5 |
|
*/ |
6 |
|
|
7 |
|
package jsr166e; |
8 |
+ |
|
9 |
|
import java.util.ArrayList; |
10 |
|
import java.util.Arrays; |
11 |
|
import java.util.Collection; |
12 |
|
import java.util.Collections; |
13 |
|
import java.util.List; |
13 |
– |
import java.util.Random; |
14 |
|
import java.util.concurrent.AbstractExecutorService; |
15 |
|
import java.util.concurrent.Callable; |
16 |
|
import java.util.concurrent.ExecutorService; |
17 |
|
import java.util.concurrent.Future; |
18 |
|
import java.util.concurrent.RejectedExecutionException; |
19 |
|
import java.util.concurrent.RunnableFuture; |
20 |
+ |
import java.util.concurrent.ThreadLocalRandom; |
21 |
|
import java.util.concurrent.TimeUnit; |
21 |
– |
import java.util.concurrent.atomic.AtomicInteger; |
22 |
– |
import java.util.concurrent.atomic.AtomicLong; |
23 |
– |
import java.util.concurrent.locks.AbstractQueuedSynchronizer; |
24 |
– |
import java.util.concurrent.locks.Condition; |
22 |
|
|
23 |
|
/** |
24 |
|
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
38 |
|
* ForkJoinPool}s may also be appropriate for use with event-style |
39 |
|
* tasks that are never joined. |
40 |
|
* |
41 |
< |
* <p>A {@code ForkJoinPool} is constructed with a given target |
42 |
< |
* parallelism level; by default, equal to the number of available |
43 |
< |
* processors. The pool attempts to maintain enough active (or |
44 |
< |
* available) threads by dynamically adding, suspending, or resuming |
45 |
< |
* internal worker threads, even if some tasks are stalled waiting to |
46 |
< |
* join others. However, no such adjustments are guaranteed in the |
47 |
< |
* face of blocked IO or other unmanaged synchronization. The nested |
48 |
< |
* {@link ManagedBlocker} interface enables extension of the kinds of |
41 |
> |
* <p>A static {@link #commonPool} is available and appropriate for |
42 |
> |
* most applications. The common pool is used by any ForkJoinTask that |
43 |
> |
* is not explicitly submitted to a specified pool. Using the common |
44 |
> |
* pool normally reduces resource usage (its threads are slowly |
45 |
> |
* reclaimed during periods of non-use, and reinstated upon subsequent |
46 |
> |
* use). |
47 |
> |
* |
48 |
> |
* <p>For applications that require separate or custom pools, a {@code |
49 |
> |
* ForkJoinPool} may be constructed with a given target parallelism |
50 |
> |
* level; by default, equal to the number of available processors. The |
51 |
> |
* pool attempts to maintain enough active (or available) threads by |
52 |
> |
* dynamically adding, suspending, or resuming internal worker |
53 |
> |
* threads, even if some tasks are stalled waiting to join |
54 |
> |
* others. However, no such adjustments are guaranteed in the face of |
55 |
> |
* blocked IO or other unmanaged synchronization. The nested {@link |
56 |
> |
* ManagedBlocker} interface enables extension of the kinds of |
57 |
|
* synchronization accommodated. |
58 |
|
* |
59 |
|
* <p>In addition to execution and lifecycle control methods, this |
98 |
|
* </tr> |
99 |
|
* </table> |
100 |
|
* |
101 |
< |
* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is |
102 |
< |
* used for all parallel task execution in a program or subsystem. |
103 |
< |
* Otherwise, use would not usually outweigh the construction and |
104 |
< |
* bookkeeping overhead of creating a large set of threads. For |
105 |
< |
* example, a common pool could be used for the {@code SortTasks} |
106 |
< |
* illustrated in {@link RecursiveAction}. Because {@code |
107 |
< |
* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon |
108 |
< |
* daemon} mode, there is typically no need to explicitly {@link |
109 |
< |
* #shutdown} such a pool upon program exit. |
105 |
< |
* |
106 |
< |
* <pre> {@code |
107 |
< |
* static final ForkJoinPool mainPool = new ForkJoinPool(); |
108 |
< |
* ... |
109 |
< |
* public void sort(long[] array) { |
110 |
< |
* mainPool.invoke(new SortTask(array, 0, array.length)); |
111 |
< |
* }}</pre> |
101 |
> |
* <p>The common pool is by default constructed with default |
102 |
> |
* parameters, but these may be controlled by setting three {@link |
103 |
> |
* System#getProperty properties} with prefix {@code |
104 |
> |
* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} -- |
105 |
> |
* an integer greater than zero, {@code threadFactory} -- the class |
106 |
> |
* name of a {@link ForkJoinWorkerThreadFactory}, and {@code |
107 |
> |
* exceptionHandler} -- the class name of a {@link |
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 |
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. Insertion of tasks in shared mode requires a |
207 |
< |
* lock (mainly to protect in the case of resizing) but we use |
208 |
< |
* only a simple spinlock (using bits in field runState), because |
209 |
< |
* submitters encountering a busy queue move on to try or create |
210 |
< |
* other queues -- they block only when creating and registering |
211 |
< |
* new queues. |
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 |
217 |
|
* ========== |
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 |
> |
* blocking 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. |
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 |
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 is |
320 |
> |
* apparently empty, they signal waiting workers (or trigger |
321 |
> |
* creation of new ones if fewer than the given parallelism level |
322 |
> |
* -- see signalWork). These primary signals are buttressed by |
323 |
> |
* signals whenever other threads scan for work or do not have a |
324 |
> |
* task to process. On most platforms, signalling (unpark) |
325 |
> |
* overhead time is noticeably long, and the time between |
326 |
> |
* signalling a thread and it actually making progress can be very |
327 |
> |
* noticeably long, so it is worth offloading these delays from |
328 |
> |
* critical paths as much as possible. |
329 |
|
* |
330 |
|
* Trimming workers. To release resources after periods of lack of |
331 |
|
* use, a worker starting to wait when the pool is quiescent will |
332 |
< |
* time out and terminate if the pool has remained quiescent for |
333 |
< |
* SHRINK_RATE nanosecs. This will slowly propagate, eventually |
334 |
< |
* terminating all workers after long periods of non-use. |
332 |
> |
* time out and terminate if the pool has remained quiescent for a |
333 |
> |
* given period -- a short period if there are more threads than |
334 |
> |
* parallelism, longer as the number of threads decreases. This |
335 |
> |
* will slowly propagate, eventually terminating all workers after |
336 |
> |
* periods of non-use. |
337 |
|
* |
338 |
|
* Shutdown and Termination. A call to shutdownNow atomically sets |
339 |
< |
* a runState bit and then (non-atomically) sets each worker's |
340 |
< |
* runState status, cancels all unprocessed tasks, and wakes up |
339 |
> |
* a plock bit and then (non-atomically) sets each worker's |
340 |
> |
* qlock status, cancels all unprocessed tasks, and wakes up |
341 |
|
* all waiting workers. Detecting whether termination should |
342 |
|
* commence after a non-abrupt shutdown() call requires more work |
343 |
|
* and bookkeeping. We need consensus about quiescence (i.e., that |
365 |
|
* method tryCompensate() may create or re-activate a spare |
366 |
|
* thread to compensate for blocked joiners until they unblock. |
367 |
|
* |
368 |
< |
* A third form (implemented in tryRemoveAndExec and |
369 |
< |
* tryPollForAndExec) amounts to helping a hypothetical |
370 |
< |
* compensator: If we can readily tell that a possible action of a |
371 |
< |
* compensator is to steal and execute the task being joined, the |
372 |
< |
* joining thread can do so directly, without the need for a |
373 |
< |
* compensation thread (although at the expense of larger run-time |
374 |
< |
* stacks, but the tradeoff is typically worthwhile). |
368 |
> |
* A third form (implemented in tryRemoveAndExec) amounts to |
369 |
> |
* helping a hypothetical compensator: If we can readily tell that |
370 |
> |
* a possible action of a compensator is to steal and execute the |
371 |
> |
* task being joined, the joining thread can do so directly, |
372 |
> |
* without the need for a compensation thread (although at the |
373 |
> |
* expense of larger run-time stacks, but the tradeoff is |
374 |
> |
* typically worthwhile). |
375 |
|
* |
376 |
|
* The ManagedBlocker extension API can't use helping so relies |
377 |
|
* only on compensation in method awaitBlocker. |
406 |
|
* to find work (see MAX_HELP) and fall back to suspending the |
407 |
|
* worker and if necessary replacing it with another. |
408 |
|
* |
409 |
+ |
* Helping actions for CountedCompleters are much simpler: Method |
410 |
+ |
* helpComplete can take and execute any task with the same root |
411 |
+ |
* as the task being waited on. However, this still entails some |
412 |
+ |
* traversal of completer chains, so is less efficient than using |
413 |
+ |
* CountedCompleters without explicit joins. |
414 |
+ |
* |
415 |
|
* It is impossible to keep exactly the target parallelism number |
416 |
|
* of threads running at any given time. Determining the |
417 |
|
* existence of conservatively safe helping targets, the |
433 |
|
* intractable) game with an opponent that may choose the worst |
434 |
|
* (for us) active thread to stall at any time. We take several |
435 |
|
* precautions to bound losses (and thus bound gains), mainly in |
436 |
< |
* methods tryCompensate and awaitJoin: (1) We only try |
437 |
< |
* compensation after attempting enough helping steps (measured |
438 |
< |
* via counting and timing) that we have already consumed the |
439 |
< |
* estimated cost of creating and activating a new thread. (2) We |
440 |
< |
* allow up to 50% of threads to be blocked before initially |
441 |
< |
* adding any others, and unless completely saturated, check that |
442 |
< |
* some work is available for a new worker before adding. Also, we |
443 |
< |
* create up to only 50% more threads until entering a mode that |
444 |
< |
* only adds a thread if all others are possibly blocked. All |
445 |
< |
* together, this means that we might be half as fast to react, |
446 |
< |
* and create half as many threads as possible in the ideal case, |
447 |
< |
* but present vastly fewer anomalies in all other cases compared |
448 |
< |
* to both more aggressive and more conservative alternatives. |
449 |
< |
* |
450 |
< |
* Style notes: There is a lot of representation-level coupling |
451 |
< |
* among classes ForkJoinPool, ForkJoinWorkerThread, and |
452 |
< |
* ForkJoinTask. The fields of WorkQueue maintain data structures |
453 |
< |
* managed by ForkJoinPool, so are directly accessed. There is |
454 |
< |
* little point trying to reduce this, since any associated future |
455 |
< |
* changes in representations will need to be accompanied by |
456 |
< |
* algorithmic changes anyway. Several methods intrinsically |
457 |
< |
* sprawl because they must accumulate sets of consistent reads of |
458 |
< |
* volatiles held in local variables. Methods signalWork() and |
459 |
< |
* scan() are the main bottlenecks, so are especially heavily |
436 |
> |
* methods tryCompensate and awaitJoin. |
437 |
> |
* |
438 |
> |
* Common Pool |
439 |
> |
* =========== |
440 |
> |
* |
441 |
> |
* The static commonPool always exists after static |
442 |
> |
* initialization. Since it (or any other created pool) need |
443 |
> |
* never be used, we minimize initial construction overhead and |
444 |
> |
* footprint to the setup of about a dozen fields, with no nested |
445 |
> |
* allocation. Most bootstrapping occurs within method |
446 |
> |
* fullExternalPush during the first submission to the pool. |
447 |
> |
* |
448 |
> |
* When external threads submit to the common pool, they can |
449 |
> |
* perform some subtask processing (see externalHelpJoin and |
450 |
> |
* related methods). We do not need to record whether these |
451 |
> |
* submissions are to the common pool -- if not, externalHelpJoin |
452 |
> |
* returns quicky (at the most helping to signal some common pool |
453 |
> |
* workers). These submitters would otherwise be blocked waiting |
454 |
> |
* for completion, so the extra effort (with liberally sprinkled |
455 |
> |
* task status checks) in inapplicable cases amounts to an odd |
456 |
> |
* form of limited spin-wait before blocking in ForkJoinTask.join. |
457 |
> |
* |
458 |
> |
* Style notes |
459 |
> |
* =========== |
460 |
> |
* |
461 |
> |
* There is a lot of representation-level coupling among classes |
462 |
> |
* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The |
463 |
> |
* fields of WorkQueue maintain data structures managed by |
464 |
> |
* ForkJoinPool, so are directly accessed. There is little point |
465 |
> |
* trying to reduce this, since any associated future changes in |
466 |
> |
* representations will need to be accompanied by algorithmic |
467 |
> |
* changes anyway. Several methods intrinsically sprawl because |
468 |
> |
* they must accumulate sets of consistent reads of volatiles held |
469 |
> |
* in local variables. Methods signalWork() and scan() are the |
470 |
> |
* main bottlenecks, so are especially heavily |
471 |
|
* micro-optimized/mangled. There are lots of inline assignments |
472 |
|
* (of form "while ((local = field) != 0)") which are usually the |
473 |
|
* simplest way to ensure the required read orderings (which are |
475 |
|
* declarations of these locals at the heads of methods or blocks. |
476 |
|
* There are several occurrences of the unusual "do {} while |
477 |
|
* (!cas...)" which is the simplest way to force an update of a |
478 |
< |
* CAS'ed variable. There are also other coding oddities that help |
478 |
> |
* CAS'ed variable. There are also other coding oddities (including |
479 |
> |
* several unnecessary-looking hoisted null checks) that help |
480 |
|
* some methods perform reasonably even when interpreted (not |
481 |
|
* compiled). |
482 |
|
* |
533 |
|
} |
534 |
|
|
535 |
|
/** |
507 |
– |
* A simple non-reentrant lock used for exclusion when managing |
508 |
– |
* queues and workers. We use a custom lock so that we can readily |
509 |
– |
* probe lock state in constructions that check among alternative |
510 |
– |
* actions. The lock is normally only very briefly held, and |
511 |
– |
* sometimes treated as a spinlock, but other usages block to |
512 |
– |
* reduce overall contention in those cases where locked code |
513 |
– |
* bodies perform allocation/resizing. |
514 |
– |
*/ |
515 |
– |
static final class Mutex extends AbstractQueuedSynchronizer { |
516 |
– |
public final boolean tryAcquire(int ignore) { |
517 |
– |
return compareAndSetState(0, 1); |
518 |
– |
} |
519 |
– |
public final boolean tryRelease(int ignore) { |
520 |
– |
setState(0); |
521 |
– |
return true; |
522 |
– |
} |
523 |
– |
public final void lock() { acquire(0); } |
524 |
– |
public final void unlock() { release(0); } |
525 |
– |
public final boolean isHeldExclusively() { return getState() == 1; } |
526 |
– |
public final Condition newCondition() { return new ConditionObject(); } |
527 |
– |
} |
528 |
– |
|
529 |
– |
/** |
536 |
|
* Class for artificial tasks that are used to replace the target |
537 |
|
* of local joins if they are removed from an interior queue slot |
538 |
|
* in WorkQueue.tryRemoveAndExec. We don't need the proxy to |
539 |
|
* actually do anything beyond having a unique identity. |
540 |
|
*/ |
541 |
|
static final class EmptyTask extends ForkJoinTask<Void> { |
542 |
+ |
private static final long serialVersionUID = -7721805057305804111L; |
543 |
|
EmptyTask() { status = ForkJoinTask.NORMAL; } // force done |
544 |
|
public final Void getRawResult() { return null; } |
545 |
|
public final void setRawResult(Void x) {} |
560 |
|
* |
561 |
|
* Field "top" is the index (mod array.length) of the next queue |
562 |
|
* slot to push to or pop from. It is written only by owner thread |
563 |
< |
* for push, or under lock for trySharedPush, and accessed by |
564 |
< |
* other threads only after reading (volatile) base. Both top and |
565 |
< |
* base are allowed to wrap around on overflow, but (top - base) |
566 |
< |
* (or more commonly -(base - top) to force volatile read of base |
567 |
< |
* before top) still estimates size. |
563 |
> |
* for push, or under lock for external/shared push, and accessed |
564 |
> |
* by other threads only after reading (volatile) base. Both top |
565 |
> |
* and base are allowed to wrap around on overflow, but (top - |
566 |
> |
* base) (or more commonly -(base - top) to force volatile read of |
567 |
> |
* base before top) still estimates size. The lock ("qlock") is |
568 |
> |
* forced to -1 on termination, causing all further lock attempts |
569 |
> |
* to fail. (Note: we don't need CAS for termination state because |
570 |
> |
* upon pool shutdown, all shared-queues will stop being used |
571 |
> |
* anyway.) Nearly all lock bodies are set up so that exceptions |
572 |
> |
* within lock bodies are "impossible" (modulo JVM errors that |
573 |
> |
* would cause failure anyway.) |
574 |
|
* |
575 |
|
* The array slots are read and written using the emulation of |
576 |
|
* volatiles/atomics provided by Unsafe. Insertions must in |
577 |
|
* general use putOrderedObject as a form of releasing store to |
578 |
|
* ensure that all writes to the task object are ordered before |
579 |
< |
* its publication in the queue. (Although we can avoid one case |
580 |
< |
* of this when locked in trySharedPush.) All removals entail a |
581 |
< |
* CAS to null. The array is always a power of two. To ensure |
582 |
< |
* safety of Unsafe array operations, all accesses perform |
570 |
< |
* explicit null checks and implicit bounds checks via |
571 |
< |
* power-of-two masking. |
579 |
> |
* its publication in the queue. All removals entail a CAS to |
580 |
> |
* null. The array is always a power of two. To ensure safety of |
581 |
> |
* Unsafe array operations, all accesses perform explicit null |
582 |
> |
* checks and implicit bounds checks via power-of-two masking. |
583 |
|
* |
584 |
|
* In addition to basic queuing support, this class contains |
585 |
|
* fields described elsewhere to control execution. It turns out |
586 |
< |
* to work better memory-layout-wise to include them in this |
587 |
< |
* class rather than a separate class. |
586 |
> |
* to work better memory-layout-wise to include them in this class |
587 |
> |
* rather than a separate class. |
588 |
|
* |
589 |
|
* Performance on most platforms is very sensitive to placement of |
590 |
|
* instances of both WorkQueues and their arrays -- we absolutely |
600 |
|
* support is in place, this padding is dependent on transient |
601 |
|
* properties of JVM field layout rules.) We also take care in |
602 |
|
* allocating, sizing and resizing the array. Non-shared queue |
603 |
< |
* arrays are initialized (via method growArray) by workers before |
604 |
< |
* use. Others are allocated on first use. |
603 |
> |
* arrays are initialized by workers before use. Others are |
604 |
> |
* allocated on first use. |
605 |
|
*/ |
606 |
|
static final class WorkQueue { |
607 |
|
/** |
624 |
|
*/ |
625 |
|
static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M |
626 |
|
|
616 |
– |
volatile long totalSteals; // cumulative number of steals |
627 |
|
int seed; // for random scanning; initialize nonzero |
628 |
|
volatile int eventCount; // encoded inactivation count; < 0 if inactive |
629 |
|
int nextWait; // encoded record of next event waiter |
620 |
– |
int rescans; // remaining scans until block |
621 |
– |
int nsteals; // top-level task executions since last idle |
630 |
|
final int mode; // lifo, fifo, or shared |
631 |
+ |
int nsteals; // cumulative number of steals |
632 |
|
int poolIndex; // index of this queue in pool (or 0) |
633 |
|
int stealHint; // index of most recent known stealer |
634 |
< |
volatile int runState; // 1: locked, -1: terminate; else 0 |
634 |
> |
volatile int qlock; // 1: locked, -1: terminate; else 0 |
635 |
|
volatile int base; // index of next slot for poll |
636 |
|
int top; // index of next slot for push |
637 |
|
ForkJoinTask<?>[] array; // the elements (initially unallocated) |
653 |
|
} |
654 |
|
|
655 |
|
/** |
647 |
– |
* Returns the approximate number of tasks in the queue. |
648 |
– |
*/ |
649 |
– |
final int queueSize() { |
650 |
– |
int n = base - top; // non-owner callers must read base first |
651 |
– |
return (n >= 0) ? 0 : -n; // ignore transient negative |
652 |
– |
} |
653 |
– |
|
654 |
– |
/** |
655 |
– |
* Provides a more accurate estimate of whether this queue has |
656 |
– |
* any tasks than does queueSize, by checking whether a |
657 |
– |
* near-empty queue has at least one unclaimed task. |
658 |
– |
*/ |
659 |
– |
final boolean isEmpty() { |
660 |
– |
ForkJoinTask<?>[] a; int m, s; |
661 |
– |
int n = base - (s = top); |
662 |
– |
return (n >= 0 || |
663 |
– |
(n == -1 && |
664 |
– |
((a = array) == null || |
665 |
– |
(m = a.length - 1) < 0 || |
666 |
– |
U.getObjectVolatile |
667 |
– |
(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null))); |
668 |
– |
} |
669 |
– |
|
670 |
– |
/** |
656 |
|
* Pushes a task. Call only by owner in unshared queues. |
657 |
+ |
* Cases needing resizing or rejection are relyaed to fullPush |
658 |
+ |
* (that also handles shared queues). |
659 |
|
* |
660 |
|
* @param task the task. Caller must ensure non-null. |
661 |
|
* @throw RejectedExecutionException if array cannot be resized |
662 |
|
*/ |
663 |
|
final void push(ForkJoinTask<?> task) { |
664 |
< |
ForkJoinTask<?>[] a; ForkJoinPool p; |
665 |
< |
int s = top, m, n; |
666 |
< |
if ((a = array) != null) { // ignore if queue removed |
664 |
> |
ForkJoinPool p; ForkJoinTask<?>[] a; |
665 |
> |
int s = top, n; |
666 |
> |
if ((a = array) != null && a.length > (n = s + 1 - base)) { |
667 |
|
U.putOrderedObject |
668 |
< |
(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task); |
669 |
< |
if ((n = (top = s + 1) - base) <= 2) { |
670 |
< |
if ((p = pool) != null) |
671 |
< |
p.signalWork(); |
685 |
< |
} |
686 |
< |
else if (n >= m) |
687 |
< |
growArray(true); |
668 |
> |
(a, (((a.length - 1) & s) << ASHIFT) + ABASE, task); |
669 |
> |
top = s + 1; |
670 |
> |
if (n <= 1 && (p = pool) != null) |
671 |
> |
p.signalWork(this, 1); |
672 |
|
} |
673 |
+ |
else |
674 |
+ |
fullPush(task, true); |
675 |
|
} |
676 |
|
|
677 |
|
/** |
678 |
|
* Pushes a task if lock is free and array is either big |
679 |
< |
* enough or can be resized to be big enough. |
679 |
> |
* enough or can be resized to be big enough. Note: a |
680 |
> |
* specialization of a common fast path of this method is in |
681 |
> |
* ForkJoinPool.externalPush. When called from a FJWT queue, |
682 |
> |
* this can fail only if the pool has been shut down or |
683 |
> |
* an out of memory error. |
684 |
|
* |
685 |
|
* @param task the task. Caller must ensure non-null. |
686 |
< |
* @return true if submitted |
686 |
> |
* @param owned if true, throw RJE on failure |
687 |
|
*/ |
688 |
< |
final boolean trySharedPush(ForkJoinTask<?> task) { |
689 |
< |
boolean submitted = false; |
690 |
< |
if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { |
691 |
< |
ForkJoinTask<?>[] a = array; |
692 |
< |
int s = top; |
693 |
< |
try { |
694 |
< |
if ((a != null && a.length > s + 1 - base) || |
695 |
< |
(a = growArray(false)) != null) { // must presize |
696 |
< |
int j = (((a.length - 1) & s) << ASHIFT) + ABASE; |
697 |
< |
U.putObject(a, (long)j, task); // don't need "ordered" |
698 |
< |
top = s + 1; |
699 |
< |
submitted = true; |
688 |
> |
final boolean fullPush(ForkJoinTask<?> task, boolean owned) { |
689 |
> |
ForkJoinPool p; ForkJoinTask<?>[] a; |
690 |
> |
if (owned) { |
691 |
> |
if (qlock < 0) // must be shutting down |
692 |
> |
throw new RejectedExecutionException(); |
693 |
> |
} |
694 |
> |
else if (!U.compareAndSwapInt(this, QLOCK, 0, 1)) |
695 |
> |
return false; |
696 |
> |
try { |
697 |
> |
int s = top, oldLen, len; |
698 |
> |
if ((a = array) == null) |
699 |
> |
a = array = new ForkJoinTask<?>[len=INITIAL_QUEUE_CAPACITY]; |
700 |
> |
else if ((oldLen = a.length) > s + 1 - base) |
701 |
> |
len = oldLen; |
702 |
> |
else if ((len = oldLen << 1) > MAXIMUM_QUEUE_CAPACITY) |
703 |
> |
throw new RejectedExecutionException("Capacity exceeded"); |
704 |
> |
else { |
705 |
> |
int oldMask, b; |
706 |
> |
ForkJoinTask<?>[] oldA = a; |
707 |
> |
a = array = new ForkJoinTask<?>[len]; |
708 |
> |
if ((oldMask = oldLen - 1) >= 0 && s - (b = base) > 0) { |
709 |
> |
int mask = len - 1; |
710 |
> |
do { |
711 |
> |
ForkJoinTask<?> x; |
712 |
> |
int oldj = ((b & oldMask) << ASHIFT) + ABASE; |
713 |
> |
int j = ((b & mask) << ASHIFT) + ABASE; |
714 |
> |
x = (ForkJoinTask<?>) |
715 |
> |
U.getObjectVolatile(oldA, oldj); |
716 |
> |
if (x != null && |
717 |
> |
U.compareAndSwapObject(oldA, oldj, x, null)) |
718 |
> |
U.putObjectVolatile(a, j, x); |
719 |
> |
} while (++b != s); |
720 |
|
} |
711 |
– |
} finally { |
712 |
– |
runState = 0; // unlock |
721 |
|
} |
722 |
+ |
U.putOrderedObject |
723 |
+ |
(a, (((len - 1) & s) << ASHIFT) + ABASE, task); |
724 |
+ |
top = s + 1; |
725 |
+ |
} finally { |
726 |
+ |
if (!owned) |
727 |
+ |
qlock = 0; |
728 |
|
} |
729 |
< |
return submitted; |
729 |
> |
if ((p = pool) != null) |
730 |
> |
p.signalWork(this, 1); |
731 |
> |
return true; |
732 |
|
} |
733 |
|
|
734 |
|
/** |
735 |
|
* Takes next task, if one exists, in LIFO order. Call only |
736 |
< |
* by owner in unshared queues. (We do not have a shared |
721 |
< |
* version of this method because it is never needed.) |
736 |
> |
* by owner in unshared queues. |
737 |
|
*/ |
738 |
|
final ForkJoinTask<?> pop() { |
739 |
|
ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m; |
788 |
|
else if (base == b) { |
789 |
|
if (b + 1 == top) |
790 |
|
break; |
791 |
< |
Thread.yield(); // wait for lagging update |
791 |
> |
Thread.yield(); // wait for lagging update (very rare) |
792 |
|
} |
793 |
|
} |
794 |
|
return null; |
815 |
|
|
816 |
|
/** |
817 |
|
* Pops the given task only if it is at the current top. |
818 |
+ |
* (A shared version is available only via FJP.tryExternalUnpush) |
819 |
|
*/ |
820 |
|
final boolean tryUnpush(ForkJoinTask<?> t) { |
821 |
|
ForkJoinTask<?>[] a; int s; |
829 |
|
} |
830 |
|
|
831 |
|
/** |
816 |
– |
* Polls the given task only if it is at the current base. |
817 |
– |
*/ |
818 |
– |
final boolean pollFor(ForkJoinTask<?> task) { |
819 |
– |
ForkJoinTask<?>[] a; int b; |
820 |
– |
if ((b = base) - top < 0 && (a = array) != null) { |
821 |
– |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
822 |
– |
if (U.getObjectVolatile(a, j) == task && base == b && |
823 |
– |
U.compareAndSwapObject(a, j, task, null)) { |
824 |
– |
base = b + 1; |
825 |
– |
return true; |
826 |
– |
} |
827 |
– |
} |
828 |
– |
return false; |
829 |
– |
} |
830 |
– |
|
831 |
– |
/** |
832 |
– |
* Initializes or doubles the capacity of array. Call either |
833 |
– |
* by owner or with lock held -- it is OK for base, but not |
834 |
– |
* top, to move while resizings are in progress. |
835 |
– |
* |
836 |
– |
* @param rejectOnFailure if true, throw exception if capacity |
837 |
– |
* exceeded (relayed ultimately to user); else return null. |
838 |
– |
*/ |
839 |
– |
final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) { |
840 |
– |
ForkJoinTask<?>[] oldA = array; |
841 |
– |
int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; |
842 |
– |
if (size <= MAXIMUM_QUEUE_CAPACITY) { |
843 |
– |
int oldMask, t, b; |
844 |
– |
ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size]; |
845 |
– |
if (oldA != null && (oldMask = oldA.length - 1) >= 0 && |
846 |
– |
(t = top) - (b = base) > 0) { |
847 |
– |
int mask = size - 1; |
848 |
– |
do { |
849 |
– |
ForkJoinTask<?> x; |
850 |
– |
int oldj = ((b & oldMask) << ASHIFT) + ABASE; |
851 |
– |
int j = ((b & mask) << ASHIFT) + ABASE; |
852 |
– |
x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj); |
853 |
– |
if (x != null && |
854 |
– |
U.compareAndSwapObject(oldA, oldj, x, null)) |
855 |
– |
U.putObjectVolatile(a, j, x); |
856 |
– |
} while (++b != t); |
857 |
– |
} |
858 |
– |
return a; |
859 |
– |
} |
860 |
– |
else if (!rejectOnFailure) |
861 |
– |
return null; |
862 |
– |
else |
863 |
– |
throw new RejectedExecutionException("Queue capacity exceeded"); |
864 |
– |
} |
865 |
– |
|
866 |
– |
/** |
832 |
|
* Removes and cancels all known tasks, ignoring any exceptions. |
833 |
|
*/ |
834 |
|
final void cancelAll() { |
852 |
|
return seed = r ^= r << 5; |
853 |
|
} |
854 |
|
|
855 |
< |
// Execution methods |
855 |
> |
/** |
856 |
> |
* Provides a more accurate estimate of size than (top - base) |
857 |
> |
* by ordering reads and checking whether a near-empty queue |
858 |
> |
* has at least one unclaimed task. |
859 |
> |
*/ |
860 |
> |
final int queueSize() { |
861 |
> |
ForkJoinTask<?>[] a; int k, s, n; |
862 |
> |
return ((n = base - (s = top)) < 0 && |
863 |
> |
(n != -1 || |
864 |
> |
((a = array) != null && (k = a.length) > 0 && |
865 |
> |
U.getObject |
866 |
> |
(a, (long)((((k - 1) & (s - 1)) << ASHIFT) + ABASE)) != null))) ? |
867 |
> |
-n : 0; |
868 |
> |
} |
869 |
> |
|
870 |
> |
// Specialized execution methods |
871 |
|
|
872 |
|
/** |
873 |
|
* Pops and runs tasks until empty. |
896 |
|
} |
897 |
|
|
898 |
|
/** |
899 |
< |
* If present, removes from queue and executes the given task, or |
900 |
< |
* any other cancelled task. Returns (true) immediately on any CAS |
899 |
> |
* If present, removes from queue and executes the given task, |
900 |
> |
* or any other cancelled task. Returns (true) on any CAS |
901 |
|
* or consistency check failure so caller can retry. |
902 |
|
* |
903 |
< |
* @return 0 if no progress can be made, else positive |
924 |
< |
* (this unusual convention simplifies use with tryHelpStealer.) |
903 |
> |
* @return false if no progress can be made, else true; |
904 |
|
*/ |
905 |
< |
final int tryRemoveAndExec(ForkJoinTask<?> task) { |
906 |
< |
int stat = 1; |
928 |
< |
boolean removed = false, empty = true; |
905 |
> |
final boolean tryRemoveAndExec(ForkJoinTask<?> task) { |
906 |
> |
boolean stat = true, removed = false, empty = true; |
907 |
|
ForkJoinTask<?>[] a; int m, s, b, n; |
908 |
|
if ((a = array) != null && (m = a.length - 1) >= 0 && |
909 |
|
(n = (s = top) - (b = base)) > 0) { |
933 |
|
} |
934 |
|
if (--n == 0) { |
935 |
|
if (!empty && base == b) |
936 |
< |
stat = 0; |
936 |
> |
stat = false; |
937 |
|
break; |
938 |
|
} |
939 |
|
} |
944 |
|
} |
945 |
|
|
946 |
|
/** |
947 |
+ |
* Polls for and executes the given task or any other task in |
948 |
+ |
* its CountedCompleter computation |
949 |
+ |
*/ |
950 |
+ |
final boolean pollAndExecCC(ForkJoinTask<?> root) { |
951 |
+ |
ForkJoinTask<?>[] a; int b; Object o; |
952 |
+ |
outer: while ((b = base) - top < 0 && (a = array) != null) { |
953 |
+ |
long j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
954 |
+ |
if ((o = U.getObject(a, j)) == null || |
955 |
+ |
!(o instanceof CountedCompleter)) |
956 |
+ |
break; |
957 |
+ |
for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) { |
958 |
+ |
if (r == root) { |
959 |
+ |
if (base == b && |
960 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
961 |
+ |
base = b + 1; |
962 |
+ |
t.doExec(); |
963 |
+ |
return true; |
964 |
+ |
} |
965 |
+ |
else |
966 |
+ |
break; // restart |
967 |
+ |
} |
968 |
+ |
if ((r = r.completer) == null) |
969 |
+ |
break outer; // not part of root computation |
970 |
+ |
} |
971 |
+ |
} |
972 |
+ |
return false; |
973 |
+ |
} |
974 |
+ |
|
975 |
+ |
/** |
976 |
|
* Executes a top-level task and any local tasks remaining |
977 |
|
* after execution. |
978 |
|
*/ |
979 |
|
final void runTask(ForkJoinTask<?> t) { |
980 |
|
if (t != null) { |
981 |
< |
currentSteal = t; |
982 |
< |
t.doExec(); |
981 |
> |
(currentSteal = t).doExec(); |
982 |
> |
currentSteal = null; |
983 |
> |
if (++nsteals < 0) { // spill on overflow |
984 |
> |
ForkJoinPool p; |
985 |
> |
if ((p = pool) != null) |
986 |
> |
p.collectStealCount(this); |
987 |
> |
} |
988 |
|
if (top != base) { // process remaining local tasks |
989 |
|
if (mode == 0) |
990 |
|
popAndExecAll(); |
991 |
|
else |
992 |
|
pollAndExecAll(); |
993 |
|
} |
982 |
– |
++nsteals; |
983 |
– |
currentSteal = null; |
994 |
|
} |
995 |
|
} |
996 |
|
|
1000 |
|
final void runSubtask(ForkJoinTask<?> t) { |
1001 |
|
if (t != null) { |
1002 |
|
ForkJoinTask<?> ps = currentSteal; |
1003 |
< |
currentSteal = t; |
994 |
< |
t.doExec(); |
1003 |
> |
(currentSteal = t).doExec(); |
1004 |
|
currentSteal = ps; |
1005 |
|
} |
1006 |
|
} |
1035 |
|
|
1036 |
|
// Unsafe mechanics |
1037 |
|
private static final sun.misc.Unsafe U; |
1038 |
< |
private static final long RUNSTATE; |
1038 |
> |
private static final long QLOCK; |
1039 |
|
private static final int ABASE; |
1040 |
|
private static final int ASHIFT; |
1041 |
|
static { |
1044 |
|
U = getUnsafe(); |
1045 |
|
Class<?> k = WorkQueue.class; |
1046 |
|
Class<?> ak = ForkJoinTask[].class; |
1047 |
< |
RUNSTATE = U.objectFieldOffset |
1048 |
< |
(k.getDeclaredField("runState")); |
1047 |
> |
QLOCK = U.objectFieldOffset |
1048 |
> |
(k.getDeclaredField("qlock")); |
1049 |
|
ABASE = U.arrayBaseOffset(ak); |
1050 |
|
s = U.arrayIndexScale(ak); |
1051 |
|
} catch (Exception e) { |
1060 |
|
/** |
1061 |
|
* Per-thread records for threads that submit to pools. Currently |
1062 |
|
* holds only pseudo-random seed / index that is used to choose |
1063 |
< |
* submission queues in method doSubmit. In the future, this may |
1063 |
> |
* submission queues in method externalPush. In the future, this may |
1064 |
|
* also incorporate a means to implement different task rejection |
1065 |
|
* and resubmission policies. |
1066 |
|
* |
1068 |
|
* the same way but are initialized and updated using slightly |
1069 |
|
* different mechanics. Both are initialized using the same |
1070 |
|
* approach as in class ThreadLocal, where successive values are |
1071 |
< |
* unlikely to collide with previous values. This is done during |
1072 |
< |
* registration for workers, but requires a separate AtomicInteger |
1073 |
< |
* for submitters. Seeds are then randomly modified upon |
1065 |
< |
* collisions using xorshifts, which requires a non-zero seed. |
1071 |
> |
* unlikely to collide with previous values. Seeds are then |
1072 |
> |
* randomly modified upon collisions using xorshifts, which |
1073 |
> |
* requires a non-zero seed. |
1074 |
|
*/ |
1075 |
|
static final class Submitter { |
1076 |
|
int seed; |
1077 |
< |
Submitter() { |
1070 |
< |
int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT); |
1071 |
< |
seed = (s == 0) ? 1 : s; // ensure non-zero |
1072 |
< |
} |
1077 |
> |
Submitter(int s) { seed = s; } |
1078 |
|
} |
1079 |
|
|
1080 |
< |
/** ThreadLocal class for Submitters */ |
1081 |
< |
static final class ThreadSubmitter extends ThreadLocal<Submitter> { |
1082 |
< |
public Submitter initialValue() { return new Submitter(); } |
1078 |
< |
} |
1080 |
> |
/** Property prefix for constructing common pool */ |
1081 |
> |
private static final String propPrefix = |
1082 |
> |
"java.util.concurrent.ForkJoinPool.common."; |
1083 |
|
|
1084 |
|
// static fields (initialized in static initializer below) |
1085 |
|
|
1091 |
|
defaultForkJoinWorkerThreadFactory; |
1092 |
|
|
1093 |
|
/** |
1094 |
< |
* Generator for assigning sequence numbers as pool names. |
1094 |
> |
* Common (static) pool. Non-null for public use unless a static |
1095 |
> |
* construction exception, but internal usages null-check on use |
1096 |
> |
* to paranoically avoid potential initialization circularities |
1097 |
> |
* as well as to simplify generated code. |
1098 |
|
*/ |
1099 |
< |
private static final AtomicInteger poolNumberGenerator; |
1093 |
< |
|
1094 |
< |
/** |
1095 |
< |
* Generator for initial hashes/seeds for submitters. Accessed by |
1096 |
< |
* Submitter class constructor. |
1097 |
< |
*/ |
1098 |
< |
static final AtomicInteger nextSubmitterSeed; |
1099 |
> |
static final ForkJoinPool commonPool; |
1100 |
|
|
1101 |
|
/** |
1102 |
|
* Permission required for callers of methods that may start or |
1108 |
|
* Per-thread submission bookkeeping. Shared across all pools |
1109 |
|
* to reduce ThreadLocal pollution and because random motion |
1110 |
|
* to avoid contention in one pool is likely to hold for others. |
1111 |
+ |
* Lazily initialized on first submission (but null-checked |
1112 |
+ |
* in other contexts to avoid unnecessary initialization). |
1113 |
+ |
*/ |
1114 |
+ |
static final ThreadLocal<Submitter> submitters; |
1115 |
+ |
|
1116 |
+ |
/** |
1117 |
+ |
* Common pool parallelism. Must equal commonPool.parallelism. |
1118 |
|
*/ |
1119 |
< |
private static final ThreadSubmitter submitters; |
1119 |
> |
static final int commonPoolParallelism; |
1120 |
> |
|
1121 |
> |
/** |
1122 |
> |
* Sequence number for creating workerNamePrefix. |
1123 |
> |
*/ |
1124 |
> |
private static int poolNumberSequence; |
1125 |
> |
|
1126 |
> |
/** |
1127 |
> |
* Return the next sequence number. We don't expect this to |
1128 |
> |
* ever contend so use simple builtin sync. |
1129 |
> |
*/ |
1130 |
> |
private static final synchronized int nextPoolId() { |
1131 |
> |
return ++poolNumberSequence; |
1132 |
> |
} |
1133 |
|
|
1134 |
|
// static constants |
1135 |
|
|
1136 |
|
/** |
1137 |
< |
* The wakeup interval (in nanoseconds) for a worker waiting for a |
1138 |
< |
* task when the pool is quiescent to instead try to shrink the |
1139 |
< |
* number of workers. The exact value does not matter too |
1140 |
< |
* much. It must be short enough to release resources during |
1141 |
< |
* sustained periods of idleness, but not so short that threads |
1142 |
< |
* are continually re-created. |
1137 |
> |
* Initial timeout value (in nanoseconds) for the thread |
1138 |
> |
* triggering quiescence to park waiting for new work. On timeout, |
1139 |
> |
* the thread will instead try to shrink the number of |
1140 |
> |
* workers. The value should be large enough to avoid overly |
1141 |
> |
* aggressive shrinkage during most transient stalls (long GCs |
1142 |
> |
* etc). |
1143 |
|
*/ |
1144 |
< |
private static final long SHRINK_RATE = |
1124 |
< |
4L * 1000L * 1000L * 1000L; // 4 seconds |
1144 |
> |
private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec |
1145 |
|
|
1146 |
|
/** |
1147 |
< |
* The timeout value for attempted shrinkage, includes |
1128 |
< |
* some slop to cope with system timer imprecision. |
1147 |
> |
* Timeout value when there are more threads than parallelism level |
1148 |
|
*/ |
1149 |
< |
private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10); |
1149 |
> |
private static final long FAST_IDLE_TIMEOUT = 200L * 1000L * 1000L; |
1150 |
|
|
1151 |
|
/** |
1152 |
|
* The maximum stolen->joining link depth allowed in method |
1153 |
< |
* tryHelpStealer. Must be a power of two. This value also |
1135 |
< |
* controls the maximum number of times to try to help join a task |
1136 |
< |
* without any apparent progress or change in pool state before |
1137 |
< |
* giving up and blocking (see awaitJoin). Depths for legitimate |
1153 |
> |
* tryHelpStealer. Must be a power of two. Depths for legitimate |
1154 |
|
* chains are unbounded, but we use a fixed constant to avoid |
1155 |
|
* (otherwise unchecked) cycles and to bound staleness of |
1156 |
|
* traversal parameters at the expense of sometimes blocking when |
1159 |
|
private static final int MAX_HELP = 64; |
1160 |
|
|
1161 |
|
/** |
1146 |
– |
* Secondary time-based bound (in nanosecs) for helping attempts |
1147 |
– |
* before trying compensated blocking in awaitJoin. Used in |
1148 |
– |
* conjunction with MAX_HELP to reduce variance due to different |
1149 |
– |
* polling rates associated with different helping options. The |
1150 |
– |
* value should roughly approximate the time required to create |
1151 |
– |
* and/or activate a worker thread. |
1152 |
– |
*/ |
1153 |
– |
private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec |
1154 |
– |
|
1155 |
– |
/** |
1162 |
|
* Increment for seed generators. See class ThreadLocal for |
1163 |
|
* explanation. |
1164 |
|
*/ |
1192 |
|
* scan for them to avoid queuing races. Note however that |
1193 |
|
* eventCount updates lag releases so usage requires care. |
1194 |
|
* |
1195 |
< |
* Field runState is an int packed with: |
1195 |
> |
* Field plock is an int packed with: |
1196 |
|
* SHUTDOWN: true if shutdown is enabled (1 bit) |
1197 |
< |
* SEQ: a sequence number updated upon (de)registering workers (30 bits) |
1198 |
< |
* INIT: set true after workQueues array construction (1 bit) |
1197 |
> |
* SEQ: a sequence lock, with PL_LOCK bit set if locked (30 bits) |
1198 |
> |
* SIGNAL: set when threads may be waiting on the lock (1 bit) |
1199 |
|
* |
1200 |
|
* The sequence number enables simple consistency checks: |
1201 |
|
* Staleness of read-only operations on the workQueues array can |
1202 |
< |
* be checked by comparing runState before vs after the reads. |
1202 |
> |
* be checked by comparing plock before vs after the reads. |
1203 |
|
*/ |
1204 |
|
|
1205 |
|
// bit positions/shifts for fields |
1211 |
|
// bounds |
1212 |
|
private static final int SMASK = 0xffff; // short bits |
1213 |
|
private static final int MAX_CAP = 0x7fff; // max #workers - 1 |
1214 |
< |
private static final int SQMASK = 0xfffe; // even short bits |
1214 |
> |
private static final int EVENMASK = 0xfffe; // even short bits |
1215 |
> |
private static final int SQMASK = 0x007e; // max 64 (even) slots |
1216 |
|
private static final int SHORT_SIGN = 1 << 15; |
1217 |
|
private static final int INT_SIGN = 1 << 31; |
1218 |
|
|
1237 |
|
private static final int E_MASK = 0x7fffffff; // no STOP_BIT |
1238 |
|
private static final int E_SEQ = 1 << EC_SHIFT; |
1239 |
|
|
1240 |
< |
// runState bits |
1240 |
> |
// plock bits |
1241 |
|
private static final int SHUTDOWN = 1 << 31; |
1242 |
+ |
private static final int PL_LOCK = 2; |
1243 |
+ |
private static final int PL_SIGNAL = 1; |
1244 |
+ |
private static final int PL_SPINS = 1 << 8; |
1245 |
|
|
1246 |
|
// access mode for WorkQueue |
1247 |
|
static final int LIFO_QUEUE = 0; |
1256 |
|
* declaration order and may differ across JVMs, but the following |
1257 |
|
* empirically works OK on current JVMs. |
1258 |
|
*/ |
1259 |
< |
|
1259 |
> |
volatile long stealCount; // collects worker counts |
1260 |
|
volatile long ctl; // main pool control |
1261 |
|
final int parallelism; // parallelism level |
1262 |
|
final int localMode; // per-worker scheduling mode |
1263 |
< |
final int submitMask; // submit queue index bound |
1264 |
< |
int nextSeed; // for initializing worker seeds |
1255 |
< |
volatile int runState; // shutdown status and seq |
1263 |
> |
volatile int indexSeed; // worker/submitter index seed |
1264 |
> |
volatile int plock; // shutdown status and seqLock |
1265 |
|
WorkQueue[] workQueues; // main registry |
1257 |
– |
final Mutex lock; // for registration |
1258 |
– |
final Condition termination; // for awaitTermination |
1266 |
|
final ForkJoinWorkerThreadFactory factory; // factory for new workers |
1267 |
|
final Thread.UncaughtExceptionHandler ueh; // per-worker UEH |
1261 |
– |
final AtomicLong stealCount; // collect counts when terminated |
1262 |
– |
final AtomicInteger nextWorkerNumber; // to create worker name string |
1268 |
|
final String workerNamePrefix; // to create worker name string |
1269 |
|
|
1270 |
< |
// Creating, registering, and deregistering workers |
1271 |
< |
|
1272 |
< |
/** |
1273 |
< |
* Tries to create and start a worker |
1274 |
< |
*/ |
1275 |
< |
private void addWorker() { |
1276 |
< |
Throwable ex = null; |
1277 |
< |
ForkJoinWorkerThread wt = null; |
1278 |
< |
try { |
1279 |
< |
if ((wt = factory.newThread(this)) != null) { |
1280 |
< |
wt.start(); |
1281 |
< |
return; |
1270 |
> |
/* |
1271 |
> |
* Acquires the plock lock to protect worker array and related |
1272 |
> |
* updates. This method is called only if an initial CAS on plock |
1273 |
> |
* fails. This acts as a spinLock for normal cases, but falls back |
1274 |
> |
* to builtin monitor to block when (rarely) needed. This would be |
1275 |
> |
* a terrible idea for a highly contended lock, but works fine as |
1276 |
> |
* a more conservative alternative to a pure spinlock. See |
1277 |
> |
* internal ConcurrentHashMap documentation for further |
1278 |
> |
* explanation of nearly the same construction. |
1279 |
> |
*/ |
1280 |
> |
private int acquirePlock() { |
1281 |
> |
int spins = PL_SPINS, r = 0, ps, nps; |
1282 |
> |
for (;;) { |
1283 |
> |
if (((ps = plock) & PL_LOCK) == 0 && |
1284 |
> |
U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK)) |
1285 |
> |
return nps; |
1286 |
> |
else if (r == 0) |
1287 |
> |
r = ThreadLocalRandom.current().nextInt(); // randomize spins |
1288 |
> |
else if (spins >= 0) { |
1289 |
> |
r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift |
1290 |
> |
if (r >= 0) |
1291 |
> |
--spins; |
1292 |
> |
} |
1293 |
> |
else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) { |
1294 |
> |
synchronized(this) { |
1295 |
> |
if ((plock & PL_SIGNAL) != 0) { |
1296 |
> |
try { |
1297 |
> |
wait(); |
1298 |
> |
} catch (InterruptedException ie) { |
1299 |
> |
try { |
1300 |
> |
Thread.currentThread().interrupt(); |
1301 |
> |
} catch (SecurityException ignore) { |
1302 |
> |
} |
1303 |
> |
} |
1304 |
> |
} |
1305 |
> |
else |
1306 |
> |
notifyAll(); |
1307 |
> |
} |
1308 |
|
} |
1278 |
– |
} catch (Throwable e) { |
1279 |
– |
ex = e; |
1309 |
|
} |
1281 |
– |
deregisterWorker(wt, ex); // adjust counts etc on failure |
1310 |
|
} |
1311 |
|
|
1312 |
|
/** |
1313 |
< |
* Callback from ForkJoinWorkerThread constructor to assign a |
1314 |
< |
* public name. This must be separate from registerWorker because |
1287 |
< |
* it is called during the "super" constructor call in |
1288 |
< |
* ForkJoinWorkerThread. |
1313 |
> |
* Unlocks and signals any thread waiting for plock. Called only |
1314 |
> |
* when CAS of seq value for unlock fails. |
1315 |
|
*/ |
1316 |
< |
final String nextWorkerName() { |
1317 |
< |
return workerNamePrefix.concat |
1318 |
< |
(Integer.toString(nextWorkerNumber.addAndGet(1))); |
1316 |
> |
private void releasePlock(int ps) { |
1317 |
> |
plock = ps; |
1318 |
> |
synchronized(this) { notifyAll(); } |
1319 |
|
} |
1320 |
|
|
1321 |
+ |
// Registering and deregistering workers |
1322 |
+ |
|
1323 |
|
/** |
1324 |
|
* Callback from ForkJoinWorkerThread constructor to establish its |
1325 |
|
* poolIndex and record its WorkQueue. To avoid scanning bias due |
1329 |
|
* |
1330 |
|
* @param w the worker's queue |
1331 |
|
*/ |
1304 |
– |
|
1332 |
|
final void registerWorker(WorkQueue w) { |
1333 |
< |
Mutex lock = this.lock; |
1334 |
< |
lock.lock(); |
1333 |
> |
int s, ps; // generate a rarely colliding candidate index seed |
1334 |
> |
do {} while (!U.compareAndSwapInt(this, INDEXSEED, |
1335 |
> |
s = indexSeed, s += SEED_INCREMENT) || |
1336 |
> |
s == 0); // skip 0 |
1337 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
1338 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1339 |
> |
ps = acquirePlock(); |
1340 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1341 |
|
try { |
1342 |
< |
WorkQueue[] ws = workQueues; |
1343 |
< |
if (w != null && ws != null) { // skip on shutdown/failure |
1344 |
< |
int rs, n = ws.length, m = n - 1; |
1345 |
< |
int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence |
1313 |
< |
w.seed = (s == 0) ? 1 : s; // ensure non-zero seed |
1342 |
> |
WorkQueue[] ws; |
1343 |
> |
if (w != null && (ws = workQueues) != null) { |
1344 |
> |
w.seed = s; |
1345 |
> |
int n = ws.length, m = n - 1; |
1346 |
|
int r = (s << 1) | 1; // use odd-numbered indices |
1347 |
|
if (ws[r &= m] != null) { // collision |
1348 |
|
int probes = 0; // step by approx half size |
1349 |
< |
int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2; |
1349 |
> |
int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2; |
1350 |
|
while (ws[r = (r + step) & m] != null) { |
1351 |
|
if (++probes >= n) { |
1352 |
|
workQueues = ws = Arrays.copyOf(ws, n <<= 1); |
1356 |
|
} |
1357 |
|
} |
1358 |
|
w.eventCount = w.poolIndex = r; // establish before recording |
1359 |
< |
ws[r] = w; // also update seq |
1328 |
< |
runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN); |
1359 |
> |
ws[r] = w; |
1360 |
|
} |
1361 |
|
} finally { |
1362 |
< |
lock.unlock(); |
1362 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1363 |
> |
releasePlock(nps); |
1364 |
|
} |
1365 |
|
} |
1366 |
|
|
1367 |
|
/** |
1368 |
|
* Final callback from terminating worker, as well as upon failure |
1369 |
< |
* to construct or start a worker in addWorker. Removes record of |
1370 |
< |
* worker from array, and adjusts counts. If pool is shutting |
1371 |
< |
* down, tries to complete termination. |
1369 |
> |
* to construct or start a worker. Removes record of worker from |
1370 |
> |
* array, and adjusts counts. If pool is shutting down, tries to |
1371 |
> |
* complete termination. |
1372 |
|
* |
1373 |
< |
* @param wt the worker thread or null if addWorker failed |
1373 |
> |
* @param wt the worker thread or null if construction failed |
1374 |
|
* @param ex the exception causing failure, or null if none |
1375 |
|
*/ |
1376 |
|
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1345 |
– |
Mutex lock = this.lock; |
1377 |
|
WorkQueue w = null; |
1378 |
|
if (wt != null && (w = wt.workQueue) != null) { |
1379 |
< |
w.runState = -1; // ensure runState is set |
1380 |
< |
stealCount.getAndAdd(w.totalSteals + w.nsteals); |
1381 |
< |
int idx = w.poolIndex; |
1382 |
< |
lock.lock(); |
1383 |
< |
try { // remove record from array |
1379 |
> |
int ps; |
1380 |
> |
collectStealCount(w); |
1381 |
> |
w.qlock = -1; // ensure set |
1382 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
1383 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1384 |
> |
ps = acquirePlock(); |
1385 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1386 |
> |
try { |
1387 |
> |
int idx = w.poolIndex; |
1388 |
|
WorkQueue[] ws = workQueues; |
1389 |
|
if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) |
1390 |
|
ws[idx] = null; |
1391 |
|
} finally { |
1392 |
< |
lock.unlock(); |
1392 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1393 |
> |
releasePlock(nps); |
1394 |
|
} |
1395 |
|
} |
1396 |
|
|
1403 |
|
if (!tryTerminate(false, false) && w != null) { |
1404 |
|
w.cancelAll(); // cancel remaining tasks |
1405 |
|
if (w.array != null) // suppress signal if never ran |
1406 |
< |
signalWork(); // wake up or create replacement |
1406 |
> |
signalWork(null, 1); // wake up or create replacement |
1407 |
|
if (ex == null) // help clean refs on way out |
1408 |
|
ForkJoinTask.helpExpungeStaleExceptions(); |
1409 |
|
} |
1410 |
|
|
1411 |
|
if (ex != null) // rethrow |
1412 |
< |
U.throwException(ex); |
1412 |
> |
ForkJoinTask.rethrow(ex); |
1413 |
|
} |
1414 |
|
|
1415 |
+ |
/** |
1416 |
+ |
* Collect worker steal count into total. Called on termination |
1417 |
+ |
* and upon int overflow of local count. (There is a possible race |
1418 |
+ |
* in the latter case vs any caller of getStealCount, which can |
1419 |
+ |
* make its results less accurate than usual.) |
1420 |
+ |
*/ |
1421 |
+ |
final void collectStealCount(WorkQueue w) { |
1422 |
+ |
if (w != null) { |
1423 |
+ |
long sc; |
1424 |
+ |
int ns = w.nsteals; |
1425 |
+ |
w.nsteals = 0; // handle overflow |
1426 |
+ |
long steals = (ns >= 0) ? ns : 1L + (long)(Integer.MAX_VALUE); |
1427 |
+ |
do {} while (!U.compareAndSwapLong(this, STEALCOUNT, |
1428 |
+ |
sc = stealCount, sc + steals)); |
1429 |
+ |
} |
1430 |
+ |
} |
1431 |
|
|
1432 |
|
// Submissions |
1433 |
|
|
1434 |
|
/** |
1435 |
|
* Unless shutting down, adds the given task to a submission queue |
1436 |
|
* at submitter's current queue index (modulo submission |
1437 |
< |
* range). If no queue exists at the index, one is created. If |
1438 |
< |
* the queue is busy, another index is randomly chosen. The |
1387 |
< |
* submitMask bounds the effective number of queues to the |
1388 |
< |
* (nearest power of two for) parallelism level. |
1437 |
> |
* range). Only the most common path is directly handled in this |
1438 |
> |
* method. All others are relayed to fullExternalPush. |
1439 |
|
* |
1440 |
|
* @param task the task. Caller must ensure non-null. |
1441 |
|
*/ |
1442 |
< |
private void doSubmit(ForkJoinTask<?> task) { |
1443 |
< |
Submitter s = submitters.get(); |
1444 |
< |
for (int r = s.seed, m = submitMask;;) { |
1445 |
< |
WorkQueue[] ws; WorkQueue q; |
1446 |
< |
int k = r & m & SQMASK; // use only even indices |
1447 |
< |
if (runState < 0 || (ws = workQueues) == null || ws.length <= k) |
1448 |
< |
throw new RejectedExecutionException(); // shutting down |
1449 |
< |
else if ((q = ws[k]) == null) { // create new queue |
1450 |
< |
WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE); |
1451 |
< |
Mutex lock = this.lock; // construct outside lock |
1452 |
< |
lock.lock(); |
1453 |
< |
try { // recheck under lock |
1454 |
< |
int rs = runState; // to update seq |
1455 |
< |
if (ws == workQueues && ws[k] == null) { |
1406 |
< |
ws[k] = nq; |
1407 |
< |
runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN)); |
1408 |
< |
} |
1409 |
< |
} finally { |
1410 |
< |
lock.unlock(); |
1411 |
< |
} |
1412 |
< |
} |
1413 |
< |
else if (q.trySharedPush(task)) { |
1414 |
< |
signalWork(); |
1442 |
> |
final void externalPush(ForkJoinTask<?> task) { |
1443 |
> |
WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a; |
1444 |
> |
if ((z = submitters.get()) != null && plock > 0 && |
1445 |
> |
(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 && |
1446 |
> |
(q = ws[m & z.seed & SQMASK]) != null && |
1447 |
> |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock |
1448 |
> |
int s = q.top, n; |
1449 |
> |
if ((a = q.array) != null && a.length > (n = s + 1 - q.base)) { |
1450 |
> |
U.putObject(a, (long)(((a.length - 1) & s) << ASHIFT) + ABASE, |
1451 |
> |
task); |
1452 |
> |
q.top = s + 1; // push on to deque |
1453 |
> |
q.qlock = 0; |
1454 |
> |
if (n <= 1) |
1455 |
> |
signalWork(q, 1); |
1456 |
|
return; |
1457 |
|
} |
1458 |
< |
else if (m > 1) { // move to a different index |
1458 |
> |
q.qlock = 0; |
1459 |
> |
} |
1460 |
> |
fullExternalPush(task); |
1461 |
> |
} |
1462 |
> |
|
1463 |
> |
/** |
1464 |
> |
* Full version of externalPush. This method is called, among |
1465 |
> |
* other times, upon the first submission of the first task to the |
1466 |
> |
* pool, so must perform secondary initialization: creating |
1467 |
> |
* workQueue array and setting plock to a valid value. It also |
1468 |
> |
* detects first submission by an external thread by looking up |
1469 |
> |
* its ThreadLocal, and creates a new shared queue if the one at |
1470 |
> |
* index if empty or contended. The lock bodies must be |
1471 |
> |
* exception-free (so no try/finally) so we optimistically |
1472 |
> |
* allocate new queues/arrays outside the locks and throw them |
1473 |
> |
* away if (very rarely) not needed. Note that the plock seq value |
1474 |
> |
* can eventually wrap around zero, but if so harmlessly fails to |
1475 |
> |
* reinitialize. |
1476 |
> |
*/ |
1477 |
> |
private void fullExternalPush(ForkJoinTask<?> task) { |
1478 |
> |
for (Submitter z = null;;) { |
1479 |
> |
WorkQueue[] ws; WorkQueue q; int ps, m, r, s; |
1480 |
> |
if ((ps = plock) < 0) |
1481 |
> |
throw new RejectedExecutionException(); |
1482 |
> |
else if ((ws = workQueues) == null || (m = ws.length - 1) < 0) { |
1483 |
> |
int n = parallelism - 1; n |= n >>> 1; n |= n >>> 2; |
1484 |
> |
n |= n >>> 4; n |= n >>> 8; n |= n >>> 16; |
1485 |
> |
WorkQueue[] nws = new WorkQueue[(n + 1) << 1]; // power of two |
1486 |
> |
if ((ps & PL_LOCK) != 0 || |
1487 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1488 |
> |
ps = acquirePlock(); |
1489 |
> |
if ((ws = workQueues) == null) |
1490 |
> |
workQueues = nws; |
1491 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1492 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1493 |
> |
releasePlock(nps); |
1494 |
> |
} |
1495 |
> |
else if (z == null && (z = submitters.get()) == null) { |
1496 |
> |
if (U.compareAndSwapInt(this, INDEXSEED, |
1497 |
> |
s = indexSeed, s += SEED_INCREMENT) && |
1498 |
> |
s != 0) // skip 0 |
1499 |
> |
submitters.set(z = new Submitter(s)); |
1500 |
> |
} |
1501 |
> |
else { |
1502 |
> |
int k = (r = z.seed) & m & SQMASK; |
1503 |
> |
if ((q = ws[k]) == null && (ps & PL_LOCK) == 0) { |
1504 |
> |
(q = new WorkQueue(this, null, SHARED_QUEUE)).poolIndex = k; |
1505 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
1506 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1507 |
> |
ps = acquirePlock(); |
1508 |
> |
WorkQueue w = null; |
1509 |
> |
if ((ws = workQueues) != null && k < ws.length && |
1510 |
> |
(w = ws[k]) == null) |
1511 |
> |
ws[k] = q; |
1512 |
> |
else |
1513 |
> |
q = w; |
1514 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1515 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1516 |
> |
releasePlock(nps); |
1517 |
> |
} |
1518 |
> |
if (q != null && q.qlock == 0 && q.fullPush(task, false)) |
1519 |
> |
return; |
1520 |
|
r ^= r << 13; // same xorshift as WorkQueues |
1521 |
|
r ^= r >>> 17; |
1522 |
< |
s.seed = r ^= r << 5; |
1522 |
> |
z.seed = r ^= r << 5; // move to a different index |
1523 |
|
} |
1422 |
– |
else |
1423 |
– |
Thread.yield(); // yield if no alternatives |
1524 |
|
} |
1525 |
|
} |
1526 |
|
|
1535 |
|
} |
1536 |
|
|
1537 |
|
/** |
1538 |
< |
* Tries to activate or create a worker if too few are active. |
1539 |
< |
*/ |
1540 |
< |
final void signalWork() { |
1541 |
< |
long c; int u; |
1542 |
< |
while ((u = (int)((c = ctl) >>> 32)) < 0) { // too few active |
1543 |
< |
WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p; |
1544 |
< |
if ((e = (int)c) > 0) { // at least one waiting |
1545 |
< |
if (ws != null && (i = e & SMASK) < ws.length && |
1538 |
> |
* Tries to create (at most one) or activate (possibly several) |
1539 |
> |
* workers if too few are active. On contention failure, continues |
1540 |
> |
* until at least one worker is signalled or the given queue is |
1541 |
> |
* empty or all workers are active. |
1542 |
> |
* |
1543 |
> |
* @param q if non-null, the queue holding tasks to be signalled |
1544 |
> |
* @param signals the target number of signals. |
1545 |
> |
*/ |
1546 |
> |
final void signalWork(WorkQueue q, int signals) { |
1547 |
> |
long c; int e, u, i; WorkQueue[] ws; WorkQueue w; Thread p; |
1548 |
> |
while ((u = (int)((c = ctl) >>> 32)) < 0) { |
1549 |
> |
if ((e = (int)c) > 0) { |
1550 |
> |
if ((ws = workQueues) != null && ws.length > (i = e & SMASK) && |
1551 |
|
(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { |
1552 |
|
long nc = (((long)(w.nextWait & E_MASK)) | |
1553 |
|
((long)(u + UAC_UNIT) << 32)); |
1554 |
|
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1555 |
|
w.eventCount = (e + E_SEQ) & E_MASK; |
1556 |
|
if ((p = w.parker) != null) |
1557 |
< |
U.unpark(p); // activate and release |
1558 |
< |
break; |
1557 |
> |
U.unpark(p); |
1558 |
> |
if (--signals <= 0) |
1559 |
> |
break; |
1560 |
|
} |
1561 |
+ |
else |
1562 |
+ |
signals = 1; |
1563 |
+ |
if ((q != null && q.queueSize() == 0)) |
1564 |
+ |
break; |
1565 |
|
} |
1566 |
|
else |
1567 |
|
break; |
1568 |
|
} |
1569 |
< |
else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total |
1569 |
> |
else if (e == 0 && (u & SHORT_SIGN) != 0) { |
1570 |
|
long nc = (long)(((u + UTC_UNIT) & UTC_MASK) | |
1571 |
|
((u + UAC_UNIT) & UAC_MASK)) << 32; |
1572 |
|
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1573 |
< |
addWorker(); |
1573 |
> |
ForkJoinWorkerThread wt = null; |
1574 |
> |
Throwable ex = null; |
1575 |
> |
boolean started = false; |
1576 |
> |
try { |
1577 |
> |
ForkJoinWorkerThreadFactory fac; |
1578 |
> |
if ((fac = factory) != null && |
1579 |
> |
(wt = fac.newThread(this)) != null) { |
1580 |
> |
wt.start(); |
1581 |
> |
started = true; |
1582 |
> |
} |
1583 |
> |
} catch (Throwable rex) { |
1584 |
> |
ex = rex; |
1585 |
> |
} |
1586 |
> |
if (!started) |
1587 |
> |
deregisterWorker(wt, ex); // adjust counts on failure |
1588 |
|
break; |
1589 |
|
} |
1590 |
|
} |
1599 |
|
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1600 |
|
*/ |
1601 |
|
final void runWorker(WorkQueue w) { |
1602 |
< |
w.growArray(false); // initialize queue array in this thread |
1603 |
< |
do { w.runTask(scan(w)); } while (w.runState >= 0); |
1602 |
> |
// initialize queue array in this thread |
1603 |
> |
w.array = new ForkJoinTask<?>[WorkQueue.INITIAL_QUEUE_CAPACITY]; |
1604 |
> |
do { w.runTask(scan(w)); } while (w.qlock >= 0); |
1605 |
|
} |
1606 |
|
|
1607 |
|
/** |
1617 |
|
* relative prime, checking each at least once). The scan |
1618 |
|
* terminates upon either finding a non-empty queue, or completing |
1619 |
|
* the sweep. If the worker is not inactivated, it takes and |
1620 |
< |
* returns a task from this queue. On failure to find a task, we |
1620 |
> |
* returns a task from this queue. Otherwise, if not activated, it |
1621 |
> |
* signals workers (that may include itself) and returns so caller |
1622 |
> |
* can retry. Also returns for trtry if the worker array may have |
1623 |
> |
* changed during an empty scan. On failure to find a task, we |
1624 |
|
* take one of the following actions, after which the caller will |
1625 |
|
* retry calling this method unless terminated. |
1626 |
|
* |
1627 |
|
* * If pool is terminating, terminate the worker. |
1628 |
|
* |
1501 |
– |
* * If not a complete sweep, try to release a waiting worker. If |
1502 |
– |
* the scan terminated because the worker is inactivated, then the |
1503 |
– |
* released worker will often be the calling worker, and it can |
1504 |
– |
* succeed obtaining a task on the next call. Or maybe it is |
1505 |
– |
* another worker, but with same net effect. Releasing in other |
1506 |
– |
* cases as well ensures that we have enough workers running. |
1507 |
– |
* |
1629 |
|
* * If not already enqueued, try to inactivate and enqueue the |
1630 |
|
* worker on wait queue. Or, if inactivating has caused the pool |
1631 |
|
* to be quiescent, relay to idleAwaitWork to check for |
1632 |
|
* termination and possibly shrink pool. |
1633 |
|
* |
1634 |
< |
* * If already inactive, and the caller has run a task since the |
1635 |
< |
* last empty scan, return (to allow rescan) unless others are |
1636 |
< |
* also inactivated. Field WorkQueue.rescans counts down on each |
1516 |
< |
* scan to ensure eventual inactivation and blocking. |
1517 |
< |
* |
1518 |
< |
* * If already enqueued and none of the above apply, park |
1519 |
< |
* awaiting signal, |
1634 |
> |
* * If already enqueued and none of the above apply, possibly |
1635 |
> |
* (with 1/2 probablility) park awaiting signal, else lingering to |
1636 |
> |
* help scan and signal. |
1637 |
|
* |
1638 |
|
* @param w the worker (via its WorkQueue) |
1639 |
< |
* @return a task or null of none found |
1639 |
> |
* @return a task or null if none found |
1640 |
|
*/ |
1641 |
|
private final ForkJoinTask<?> scan(WorkQueue w) { |
1642 |
< |
WorkQueue[] ws; // first update random seed |
1642 |
> |
WorkQueue[] ws; WorkQueue q; // first update random seed |
1643 |
|
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
1644 |
< |
int rs = runState, m; // volatile read order matters |
1644 |
> |
int ps = plock, m; // volatile read order matters |
1645 |
|
if ((ws = workQueues) != null && (m = ws.length - 1) > 0) { |
1646 |
< |
int ec = w.eventCount; // ec is negative if inactive |
1647 |
< |
int step = (r >>> 16) | 1; // relative prime |
1648 |
< |
for (int j = (m + 1) << 2; ; r += step) { |
1649 |
< |
WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b; |
1646 |
> |
int ec = w.eventCount; // ec is negative if inactive |
1647 |
> |
int step = (r >>> 16) | 1; // relatively prime |
1648 |
> |
for (int j = (m + 1) << 2; ; --j, r += step) { |
1649 |
> |
ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b, n; |
1650 |
|
if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 && |
1651 |
< |
(a = q.array) != null) { // probably nonempty |
1651 |
> |
(a = q.array) != null) { // probably nonempty |
1652 |
|
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1653 |
|
t = (ForkJoinTask<?>)U.getObjectVolatile(a, i); |
1654 |
|
if (q.base == b && ec >= 0 && t != null && |
1655 |
|
U.compareAndSwapObject(a, i, t, null)) { |
1656 |
< |
if (q.top - (q.base = b + 1) > 1) |
1657 |
< |
signalWork(); // help pushes signal |
1658 |
< |
return t; |
1659 |
< |
} |
1660 |
< |
else if (ec < 0 || j <= m) { |
1661 |
< |
rs = 0; // mark scan as imcomplete |
1662 |
< |
break; // caller can retry after release |
1656 |
> |
if ((n = q.top - (q.base = b + 1)) > 0) |
1657 |
> |
signalWork(q, n); |
1658 |
> |
return t; // taken |
1659 |
> |
} |
1660 |
> |
if (j < m || (ec < 0 && (ec = w.eventCount) < 0)) { |
1661 |
> |
if ((n = q.queueSize() - 1) > 0) |
1662 |
> |
signalWork(q, n); |
1663 |
> |
break; // let caller retry after signal |
1664 |
|
} |
1665 |
|
} |
1666 |
< |
if (--j < 0) |
1667 |
< |
break; |
1668 |
< |
} |
1669 |
< |
|
1670 |
< |
long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns; |
1671 |
< |
if (e < 0) // decode ctl on empty scan |
1672 |
< |
w.runState = -1; // pool is terminating |
1673 |
< |
else if (rs == 0 || rs != runState) { // incomplete scan |
1674 |
< |
WorkQueue v; Thread p; // try to release a waiter |
1675 |
< |
if (e > 0 && a < 0 && w.eventCount == ec && |
1676 |
< |
(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) { |
1677 |
< |
long nc = ((long)(v.nextWait & E_MASK) | |
1678 |
< |
((c + AC_UNIT) & (AC_MASK|TC_MASK))); |
1679 |
< |
if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) { |
1680 |
< |
v.eventCount = (e + E_SEQ) & E_MASK; |
1681 |
< |
if ((p = v.parker) != null) |
1682 |
< |
U.unpark(p); |
1666 |
> |
else if (j < 0) { // end of scan |
1667 |
> |
long c = ctl; int e; |
1668 |
> |
if (plock != ps) // incomplete sweep |
1669 |
> |
break; |
1670 |
> |
if ((e = (int)c) < 0) // pool is terminating |
1671 |
> |
w.qlock = -1; |
1672 |
> |
else if (ec >= 0) { // try to enqueue/inactivate |
1673 |
> |
long nc = ((long)ec | |
1674 |
> |
((c - AC_UNIT) & (AC_MASK|TC_MASK))); |
1675 |
> |
w.nextWait = e; |
1676 |
> |
w.eventCount = ec | INT_SIGN; // mark as inactive |
1677 |
> |
if (ctl != c || |
1678 |
> |
!U.compareAndSwapLong(this, CTL, c, nc)) |
1679 |
> |
w.eventCount = ec; // unmark on CAS failure |
1680 |
> |
else if ((int)(c >> AC_SHIFT) == 1 - parallelism) |
1681 |
> |
idleAwaitWork(w, nc, c); // quiescent |
1682 |
> |
} |
1683 |
> |
else if (w.seed >= 0 && w.eventCount < 0) { |
1684 |
> |
Thread wt = Thread.currentThread(); |
1685 |
> |
Thread.interrupted(); // clear status |
1686 |
> |
U.putObject(wt, PARKBLOCKER, this); |
1687 |
> |
w.parker = wt; // emulate LockSupport.park |
1688 |
> |
if (w.eventCount < 0) // recheck |
1689 |
> |
U.park(false, 0L); |
1690 |
> |
w.parker = null; |
1691 |
> |
U.putObject(wt, PARKBLOCKER, null); |
1692 |
|
} |
1693 |
< |
} |
1567 |
< |
} |
1568 |
< |
else if (ec >= 0) { // try to enqueue/inactivate |
1569 |
< |
long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); |
1570 |
< |
w.nextWait = e; |
1571 |
< |
w.eventCount = ec | INT_SIGN; // mark as inactive |
1572 |
< |
if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc)) |
1573 |
< |
w.eventCount = ec; // unmark on CAS failure |
1574 |
< |
else { |
1575 |
< |
if ((ns = w.nsteals) != 0) { |
1576 |
< |
w.nsteals = 0; // set rescans if ran task |
1577 |
< |
w.rescans = (a > 0) ? 0 : a + parallelism; |
1578 |
< |
w.totalSteals += ns; |
1579 |
< |
} |
1580 |
< |
if (a == 1 - parallelism) // quiescent |
1581 |
< |
idleAwaitWork(w, nc, c); |
1582 |
< |
} |
1583 |
< |
} |
1584 |
< |
else if (w.eventCount < 0) { // already queued |
1585 |
< |
if ((nr = w.rescans) > 0) { // continue rescanning |
1586 |
< |
int ac = a + parallelism; |
1587 |
< |
if (((w.rescans = (ac < nr) ? ac : nr - 1) & 3) == 0) |
1588 |
< |
Thread.yield(); // yield before block |
1589 |
< |
} |
1590 |
< |
else { |
1591 |
< |
Thread.interrupted(); // clear status |
1592 |
< |
Thread wt = Thread.currentThread(); |
1593 |
< |
U.putObject(wt, PARKBLOCKER, this); |
1594 |
< |
w.parker = wt; // emulate LockSupport.park |
1595 |
< |
if (w.eventCount < 0) // recheck |
1596 |
< |
U.park(false, 0L); |
1597 |
< |
w.parker = null; |
1598 |
< |
U.putObject(wt, PARKBLOCKER, null); |
1693 |
> |
break; |
1694 |
|
} |
1695 |
|
} |
1696 |
|
} |
1700 |
|
/** |
1701 |
|
* If inactivating worker w has caused the pool to become |
1702 |
|
* quiescent, checks for pool termination, and, so long as this is |
1703 |
< |
* not the only worker, waits for event for up to SHRINK_RATE |
1704 |
< |
* nanosecs. On timeout, if ctl has not changed, terminates the |
1703 |
> |
* not the only worker, waits for event for up to a given |
1704 |
> |
* duration. On timeout, if ctl has not changed, terminates the |
1705 |
|
* worker, which will in turn wake up another worker to possibly |
1706 |
|
* repeat this process. |
1707 |
|
* |
1710 |
|
* @param prevCtl the ctl value to restore if thread is terminated |
1711 |
|
*/ |
1712 |
|
private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) { |
1713 |
< |
if (w.eventCount < 0 && !tryTerminate(false, false) && |
1714 |
< |
(int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) { |
1713 |
> |
if (w.eventCount < 0 && |
1714 |
> |
(this == commonPool || !tryTerminate(false, false)) && |
1715 |
> |
(int)prevCtl != 0) { |
1716 |
> |
int dc = -(short)(currentCtl >>> TC_SHIFT); |
1717 |
> |
long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT; |
1718 |
> |
long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop |
1719 |
|
Thread wt = Thread.currentThread(); |
1621 |
– |
Thread.yield(); // yield before block |
1720 |
|
while (ctl == currentCtl) { |
1623 |
– |
long startTime = System.nanoTime(); |
1721 |
|
Thread.interrupted(); // timed variant of version in scan() |
1722 |
|
U.putObject(wt, PARKBLOCKER, this); |
1723 |
|
w.parker = wt; |
1724 |
|
if (ctl == currentCtl) |
1725 |
< |
U.park(false, SHRINK_RATE); |
1725 |
> |
U.park(false, parkTime); |
1726 |
|
w.parker = null; |
1727 |
|
U.putObject(wt, PARKBLOCKER, null); |
1728 |
|
if (ctl != currentCtl) |
1729 |
|
break; |
1730 |
< |
if (System.nanoTime() - startTime >= SHRINK_TIMEOUT && |
1730 |
> |
if (deadline - System.nanoTime() <= 0L && |
1731 |
|
U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) { |
1732 |
|
w.eventCount = (w.eventCount + E_SEQ) | E_MASK; |
1733 |
< |
w.runState = -1; // shrink |
1733 |
> |
w.qlock = -1; // shrink |
1734 |
|
break; |
1735 |
|
} |
1736 |
|
} |
1738 |
|
} |
1739 |
|
|
1740 |
|
/** |
1741 |
+ |
* Scans through queues looking for work while joining a task; |
1742 |
+ |
* if any are present, signals. |
1743 |
+ |
* |
1744 |
+ |
* @param task to return early if done |
1745 |
+ |
* @param origin an index to start scan |
1746 |
+ |
*/ |
1747 |
+ |
final int helpSignal(ForkJoinTask<?> task, int origin) { |
1748 |
+ |
WorkQueue[] ws; WorkQueue q; int m, n, s; |
1749 |
+ |
if (task != null && (ws = workQueues) != null && |
1750 |
+ |
(m = ws.length - 1) >= 0) { |
1751 |
+ |
for (int i = 0; i <= m; ++i) { |
1752 |
+ |
if ((s = task.status) < 0) |
1753 |
+ |
return s; |
1754 |
+ |
if ((q = ws[(i + origin) & m]) != null && |
1755 |
+ |
(n = q.queueSize()) > 0) { |
1756 |
+ |
signalWork(q, n); |
1757 |
+ |
if ((int)(ctl >> AC_SHIFT) >= 0) |
1758 |
+ |
break; |
1759 |
+ |
} |
1760 |
+ |
} |
1761 |
+ |
} |
1762 |
+ |
return 0; |
1763 |
+ |
} |
1764 |
+ |
|
1765 |
+ |
/** |
1766 |
|
* Tries to locate and execute tasks for a stealer of the given |
1767 |
|
* task, or in turn one of its stealers, Traces currentSteal -> |
1768 |
|
* currentJoin links looking for a thread working on a descendant |
1849 |
|
} |
1850 |
|
|
1851 |
|
/** |
1852 |
< |
* If task is at base of some steal queue, steals and executes it. |
1852 |
> |
* Analog of tryHelpStealer for CountedCompleters. Tries to steal |
1853 |
> |
* and run tasks within the target's computation |
1854 |
> |
* |
1855 |
> |
* @param task the task to join |
1856 |
> |
* @param mode if shared, exit upon completing any task |
1857 |
> |
* if all workers are active |
1858 |
|
* |
1732 |
– |
* @param joiner the joining worker |
1733 |
– |
* @param task the task |
1859 |
|
*/ |
1860 |
< |
private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) { |
1861 |
< |
WorkQueue[] ws; |
1862 |
< |
if ((ws = workQueues) != null) { |
1863 |
< |
for (int j = 1; j < ws.length && task.status >= 0; j += 2) { |
1864 |
< |
WorkQueue q = ws[j]; |
1865 |
< |
if (q != null && q.pollFor(task)) { |
1866 |
< |
joiner.runSubtask(task); |
1867 |
< |
break; |
1860 |
> |
private int helpComplete(ForkJoinTask<?> task, int mode) { |
1861 |
> |
WorkQueue[] ws; WorkQueue q; int m, n, s; |
1862 |
> |
if (task != null && (ws = workQueues) != null && |
1863 |
> |
(m = ws.length - 1) >= 0) { |
1864 |
> |
for (int j = 1, origin = j;;) { |
1865 |
> |
if ((s = task.status) < 0) |
1866 |
> |
return s; |
1867 |
> |
if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) { |
1868 |
> |
origin = j; |
1869 |
> |
if (mode == SHARED_QUEUE && (int)(ctl >> AC_SHIFT) >= 0) |
1870 |
> |
break; |
1871 |
|
} |
1872 |
+ |
else if ((j = (j + 2) & m) == origin) |
1873 |
+ |
break; |
1874 |
|
} |
1875 |
|
} |
1876 |
+ |
return 0; |
1877 |
|
} |
1878 |
|
|
1879 |
|
/** |
1880 |
|
* Tries to decrement active count (sometimes implicitly) and |
1881 |
|
* possibly release or create a compensating worker in preparation |
1882 |
|
* for blocking. Fails on contention or termination. Otherwise, |
1883 |
< |
* adds a new thread if no idle workers are available and either |
1884 |
< |
* pool would become completely starved or: (at least half |
1754 |
< |
* starved, and fewer than 50% spares exist, and there is at least |
1755 |
< |
* one task apparently available). Even though the availability |
1756 |
< |
* check requires a full scan, it is worthwhile in reducing false |
1757 |
< |
* alarms. |
1758 |
< |
* |
1759 |
< |
* @param task if non-null, a task being waited for |
1760 |
< |
* @param blocker if non-null, a blocker being waited for |
1761 |
< |
* @return true if the caller can block, else should recheck and retry |
1883 |
> |
* adds a new thread if no idle workers are available and pool |
1884 |
> |
* may become starved. |
1885 |
|
*/ |
1886 |
< |
final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) { |
1887 |
< |
int pc = parallelism, e; |
1888 |
< |
long c = ctl; |
1889 |
< |
WorkQueue[] ws = workQueues; |
1890 |
< |
if ((e = (int)c) >= 0 && ws != null) { |
1891 |
< |
int u, a, ac, hc; |
1892 |
< |
int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc; |
1893 |
< |
boolean replace = false; |
1894 |
< |
if ((a = u >> UAC_SHIFT) <= 0) { |
1895 |
< |
if ((ac = a + pc) <= 1) |
1896 |
< |
replace = true; |
1897 |
< |
else if ((e > 0 || (task != null && |
1898 |
< |
ac <= (hc = pc >>> 1) && tc < pc + hc))) { |
1776 |
< |
WorkQueue w; |
1777 |
< |
for (int j = 0; j < ws.length; ++j) { |
1778 |
< |
if ((w = ws[j]) != null && !w.isEmpty()) { |
1779 |
< |
replace = true; |
1780 |
< |
break; // in compensation range and tasks available |
1781 |
< |
} |
1782 |
< |
} |
1886 |
> |
final boolean tryCompensate() { |
1887 |
> |
int pc = parallelism, e, u, i, tc; long c; |
1888 |
> |
WorkQueue[] ws; WorkQueue w; Thread p; |
1889 |
> |
if ((e = (int)(c = ctl)) >= 0 && (ws = workQueues) != null) { |
1890 |
> |
if (e != 0 && (i = e & SMASK) < ws.length && |
1891 |
> |
(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { |
1892 |
> |
long nc = ((long)(w.nextWait & E_MASK) | |
1893 |
> |
(c & (AC_MASK|TC_MASK))); |
1894 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1895 |
> |
w.eventCount = (e + E_SEQ) & E_MASK; |
1896 |
> |
if ((p = w.parker) != null) |
1897 |
> |
U.unpark(p); |
1898 |
> |
return true; // replace with idle worker |
1899 |
|
} |
1900 |
|
} |
1901 |
< |
if ((task == null || task.status >= 0) && // recheck need to block |
1902 |
< |
(blocker == null || !blocker.isReleasable()) && ctl == c) { |
1903 |
< |
if (!replace) { // no compensation |
1904 |
< |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1905 |
< |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1906 |
< |
return true; |
1907 |
< |
} |
1908 |
< |
else if (e != 0) { // release an idle worker |
1909 |
< |
WorkQueue w; Thread p; int i; |
1910 |
< |
if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) { |
1911 |
< |
long nc = ((long)(w.nextWait & E_MASK) | |
1912 |
< |
(c & (AC_MASK|TC_MASK))); |
1913 |
< |
if (w.eventCount == (e | INT_SIGN) && |
1914 |
< |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1915 |
< |
w.eventCount = (e + E_SEQ) & E_MASK; |
1916 |
< |
if ((p = w.parker) != null) |
1801 |
< |
U.unpark(p); |
1901 |
> |
else if ((short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) >= 0 && |
1902 |
> |
(u >> UAC_SHIFT) + pc > 1) { |
1903 |
> |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1904 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1905 |
> |
return true; // no compensation |
1906 |
> |
} |
1907 |
> |
else if ((tc = u + pc) < MAX_CAP) { |
1908 |
> |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1909 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1910 |
> |
Throwable ex = null; |
1911 |
> |
ForkJoinWorkerThread wt = null; |
1912 |
> |
try { |
1913 |
> |
ForkJoinWorkerThreadFactory fac; |
1914 |
> |
if ((fac = factory) != null && |
1915 |
> |
(wt = fac.newThread(this)) != null) { |
1916 |
> |
wt.start(); |
1917 |
|
return true; |
1918 |
|
} |
1919 |
+ |
} catch (Throwable rex) { |
1920 |
+ |
ex = rex; |
1921 |
|
} |
1922 |
< |
} |
1806 |
< |
else if (tc < MAX_CAP) { // create replacement |
1807 |
< |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1808 |
< |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1809 |
< |
addWorker(); |
1810 |
< |
return true; |
1811 |
< |
} |
1922 |
> |
deregisterWorker(wt, ex); // adjust counts etc |
1923 |
|
} |
1924 |
|
} |
1925 |
|
} |
1934 |
|
* @return task status on exit |
1935 |
|
*/ |
1936 |
|
final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) { |
1937 |
< |
int s; |
1938 |
< |
if ((s = task.status) >= 0) { |
1937 |
> |
int s = 0; |
1938 |
> |
if (joiner != null && task != null && (s = task.status) >= 0) { |
1939 |
|
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
1940 |
|
joiner.currentJoin = task; |
1941 |
< |
long startTime = 0L; |
1942 |
< |
for (int k = 0;;) { |
1943 |
< |
if ((s = (joiner.isEmpty() ? // try to help |
1944 |
< |
tryHelpStealer(joiner, task) : |
1945 |
< |
joiner.tryRemoveAndExec(task))) == 0 && |
1946 |
< |
(s = task.status) >= 0) { |
1947 |
< |
if (k == 0) { |
1948 |
< |
startTime = System.nanoTime(); |
1949 |
< |
tryPollForAndExec(joiner, task); // check uncommon case |
1950 |
< |
} |
1951 |
< |
else if ((k & (MAX_HELP - 1)) == 0 && |
1952 |
< |
System.nanoTime() - startTime >= |
1953 |
< |
COMPENSATION_DELAY && |
1954 |
< |
tryCompensate(task, null)) { |
1955 |
< |
if (task.trySetSignal()) { |
1956 |
< |
synchronized (task) { |
1957 |
< |
if (task.status >= 0) { |
1847 |
< |
try { // see ForkJoinTask |
1848 |
< |
task.wait(); // for explanation |
1849 |
< |
} catch (InterruptedException ie) { |
1850 |
< |
} |
1941 |
> |
do {} while ((s = task.status) >= 0 && |
1942 |
> |
joiner.queueSize() > 0 && |
1943 |
> |
joiner.tryRemoveAndExec(task)); // process local tasks |
1944 |
> |
if (s >= 0 && (s = task.status) >= 0 && |
1945 |
> |
(s = helpSignal(task, joiner.poolIndex)) >= 0 && |
1946 |
> |
(task instanceof CountedCompleter)) |
1947 |
> |
s = helpComplete(task, LIFO_QUEUE); |
1948 |
> |
while (s >= 0 && (s = task.status) >= 0) { |
1949 |
> |
if ((joiner.queueSize() > 0 || // try helping |
1950 |
> |
(s = tryHelpStealer(joiner, task)) == 0) && |
1951 |
> |
(s = task.status) >= 0 && tryCompensate()) { |
1952 |
> |
if (task.trySetSignal() && (s = task.status) >= 0) { |
1953 |
> |
synchronized (task) { |
1954 |
> |
if (task.status >= 0) { |
1955 |
> |
try { // see ForkJoinTask |
1956 |
> |
task.wait(); // for explanation |
1957 |
> |
} catch (InterruptedException ie) { |
1958 |
|
} |
1852 |
– |
else |
1853 |
– |
task.notifyAll(); |
1959 |
|
} |
1960 |
+ |
else |
1961 |
+ |
task.notifyAll(); |
1962 |
|
} |
1856 |
– |
long c; // re-activate |
1857 |
– |
do {} while (!U.compareAndSwapLong |
1858 |
– |
(this, CTL, c = ctl, c + AC_UNIT)); |
1963 |
|
} |
1964 |
+ |
long c; // re-activate |
1965 |
+ |
do {} while (!U.compareAndSwapLong |
1966 |
+ |
(this, CTL, c = ctl, c + AC_UNIT)); |
1967 |
|
} |
1861 |
– |
if (s < 0 || (s = task.status) < 0) { |
1862 |
– |
joiner.currentJoin = prevJoin; |
1863 |
– |
break; |
1864 |
– |
} |
1865 |
– |
else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1) |
1866 |
– |
Thread.yield(); // for politeness |
1968 |
|
} |
1969 |
+ |
joiner.currentJoin = prevJoin; |
1970 |
|
} |
1971 |
|
return s; |
1972 |
|
} |
1978 |
|
* |
1979 |
|
* @param joiner the joining worker |
1980 |
|
* @param task the task |
1879 |
– |
* @return task status on exit |
1981 |
|
*/ |
1982 |
< |
final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) { |
1982 |
> |
final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) { |
1983 |
|
int s; |
1984 |
< |
while ((s = task.status) >= 0 && |
1985 |
< |
(joiner.isEmpty() ? |
1986 |
< |
tryHelpStealer(joiner, task) : |
1987 |
< |
joiner.tryRemoveAndExec(task)) != 0) |
1988 |
< |
; |
1989 |
< |
return s; |
1984 |
> |
if (joiner != null && task != null && (s = task.status) >= 0) { |
1985 |
> |
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
1986 |
> |
joiner.currentJoin = task; |
1987 |
> |
do {} while ((s = task.status) >= 0 && |
1988 |
> |
joiner.queueSize() > 0 && |
1989 |
> |
joiner.tryRemoveAndExec(task)); |
1990 |
> |
if (s >= 0 && (s = task.status) >= 0 && |
1991 |
> |
(s = helpSignal(task, joiner.poolIndex)) >= 0 && |
1992 |
> |
(task instanceof CountedCompleter)) |
1993 |
> |
s = helpComplete(task, LIFO_QUEUE); |
1994 |
> |
if (s >= 0 && joiner.queueSize() == 0) { |
1995 |
> |
do {} while (task.status >= 0 && |
1996 |
> |
tryHelpStealer(joiner, task) > 0); |
1997 |
> |
} |
1998 |
> |
joiner.currentJoin = prevJoin; |
1999 |
> |
} |
2000 |
|
} |
2001 |
|
|
2002 |
|
/** |
2004 |
|
* during a random, then cyclic scan, else null. This method must |
2005 |
|
* be retried by caller if, by the time it tries to use the queue, |
2006 |
|
* it is empty. |
2007 |
+ |
* @param r a (random) seed for scanning |
2008 |
|
*/ |
2009 |
< |
private WorkQueue findNonEmptyStealQueue(WorkQueue w) { |
1898 |
< |
// Similar to loop in scan(), but ignoring submissions |
1899 |
< |
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
2009 |
> |
private WorkQueue findNonEmptyStealQueue(int r) { |
2010 |
|
int step = (r >>> 16) | 1; |
2011 |
|
for (WorkQueue[] ws;;) { |
2012 |
< |
int rs = runState, m; |
2012 |
> |
int ps = plock, m; |
2013 |
|
if ((ws = workQueues) == null || (m = ws.length - 1) < 1) |
2014 |
|
return null; |
2015 |
|
for (int j = (m + 1) << 2; ; r += step) { |
2016 |
|
WorkQueue q = ws[((r << 1) | 1) & m]; |
2017 |
< |
if (q != null && !q.isEmpty()) |
2017 |
> |
if (q != null && q.queueSize() > 0) |
2018 |
|
return q; |
2019 |
|
else if (--j < 0) { |
2020 |
< |
if (runState == rs) |
2020 |
> |
if (plock == ps) |
2021 |
|
return null; |
2022 |
|
break; |
2023 |
|
} |
2025 |
|
} |
2026 |
|
} |
2027 |
|
|
1918 |
– |
|
2028 |
|
/** |
2029 |
|
* Runs tasks until {@code isQuiescent()}. We piggyback on |
2030 |
|
* active count ctl maintenance, but rather than blocking |
2036 |
|
ForkJoinTask<?> localTask; // exhaust local queue |
2037 |
|
while ((localTask = w.nextLocalTask()) != null) |
2038 |
|
localTask.doExec(); |
2039 |
< |
WorkQueue q = findNonEmptyStealQueue(w); |
2039 |
> |
// Similar to loop in scan(), but ignoring submissions |
2040 |
> |
WorkQueue q = findNonEmptyStealQueue(w.nextSeed()); |
2041 |
|
if (q != null) { |
2042 |
|
ForkJoinTask<?> t; int b; |
2043 |
|
if (!active) { // re-establish active count |
2077 |
|
WorkQueue q; int b; |
2078 |
|
if ((t = w.nextLocalTask()) != null) |
2079 |
|
return t; |
2080 |
< |
if ((q = findNonEmptyStealQueue(w)) == null) |
2080 |
> |
if ((q = findNonEmptyStealQueue(w.nextSeed())) == null) |
2081 |
|
return null; |
2082 |
|
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
2083 |
|
return t; |
2085 |
|
} |
2086 |
|
|
2087 |
|
/** |
2088 |
< |
* Returns the approximate (non-atomic) number of idle threads per |
2089 |
< |
* active thread to offset steal queue size for method |
2090 |
< |
* ForkJoinTask.getSurplusQueuedTaskCount(). |
2091 |
< |
*/ |
2092 |
< |
final int idlePerActive() { |
2093 |
< |
// Approximate at powers of two for small values, saturate past 4 |
2094 |
< |
int p = parallelism; |
2095 |
< |
int a = p + (int)(ctl >> AC_SHIFT); |
2096 |
< |
return (a > (p >>>= 1) ? 0 : |
2097 |
< |
a > (p >>>= 1) ? 1 : |
2098 |
< |
a > (p >>>= 1) ? 2 : |
2099 |
< |
a > (p >>>= 1) ? 4 : |
2100 |
< |
8); |
2088 |
> |
* Returns a cheap heuristic guide for task partitioning when |
2089 |
> |
* programmers, frameworks, tools, or languages have little or no |
2090 |
> |
* idea about task granularity. In essence by offering this |
2091 |
> |
* method, we ask users only about tradeoffs in overhead vs |
2092 |
> |
* expected throughput and its variance, rather than how finely to |
2093 |
> |
* partition tasks. |
2094 |
> |
* |
2095 |
> |
* In a steady state strict (tree-structured) computation, each |
2096 |
> |
* thread makes available for stealing enough tasks for other |
2097 |
> |
* threads to remain active. Inductively, if all threads play by |
2098 |
> |
* the same rules, each thread should make available only a |
2099 |
> |
* constant number of tasks. |
2100 |
> |
* |
2101 |
> |
* The minimum useful constant is just 1. But using a value of 1 |
2102 |
> |
* would require immediate replenishment upon each steal to |
2103 |
> |
* maintain enough tasks, which is infeasible. Further, |
2104 |
> |
* partitionings/granularities of offered tasks should minimize |
2105 |
> |
* steal rates, which in general means that threads nearer the top |
2106 |
> |
* of computation tree should generate more than those nearer the |
2107 |
> |
* bottom. In perfect steady state, each thread is at |
2108 |
> |
* approximately the same level of computation tree. However, |
2109 |
> |
* producing extra tasks amortizes the uncertainty of progress and |
2110 |
> |
* diffusion assumptions. |
2111 |
> |
* |
2112 |
> |
* So, users will want to use values larger, but not much larger |
2113 |
> |
* than 1 to both smooth over transient shortages and hedge |
2114 |
> |
* against uneven progress; as traded off against the cost of |
2115 |
> |
* extra task overhead. We leave the user to pick a threshold |
2116 |
> |
* value to compare with the results of this call to guide |
2117 |
> |
* decisions, but recommend values such as 3. |
2118 |
> |
* |
2119 |
> |
* When all threads are active, it is on average OK to estimate |
2120 |
> |
* surplus strictly locally. In steady-state, if one thread is |
2121 |
> |
* maintaining say 2 surplus tasks, then so are others. So we can |
2122 |
> |
* just use estimated queue length. However, this strategy alone |
2123 |
> |
* leads to serious mis-estimates in some non-steady-state |
2124 |
> |
* conditions (ramp-up, ramp-down, other stalls). We can detect |
2125 |
> |
* many of these by further considering the number of "idle" |
2126 |
> |
* threads, that are known to have zero queued tasks, so |
2127 |
> |
* compensate by a factor of (#idle/#active) threads. |
2128 |
> |
* |
2129 |
> |
* Note: The approximation of #busy workers as #active workers is |
2130 |
> |
* not very good under current signalling scheme, and should be |
2131 |
> |
* improved. |
2132 |
> |
*/ |
2133 |
> |
static int getSurplusQueuedTaskCount() { |
2134 |
> |
Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; |
2135 |
> |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) { |
2136 |
> |
int b = (q = (wt = (ForkJoinWorkerThread)t).workQueue).base; |
2137 |
> |
int p = (pool = wt.pool).parallelism; |
2138 |
> |
int a = (int)(pool.ctl >> AC_SHIFT) + p; |
2139 |
> |
return q.top - b - (a > (p >>>= 1) ? 0 : |
2140 |
> |
a > (p >>>= 1) ? 1 : |
2141 |
> |
a > (p >>>= 1) ? 2 : |
2142 |
> |
a > (p >>>= 1) ? 4 : |
2143 |
> |
8); |
2144 |
> |
} |
2145 |
> |
return 0; |
2146 |
|
} |
2147 |
|
|
2148 |
|
// Termination |
2162 |
|
* @return true if now terminating or terminated |
2163 |
|
*/ |
2164 |
|
private boolean tryTerminate(boolean now, boolean enable) { |
2165 |
< |
Mutex lock = this.lock; |
2165 |
> |
if (this == commonPool) // cannot shut down |
2166 |
> |
return false; |
2167 |
|
for (long c;;) { |
2168 |
|
if (((c = ctl) & STOP_BIT) != 0) { // already terminating |
2169 |
|
if ((short)(c >>> TC_SHIFT) == -parallelism) { |
2170 |
< |
lock.lock(); // don't need try/finally |
2171 |
< |
termination.signalAll(); // signal when 0 workers |
2172 |
< |
lock.unlock(); |
2170 |
> |
synchronized (this) { |
2171 |
> |
notifyAll(); // signal when 0 workers |
2172 |
> |
} |
2173 |
|
} |
2174 |
|
return true; |
2175 |
|
} |
2176 |
< |
if (runState >= 0) { // not yet enabled |
2176 |
> |
if (plock >= 0) { // not yet enabled |
2177 |
> |
int ps; |
2178 |
|
if (!enable) |
2179 |
|
return false; |
2180 |
< |
lock.lock(); |
2181 |
< |
runState |= SHUTDOWN; |
2182 |
< |
lock.unlock(); |
2180 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
2181 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
2182 |
> |
ps = acquirePlock(); |
2183 |
> |
int nps = SHUTDOWN; |
2184 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
2185 |
> |
releasePlock(nps); |
2186 |
|
} |
2187 |
|
if (!now) { // check if idle & no tasks |
2188 |
|
if ((int)(c >> AC_SHIFT) != -parallelism || |
2205 |
|
int n = ws.length; |
2206 |
|
for (int i = 0; i < n; ++i) { |
2207 |
|
if ((w = ws[i]) != null) { |
2208 |
< |
w.runState = -1; |
2208 |
> |
w.qlock = -1; |
2209 |
|
if (pass > 0) { |
2210 |
|
w.cancelAll(); |
2211 |
|
if (pass > 1) |
2224 |
|
if (w.eventCount == (e | INT_SIGN) && |
2225 |
|
U.compareAndSwapLong(this, CTL, cc, nc)) { |
2226 |
|
w.eventCount = (e + E_SEQ) & E_MASK; |
2227 |
< |
w.runState = -1; |
2227 |
> |
w.qlock = -1; |
2228 |
|
if ((p = w.parker) != null) |
2229 |
|
U.unpark(p); |
2230 |
|
} |
2235 |
|
} |
2236 |
|
} |
2237 |
|
|
2238 |
+ |
// external operations on common pool |
2239 |
+ |
|
2240 |
+ |
/** |
2241 |
+ |
* Returns common pool queue for a thread that has submitted at |
2242 |
+ |
* least one task. |
2243 |
+ |
*/ |
2244 |
+ |
static WorkQueue commonSubmitterQueue() { |
2245 |
+ |
ForkJoinPool p; WorkQueue[] ws; int m; Submitter z; |
2246 |
+ |
return ((z = submitters.get()) != null && |
2247 |
+ |
(p = commonPool) != null && |
2248 |
+ |
(ws = p.workQueues) != null && |
2249 |
+ |
(m = ws.length - 1) >= 0) ? |
2250 |
+ |
ws[m & z.seed & SQMASK] : null; |
2251 |
+ |
} |
2252 |
+ |
|
2253 |
+ |
/** |
2254 |
+ |
* Tries to pop the given task from submitter's queue in common pool. |
2255 |
+ |
*/ |
2256 |
+ |
static boolean tryExternalUnpush(ForkJoinTask<?> t) { |
2257 |
+ |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z; |
2258 |
+ |
ForkJoinTask<?>[] a; int m, s; long j; |
2259 |
+ |
if ((z = submitters.get()) != null && |
2260 |
+ |
(p = commonPool) != null && |
2261 |
+ |
(ws = p.workQueues) != null && |
2262 |
+ |
(m = ws.length - 1) >= 0 && |
2263 |
+ |
(q = ws[m & z.seed & SQMASK]) != null && |
2264 |
+ |
(s = q.top) != q.base && |
2265 |
+ |
(a = q.array) != null && |
2266 |
+ |
U.getObjectVolatile |
2267 |
+ |
(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t && |
2268 |
+ |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2269 |
+ |
if (q.array == a && q.top == s && // recheck |
2270 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
2271 |
+ |
q.top = s - 1; |
2272 |
+ |
q.qlock = 0; |
2273 |
+ |
return true; |
2274 |
+ |
} |
2275 |
+ |
q.qlock = 0; |
2276 |
+ |
} |
2277 |
+ |
return false; |
2278 |
+ |
} |
2279 |
+ |
|
2280 |
+ |
/** |
2281 |
+ |
* Tries to pop and run local tasks within the same computation |
2282 |
+ |
* as the given root. On failure, tries to help complete from |
2283 |
+ |
* other queues via helpComplete. |
2284 |
+ |
*/ |
2285 |
+ |
private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) { |
2286 |
+ |
ForkJoinTask<?>[] a; int m; |
2287 |
+ |
if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 && |
2288 |
+ |
root != null && root.status >= 0) { |
2289 |
+ |
for (;;) { |
2290 |
+ |
int s; Object o; CountedCompleter<?> task = null; |
2291 |
+ |
if ((s = q.top) - q.base > 0) { |
2292 |
+ |
long j = ((m & (s - 1)) << ASHIFT) + ABASE; |
2293 |
+ |
if ((o = U.getObject(a, j)) != null && |
2294 |
+ |
(o instanceof CountedCompleter)) { |
2295 |
+ |
CountedCompleter<?> t = (CountedCompleter<?>)o, r = t; |
2296 |
+ |
do { |
2297 |
+ |
if (r == root) { |
2298 |
+ |
if (U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2299 |
+ |
if (q.array == a && q.top == s && |
2300 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
2301 |
+ |
q.top = s - 1; |
2302 |
+ |
task = t; |
2303 |
+ |
} |
2304 |
+ |
q.qlock = 0; |
2305 |
+ |
} |
2306 |
+ |
break; |
2307 |
+ |
} |
2308 |
+ |
} while((r = r.completer) != null); |
2309 |
+ |
} |
2310 |
+ |
} |
2311 |
+ |
if (task != null) |
2312 |
+ |
task.doExec(); |
2313 |
+ |
if (root.status < 0 || (int)(ctl >> AC_SHIFT) >= 0) |
2314 |
+ |
break; |
2315 |
+ |
if (task == null) { |
2316 |
+ |
if (helpSignal(root, q.poolIndex) >= 0) |
2317 |
+ |
helpComplete(root, SHARED_QUEUE); |
2318 |
+ |
break; |
2319 |
+ |
} |
2320 |
+ |
} |
2321 |
+ |
} |
2322 |
+ |
} |
2323 |
+ |
|
2324 |
+ |
/** |
2325 |
+ |
* Tries to help execute or signal availability of the given task |
2326 |
+ |
* from submitter's queue in common pool. |
2327 |
+ |
*/ |
2328 |
+ |
static void externalHelpJoin(ForkJoinTask<?> t) { |
2329 |
+ |
// Some hard-to-avoid overlap with tryExternalUnpush |
2330 |
+ |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z; |
2331 |
+ |
ForkJoinTask<?>[] a; int m, s, n; long j; |
2332 |
+ |
if (t != null && t.status >= 0 && |
2333 |
+ |
(z = submitters.get()) != null && |
2334 |
+ |
(p = commonPool) != null && |
2335 |
+ |
(ws = p.workQueues) != null && |
2336 |
+ |
(m = ws.length - 1) >= 0 && |
2337 |
+ |
(q = ws[m & z.seed & SQMASK]) != null && |
2338 |
+ |
(a = q.array) != null) { |
2339 |
+ |
if ((s = q.top) != q.base && |
2340 |
+ |
U.getObjectVolatile |
2341 |
+ |
(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t && |
2342 |
+ |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2343 |
+ |
if (q.array == a && q.top == s && |
2344 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
2345 |
+ |
q.top = s - 1; |
2346 |
+ |
q.qlock = 0; |
2347 |
+ |
t.doExec(); |
2348 |
+ |
} |
2349 |
+ |
else |
2350 |
+ |
q.qlock = 0; |
2351 |
+ |
} |
2352 |
+ |
if (t.status >= 0) { |
2353 |
+ |
if (t instanceof CountedCompleter) |
2354 |
+ |
p.externalHelpComplete(q, t); |
2355 |
+ |
else |
2356 |
+ |
p.helpSignal(t, q.poolIndex); |
2357 |
+ |
} |
2358 |
+ |
} |
2359 |
+ |
} |
2360 |
+ |
|
2361 |
+ |
/** |
2362 |
+ |
* Restricted version of helpQuiescePool for external callers |
2363 |
+ |
*/ |
2364 |
+ |
static void externalHelpQuiescePool() { |
2365 |
+ |
ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b; |
2366 |
+ |
int r = ThreadLocalRandom.current().nextInt(); |
2367 |
+ |
if ((p = commonPool) != null && |
2368 |
+ |
(q = p.findNonEmptyStealQueue(r)) != null && |
2369 |
+ |
(b = q.base) - q.top < 0 && |
2370 |
+ |
(t = q.pollAt(b)) != null) |
2371 |
+ |
t.doExec(); |
2372 |
+ |
} |
2373 |
+ |
|
2374 |
|
// Exported methods |
2375 |
|
|
2376 |
|
// Constructors |
2448 |
|
this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE; |
2449 |
|
long np = (long)(-parallelism); // offset ctl counts |
2450 |
|
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2451 |
< |
// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2. |
2156 |
< |
int n = parallelism - 1; |
2157 |
< |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16; |
2158 |
< |
int size = (n + 1) << 1; // #slots = 2*#workers |
2159 |
< |
this.submitMask = size - 1; // room for max # of submit queues |
2160 |
< |
this.workQueues = new WorkQueue[size]; |
2161 |
< |
this.termination = (this.lock = new Mutex()).newCondition(); |
2162 |
< |
this.stealCount = new AtomicLong(); |
2163 |
< |
this.nextWorkerNumber = new AtomicInteger(); |
2164 |
< |
int pn = poolNumberGenerator.incrementAndGet(); |
2451 |
> |
int pn = nextPoolId(); |
2452 |
|
StringBuilder sb = new StringBuilder("ForkJoinPool-"); |
2453 |
|
sb.append(Integer.toString(pn)); |
2454 |
|
sb.append("-worker-"); |
2455 |
|
this.workerNamePrefix = sb.toString(); |
2456 |
< |
lock.lock(); |
2457 |
< |
this.runState = 1; // set init flag |
2458 |
< |
lock.unlock(); |
2456 |
> |
} |
2457 |
> |
|
2458 |
> |
/** |
2459 |
> |
* Constructor for common pool, suitable only for static initialization. |
2460 |
> |
* Basically the same as above, but uses smallest possible initial footprint. |
2461 |
> |
*/ |
2462 |
> |
ForkJoinPool(int parallelism, long ctl, |
2463 |
> |
ForkJoinWorkerThreadFactory factory, |
2464 |
> |
Thread.UncaughtExceptionHandler handler) { |
2465 |
> |
this.parallelism = parallelism; |
2466 |
> |
this.ctl = ctl; |
2467 |
> |
this.factory = factory; |
2468 |
> |
this.ueh = handler; |
2469 |
> |
this.localMode = LIFO_QUEUE; |
2470 |
> |
this.workerNamePrefix = "ForkJoinPool.commonPool-worker-"; |
2471 |
> |
} |
2472 |
> |
|
2473 |
> |
/** |
2474 |
> |
* Returns the common pool instance. |
2475 |
> |
* |
2476 |
> |
* @return the common pool instance |
2477 |
> |
*/ |
2478 |
> |
public static ForkJoinPool commonPool() { |
2479 |
> |
return commonPool; // cannot be null (if so, a static init error) |
2480 |
|
} |
2481 |
|
|
2482 |
|
// Execution methods |
2500 |
|
public <T> T invoke(ForkJoinTask<T> task) { |
2501 |
|
if (task == null) |
2502 |
|
throw new NullPointerException(); |
2503 |
< |
doSubmit(task); |
2503 |
> |
externalPush(task); |
2504 |
|
return task.join(); |
2505 |
|
} |
2506 |
|
|
2515 |
|
public void execute(ForkJoinTask<?> task) { |
2516 |
|
if (task == null) |
2517 |
|
throw new NullPointerException(); |
2518 |
< |
doSubmit(task); |
2518 |
> |
externalPush(task); |
2519 |
|
} |
2520 |
|
|
2521 |
|
// AbstractExecutorService methods |
2533 |
|
job = (ForkJoinTask<?>) task; |
2534 |
|
else |
2535 |
|
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2536 |
< |
doSubmit(job); |
2536 |
> |
externalPush(job); |
2537 |
|
} |
2538 |
|
|
2539 |
|
/** |
2548 |
|
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2549 |
|
if (task == null) |
2550 |
|
throw new NullPointerException(); |
2551 |
< |
doSubmit(task); |
2551 |
> |
externalPush(task); |
2552 |
|
return task; |
2553 |
|
} |
2554 |
|
|
2559 |
|
*/ |
2560 |
|
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2561 |
|
ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task); |
2562 |
< |
doSubmit(job); |
2562 |
> |
externalPush(job); |
2563 |
|
return job; |
2564 |
|
} |
2565 |
|
|
2570 |
|
*/ |
2571 |
|
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2572 |
|
ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result); |
2573 |
< |
doSubmit(job); |
2573 |
> |
externalPush(job); |
2574 |
|
return job; |
2575 |
|
} |
2576 |
|
|
2587 |
|
job = (ForkJoinTask<?>) task; |
2588 |
|
else |
2589 |
|
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2590 |
< |
doSubmit(job); |
2590 |
> |
externalPush(job); |
2591 |
|
return job; |
2592 |
|
} |
2593 |
|
|
2609 |
|
try { |
2610 |
|
for (Callable<T> t : tasks) { |
2611 |
|
ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t); |
2612 |
< |
doSubmit(f); |
2612 |
> |
externalPush(f); |
2613 |
|
fs.add(f); |
2614 |
|
} |
2615 |
|
for (ForkJoinTask<T> f : fs) |
2652 |
|
} |
2653 |
|
|
2654 |
|
/** |
2655 |
+ |
* Returns the targeted parallelism level of the common pool. |
2656 |
+ |
* |
2657 |
+ |
* @return the targeted parallelism level of the common pool |
2658 |
+ |
*/ |
2659 |
+ |
public static int getCommonPoolParallelism() { |
2660 |
+ |
return commonPoolParallelism; |
2661 |
+ |
} |
2662 |
+ |
|
2663 |
+ |
/** |
2664 |
|
* Returns the number of worker threads that have started but not |
2665 |
|
* yet terminated. The result returned by this method may differ |
2666 |
|
* from {@link #getParallelism} when threads are created to |
2741 |
|
* @return the number of steals |
2742 |
|
*/ |
2743 |
|
public long getStealCount() { |
2744 |
< |
long count = stealCount.get(); |
2744 |
> |
long count = stealCount; |
2745 |
|
WorkQueue[] ws; WorkQueue w; |
2746 |
|
if ((ws = workQueues) != null) { |
2747 |
|
for (int i = 1; i < ws.length; i += 2) { |
2748 |
|
if ((w = ws[i]) != null) |
2749 |
< |
count += w.totalSteals; |
2749 |
> |
count += w.nsteals; |
2750 |
|
} |
2751 |
|
} |
2752 |
|
return count; |
2803 |
|
WorkQueue[] ws; WorkQueue w; |
2804 |
|
if ((ws = workQueues) != null) { |
2805 |
|
for (int i = 0; i < ws.length; i += 2) { |
2806 |
< |
if ((w = ws[i]) != null && !w.isEmpty()) |
2806 |
> |
if ((w = ws[i]) != null && w.queueSize() != 0) |
2807 |
|
return true; |
2808 |
|
} |
2809 |
|
} |
2871 |
|
public String toString() { |
2872 |
|
// Use a single pass through workQueues to collect counts |
2873 |
|
long qt = 0L, qs = 0L; int rc = 0; |
2874 |
< |
long st = stealCount.get(); |
2874 |
> |
long st = stealCount; |
2875 |
|
long c = ctl; |
2876 |
|
WorkQueue[] ws; WorkQueue w; |
2877 |
|
if ((ws = workQueues) != null) { |
2882 |
|
qs += size; |
2883 |
|
else { |
2884 |
|
qt += size; |
2885 |
< |
st += w.totalSteals; |
2885 |
> |
st += w.nsteals; |
2886 |
|
if (w.isApparentlyUnblocked()) |
2887 |
|
++rc; |
2888 |
|
} |
2898 |
|
if ((c & STOP_BIT) != 0) |
2899 |
|
level = (tc == 0) ? "Terminated" : "Terminating"; |
2900 |
|
else |
2901 |
< |
level = runState < 0 ? "Shutting down" : "Running"; |
2901 |
> |
level = plock < 0 ? "Shutting down" : "Running"; |
2902 |
|
return super.toString() + |
2903 |
|
"[" + level + |
2904 |
|
", parallelism = " + pc + |
2912 |
|
} |
2913 |
|
|
2914 |
|
/** |
2915 |
< |
* Initiates an orderly shutdown in which previously submitted |
2916 |
< |
* tasks are executed, but no new tasks will be accepted. |
2917 |
< |
* Invocation has no additional effect if already shut down. |
2918 |
< |
* Tasks that are in the process of being submitted concurrently |
2919 |
< |
* during the course of this method may or may not be rejected. |
2915 |
> |
* Possibly initiates an orderly shutdown in which previously |
2916 |
> |
* submitted tasks are executed, but no new tasks will be |
2917 |
> |
* accepted. Invocation has no effect on execution state if this |
2918 |
> |
* is the {@link #commonPool}, and no additional effect if |
2919 |
> |
* already shut down. Tasks that are in the process of being |
2920 |
> |
* submitted concurrently during the course of this method may or |
2921 |
> |
* may not be rejected. |
2922 |
|
* |
2923 |
|
* @throws SecurityException if a security manager exists and |
2924 |
|
* the caller is not permitted to modify threads |
2931 |
|
} |
2932 |
|
|
2933 |
|
/** |
2934 |
< |
* Attempts to cancel and/or stop all tasks, and reject all |
2935 |
< |
* subsequently submitted tasks. Tasks that are in the process of |
2936 |
< |
* being submitted or executed concurrently during the course of |
2937 |
< |
* this method may or may not be rejected. This method cancels |
2938 |
< |
* both existing and unexecuted tasks, in order to permit |
2939 |
< |
* termination in the presence of task dependencies. So the method |
2940 |
< |
* always returns an empty list (unlike the case for some other |
2941 |
< |
* Executors). |
2934 |
> |
* Possibly attempts to cancel and/or stop all tasks, and reject |
2935 |
> |
* all subsequently submitted tasks. Invocation has no effect on |
2936 |
> |
* execution state if this is the {@link #commonPool}, and no |
2937 |
> |
* additional effect if already shut down. Otherwise, tasks that |
2938 |
> |
* are in the process of being submitted or executed concurrently |
2939 |
> |
* during the course of this method may or may not be |
2940 |
> |
* rejected. This method cancels both existing and unexecuted |
2941 |
> |
* tasks, in order to permit termination in the presence of task |
2942 |
> |
* dependencies. So the method always returns an empty list |
2943 |
> |
* (unlike the case for some other Executors). |
2944 |
|
* |
2945 |
|
* @return an empty list |
2946 |
|
* @throws SecurityException if a security manager exists and |
2990 |
|
* @return {@code true} if this pool has been shut down |
2991 |
|
*/ |
2992 |
|
public boolean isShutdown() { |
2993 |
< |
return runState < 0; |
2993 |
> |
return plock < 0; |
2994 |
|
} |
2995 |
|
|
2996 |
|
/** |
2997 |
< |
* Blocks until all tasks have completed execution after a shutdown |
2998 |
< |
* request, or the timeout occurs, or the current thread is |
2999 |
< |
* interrupted, whichever happens first. |
2997 |
> |
* Blocks until all tasks have completed execution after a |
2998 |
> |
* shutdown request, or the timeout occurs, or the current thread |
2999 |
> |
* is interrupted, whichever happens first. Note that the {@link |
3000 |
> |
* #commonPool()} never terminates until program shutdown so |
3001 |
> |
* this method will always time out. |
3002 |
|
* |
3003 |
|
* @param timeout the maximum time to wait |
3004 |
|
* @param unit the time unit of the timeout argument |
3009 |
|
public boolean awaitTermination(long timeout, TimeUnit unit) |
3010 |
|
throws InterruptedException { |
3011 |
|
long nanos = unit.toNanos(timeout); |
3012 |
< |
final Mutex lock = this.lock; |
3013 |
< |
lock.lock(); |
3014 |
< |
try { |
3015 |
< |
for (;;) { |
3016 |
< |
if (isTerminated()) |
3017 |
< |
return true; |
3018 |
< |
if (nanos <= 0) |
3019 |
< |
return false; |
3020 |
< |
nanos = termination.awaitNanos(nanos); |
3012 |
> |
if (isTerminated()) |
3013 |
> |
return true; |
3014 |
> |
long startTime = System.nanoTime(); |
3015 |
> |
boolean terminated = false; |
3016 |
> |
synchronized (this) { |
3017 |
> |
for (long waitTime = nanos, millis = 0L;;) { |
3018 |
> |
if (terminated = isTerminated() || |
3019 |
> |
waitTime <= 0L || |
3020 |
> |
(millis = unit.toMillis(waitTime)) <= 0L) |
3021 |
> |
break; |
3022 |
> |
wait(millis); |
3023 |
> |
waitTime = nanos - (System.nanoTime() - startTime); |
3024 |
|
} |
2699 |
– |
} finally { |
2700 |
– |
lock.unlock(); |
3025 |
|
} |
3026 |
+ |
return terminated; |
3027 |
|
} |
3028 |
|
|
3029 |
|
/** |
3122 |
|
public static void managedBlock(ManagedBlocker blocker) |
3123 |
|
throws InterruptedException { |
3124 |
|
Thread t = Thread.currentThread(); |
3125 |
< |
ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ? |
3126 |
< |
((ForkJoinWorkerThread)t).pool : null); |
3127 |
< |
while (!blocker.isReleasable()) { |
3128 |
< |
if (p == null || p.tryCompensate(null, blocker)) { |
3129 |
< |
try { |
3130 |
< |
do {} while (!blocker.isReleasable() && !blocker.block()); |
3131 |
< |
} finally { |
3132 |
< |
if (p != null) |
3125 |
> |
if (t instanceof ForkJoinWorkerThread) { |
3126 |
> |
ForkJoinPool p = ((ForkJoinWorkerThread)t).pool; |
3127 |
> |
while (!blocker.isReleasable()) { // variant of helpSignal |
3128 |
> |
WorkQueue[] ws; WorkQueue q; int m, n; |
3129 |
> |
if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) { |
3130 |
> |
for (int i = 0; i <= m; ++i) { |
3131 |
> |
if (blocker.isReleasable()) |
3132 |
> |
return; |
3133 |
> |
if ((q = ws[i]) != null && (n = q.queueSize()) > 0) { |
3134 |
> |
p.signalWork(q, n); |
3135 |
> |
if ((int)(p.ctl >> AC_SHIFT) >= 0) |
3136 |
> |
break; |
3137 |
> |
} |
3138 |
> |
} |
3139 |
> |
} |
3140 |
> |
if (p.tryCompensate()) { |
3141 |
> |
try { |
3142 |
> |
do {} while (!blocker.isReleasable() && |
3143 |
> |
!blocker.block()); |
3144 |
> |
} finally { |
3145 |
|
p.incrementActiveCount(); |
3146 |
+ |
} |
3147 |
+ |
break; |
3148 |
|
} |
2810 |
– |
break; |
3149 |
|
} |
3150 |
|
} |
3151 |
+ |
else { |
3152 |
+ |
do {} while (!blocker.isReleasable() && |
3153 |
+ |
!blocker.block()); |
3154 |
+ |
} |
3155 |
|
} |
3156 |
|
|
3157 |
|
// AbstractExecutorService overrides. These rely on undocumented |
3172 |
|
private static final long PARKBLOCKER; |
3173 |
|
private static final int ABASE; |
3174 |
|
private static final int ASHIFT; |
3175 |
+ |
private static final long STEALCOUNT; |
3176 |
+ |
private static final long PLOCK; |
3177 |
+ |
private static final long INDEXSEED; |
3178 |
+ |
private static final long QLOCK; |
3179 |
|
|
3180 |
|
static { |
3181 |
< |
poolNumberGenerator = new AtomicInteger(); |
3182 |
< |
nextSubmitterSeed = new AtomicInteger(0x55555555); |
3183 |
< |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3184 |
< |
defaultForkJoinWorkerThreadFactory = |
3185 |
< |
new DefaultForkJoinWorkerThreadFactory(); |
3186 |
< |
submitters = new ThreadSubmitter(); |
3187 |
< |
int s; |
3181 |
> |
// Establish common pool parameters |
3182 |
> |
// TBD: limit or report ignored exceptions? |
3183 |
> |
|
3184 |
> |
int par = 0; |
3185 |
> |
ForkJoinWorkerThreadFactory fac = null; |
3186 |
> |
Thread.UncaughtExceptionHandler handler = null; |
3187 |
> |
try { |
3188 |
> |
String pp = System.getProperty(propPrefix + "parallelism"); |
3189 |
> |
String hp = System.getProperty(propPrefix + "exceptionHandler"); |
3190 |
> |
String fp = System.getProperty(propPrefix + "threadFactory"); |
3191 |
> |
if (fp != null) |
3192 |
> |
fac = ((ForkJoinWorkerThreadFactory)ClassLoader. |
3193 |
> |
getSystemClassLoader().loadClass(fp).newInstance()); |
3194 |
> |
if (hp != null) |
3195 |
> |
handler = ((Thread.UncaughtExceptionHandler)ClassLoader. |
3196 |
> |
getSystemClassLoader().loadClass(hp).newInstance()); |
3197 |
> |
if (pp != null) |
3198 |
> |
par = Integer.parseInt(pp); |
3199 |
> |
} catch(Exception ignore) { |
3200 |
> |
} |
3201 |
> |
|
3202 |
> |
int s; // initialize field offsets for CAS etc |
3203 |
|
try { |
3204 |
|
U = getUnsafe(); |
3205 |
|
Class<?> k = ForkJoinPool.class; |
2845 |
– |
Class<?> ak = ForkJoinTask[].class; |
3206 |
|
CTL = U.objectFieldOffset |
3207 |
|
(k.getDeclaredField("ctl")); |
3208 |
+ |
STEALCOUNT = U.objectFieldOffset |
3209 |
+ |
(k.getDeclaredField("stealCount")); |
3210 |
+ |
PLOCK = U.objectFieldOffset |
3211 |
+ |
(k.getDeclaredField("plock")); |
3212 |
+ |
INDEXSEED = U.objectFieldOffset |
3213 |
+ |
(k.getDeclaredField("indexSeed")); |
3214 |
|
Class<?> tk = Thread.class; |
3215 |
|
PARKBLOCKER = U.objectFieldOffset |
3216 |
|
(tk.getDeclaredField("parkBlocker")); |
3217 |
+ |
Class<?> wk = WorkQueue.class; |
3218 |
+ |
QLOCK = U.objectFieldOffset |
3219 |
+ |
(wk.getDeclaredField("qlock")); |
3220 |
+ |
Class<?> ak = ForkJoinTask[].class; |
3221 |
|
ABASE = U.arrayBaseOffset(ak); |
3222 |
|
s = U.arrayIndexScale(ak); |
3223 |
+ |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
3224 |
|
} catch (Exception e) { |
3225 |
|
throw new Error(e); |
3226 |
|
} |
3227 |
|
if ((s & (s-1)) != 0) |
3228 |
|
throw new Error("data type scale not a power of two"); |
3229 |
< |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
3229 |
> |
|
3230 |
> |
/* |
3231 |
> |
* For extra caution, computations to set up pool state are |
3232 |
> |
* here; the constructor just assigns these values to fields. |
3233 |
> |
*/ |
3234 |
> |
ForkJoinWorkerThreadFactory defaultFac = |
3235 |
> |
defaultForkJoinWorkerThreadFactory = |
3236 |
> |
new DefaultForkJoinWorkerThreadFactory(); |
3237 |
> |
if (fac == null) |
3238 |
> |
fac = defaultFac; |
3239 |
> |
if (par <= 0) |
3240 |
> |
par = Runtime.getRuntime().availableProcessors(); |
3241 |
> |
if (par > MAX_CAP) |
3242 |
> |
par = MAX_CAP; |
3243 |
> |
long np = (long)(-par); // precompute initial ctl value |
3244 |
> |
long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
3245 |
> |
|
3246 |
> |
commonPoolParallelism = par; |
3247 |
> |
commonPool = new ForkJoinPool(par, ct, fac, handler); |
3248 |
> |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3249 |
> |
submitters = new ThreadLocal<Submitter>(); |
3250 |
|
} |
3251 |
|
|
3252 |
|
/** |