 |
> |
// Wake up workers parked on event queue |
2279 |
> |
int i, e; long cc; Thread p; |
2280 |
> |
while ((e = (int)(cc = ctl) & E_MASK) != 0 && |
2281 |
> |
(i = e & SMASK) < n && |
2282 |
> |
(w = ws[i]) != null) { |
2283 |
> |
long nc = ((long)(w.nextWait & E_MASK) | |
2284 |
> |
((cc + AC_UNIT) & AC_MASK) | |
2285 |
> |
(cc & (TC_MASK|STOP_BIT))); |
2286 |
> |
if (w.eventCount == (e | INT_SIGN) && |
2287 |
> |
U.compareAndSwapLong(this, CTL, cc, nc)) { |
2288 |
> |
w.eventCount = (e + E_SEQ) & E_MASK; |
2289 |
> |
w.qlock = -1; |
2290 |
> |
if ((p = w.parker) != null) |
2291 |
> |
U.unpark(p); |
2292 |
> |
} |
2293 |
> |
} |
2294 |
> |
} |
2295 |
> |
} |
2296 |
> |
} |
2297 |
|
} |
2298 |
|
} |
2299 |
|
|
2300 |
+ |
// external operations on common pool |
2301 |
+ |
|
2302 |
|
/** |
2303 |
< |
* Initiates termination: Runs three passes through workQueues: |
2304 |
< |
* (0) Setting termination status, followed by wakeups of queued |
1836 |
< |
* workers; (1) cancelling all tasks; (2) interrupting lagging |
1837 |
< |
* threads (likely in external tasks, but possibly also blocked in |
1838 |
< |
* joins). Each pass repeats previous steps because of potential |
1839 |
< |
* lagging thread creation. |
2303 |
> |
* Returns common pool queue for a thread that has submitted at |
2304 |
> |
* least one task. |
2305 |
|
*/ |
2306 |
< |
private void startTerminating() { |
2307 |
< |
for (int pass = 0; pass < 3; ++pass) { |
2308 |
< |
WorkQueue[] ws = workQueues; |
2309 |
< |
if (ws != null) { |
2310 |
< |
WorkQueue w; Thread wt; |
2311 |
< |
int n = ws.length; |
2312 |
< |
for (int j = 0; j < n; ++j) { |
2313 |
< |
if ((w = ws[j]) != null) { |
2314 |
< |
w.runState = -1; |
2315 |
< |
if (pass > 0) { |
2316 |
< |
w.cancelAll(); |
2317 |
< |
if (pass > 1 && (wt = w.owner) != null && |
2318 |
< |
!wt.isInterrupted()) { |
2319 |
< |
try { |
2320 |
< |
wt.interrupt(); |
2321 |
< |
} catch (SecurityException ignore) { |
2306 |
> |
static WorkQueue commonSubmitterQueue() { |
2307 |
> |
ForkJoinPool p; WorkQueue[] ws; int m; Submitter z; |
2308 |
> |
return ((z = submitters.get()) != null && |
2309 |
> |
(p = commonPool) != null && |
2310 |
> |
(ws = p.workQueues) != null && |
2311 |
> |
(m = ws.length - 1) >= 0) ? |
2312 |
> |
ws[m & z.seed & SQMASK] : null; |
2313 |
> |
} |
2314 |
> |
|
2315 |
> |
/** |
2316 |
> |
* Tries to pop the given task from submitter's queue in common pool. |
2317 |
> |
*/ |
2318 |
> |
static boolean tryExternalUnpush(ForkJoinTask<?> t) { |
2319 |
> |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z; |
2320 |
> |
ForkJoinTask<?>[] a; int m, s; |
2321 |
> |
if (t != null && |
2322 |
> |
(z = submitters.get()) != null && |
2323 |
> |
(p = commonPool) != null && |
2324 |
> |
(ws = p.workQueues) != null && |
2325 |
> |
(m = ws.length - 1) >= 0 && |
2326 |
> |
(q = ws[m & z.seed & SQMASK]) != null && |
2327 |
> |
(s = q.top) != q.base && |
2328 |
> |
(a = q.array) != null) { |
2329 |
> |
long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE; |
2330 |
> |
if (U.getObject(a, j) == t && |
2331 |
> |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2332 |
> |
if (q.array == a && q.top == s && // recheck |
2333 |
> |
U.compareAndSwapObject(a, j, t, null)) { |
2334 |
> |
q.top = s - 1; |
2335 |
> |
q.qlock = 0; |
2336 |
> |
return true; |
2337 |
> |
} |
2338 |
> |
q.qlock = 0; |
2339 |
> |
} |
2340 |
> |
} |
2341 |
> |
return false; |
2342 |
> |
} |
2343 |
> |
|
2344 |
> |
/** |
2345 |
> |
* Tries to pop and run local tasks within the same computation |
2346 |
> |
* as the given root. On failure, tries to help complete from |
2347 |
> |
* other queues via helpComplete. |
2348 |
> |
*/ |
2349 |
> |
private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) { |
2350 |
> |
ForkJoinTask<?>[] a; int m; |
2351 |
> |
if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 && |
2352 |
> |
root != null && root.status >= 0) { |
2353 |
> |
for (;;) { |
2354 |
> |
int s, u; Object o; CountedCompleter<?> task = null; |
2355 |
> |
if ((s = q.top) - q.base > 0) { |
2356 |
> |
long j = ((m & (s - 1)) << ASHIFT) + ABASE; |
2357 |
> |
if ((o = U.getObject(a, j)) != null && |
2358 |
> |
(o instanceof CountedCompleter)) { |
2359 |
> |
CountedCompleter<?> t = (CountedCompleter<?>)o, r = t; |
2360 |
> |
do { |
2361 |
> |
if (r == root) { |
2362 |
> |
if (U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2363 |
> |
if (q.array == a && q.top == s && |
2364 |
> |
U.compareAndSwapObject(a, j, t, null)) { |
2365 |
> |
q.top = s - 1; |
2366 |
> |
task = t; |
2367 |
> |
} |
2368 |
> |
q.qlock = 0; |
2369 |
|
} |
2370 |
+ |
break; |
2371 |
|
} |
2372 |
< |
} |
2372 |
> |
} while ((r = r.completer) != null); |
2373 |
|
} |
2374 |
|
} |
2375 |
< |
// Wake up workers parked on event queue |
2376 |
< |
int i, e; long c; Thread p; |
2377 |
< |
while ((i = ((~(e = (int)(c = ctl)) << 1) | 1) & SMASK) < n && |
2378 |
< |
(w = ws[i]) != null && |
2379 |
< |
w.eventCount == (e | INT_SIGN)) { |
2380 |
< |
long nc = ((long)(w.nextWait & E_MASK) | |
2381 |
< |
((c + AC_UNIT) & AC_MASK) | |
2382 |
< |
(c & (TC_MASK|STOP_BIT))); |
2383 |
< |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
2384 |
< |
w.eventCount = (e + E_SEQ) & E_MASK; |
2385 |
< |
if ((p = w.parker) != null) |
2386 |
< |
U.unpark(p); |
2375 |
> |
if (task != null) |
2376 |
> |
task.doExec(); |
2377 |
> |
if (root.status < 0 || |
2378 |
> |
(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0) |
2379 |
> |
break; |
2380 |
> |
if (task == null) { |
2381 |
> |
helpSignal(root, q.poolIndex, false); |
2382 |
> |
if (root.status >= 0) |
2383 |
> |
helpComplete(root, SHARED_QUEUE); |
2384 |
> |
break; |
2385 |
> |
} |
2386 |
> |
} |
2387 |
> |
} |
2388 |
> |
} |
2389 |
> |
|
2390 |
> |
/** |
2391 |
> |
* Tries to help execute or signal availability of the given task |
2392 |
> |
* from submitter's queue in common pool. |
2393 |
> |
*/ |
2394 |
> |
static void externalHelpJoin(ForkJoinTask<?> t) { |
2395 |
> |
// Some hard-to-avoid overlap with tryExternalUnpush |
2396 |
> |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z; |
2397 |
> |
ForkJoinTask<?>[] a; int m, s, n; |
2398 |
> |
if (t != null && |
2399 |
> |
(z = submitters.get()) != null && |
2400 |
> |
(p = commonPool) != null && |
2401 |
> |
(ws = p.workQueues) != null && |
2402 |
> |
(m = ws.length - 1) >= 0 && |
2403 |
> |
(q = ws[m & z.seed & SQMASK]) != null && |
2404 |
> |
(a = q.array) != null) { |
2405 |
> |
int am = a.length - 1; |
2406 |
> |
if ((s = q.top) != q.base) { |
2407 |
> |
long j = ((am & (s - 1)) << ASHIFT) + ABASE; |
2408 |
> |
if (U.getObject(a, j) == t && |
2409 |
> |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2410 |
> |
if (q.array == a && q.top == s && |
2411 |
> |
U.compareAndSwapObject(a, j, t, null)) { |
2412 |
> |
q.top = s - 1; |
2413 |
> |
q.qlock = 0; |
2414 |
> |
t.doExec(); |
2415 |
|
} |
2416 |
+ |
else |
2417 |
+ |
q.qlock = 0; |
2418 |
|
} |
2419 |
|
} |
2420 |
+ |
if (t.status >= 0) { |
2421 |
+ |
if (t instanceof CountedCompleter) |
2422 |
+ |
p.externalHelpComplete(q, t); |
2423 |
+ |
else |
2424 |
+ |
p.helpSignal(t, q.poolIndex, false); |
2425 |
+ |
} |
2426 |
|
} |
2427 |
|
} |
2428 |
|
|
2429 |
+ |
/** |
2430 |
+ |
* Restricted version of helpQuiescePool for external callers |
2431 |
+ |
*/ |
2432 |
+ |
static void externalHelpQuiescePool() { |
2433 |
+ |
ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b; |
2434 |
+ |
if ((p = commonPool) != null && |
2435 |
+ |
(q = p.findNonEmptyStealQueue(1)) != null && |
2436 |
+ |
(b = q.base) - q.top < 0 && |
2437 |
+ |
(t = q.pollAt(b)) != null) |
2438 |
+ |
t.doExec(); |
2439 |
+ |
} |
2440 |
+ |
|
2441 |
|
// Exported methods |
2442 |
|
|
2443 |
|
// Constructors |
2507 |
|
checkPermission(); |
2508 |
|
if (factory == null) |
2509 |
|
throw new NullPointerException(); |
2510 |
< |
if (parallelism <= 0 || parallelism > MAX_ID) |
2510 |
> |
if (parallelism <= 0 || parallelism > MAX_CAP) |
2511 |
|
throw new IllegalArgumentException(); |
1951 |
– |
this.parallelism = parallelism; |
2512 |
|
this.factory = factory; |
2513 |
|
this.ueh = handler; |
2514 |
< |
this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE; |
1955 |
< |
this.nextPoolIndex = 1; |
2514 |
> |
this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0); |
2515 |
|
long np = (long)(-parallelism); // offset ctl counts |
2516 |
|
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2517 |
< |
// initialize workQueues array with room for 2*parallelism if possible |
1959 |
< |
int n = parallelism << 1; |
1960 |
< |
if (n >= MAX_ID) |
1961 |
< |
n = MAX_ID; |
1962 |
< |
else { // See Hackers Delight, sec 3.2, where n < (1 << 16) |
1963 |
< |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; |
1964 |
< |
} |
1965 |
< |
this.workQueues = new WorkQueue[(n + 1) << 1]; |
1966 |
< |
ReentrantLock lck = this.lock = new ReentrantLock(); |
1967 |
< |
this.termination = lck.newCondition(); |
1968 |
< |
this.stealCount = new AtomicLong(); |
1969 |
< |
this.nextWorkerNumber = new AtomicInteger(); |
2517 |
> |
int pn = nextPoolId(); |
2518 |
|
StringBuilder sb = new StringBuilder("ForkJoinPool-"); |
2519 |
< |
sb.append(poolNumberGenerator.incrementAndGet()); |
2519 |
> |
sb.append(Integer.toString(pn)); |
2520 |
|
sb.append("-worker-"); |
2521 |
|
this.workerNamePrefix = sb.toString(); |
2522 |
< |
// Create initial submission queue |
2523 |
< |
WorkQueue sq = tryAddSharedQueue(0); |
2524 |
< |
if (sq != null) |
2525 |
< |
sq.growArray(false); |
2522 |
> |
} |
2523 |
> |
|
2524 |
> |
/** |
2525 |
> |
* Constructor for common pool, suitable only for static initialization. |
2526 |
> |
* Basically the same as above, but uses smallest possible initial footprint. |
2527 |
> |
*/ |
2528 |
> |
ForkJoinPool(int parallelism, long ctl, |
2529 |
> |
ForkJoinWorkerThreadFactory factory, |
2530 |
> |
Thread.UncaughtExceptionHandler handler) { |
2531 |
> |
this.config = parallelism; |
2532 |
> |
this.ctl = ctl; |
2533 |
> |
this.factory = factory; |
2534 |
> |
this.ueh = handler; |
2535 |
> |
this.workerNamePrefix = "ForkJoinPool.commonPool-worker-"; |
2536 |
> |
} |
2537 |
> |
|
2538 |
> |
/** |
2539 |
> |
* Returns the common pool instance. |
2540 |
> |
* |
2541 |
> |
* @return the common pool instance |
2542 |
> |
*/ |
2543 |
> |
public static ForkJoinPool commonPool() { |
2544 |
> |
// assert commonPool != null : "static init error"; |
2545 |
> |
return commonPool; |
2546 |
|
} |
2547 |
|
|
2548 |
|
// Execution methods |
2564 |
|
* scheduled for execution |
2565 |
|
*/ |
2566 |
|
public <T> T invoke(ForkJoinTask<T> task) { |
2567 |
< |
doSubmit(task); |
2567 |
> |
if (task == null) |
2568 |
> |
throw new NullPointerException(); |
2569 |
> |
externalPush(task); |
2570 |
|
return task.join(); |
2571 |
|
} |
2572 |
|
|
2579 |
|
* scheduled for execution |
2580 |
|
*/ |
2581 |
|
public void execute(ForkJoinTask<?> task) { |
2582 |
< |
doSubmit(task); |
2582 |
> |
if (task == null) |
2583 |
> |
throw new NullPointerException(); |
2584 |
> |
externalPush(task); |
2585 |
|
} |
2586 |
|
|
2587 |
|
// AbstractExecutorService methods |
2598 |
|
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2599 |
|
job = (ForkJoinTask<?>) task; |
2600 |
|
else |
2601 |
< |
job = ForkJoinTask.adapt(task, null); |
2602 |
< |
doSubmit(job); |
2601 |
> |
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2602 |
> |
externalPush(job); |
2603 |
|
} |
2604 |
|
|
2605 |
|
/** |
2612 |
|
* scheduled for execution |
2613 |
|
*/ |
2614 |
|
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2615 |
< |
doSubmit(task); |
2615 |
> |
if (task == null) |
2616 |
> |
throw new NullPointerException(); |
2617 |
> |
externalPush(task); |
2618 |
|
return task; |
2619 |
|
} |
2620 |
|
|
2624 |
|
* scheduled for execution |
2625 |
|
*/ |
2626 |
|
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2627 |
< |
if (task == null) |
2628 |
< |
throw new NullPointerException(); |
2055 |
< |
ForkJoinTask<T> job = ForkJoinTask.adapt(task); |
2056 |
< |
doSubmit(job); |
2627 |
> |
ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task); |
2628 |
> |
externalPush(job); |
2629 |
|
return job; |
2630 |
|
} |
2631 |
|
|
2635 |
|
* scheduled for execution |
2636 |
|
*/ |
2637 |
|
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2638 |
< |
if (task == null) |
2639 |
< |
throw new NullPointerException(); |
2068 |
< |
ForkJoinTask<T> job = ForkJoinTask.adapt(task, result); |
2069 |
< |
doSubmit(job); |
2638 |
> |
ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result); |
2639 |
> |
externalPush(job); |
2640 |
|
return job; |
2641 |
|
} |
2642 |
|
|
2652 |
|
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2653 |
|
job = (ForkJoinTask<?>) task; |
2654 |
|
else |
2655 |
< |
job = ForkJoinTask.adapt(task, null); |
2656 |
< |
doSubmit(job); |
2655 |
> |
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2656 |
> |
externalPush(job); |
2657 |
|
return job; |
2658 |
|
} |
2659 |
|
|
2662 |
|
* @throws RejectedExecutionException {@inheritDoc} |
2663 |
|
*/ |
2664 |
|
public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) { |
2665 |
< |
ArrayList<ForkJoinTask<T>> forkJoinTasks = |
2666 |
< |
new ArrayList<ForkJoinTask<T>>(tasks.size()); |
2667 |
< |
for (Callable<T> task : tasks) |
2668 |
< |
forkJoinTasks.add(ForkJoinTask.adapt(task)); |
2669 |
< |
invoke(new InvokeAll<T>(forkJoinTasks)); |
2670 |
< |
|
2665 |
> |
// In previous versions of this class, this method constructed |
2666 |
> |
// a task to run ForkJoinTask.invokeAll, but now external |
2667 |
> |
// invocation of multiple tasks is at least as efficient. |
2668 |
> |
List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size()); |
2669 |
> |
// Workaround needed because method wasn't declared with |
2670 |
> |
// wildcards in return type but should have been. |
2671 |
|
@SuppressWarnings({"unchecked", "rawtypes"}) |
2672 |
< |
List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks; |
2103 |
< |
return futures; |
2104 |
< |
} |
2672 |
> |
List<Future<T>> futures = (List<Future<T>>) (List) fs; |
2673 |
|
|
2674 |
< |
static final class InvokeAll<T> extends RecursiveAction { |
2675 |
< |
final ArrayList<ForkJoinTask<T>> tasks; |
2676 |
< |
InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; } |
2677 |
< |
public void compute() { |
2678 |
< |
try { invokeAll(tasks); } |
2679 |
< |
catch (Exception ignore) {} |
2674 |
> |
boolean done = false; |
2675 |
> |
try { |
2676 |
> |
for (Callable<T> t : tasks) { |
2677 |
> |
ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t); |
2678 |
> |
externalPush(f); |
2679 |
> |
fs.add(f); |
2680 |
> |
} |
2681 |
> |
for (ForkJoinTask<T> f : fs) |
2682 |
> |
f.quietlyJoin(); |
2683 |
> |
done = true; |
2684 |
> |
return futures; |
2685 |
> |
} finally { |
2686 |
> |
if (!done) |
2687 |
> |
for (ForkJoinTask<T> f : fs) |
2688 |
> |
f.cancel(false); |
2689 |
|
} |
2113 |
– |
private static final long serialVersionUID = -7914297376763021607L; |
2690 |
|
} |
2691 |
|
|
2692 |
|
/** |
2714 |
|
* @return the targeted parallelism level of this pool |
2715 |
|
*/ |
2716 |
|
public int getParallelism() { |
2717 |
< |
return parallelism; |
2717 |
> |
return config & SMASK; |
2718 |
> |
} |
2719 |
> |
|
2720 |
> |
/** |
2721 |
> |
* Returns the targeted parallelism level of the common pool. |
2722 |
> |
* |
2723 |
> |
* @return the targeted parallelism level of the common pool |
2724 |
> |
*/ |
2725 |
> |
public static int getCommonPoolParallelism() { |
2726 |
> |
return commonPoolParallelism; |
2727 |
|
} |
2728 |
|
|
2729 |
|
/** |
2735 |
|
* @return the number of worker threads |
2736 |
|
*/ |
2737 |
|
public int getPoolSize() { |
2738 |
< |
return parallelism + (short)(ctl >>> TC_SHIFT); |
2738 |
> |
return (config & SMASK) + (short)(ctl >>> TC_SHIFT); |
2739 |
|
} |
2740 |
|
|
2741 |
|
/** |
2745 |
|
* @return {@code true} if this pool uses async mode |
2746 |
|
*/ |
2747 |
|
public boolean getAsyncMode() { |
2748 |
< |
return localMode != 0; |
2748 |
> |
return (config >>> 16) == FIFO_QUEUE; |
2749 |
|
} |
2750 |
|
|
2751 |
|
/** |
2760 |
|
int rc = 0; |
2761 |
|
WorkQueue[] ws; WorkQueue w; |
2762 |
|
if ((ws = workQueues) != null) { |
2763 |
< |
int n = ws.length; |
2764 |
< |
for (int i = 1; i < n; i += 2) { |
2180 |
< |
Thread.State s; ForkJoinWorkerThread wt; |
2181 |
< |
if ((w = ws[i]) != null && (wt = w.owner) != null && |
2182 |
< |
w.eventCount >= 0 && |
2183 |
< |
(s = wt.getState()) != Thread.State.BLOCKED && |
2184 |
< |
s != Thread.State.WAITING && |
2185 |
< |
s != Thread.State.TIMED_WAITING) |
2763 |
> |
for (int i = 1; i < ws.length; i += 2) { |
2764 |
> |
if ((w = ws[i]) != null && w.isApparentlyUnblocked()) |
2765 |
|
++rc; |
2766 |
|
} |
2767 |
|
} |
2776 |
|
* @return the number of active threads |
2777 |
|
*/ |
2778 |
|
public int getActiveThreadCount() { |
2779 |
< |
int r = parallelism + (int)(ctl >> AC_SHIFT); |
2779 |
> |
int r = (config & SMASK) + (int)(ctl >> AC_SHIFT); |
2780 |
|
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
2781 |
|
} |
2782 |
|
|
2792 |
|
* @return {@code true} if all threads are currently idle |
2793 |
|
*/ |
2794 |
|
public boolean isQuiescent() { |
2795 |
< |
return (int)(ctl >> AC_SHIFT) + parallelism == 0; |
2795 |
> |
return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0; |
2796 |
|
} |
2797 |
|
|
2798 |
|
/** |
2807 |
|
* @return the number of steals |
2808 |
|
*/ |
2809 |
|
public long getStealCount() { |
2810 |
< |
long count = stealCount.get(); |
2810 |
> |
long count = stealCount; |
2811 |
|
WorkQueue[] ws; WorkQueue w; |
2812 |
|
if ((ws = workQueues) != null) { |
2813 |
< |
int n = ws.length; |
2235 |
< |
for (int i = 1; i < n; i += 2) { |
2813 |
> |
for (int i = 1; i < ws.length; i += 2) { |
2814 |
|
if ((w = ws[i]) != null) |
2815 |
< |
count += w.totalSteals; |
2815 |
> |
count += w.nsteals; |
2816 |
|
} |
2817 |
|
} |
2818 |
|
return count; |
2832 |
|
long count = 0; |
2833 |
|
WorkQueue[] ws; WorkQueue w; |
2834 |
|
if ((ws = workQueues) != null) { |
2835 |
< |
int n = ws.length; |
2258 |
< |
for (int i = 1; i < n; i += 2) { |
2835 |
> |
for (int i = 1; i < ws.length; i += 2) { |
2836 |
|
if ((w = ws[i]) != null) |
2837 |
|
count += w.queueSize(); |
2838 |
|
} |
2851 |
|
int count = 0; |
2852 |
|
WorkQueue[] ws; WorkQueue w; |
2853 |
|
if ((ws = workQueues) != null) { |
2854 |
< |
int n = ws.length; |
2278 |
< |
for (int i = 0; i < n; i += 2) { |
2854 |
> |
for (int i = 0; i < ws.length; i += 2) { |
2855 |
|
if ((w = ws[i]) != null) |
2856 |
|
count += w.queueSize(); |
2857 |
|
} |
2868 |
|
public boolean hasQueuedSubmissions() { |
2869 |
|
WorkQueue[] ws; WorkQueue w; |
2870 |
|
if ((ws = workQueues) != null) { |
2871 |
< |
int n = ws.length; |
2872 |
< |
for (int i = 0; i < n; i += 2) { |
2297 |
< |
if ((w = ws[i]) != null && w.queueSize() != 0) |
2871 |
> |
for (int i = 0; i < ws.length; i += 2) { |
2872 |
> |
if ((w = ws[i]) != null && !w.isEmpty()) |
2873 |
|
return true; |
2874 |
|
} |
2875 |
|
} |
2886 |
|
protected ForkJoinTask<?> pollSubmission() { |
2887 |
|
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2888 |
|
if ((ws = workQueues) != null) { |
2889 |
< |
int n = ws.length; |
2315 |
< |
for (int i = 0; i < n; i += 2) { |
2889 |
> |
for (int i = 0; i < ws.length; i += 2) { |
2890 |
|
if ((w = ws[i]) != null && (t = w.poll()) != null) |
2891 |
|
return t; |
2892 |
|
} |
2915 |
|
int count = 0; |
2916 |
|
WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t; |
2917 |
|
if ((ws = workQueues) != null) { |
2918 |
< |
int n = ws.length; |
2345 |
< |
for (int i = 0; i < n; ++i) { |
2918 |
> |
for (int i = 0; i < ws.length; ++i) { |
2919 |
|
if ((w = ws[i]) != null) { |
2920 |
|
while ((t = w.poll()) != null) { |
2921 |
|
c.add(t); |
2935 |
|
* @return a string identifying this pool, as well as its state |
2936 |
|
*/ |
2937 |
|
public String toString() { |
2938 |
< |
long st = getStealCount(); |
2939 |
< |
long qt = getQueuedTaskCount(); |
2940 |
< |
long qs = getQueuedSubmissionCount(); |
2368 |
< |
int rc = getRunningThreadCount(); |
2369 |
< |
int pc = parallelism; |
2938 |
> |
// Use a single pass through workQueues to collect counts |
2939 |
> |
long qt = 0L, qs = 0L; int rc = 0; |
2940 |
> |
long st = stealCount; |
2941 |
|
long c = ctl; |
2942 |
+ |
WorkQueue[] ws; WorkQueue w; |
2943 |
+ |
if ((ws = workQueues) != null) { |
2944 |
+ |
for (int i = 0; i < ws.length; ++i) { |
2945 |
+ |
if ((w = ws[i]) != null) { |
2946 |
+ |
int size = w.queueSize(); |
2947 |
+ |
if ((i & 1) == 0) |
2948 |
+ |
qs += size; |
2949 |
+ |
else { |
2950 |
+ |
qt += size; |
2951 |
+ |
st += w.nsteals; |
2952 |
+ |
if (w.isApparentlyUnblocked()) |
2953 |
+ |
++rc; |
2954 |
+ |
} |
2955 |
+ |
} |
2956 |
+ |
} |
2957 |
+ |
} |
2958 |
+ |
int pc = (config & SMASK); |
2959 |
|
int tc = pc + (short)(c >>> TC_SHIFT); |
2960 |
|
int ac = pc + (int)(c >> AC_SHIFT); |
2961 |
|
if (ac < 0) // ignore transient negative |
2964 |
|
if ((c & STOP_BIT) != 0) |
2965 |
|
level = (tc == 0) ? "Terminated" : "Terminating"; |
2966 |
|
else |
2967 |
< |
level = runState < 0 ? "Shutting down" : "Running"; |
2967 |
> |
level = plock < 0 ? "Shutting down" : "Running"; |
2968 |
|
return super.toString() + |
2969 |
|
"[" + level + |
2970 |
|
", parallelism = " + pc + |
2978 |
|
} |
2979 |
|
|
2980 |
|
/** |
2981 |
< |
* Initiates an orderly shutdown in which previously submitted |
2982 |
< |
* tasks are executed, but no new tasks will be accepted. |
2983 |
< |
* Invocation has no additional effect if already shut down. |
2984 |
< |
* Tasks that are in the process of being submitted concurrently |
2985 |
< |
* during the course of this method may or may not be rejected. |
2981 |
> |
* Possibly initiates an orderly shutdown in which previously |
2982 |
> |
* submitted tasks are executed, but no new tasks will be |
2983 |
> |
* accepted. Invocation has no effect on execution state if this |
2984 |
> |
* is the {@link #commonPool}, and no additional effect if |
2985 |
> |
* already shut down. Tasks that are in the process of being |
2986 |
> |
* submitted concurrently during the course of this method may or |
2987 |
> |
* may not be rejected. |
2988 |
|
* |
2989 |
|
* @throws SecurityException if a security manager exists and |
2990 |
|
* the caller is not permitted to modify threads |
2993 |
|
*/ |
2994 |
|
public void shutdown() { |
2995 |
|
checkPermission(); |
2996 |
< |
enableShutdown(); |
2407 |
< |
tryTerminate(false); |
2996 |
> |
tryTerminate(false, true); |
2997 |
|
} |
2998 |
|
|
2999 |
|
/** |
3000 |
< |
* Attempts to cancel and/or stop all tasks, and reject all |
3001 |
< |
* subsequently submitted tasks. Tasks that are in the process of |
3002 |
< |
* being submitted or executed concurrently during the course of |
3003 |
< |
* this method may or may not be rejected. This method cancels |
3004 |
< |
* both existing and unexecuted tasks, in order to permit |
3005 |
< |
* termination in the presence of task dependencies. So the method |
3006 |
< |
* always returns an empty list (unlike the case for some other |
3007 |
< |
* Executors). |
3000 |
> |
* Possibly attempts to cancel and/or stop all tasks, and reject |
3001 |
> |
* all subsequently submitted tasks. Invocation has no effect on |
3002 |
> |
* execution state if this is the {@link #commonPool}, and no |
3003 |
> |
* additional effect if already shut down. Otherwise, tasks that |
3004 |
> |
* are in the process of being submitted or executed concurrently |
3005 |
> |
* during the course of this method may or may not be |
3006 |
> |
* rejected. This method cancels both existing and unexecuted |
3007 |
> |
* tasks, in order to permit termination in the presence of task |
3008 |
> |
* dependencies. So the method always returns an empty list |
3009 |
> |
* (unlike the case for some other Executors). |
3010 |
|
* |
3011 |
|
* @return an empty list |
3012 |
|
* @throws SecurityException if a security manager exists and |
3016 |
|
*/ |
3017 |
|
public List<Runnable> shutdownNow() { |
3018 |
|
checkPermission(); |
3019 |
< |
enableShutdown(); |
2429 |
< |
tryTerminate(true); |
3019 |
> |
tryTerminate(true, true); |
3020 |
|
return Collections.emptyList(); |
3021 |
|
} |
3022 |
|
|
3028 |
|
public boolean isTerminated() { |
3029 |
|
long c = ctl; |
3030 |
|
return ((c & STOP_BIT) != 0L && |
3031 |
< |
(short)(c >>> TC_SHIFT) == -parallelism); |
3031 |
> |
(short)(c >>> TC_SHIFT) == -(config & SMASK)); |
3032 |
|
} |
3033 |
|
|
3034 |
|
/** |
3047 |
|
public boolean isTerminating() { |
3048 |
|
long c = ctl; |
3049 |
|
return ((c & STOP_BIT) != 0L && |
3050 |
< |
(short)(c >>> TC_SHIFT) != -parallelism); |
3050 |
> |
(short)(c >>> TC_SHIFT) != -(config & SMASK)); |
3051 |
|
} |
3052 |
|
|
3053 |
|
/** |
3056 |
|
* @return {@code true} if this pool has been shut down |
3057 |
|
*/ |
3058 |
|
public boolean isShutdown() { |
3059 |
< |
return runState < 0; |
3059 |
> |
return plock < 0; |
3060 |
|
} |
3061 |
|
|
3062 |
|
/** |
3063 |
< |
* Blocks until all tasks have completed execution after a shutdown |
3064 |
< |
* request, or the timeout occurs, or the current thread is |
3065 |
< |
* interrupted, whichever happens first. |
3063 |
> |
* Blocks until all tasks have completed execution after a |
3064 |
> |
* shutdown request, or the timeout occurs, or the current thread |
3065 |
> |
* is interrupted, whichever happens first. Note that the {@link |
3066 |
> |
* #commonPool()} never terminates until program shutdown so |
3067 |
> |
* this method will always time out. |
3068 |
|
* |
3069 |
|
* @param timeout the maximum time to wait |
3070 |
|
* @param unit the time unit of the timeout argument |
3075 |
|
public boolean awaitTermination(long timeout, TimeUnit unit) |
3076 |
|
throws InterruptedException { |
3077 |
|
long nanos = unit.toNanos(timeout); |
3078 |
< |
final ReentrantLock lock = this.lock; |
3079 |
< |
lock.lock(); |
3080 |
< |
try { |
3081 |
< |
for (;;) { |
3082 |
< |
if (isTerminated()) |
3083 |
< |
return true; |
3084 |
< |
if (nanos <= 0) |
3085 |
< |
return false; |
3086 |
< |
nanos = termination.awaitNanos(nanos); |
3078 |
> |
if (isTerminated()) |
3079 |
> |
return true; |
3080 |
> |
long startTime = System.nanoTime(); |
3081 |
> |
boolean terminated = false; |
3082 |
> |
synchronized (this) { |
3083 |
> |
for (long waitTime = nanos, millis = 0L;;) { |
3084 |
> |
if (terminated = isTerminated() || |
3085 |
> |
waitTime <= 0L || |
3086 |
> |
(millis = unit.toMillis(waitTime)) <= 0L) |
3087 |
> |
break; |
3088 |
> |
wait(millis); |
3089 |
> |
waitTime = nanos - (System.nanoTime() - startTime); |
3090 |
|
} |
2496 |
– |
} finally { |
2497 |
– |
lock.unlock(); |
3091 |
|
} |
3092 |
+ |
return terminated; |
3093 |
|
} |
3094 |
|
|
3095 |
|
/** |
3188 |
|
public static void managedBlock(ManagedBlocker blocker) |
3189 |
|
throws InterruptedException { |
3190 |
|
Thread t = Thread.currentThread(); |
3191 |
< |
ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ? |
3192 |
< |
((ForkJoinWorkerThread)t).pool : null); |
3193 |
< |
while (!blocker.isReleasable()) { |
3194 |
< |
if (p == null || p.tryCompensate()) { |
3195 |
< |
try { |
3196 |
< |
do {} while (!blocker.isReleasable() && !blocker.block()); |
3197 |
< |
} finally { |
3198 |
< |
if (p != null) |
3191 |
> |
if (t instanceof ForkJoinWorkerThread) { |
3192 |
> |
ForkJoinPool p = ((ForkJoinWorkerThread)t).pool; |
3193 |
> |
while (!blocker.isReleasable()) { // variant of helpSignal |
3194 |
> |
WorkQueue[] ws; WorkQueue q; int m, u; |
3195 |
> |
if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) { |
3196 |
> |
for (int i = 0; i <= m; ++i) { |
3197 |
> |
if (blocker.isReleasable()) |
3198 |
> |
return; |
3199 |
> |
if ((q = ws[i]) != null && q.base - q.top < 0) { |
3200 |
> |
p.signalWork(q); |
3201 |
> |
if ((u = (int)(p.ctl >>> 32)) >= 0 || |
3202 |
> |
(u >> UAC_SHIFT) >= 0) |
3203 |
> |
break; |
3204 |
> |
} |
3205 |
> |
} |
3206 |
> |
} |
3207 |
> |
if (p.tryCompensate()) { |
3208 |
> |
try { |
3209 |
> |
do {} while (!blocker.isReleasable() && |
3210 |
> |
!blocker.block()); |
3211 |
> |
} finally { |
3212 |
|
p.incrementActiveCount(); |
3213 |
+ |
} |
3214 |
+ |
break; |
3215 |
|
} |
2607 |
– |
break; |
3216 |
|
} |
3217 |
|
} |
3218 |
+ |
else { |
3219 |
+ |
do {} while (!blocker.isReleasable() && |
3220 |
+ |
!blocker.block()); |
3221 |
+ |
} |
3222 |
|
} |
3223 |
|
|
3224 |
|
// AbstractExecutorService overrides. These rely on undocumented |
3226 |
|
// implement RunnableFuture. |
3227 |
|
|
3228 |
|
protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) { |
3229 |
< |
return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value); |
3229 |
> |
return new ForkJoinTask.AdaptedRunnable<T>(runnable, value); |
3230 |
|
} |
3231 |
|
|
3232 |
|
protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) { |
3233 |
< |
return (RunnableFuture<T>) ForkJoinTask.adapt(callable); |
3233 |
> |
return new ForkJoinTask.AdaptedCallable<T>(callable); |
3234 |
|
} |
3235 |
|
|
3236 |
|
// Unsafe mechanics |
3237 |
|
private static final sun.misc.Unsafe U; |
3238 |
|
private static final long CTL; |
2627 |
– |
private static final long RUNSTATE; |
3239 |
|
private static final long PARKBLOCKER; |
3240 |
+ |
private static final int ABASE; |
3241 |
+ |
private static final int ASHIFT; |
3242 |
+ |
private static final long STEALCOUNT; |
3243 |
+ |
private static final long PLOCK; |
3244 |
+ |
private static final long INDEXSEED; |
3245 |
+ |
private static final long QLOCK; |
3246 |
|
|
3247 |
|
static { |
3248 |
< |
poolNumberGenerator = new AtomicInteger(); |
2632 |
< |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
2633 |
< |
defaultForkJoinWorkerThreadFactory = |
2634 |
< |
new DefaultForkJoinWorkerThreadFactory(); |
2635 |
< |
int s; |
3248 |
> |
int s; // initialize field offsets for CAS etc |
3249 |
|
try { |
3250 |
|
U = getUnsafe(); |
3251 |
|
Class<?> k = ForkJoinPool.class; |
2639 |
– |
Class<?> tk = Thread.class; |
3252 |
|
CTL = U.objectFieldOffset |
3253 |
|
(k.getDeclaredField("ctl")); |
3254 |
< |
RUNSTATE = U.objectFieldOffset |
3255 |
< |
(k.getDeclaredField("runState")); |
3254 |
> |
STEALCOUNT = U.objectFieldOffset |
3255 |
> |
(k.getDeclaredField("stealCount")); |
3256 |
> |
PLOCK = U.objectFieldOffset |
3257 |
> |
(k.getDeclaredField("plock")); |
3258 |
> |
INDEXSEED = U.objectFieldOffset |
3259 |
> |
(k.getDeclaredField("indexSeed")); |
3260 |
> |
Class<?> tk = Thread.class; |
3261 |
|
PARKBLOCKER = U.objectFieldOffset |
3262 |
|
(tk.getDeclaredField("parkBlocker")); |
3263 |
+ |
Class<?> wk = WorkQueue.class; |
3264 |
+ |
QLOCK = U.objectFieldOffset |
3265 |
+ |
(wk.getDeclaredField("qlock")); |
3266 |
+ |
Class<?> ak = ForkJoinTask[].class; |
3267 |
+ |
ABASE = U.arrayBaseOffset(ak); |
3268 |
+ |
s = U.arrayIndexScale(ak); |
3269 |
+ |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
3270 |
|
} catch (Exception e) { |
3271 |
|
throw new Error(e); |
3272 |
|
} |
3273 |
+ |
if ((s & (s-1)) != 0) |
3274 |
+ |
throw new Error("data type scale not a power of two"); |
3275 |
+ |
|
3276 |
+ |
submitters = new ThreadLocal<Submitter>(); |
3277 |
+ |
ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory = |
3278 |
+ |
new DefaultForkJoinWorkerThreadFactory(); |
3279 |
+ |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3280 |
+ |
|
3281 |
+ |
/* |
3282 |
+ |
* Establish common pool parameters. For extra caution, |
3283 |
+ |
* computations to set up common pool state are here; the |
3284 |
+ |
* constructor just assigns these values to fields. |
3285 |
+ |
*/ |
3286 |
+ |
|
3287 |
+ |
int par = 0; |
3288 |
+ |
Thread.UncaughtExceptionHandler handler = null; |
3289 |
+ |
try { // TBD: limit or report ignored exceptions? |
3290 |
+ |
String pp = System.getProperty |
3291 |
+ |
("java.util.concurrent.ForkJoinPool.common.parallelism"); |
3292 |
+ |
String hp = System.getProperty |
3293 |
+ |
("java.util.concurrent.ForkJoinPool.common.exceptionHandler"); |
3294 |
+ |
String fp = System.getProperty |
3295 |
+ |
("java.util.concurrent.ForkJoinPool.common.threadFactory"); |
3296 |
+ |
if (fp != null) |
3297 |
+ |
fac = ((ForkJoinWorkerThreadFactory)ClassLoader. |
3298 |
+ |
getSystemClassLoader().loadClass(fp).newInstance()); |
3299 |
+ |
if (hp != null) |
3300 |
+ |
handler = ((Thread.UncaughtExceptionHandler)ClassLoader. |
3301 |
+ |
getSystemClassLoader().loadClass(hp).newInstance()); |
3302 |
+ |
if (pp != null) |
3303 |
+ |
par = Integer.parseInt(pp); |
3304 |
+ |
} catch (Exception ignore) { |
3305 |
+ |
} |
3306 |
+ |
|
3307 |
+ |
if (par <= 0) |
3308 |
+ |
par = Runtime.getRuntime().availableProcessors(); |
3309 |
+ |
if (par > MAX_CAP) |
3310 |
+ |
par = MAX_CAP; |
3311 |
+ |
commonPoolParallelism = par; |
3312 |
+ |
long np = (long)(-par); // precompute initial ctl value |
3313 |
+ |
long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
3314 |
+ |
|
3315 |
+ |
commonPool = new ForkJoinPool(par, ct, fac, handler); |
3316 |
|
} |
3317 |
|
|
3318 |
|
/** |