11 |
|
import java.util.Collection; |
12 |
|
import java.util.Collections; |
13 |
|
import java.util.List; |
14 |
– |
import java.util.Random; |
14 |
|
import java.util.concurrent.AbstractExecutorService; |
15 |
|
import java.util.concurrent.Callable; |
16 |
|
import java.util.concurrent.ExecutorService; |
17 |
|
import java.util.concurrent.Future; |
18 |
|
import java.util.concurrent.RejectedExecutionException; |
19 |
|
import java.util.concurrent.RunnableFuture; |
21 |
– |
import java.util.concurrent.ThreadLocalRandom; |
20 |
|
import java.util.concurrent.TimeUnit; |
23 |
– |
import java.util.concurrent.atomic.AtomicInteger; |
24 |
– |
import java.util.concurrent.atomic.AtomicLong; |
25 |
– |
import java.util.concurrent.locks.AbstractQueuedSynchronizer; |
26 |
– |
import java.util.concurrent.locks.Condition; |
21 |
|
|
22 |
|
/** |
23 |
|
* An {@link ExecutorService} for running {@link ForkJoinTask}s. |
37 |
|
* ForkJoinPool}s may also be appropriate for use with event-style |
38 |
|
* tasks that are never joined. |
39 |
|
* |
40 |
< |
* <p>A static {@link #commonPool} is available and appropriate for |
40 |
> |
* <p>A static {@link #commonPool()} is available and appropriate for |
41 |
|
* most applications. The common pool is used by any ForkJoinTask that |
42 |
|
* is not explicitly submitted to a specified pool. Using the common |
43 |
|
* pool normally reduces resource usage (its threads are slowly |
44 |
|
* reclaimed during periods of non-use, and reinstated upon subsequent |
45 |
< |
* use). The common pool is by default constructed with default |
52 |
< |
* parameters, but these may be controlled by setting any or all of |
53 |
< |
* the three properties {@code |
54 |
< |
* java.util.concurrent.ForkJoinPool.common.{parallelism, |
55 |
< |
* threadFactory, exceptionHandler}}. |
45 |
> |
* use). |
46 |
|
* |
47 |
|
* <p>For applications that require separate or custom pools, a {@code |
48 |
|
* ForkJoinPool} may be constructed with a given target parallelism |
51 |
|
* dynamically adding, suspending, or resuming internal worker |
52 |
|
* threads, even if some tasks are stalled waiting to join |
53 |
|
* others. However, no such adjustments are guaranteed in the face of |
54 |
< |
* blocked IO or other unmanaged synchronization. The nested {@link |
54 |
> |
* blocked I/O or other unmanaged synchronization. The nested {@link |
55 |
|
* ManagedBlocker} interface enables extension of the kinds of |
56 |
|
* synchronization accommodated. |
57 |
|
* |
62 |
|
* {@link #toString} returns indications of pool state in a |
63 |
|
* convenient form for informal monitoring. |
64 |
|
* |
65 |
< |
* <p> As is the case with other ExecutorServices, there are three |
65 |
> |
* <p>As is the case with other ExecutorServices, there are three |
66 |
|
* main task execution methods summarized in the following table. |
67 |
|
* These are designed to be used primarily by clients not already |
68 |
|
* engaged in fork/join computations in the current pool. The main |
97 |
|
* </tr> |
98 |
|
* </table> |
99 |
|
* |
100 |
+ |
* <p>The common pool is by default constructed with default |
101 |
+ |
* parameters, but these may be controlled by setting three {@link |
102 |
+ |
* System#getProperty system properties} with prefix {@code |
103 |
+ |
* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} -- |
104 |
+ |
* an integer greater than zero, {@code threadFactory} -- the class |
105 |
+ |
* name of a {@link ForkJoinWorkerThreadFactory}, and {@code |
106 |
+ |
* exceptionHandler} -- the class name of a {@link |
107 |
+ |
* java.lang.Thread.UncaughtExceptionHandler |
108 |
+ |
* Thread.UncaughtExceptionHandler}. Upon any error in establishing |
109 |
+ |
* these settings, default parameters are used. |
110 |
+ |
* |
111 |
|
* <p><b>Implementation notes</b>: This implementation restricts the |
112 |
|
* maximum number of running threads to 32767. Attempts to create |
113 |
|
* pools with greater than the maximum number result in |
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 |
> |
* block when needed. |
243 |
|
* |
244 |
|
* Recording WorkQueues. WorkQueues are recorded in the |
245 |
|
* "workQueues" array that is created upon first use and expanded |
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. Shared (submission) queues are at even |
252 |
< |
* indices, worker queues at odd indices. Grouping them together |
253 |
< |
* in this way simplifies and speeds up task scanning. |
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 has fewer |
320 |
> |
* than two tasks, they signal waiting workers (or trigger |
321 |
> |
* creation of new ones if fewer than the given parallelism level |
322 |
> |
* -- signalWork), and may leave a hint to the unparked worker to |
323 |
> |
* help signal others upon wakeup). These primary signals are |
324 |
> |
* buttressed by others (see method helpSignal) whenever other |
325 |
> |
* threads scan for work or do not have a task to process. On |
326 |
> |
* most platforms, signalling (unpark) overhead time is noticeably |
327 |
> |
* long, and the time between signalling a thread and it actually |
328 |
> |
* making progress can be very noticeably long, so it is worth |
329 |
> |
* offloading these delays from critical paths as much as |
330 |
> |
* possible. |
331 |
|
* |
332 |
|
* Trimming workers. To release resources after periods of lack of |
333 |
|
* use, a worker starting to wait when the pool is quiescent will |
338 |
|
* periods of non-use. |
339 |
|
* |
340 |
|
* Shutdown and Termination. A call to shutdownNow atomically sets |
341 |
< |
* a runState bit and then (non-atomically) sets each worker's |
342 |
< |
* runState status, cancels all unprocessed tasks, and wakes up |
341 |
> |
* a plock bit and then (non-atomically) sets each worker's |
342 |
> |
* qlock status, cancels all unprocessed tasks, and wakes up |
343 |
|
* all waiting workers. Detecting whether termination should |
344 |
|
* commence after a non-abrupt shutdown() call requires more work |
345 |
|
* and bookkeeping. We need consensus about quiescence (i.e., that |
367 |
|
* method tryCompensate() may create or re-activate a spare |
368 |
|
* thread to compensate for blocked joiners until they unblock. |
369 |
|
* |
370 |
< |
* A third form (implemented in tryRemoveAndExec and |
371 |
< |
* tryPollForAndExec) amounts to helping a hypothetical |
372 |
< |
* compensator: If we can readily tell that a possible action of a |
373 |
< |
* compensator is to steal and execute the task being joined, the |
374 |
< |
* joining thread can do so directly, without the need for a |
375 |
< |
* compensation thread (although at the expense of larger run-time |
376 |
< |
* stacks, but the tradeoff is typically worthwhile). |
370 |
> |
* A third form (implemented in tryRemoveAndExec) amounts to |
371 |
> |
* helping a hypothetical compensator: If we can readily tell that |
372 |
> |
* a possible action of a compensator is to steal and execute the |
373 |
> |
* task being joined, the joining thread can do so directly, |
374 |
> |
* without the need for a compensation thread (although at the |
375 |
> |
* expense of larger run-time stacks, but the tradeoff is |
376 |
> |
* typically worthwhile). |
377 |
|
* |
378 |
|
* The ManagedBlocker extension API can't use helping so relies |
379 |
|
* only on compensation in method awaitBlocker. |
395 |
|
* steals, rather than use per-task bookkeeping. This sometimes |
396 |
|
* requires a linear scan of workQueues array to locate stealers, |
397 |
|
* but often doesn't because stealers leave hints (that may become |
398 |
< |
* stale/wrong) of where to locate them. A stealHint is only a |
399 |
< |
* hint because a worker might have had multiple steals and the |
400 |
< |
* hint records only one of them (usually the most current). |
401 |
< |
* Hinting isolates cost to when it is needed, rather than adding |
402 |
< |
* to per-task overhead. (2) It is "shallow", ignoring nesting |
403 |
< |
* and potentially cyclic mutual steals. (3) It is intentionally |
398 |
> |
* stale/wrong) of where to locate them. It is only a hint |
399 |
> |
* because a worker might have had multiple steals and the hint |
400 |
> |
* records only one of them (usually the most current). Hinting |
401 |
> |
* isolates cost to when it is needed, rather than adding to |
402 |
> |
* per-task overhead. (2) It is "shallow", ignoring nesting and |
403 |
> |
* potentially cyclic mutual steals. (3) It is intentionally |
404 |
|
* racy: field currentJoin is updated only while actively joining, |
405 |
|
* which means that we miss links in the chain during long-lived |
406 |
|
* tasks, GC stalls etc (which is OK since blocking in such cases |
408 |
|
* to find work (see MAX_HELP) and fall back to suspending the |
409 |
|
* worker and if necessary replacing it with another. |
410 |
|
* |
411 |
+ |
* Helping actions for CountedCompleters are much simpler: Method |
412 |
+ |
* helpComplete can take and execute any task with the same root |
413 |
+ |
* as the task being waited on. However, this still entails some |
414 |
+ |
* traversal of completer chains, so is less efficient than using |
415 |
+ |
* CountedCompleters without explicit joins. |
416 |
+ |
* |
417 |
|
* It is impossible to keep exactly the target parallelism number |
418 |
|
* of threads running at any given time. Determining the |
419 |
|
* existence of conservatively safe helping targets, the |
435 |
|
* intractable) game with an opponent that may choose the worst |
436 |
|
* (for us) active thread to stall at any time. We take several |
437 |
|
* precautions to bound losses (and thus bound gains), mainly in |
438 |
< |
* methods tryCompensate and awaitJoin: (1) We only try |
439 |
< |
* compensation after attempting enough helping steps (measured |
440 |
< |
* via counting and timing) that we have already consumed the |
441 |
< |
* estimated cost of creating and activating a new thread. (2) We |
442 |
< |
* allow up to 50% of threads to be blocked before initially |
443 |
< |
* adding any others, and unless completely saturated, check that |
444 |
< |
* some work is available for a new worker before adding. Also, we |
445 |
< |
* create up to only 50% more threads until entering a mode that |
446 |
< |
* only adds a thread if all others are possibly blocked. All |
447 |
< |
* together, this means that we might be half as fast to react, |
448 |
< |
* and create half as many threads as possible in the ideal case, |
449 |
< |
* but present vastly fewer anomalies in all other cases compared |
450 |
< |
* to both more aggressive and more conservative alternatives. |
451 |
< |
* |
452 |
< |
* Style notes: There is a lot of representation-level coupling |
453 |
< |
* among classes ForkJoinPool, ForkJoinWorkerThread, and |
454 |
< |
* ForkJoinTask. The fields of WorkQueue maintain data structures |
455 |
< |
* managed by ForkJoinPool, so are directly accessed. There is |
456 |
< |
* little point trying to reduce this, since any associated future |
457 |
< |
* changes in representations will need to be accompanied by |
458 |
< |
* algorithmic changes anyway. Several methods intrinsically |
459 |
< |
* sprawl because they must accumulate sets of consistent reads of |
460 |
< |
* volatiles held in local variables. Methods signalWork() and |
461 |
< |
* scan() are the main bottlenecks, so are especially heavily |
438 |
> |
* methods tryCompensate and awaitJoin. |
439 |
> |
* |
440 |
> |
* Common Pool |
441 |
> |
* =========== |
442 |
> |
* |
443 |
> |
* The static common Pool always exists after static |
444 |
> |
* initialization. Since it (or any other created pool) need |
445 |
> |
* never be used, we minimize initial construction overhead and |
446 |
> |
* footprint to the setup of about a dozen fields, with no nested |
447 |
> |
* allocation. Most bootstrapping occurs within method |
448 |
> |
* fullExternalPush during the first submission to the pool. |
449 |
> |
* |
450 |
> |
* When external threads submit to the common pool, they can |
451 |
> |
* perform some subtask processing (see externalHelpJoin and |
452 |
> |
* related methods). We do not need to record whether these |
453 |
> |
* submissions are to the common pool -- if not, externalHelpJoin |
454 |
> |
* returns quickly (at the most helping to signal some common pool |
455 |
> |
* workers). These submitters would otherwise be blocked waiting |
456 |
> |
* for completion, so the extra effort (with liberally sprinkled |
457 |
> |
* task status checks) in inapplicable cases amounts to an odd |
458 |
> |
* form of limited spin-wait before blocking in ForkJoinTask.join. |
459 |
> |
* |
460 |
> |
* Style notes |
461 |
> |
* =========== |
462 |
> |
* |
463 |
> |
* There is a lot of representation-level coupling among classes |
464 |
> |
* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The |
465 |
> |
* fields of WorkQueue maintain data structures managed by |
466 |
> |
* ForkJoinPool, so are directly accessed. There is little point |
467 |
> |
* trying to reduce this, since any associated future changes in |
468 |
> |
* representations will need to be accompanied by algorithmic |
469 |
> |
* changes anyway. Several methods intrinsically sprawl because |
470 |
> |
* they must accumulate sets of consistent reads of volatiles held |
471 |
> |
* in local variables. Methods signalWork() and scan() are the |
472 |
> |
* main bottlenecks, so are especially heavily |
473 |
|
* micro-optimized/mangled. There are lots of inline assignments |
474 |
|
* (of form "while ((local = field) != 0)") which are usually the |
475 |
|
* simplest way to ensure the required read orderings (which are |
477 |
|
* declarations of these locals at the heads of methods or blocks. |
478 |
|
* There are several occurrences of the unusual "do {} while |
479 |
|
* (!cas...)" which is the simplest way to force an update of a |
480 |
< |
* CAS'ed variable. There are also other coding oddities that help |
480 |
> |
* CAS'ed variable. There are also other coding oddities (including |
481 |
> |
* several unnecessary-looking hoisted null checks) that help |
482 |
|
* some methods perform reasonably even when interpreted (not |
483 |
|
* compiled). |
484 |
|
* |
527 |
|
* Default ForkJoinWorkerThreadFactory implementation; creates a |
528 |
|
* new ForkJoinWorkerThread. |
529 |
|
*/ |
530 |
< |
static class DefaultForkJoinWorkerThreadFactory |
530 |
> |
static final class DefaultForkJoinWorkerThreadFactory |
531 |
|
implements ForkJoinWorkerThreadFactory { |
532 |
< |
public ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
532 |
> |
public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { |
533 |
|
return new ForkJoinWorkerThread(pool); |
534 |
|
} |
535 |
|
} |
536 |
|
|
537 |
|
/** |
538 |
+ |
* Per-thread records for threads that submit to pools. Currently |
539 |
+ |
* holds only pseudo-random seed / index that is used to choose |
540 |
+ |
* submission queues in method externalPush. In the future, this may |
541 |
+ |
* also incorporate a means to implement different task rejection |
542 |
+ |
* and resubmission policies. |
543 |
+ |
* |
544 |
+ |
* Seeds for submitters and workers/workQueues work in basically |
545 |
+ |
* the same way but are initialized and updated using slightly |
546 |
+ |
* different mechanics. Both are initialized using the same |
547 |
+ |
* approach as in class ThreadLocal, where successive values are |
548 |
+ |
* unlikely to collide with previous values. Seeds are then |
549 |
+ |
* randomly modified upon collisions using xorshifts, which |
550 |
+ |
* requires a non-zero seed. |
551 |
+ |
*/ |
552 |
+ |
static final class Submitter { |
553 |
+ |
int seed; |
554 |
+ |
Submitter(int s) { seed = s; } |
555 |
+ |
} |
556 |
+ |
|
557 |
+ |
/** |
558 |
|
* Class for artificial tasks that are used to replace the target |
559 |
|
* of local joins if they are removed from an interior queue slot |
560 |
|
* in WorkQueue.tryRemoveAndExec. We don't need the proxy to |
561 |
|
* actually do anything beyond having a unique identity. |
562 |
|
*/ |
563 |
|
static final class EmptyTask extends ForkJoinTask<Void> { |
564 |
+ |
private static final long serialVersionUID = -7721805057305804111L; |
565 |
|
EmptyTask() { status = ForkJoinTask.NORMAL; } // force done |
566 |
|
public final Void getRawResult() { return null; } |
567 |
|
public final void setRawResult(Void x) {} |
582 |
|
* |
583 |
|
* Field "top" is the index (mod array.length) of the next queue |
584 |
|
* slot to push to or pop from. It is written only by owner thread |
585 |
< |
* for push, or under lock for trySharedPush, and accessed by |
586 |
< |
* other threads only after reading (volatile) base. Both top and |
587 |
< |
* base are allowed to wrap around on overflow, but (top - base) |
588 |
< |
* (or more commonly -(base - top) to force volatile read of base |
589 |
< |
* before top) still estimates size. |
585 |
> |
* for push, or under lock for external/shared push, and accessed |
586 |
> |
* by other threads only after reading (volatile) base. Both top |
587 |
> |
* and base are allowed to wrap around on overflow, but (top - |
588 |
> |
* base) (or more commonly -(base - top) to force volatile read of |
589 |
> |
* base before top) still estimates size. The lock ("qlock") is |
590 |
> |
* forced to -1 on termination, causing all further lock attempts |
591 |
> |
* to fail. (Note: we don't need CAS for termination state because |
592 |
> |
* upon pool shutdown, all shared-queues will stop being used |
593 |
> |
* anyway.) Nearly all lock bodies are set up so that exceptions |
594 |
> |
* within lock bodies are "impossible" (modulo JVM errors that |
595 |
> |
* would cause failure anyway.) |
596 |
|
* |
597 |
|
* The array slots are read and written using the emulation of |
598 |
|
* volatiles/atomics provided by Unsafe. Insertions must in |
599 |
|
* general use putOrderedObject as a form of releasing store to |
600 |
|
* ensure that all writes to the task object are ordered before |
601 |
< |
* its publication in the queue. (Although we can avoid one case |
602 |
< |
* of this when locked in trySharedPush.) All removals entail a |
603 |
< |
* CAS to null. The array is always a power of two. To ensure |
604 |
< |
* safety of Unsafe array operations, all accesses perform |
545 |
< |
* explicit null checks and implicit bounds checks via |
546 |
< |
* power-of-two masking. |
601 |
> |
* its publication in the queue. All removals entail a CAS to |
602 |
> |
* null. The array is always a power of two. To ensure safety of |
603 |
> |
* Unsafe array operations, all accesses perform explicit null |
604 |
> |
* checks and implicit bounds checks via power-of-two masking. |
605 |
|
* |
606 |
|
* In addition to basic queuing support, this class contains |
607 |
|
* fields described elsewhere to control execution. It turns out |
608 |
< |
* to work better memory-layout-wise to include them in this |
609 |
< |
* class rather than a separate class. |
608 |
> |
* to work better memory-layout-wise to include them in this class |
609 |
> |
* rather than a separate class. |
610 |
|
* |
611 |
|
* Performance on most platforms is very sensitive to placement of |
612 |
|
* instances of both WorkQueues and their arrays -- we absolutely |
620 |
|
* trades off slightly slower average field access for the sake of |
621 |
|
* avoiding really bad worst-case access. (Until better JVM |
622 |
|
* support is in place, this padding is dependent on transient |
623 |
< |
* properties of JVM field layout rules.) We also take care in |
623 |
> |
* properties of JVM field layout rules.) We also take care in |
624 |
|
* allocating, sizing and resizing the array. Non-shared queue |
625 |
< |
* arrays are initialized (via method growArray) by workers before |
626 |
< |
* use. Others are allocated on first use. |
625 |
> |
* arrays are initialized by workers before use. Others are |
626 |
> |
* allocated on first use. |
627 |
|
*/ |
628 |
|
static final class WorkQueue { |
629 |
|
/** |
646 |
|
*/ |
647 |
|
static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M |
648 |
|
|
649 |
< |
volatile long totalSteals; // cumulative number of steals |
649 |
> |
// Heuristic padding to ameliorate unfortunate memory placements |
650 |
> |
volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; |
651 |
> |
|
652 |
|
int seed; // for random scanning; initialize nonzero |
653 |
|
volatile int eventCount; // encoded inactivation count; < 0 if inactive |
654 |
|
int nextWait; // encoded record of next event waiter |
655 |
< |
int rescans; // remaining scans until block |
596 |
< |
int nsteals; // top-level task executions since last idle |
597 |
< |
final int mode; // lifo, fifo, or shared |
655 |
> |
int hint; // steal or signal hint (index) |
656 |
|
int poolIndex; // index of this queue in pool (or 0) |
657 |
< |
int stealHint; // index of most recent known stealer |
658 |
< |
volatile int runState; // 1: locked, -1: terminate; else 0 |
657 |
> |
final int mode; // 0: lifo, > 0: fifo, < 0: shared |
658 |
> |
int nsteals; // number of steals |
659 |
> |
volatile int qlock; // 1: locked, -1: terminate; else 0 |
660 |
|
volatile int base; // index of next slot for poll |
661 |
|
int top; // index of next slot for push |
662 |
|
ForkJoinTask<?>[] array; // the elements (initially unallocated) |
665 |
|
volatile Thread parker; // == owner during call to park; else null |
666 |
|
volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin |
667 |
|
ForkJoinTask<?> currentSteal; // current non-local task being executed |
609 |
– |
// Heuristic padding to ameliorate unfortunate memory placements |
610 |
– |
Object p00, p01, p02, p03, p04, p05, p06, p07; |
611 |
– |
Object p08, p09, p0a, p0b, p0c, p0d, p0e; |
668 |
|
|
669 |
< |
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) { |
670 |
< |
this.mode = mode; |
669 |
> |
volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; |
670 |
> |
volatile Object pad18, pad19, pad1a, pad1b, pad1c, pad1d; |
671 |
> |
|
672 |
> |
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode, |
673 |
> |
int seed) { |
674 |
|
this.pool = pool; |
675 |
|
this.owner = owner; |
676 |
+ |
this.mode = mode; |
677 |
+ |
this.seed = seed; |
678 |
|
// Place indices in the center of array (that is not yet allocated) |
679 |
|
base = top = INITIAL_QUEUE_CAPACITY >>> 1; |
680 |
|
} |
687 |
|
return (n >= 0) ? 0 : -n; // ignore transient negative |
688 |
|
} |
689 |
|
|
690 |
< |
/** |
690 |
> |
/** |
691 |
|
* Provides a more accurate estimate of whether this queue has |
692 |
|
* any tasks than does queueSize, by checking whether a |
693 |
|
* near-empty queue has at least one unclaimed task. |
699 |
|
(n == -1 && |
700 |
|
((a = array) == null || |
701 |
|
(m = a.length - 1) < 0 || |
702 |
< |
U.getObjectVolatile |
703 |
< |
(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null))); |
702 |
> |
U.getObject |
703 |
> |
(a, (long)((m & (s - 1)) << ASHIFT) + ABASE) == null))); |
704 |
|
} |
705 |
|
|
706 |
|
/** |
707 |
< |
* Pushes a task. Call only by owner in unshared queues. |
707 |
> |
* Pushes a task. Call only by owner in unshared queues. (The |
708 |
> |
* shared-queue version is embedded in method externalPush.) |
709 |
|
* |
710 |
|
* @param task the task. Caller must ensure non-null. |
711 |
|
* @throw RejectedExecutionException if array cannot be resized |
714 |
|
ForkJoinTask<?>[] a; ForkJoinPool p; |
715 |
|
int s = top, m, n; |
716 |
|
if ((a = array) != null) { // ignore if queue removed |
717 |
< |
U.putOrderedObject |
718 |
< |
(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task); |
717 |
> |
int j = (((m = a.length - 1) & s) << ASHIFT) + ABASE; |
718 |
> |
U.putOrderedObject(a, j, task); |
719 |
|
if ((n = (top = s + 1) - base) <= 2) { |
720 |
|
if ((p = pool) != null) |
721 |
< |
p.signalWork(); |
721 |
> |
p.signalWork(this); |
722 |
|
} |
723 |
|
else if (n >= m) |
724 |
< |
growArray(true); |
724 |
> |
growArray(); |
725 |
|
} |
726 |
|
} |
727 |
|
|
728 |
< |
/** |
729 |
< |
* Pushes a task if lock is free and array is either big |
730 |
< |
* enough or can be resized to be big enough. |
731 |
< |
* |
670 |
< |
* @param task the task. Caller must ensure non-null. |
671 |
< |
* @return true if submitted |
728 |
> |
/** |
729 |
> |
* Initializes or doubles the capacity of array. Call either |
730 |
> |
* by owner or with lock held -- it is OK for base, but not |
731 |
> |
* top, to move while resizings are in progress. |
732 |
|
*/ |
733 |
< |
final boolean trySharedPush(ForkJoinTask<?> task) { |
734 |
< |
boolean submitted = false; |
735 |
< |
if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { |
736 |
< |
ForkJoinTask<?>[] a = array; |
737 |
< |
int s = top; |
738 |
< |
try { |
739 |
< |
if ((a != null && a.length > s + 1 - base) || |
740 |
< |
(a = growArray(false)) != null) { // must presize |
741 |
< |
int j = (((a.length - 1) & s) << ASHIFT) + ABASE; |
742 |
< |
U.putObject(a, (long)j, task); // don't need "ordered" |
743 |
< |
top = s + 1; |
744 |
< |
submitted = true; |
745 |
< |
} |
746 |
< |
} finally { |
747 |
< |
runState = 0; // unlock |
748 |
< |
} |
733 |
> |
final ForkJoinTask<?>[] growArray() { |
734 |
> |
ForkJoinTask<?>[] oldA = array; |
735 |
> |
int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; |
736 |
> |
if (size > MAXIMUM_QUEUE_CAPACITY) |
737 |
> |
throw new RejectedExecutionException("Queue capacity exceeded"); |
738 |
> |
int oldMask, t, b; |
739 |
> |
ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size]; |
740 |
> |
if (oldA != null && (oldMask = oldA.length - 1) >= 0 && |
741 |
> |
(t = top) - (b = base) > 0) { |
742 |
> |
int mask = size - 1; |
743 |
> |
do { |
744 |
> |
ForkJoinTask<?> x; |
745 |
> |
int oldj = ((b & oldMask) << ASHIFT) + ABASE; |
746 |
> |
int j = ((b & mask) << ASHIFT) + ABASE; |
747 |
> |
x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj); |
748 |
> |
if (x != null && |
749 |
> |
U.compareAndSwapObject(oldA, oldj, x, null)) |
750 |
> |
U.putObjectVolatile(a, j, x); |
751 |
> |
} while (++b != t); |
752 |
|
} |
753 |
< |
return submitted; |
753 |
> |
return a; |
754 |
|
} |
755 |
|
|
756 |
|
/** |
773 |
|
return null; |
774 |
|
} |
775 |
|
|
713 |
– |
final ForkJoinTask<?> sharedPop() { |
714 |
– |
ForkJoinTask<?> task = null; |
715 |
– |
if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { |
716 |
– |
try { |
717 |
– |
ForkJoinTask<?>[] a; int m; |
718 |
– |
if ((a = array) != null && (m = a.length - 1) >= 0) { |
719 |
– |
for (int s; (s = top - 1) - base >= 0;) { |
720 |
– |
long j = ((m & s) << ASHIFT) + ABASE; |
721 |
– |
ForkJoinTask<?> t = |
722 |
– |
(ForkJoinTask<?>)U.getObject(a, j); |
723 |
– |
if (t == null) |
724 |
– |
break; |
725 |
– |
if (U.compareAndSwapObject(a, j, t, null)) { |
726 |
– |
top = s; |
727 |
– |
task = t; |
728 |
– |
break; |
729 |
– |
} |
730 |
– |
} |
731 |
– |
} |
732 |
– |
} finally { |
733 |
– |
runState = 0; |
734 |
– |
} |
735 |
– |
} |
736 |
– |
return task; |
737 |
– |
} |
738 |
– |
|
739 |
– |
/** |
740 |
– |
* Version of pop that takes top element only if it |
741 |
– |
* its root is the given CountedCompleter. |
742 |
– |
*/ |
743 |
– |
final ForkJoinTask<?> popCC(CountedCompleter<?> root) { |
744 |
– |
ForkJoinTask<?>[] a; int m; |
745 |
– |
if (root != null && (a = array) != null && (m = a.length - 1) >= 0) { |
746 |
– |
for (int s; (s = top - 1) - base >= 0;) { |
747 |
– |
long j = ((m & s) << ASHIFT) + ABASE; |
748 |
– |
ForkJoinTask<?> t = |
749 |
– |
(ForkJoinTask<?>)U.getObject(a, j); |
750 |
– |
if (t == null || !(t instanceof CountedCompleter) || |
751 |
– |
((CountedCompleter<?>)t).getRoot() != root) |
752 |
– |
break; |
753 |
– |
if (U.compareAndSwapObject(a, j, t, null)) { |
754 |
– |
top = s; |
755 |
– |
return t; |
756 |
– |
} |
757 |
– |
if (root.status < 0) |
758 |
– |
break; |
759 |
– |
} |
760 |
– |
} |
761 |
– |
return null; |
762 |
– |
} |
763 |
– |
|
764 |
– |
/** |
765 |
– |
* Shared version of popCC |
766 |
– |
*/ |
767 |
– |
final ForkJoinTask<?> sharedPopCC(CountedCompleter<?> root) { |
768 |
– |
ForkJoinTask<?> task = null; |
769 |
– |
if (root != null && |
770 |
– |
runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { |
771 |
– |
try { |
772 |
– |
ForkJoinTask<?>[] a; int m; |
773 |
– |
if ((a = array) != null && (m = a.length - 1) >= 0) { |
774 |
– |
for (int s; (s = top - 1) - base >= 0;) { |
775 |
– |
long j = ((m & s) << ASHIFT) + ABASE; |
776 |
– |
ForkJoinTask<?> t = |
777 |
– |
(ForkJoinTask<?>)U.getObject(a, j); |
778 |
– |
if (t == null || !(t instanceof CountedCompleter) || |
779 |
– |
((CountedCompleter<?>)t).getRoot() != root) |
780 |
– |
break; |
781 |
– |
if (U.compareAndSwapObject(a, j, t, null)) { |
782 |
– |
top = s; |
783 |
– |
task = t; |
784 |
– |
break; |
785 |
– |
} |
786 |
– |
if (root.status < 0) |
787 |
– |
break; |
788 |
– |
} |
789 |
– |
} |
790 |
– |
} finally { |
791 |
– |
runState = 0; |
792 |
– |
} |
793 |
– |
} |
794 |
– |
return task; |
795 |
– |
} |
796 |
– |
|
776 |
|
/** |
777 |
|
* Takes a task in FIFO order if b is base of queue and a task |
778 |
|
* can be claimed without contention. Specialized versions |
810 |
|
else if (base == b) { |
811 |
|
if (b + 1 == top) |
812 |
|
break; |
813 |
< |
Thread.yield(); // wait for lagging update |
813 |
> |
Thread.yield(); // wait for lagging update (very rare) |
814 |
|
} |
815 |
|
} |
816 |
|
return null; |
837 |
|
|
838 |
|
/** |
839 |
|
* Pops the given task only if it is at the current top. |
840 |
+ |
* (A shared version is available only via FJP.tryExternalUnpush) |
841 |
|
*/ |
842 |
|
final boolean tryUnpush(ForkJoinTask<?> t) { |
843 |
|
ForkJoinTask<?>[] a; int s; |
851 |
|
} |
852 |
|
|
853 |
|
/** |
874 |
– |
* Version of tryUnpush for shared queues; called by non-FJ |
875 |
– |
* submitters after prechecking that task probably exists. |
876 |
– |
*/ |
877 |
– |
final boolean trySharedUnpush(ForkJoinTask<?> t) { |
878 |
– |
boolean success = false; |
879 |
– |
if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) { |
880 |
– |
try { |
881 |
– |
ForkJoinTask<?>[] a; int s; |
882 |
– |
if ((a = array) != null && (s = top) != base && |
883 |
– |
U.compareAndSwapObject |
884 |
– |
(a, (((a.length - 1) & --s) << ASHIFT) + ABASE, t, null)) { |
885 |
– |
top = s; |
886 |
– |
success = true; |
887 |
– |
} |
888 |
– |
} finally { |
889 |
– |
runState = 0; // unlock |
890 |
– |
} |
891 |
– |
} |
892 |
– |
return success; |
893 |
– |
} |
894 |
– |
|
895 |
– |
/** |
896 |
– |
* Polls the given task only if it is at the current base. |
897 |
– |
*/ |
898 |
– |
final boolean pollFor(ForkJoinTask<?> task) { |
899 |
– |
ForkJoinTask<?>[] a; int b; |
900 |
– |
if ((b = base) - top < 0 && (a = array) != null) { |
901 |
– |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
902 |
– |
if (U.getObjectVolatile(a, j) == task && base == b && |
903 |
– |
U.compareAndSwapObject(a, j, task, null)) { |
904 |
– |
base = b + 1; |
905 |
– |
return true; |
906 |
– |
} |
907 |
– |
} |
908 |
– |
return false; |
909 |
– |
} |
910 |
– |
|
911 |
– |
/** |
912 |
– |
* Initializes or doubles the capacity of array. Call either |
913 |
– |
* by owner or with lock held -- it is OK for base, but not |
914 |
– |
* top, to move while resizings are in progress. |
915 |
– |
* |
916 |
– |
* @param rejectOnFailure if true, throw exception if capacity |
917 |
– |
* exceeded (relayed ultimately to user); else return null. |
918 |
– |
*/ |
919 |
– |
final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) { |
920 |
– |
ForkJoinTask<?>[] oldA = array; |
921 |
– |
int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY; |
922 |
– |
if (size <= MAXIMUM_QUEUE_CAPACITY) { |
923 |
– |
int oldMask, t, b; |
924 |
– |
ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size]; |
925 |
– |
if (oldA != null && (oldMask = oldA.length - 1) >= 0 && |
926 |
– |
(t = top) - (b = base) > 0) { |
927 |
– |
int mask = size - 1; |
928 |
– |
do { |
929 |
– |
ForkJoinTask<?> x; |
930 |
– |
int oldj = ((b & oldMask) << ASHIFT) + ABASE; |
931 |
– |
int j = ((b & mask) << ASHIFT) + ABASE; |
932 |
– |
x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj); |
933 |
– |
if (x != null && |
934 |
– |
U.compareAndSwapObject(oldA, oldj, x, null)) |
935 |
– |
U.putObjectVolatile(a, j, x); |
936 |
– |
} while (++b != t); |
937 |
– |
} |
938 |
– |
return a; |
939 |
– |
} |
940 |
– |
else if (!rejectOnFailure) |
941 |
– |
return null; |
942 |
– |
else |
943 |
– |
throw new RejectedExecutionException("Queue capacity exceeded"); |
944 |
– |
} |
945 |
– |
|
946 |
– |
/** |
854 |
|
* Removes and cancels all known tasks, ignoring any exceptions. |
855 |
|
*/ |
856 |
|
final void cancelAll() { |
874 |
|
return seed = r ^= r << 5; |
875 |
|
} |
876 |
|
|
877 |
< |
// Execution methods |
877 |
> |
// Specialized execution methods |
878 |
|
|
879 |
|
/** |
880 |
|
* Pops and runs tasks until empty. |
903 |
|
} |
904 |
|
|
905 |
|
/** |
906 |
< |
* If present, removes from queue and executes the given task, or |
907 |
< |
* any other cancelled task. Returns (true) immediately on any CAS |
906 |
> |
* If present, removes from queue and executes the given task, |
907 |
> |
* or any other cancelled task. Returns (true) on any CAS |
908 |
|
* or consistency check failure so caller can retry. |
909 |
|
* |
910 |
< |
* @return 0 if no progress can be made, else positive |
1004 |
< |
* (this unusual convention simplifies use with tryHelpStealer.) |
910 |
> |
* @return false if no progress can be made, else true; |
911 |
|
*/ |
912 |
< |
final int tryRemoveAndExec(ForkJoinTask<?> task) { |
913 |
< |
int stat = 1; |
1008 |
< |
boolean removed = false, empty = true; |
912 |
> |
final boolean tryRemoveAndExec(ForkJoinTask<?> task) { |
913 |
> |
boolean stat = true, removed = false, empty = true; |
914 |
|
ForkJoinTask<?>[] a; int m, s, b, n; |
915 |
|
if ((a = array) != null && (m = a.length - 1) >= 0 && |
916 |
|
(n = (s = top) - (b = base)) > 0) { |
940 |
|
} |
941 |
|
if (--n == 0) { |
942 |
|
if (!empty && base == b) |
943 |
< |
stat = 0; |
943 |
> |
stat = false; |
944 |
|
break; |
945 |
|
} |
946 |
|
} |
951 |
|
} |
952 |
|
|
953 |
|
/** |
954 |
+ |
* Polls for and executes the given task or any other task in |
955 |
+ |
* its CountedCompleter computation |
956 |
+ |
*/ |
957 |
+ |
final boolean pollAndExecCC(ForkJoinTask<?> root) { |
958 |
+ |
ForkJoinTask<?>[] a; int b; Object o; |
959 |
+ |
outer: while ((b = base) - top < 0 && (a = array) != null) { |
960 |
+ |
long j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
961 |
+ |
if ((o = U.getObject(a, j)) == null || |
962 |
+ |
!(o instanceof CountedCompleter)) |
963 |
+ |
break; |
964 |
+ |
for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) { |
965 |
+ |
if (r == root) { |
966 |
+ |
if (base == b && |
967 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
968 |
+ |
base = b + 1; |
969 |
+ |
t.doExec(); |
970 |
+ |
return true; |
971 |
+ |
} |
972 |
+ |
else |
973 |
+ |
break; // restart |
974 |
+ |
} |
975 |
+ |
if ((r = r.completer) == null) |
976 |
+ |
break outer; // not part of root computation |
977 |
+ |
} |
978 |
+ |
} |
979 |
+ |
return false; |
980 |
+ |
} |
981 |
+ |
|
982 |
+ |
/** |
983 |
|
* Executes a top-level task and any local tasks remaining |
984 |
|
* after execution. |
985 |
|
*/ |
986 |
|
final void runTask(ForkJoinTask<?> t) { |
987 |
|
if (t != null) { |
988 |
< |
currentSteal = t; |
989 |
< |
t.doExec(); |
990 |
< |
if (top != base) { // process remaining local tasks |
988 |
> |
(currentSteal = t).doExec(); |
989 |
> |
currentSteal = null; |
990 |
> |
++nsteals; |
991 |
> |
if (base - top < 0) { // process remaining local tasks |
992 |
|
if (mode == 0) |
993 |
|
popAndExecAll(); |
994 |
|
else |
995 |
|
pollAndExecAll(); |
996 |
|
} |
1062 |
– |
++nsteals; |
1063 |
– |
currentSteal = null; |
997 |
|
} |
998 |
|
} |
999 |
|
|
1003 |
|
final void runSubtask(ForkJoinTask<?> t) { |
1004 |
|
if (t != null) { |
1005 |
|
ForkJoinTask<?> ps = currentSteal; |
1006 |
< |
currentSteal = t; |
1074 |
< |
t.doExec(); |
1006 |
> |
(currentSteal = t).doExec(); |
1007 |
|
currentSteal = ps; |
1008 |
|
} |
1009 |
|
} |
1020 |
|
s != Thread.State.TIMED_WAITING); |
1021 |
|
} |
1022 |
|
|
1091 |
– |
/** |
1092 |
– |
* If this owned and is not already interrupted, try to |
1093 |
– |
* interrupt and/or unpark, ignoring exceptions. |
1094 |
– |
*/ |
1095 |
– |
final void interruptOwner() { |
1096 |
– |
Thread wt, p; |
1097 |
– |
if ((wt = owner) != null && !wt.isInterrupted()) { |
1098 |
– |
try { |
1099 |
– |
wt.interrupt(); |
1100 |
– |
} catch (SecurityException ignore) { |
1101 |
– |
} |
1102 |
– |
} |
1103 |
– |
if ((p = parker) != null) |
1104 |
– |
U.unpark(p); |
1105 |
– |
} |
1106 |
– |
|
1023 |
|
// Unsafe mechanics |
1024 |
|
private static final sun.misc.Unsafe U; |
1025 |
< |
private static final long RUNSTATE; |
1025 |
> |
private static final long QLOCK; |
1026 |
|
private static final int ABASE; |
1027 |
|
private static final int ASHIFT; |
1028 |
|
static { |
1031 |
|
U = getUnsafe(); |
1032 |
|
Class<?> k = WorkQueue.class; |
1033 |
|
Class<?> ak = ForkJoinTask[].class; |
1034 |
< |
RUNSTATE = U.objectFieldOffset |
1035 |
< |
(k.getDeclaredField("runState")); |
1034 |
> |
QLOCK = U.objectFieldOffset |
1035 |
> |
(k.getDeclaredField("qlock")); |
1036 |
|
ABASE = U.arrayBaseOffset(ak); |
1037 |
|
s = U.arrayIndexScale(ak); |
1038 |
|
} catch (Exception e) { |
1044 |
|
} |
1045 |
|
} |
1046 |
|
|
1131 |
– |
/** |
1132 |
– |
* Per-thread records for threads that submit to pools. Currently |
1133 |
– |
* holds only pseudo-random seed / index that is used to choose |
1134 |
– |
* submission queues in method doSubmit. In the future, this may |
1135 |
– |
* also incorporate a means to implement different task rejection |
1136 |
– |
* and resubmission policies. |
1137 |
– |
* |
1138 |
– |
* Seeds for submitters and workers/workQueues work in basically |
1139 |
– |
* the same way but are initialized and updated using slightly |
1140 |
– |
* different mechanics. Both are initialized using the same |
1141 |
– |
* approach as in class ThreadLocal, where successive values are |
1142 |
– |
* unlikely to collide with previous values. This is done during |
1143 |
– |
* registration for workers, but requires a separate AtomicInteger |
1144 |
– |
* for submitters. Seeds are then randomly modified upon |
1145 |
– |
* collisions using xorshifts, which requires a non-zero seed. |
1146 |
– |
*/ |
1147 |
– |
static final class Submitter { |
1148 |
– |
int seed; |
1149 |
– |
Submitter() { |
1150 |
– |
int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT); |
1151 |
– |
seed = (s == 0) ? 1 : s; // ensure non-zero |
1152 |
– |
} |
1153 |
– |
} |
1154 |
– |
|
1155 |
– |
/** ThreadLocal class for Submitters */ |
1156 |
– |
static final class ThreadSubmitter extends ThreadLocal<Submitter> { |
1157 |
– |
public Submitter initialValue() { return new Submitter(); } |
1158 |
– |
} |
1159 |
– |
|
1047 |
|
// static fields (initialized in static initializer below) |
1048 |
|
|
1049 |
|
/** |
1053 |
|
public static final ForkJoinWorkerThreadFactory |
1054 |
|
defaultForkJoinWorkerThreadFactory; |
1055 |
|
|
1169 |
– |
|
1170 |
– |
/** Property prefix for constructing common pool */ |
1171 |
– |
private static final String propPrefix = |
1172 |
– |
"java.util.concurrent.ForkJoinPool.common."; |
1173 |
– |
|
1056 |
|
/** |
1057 |
< |
* Common (static) pool. Non-null for public use unless a static |
1058 |
< |
* construction exception, but internal usages must null-check on |
1059 |
< |
* use. |
1057 |
> |
* Per-thread submission bookkeeping. Shared across all pools |
1058 |
> |
* to reduce ThreadLocal pollution and because random motion |
1059 |
> |
* to avoid contention in one pool is likely to hold for others. |
1060 |
> |
* Lazily initialized on first submission (but null-checked |
1061 |
> |
* in other contexts to avoid unnecessary initialization). |
1062 |
|
*/ |
1063 |
< |
static final ForkJoinPool commonPool; |
1063 |
> |
static final ThreadLocal<Submitter> submitters; |
1064 |
|
|
1065 |
|
/** |
1066 |
< |
* Common pool parallelism. Must equal commonPool.parallelism. |
1066 |
> |
* Permission required for callers of methods that may start or |
1067 |
> |
* kill threads. |
1068 |
|
*/ |
1069 |
< |
static final int commonPoolParallelism; |
1069 |
> |
private static final RuntimePermission modifyThreadPermission; |
1070 |
|
|
1071 |
|
/** |
1072 |
< |
* Generator for assigning sequence numbers as pool names. |
1072 |
> |
* Common (static) pool. Non-null for public use unless a static |
1073 |
> |
* construction exception, but internal usages null-check on use |
1074 |
> |
* to paranoically avoid potential initialization circularities |
1075 |
> |
* as well as to simplify generated code. |
1076 |
|
*/ |
1077 |
< |
private static final AtomicInteger poolNumberGenerator; |
1077 |
> |
static final ForkJoinPool common; |
1078 |
|
|
1079 |
|
/** |
1080 |
< |
* Generator for initial hashes/seeds for submitters. Accessed by |
1193 |
< |
* Submitter class constructor. |
1080 |
> |
* Common pool parallelism. Must equal common.parallelism. |
1081 |
|
*/ |
1082 |
< |
static final AtomicInteger nextSubmitterSeed; |
1082 |
> |
static final int commonParallelism; |
1083 |
|
|
1084 |
|
/** |
1085 |
< |
* Permission required for callers of methods that may start or |
1199 |
< |
* kill threads. |
1085 |
> |
* Sequence number for creating workerNamePrefix. |
1086 |
|
*/ |
1087 |
< |
private static final RuntimePermission modifyThreadPermission; |
1087 |
> |
private static int poolNumberSequence; |
1088 |
|
|
1089 |
|
/** |
1090 |
< |
* Per-thread submission bookkeeping. Shared across all pools |
1091 |
< |
* to reduce ThreadLocal pollution and because random motion |
1206 |
< |
* to avoid contention in one pool is likely to hold for others. |
1090 |
> |
* Returns the next sequence number. We don't expect this to |
1091 |
> |
* ever contend, so use simple builtin sync. |
1092 |
|
*/ |
1093 |
< |
private static final ThreadSubmitter submitters; |
1093 |
> |
private static final synchronized int nextPoolId() { |
1094 |
> |
return ++poolNumberSequence; |
1095 |
> |
} |
1096 |
|
|
1097 |
|
// static constants |
1098 |
|
|
1099 |
|
/** |
1100 |
< |
* Initial timeout value (in nanoseconds) for the thread triggering |
1101 |
< |
* quiescence to park waiting for new work. On timeout, the thread |
1102 |
< |
* will instead try to shrink the number of workers. |
1100 |
> |
* Initial timeout value (in nanoseconds) for the thread |
1101 |
> |
* triggering quiescence to park waiting for new work. On timeout, |
1102 |
> |
* the thread will instead try to shrink the number of |
1103 |
> |
* workers. The value should be large enough to avoid overly |
1104 |
> |
* aggressive shrinkage during most transient stalls (long GCs |
1105 |
> |
* etc). |
1106 |
|
*/ |
1107 |
< |
private static final long IDLE_TIMEOUT = 1000L * 1000L * 1000L; // 1sec |
1107 |
> |
private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec |
1108 |
|
|
1109 |
|
/** |
1110 |
|
* Timeout value when there are more threads than parallelism level |
1111 |
|
*/ |
1112 |
< |
private static final long FAST_IDLE_TIMEOUT = 100L * 1000L * 1000L; |
1112 |
> |
private static final long FAST_IDLE_TIMEOUT = 200L * 1000L * 1000L; |
1113 |
> |
|
1114 |
> |
/** |
1115 |
> |
* Tolerance for idle timeouts, to cope with timer undershoots |
1116 |
> |
*/ |
1117 |
> |
private static final long TIMEOUT_SLOP = 2000000L; |
1118 |
|
|
1119 |
|
/** |
1120 |
|
* The maximum stolen->joining link depth allowed in method |
1121 |
< |
* tryHelpStealer. Must be a power of two. This value also |
1227 |
< |
* controls the maximum number of times to try to help join a task |
1228 |
< |
* without any apparent progress or change in pool state before |
1229 |
< |
* giving up and blocking (see awaitJoin). Depths for legitimate |
1121 |
> |
* tryHelpStealer. Must be a power of two. Depths for legitimate |
1122 |
|
* chains are unbounded, but we use a fixed constant to avoid |
1123 |
|
* (otherwise unchecked) cycles and to bound staleness of |
1124 |
|
* traversal parameters at the expense of sometimes blocking when |
1127 |
|
private static final int MAX_HELP = 64; |
1128 |
|
|
1129 |
|
/** |
1238 |
– |
* Secondary time-based bound (in nanosecs) for helping attempts |
1239 |
– |
* before trying compensated blocking in awaitJoin. Used in |
1240 |
– |
* conjunction with MAX_HELP to reduce variance due to different |
1241 |
– |
* polling rates associated with different helping options. The |
1242 |
– |
* value should roughly approximate the time required to create |
1243 |
– |
* and/or activate a worker thread. |
1244 |
– |
*/ |
1245 |
– |
private static final long COMPENSATION_DELAY = 1L << 18; // ~0.25 millisec |
1246 |
– |
|
1247 |
– |
/** |
1130 |
|
* Increment for seed generators. See class ThreadLocal for |
1131 |
|
* explanation. |
1132 |
|
*/ |
1160 |
|
* scan for them to avoid queuing races. Note however that |
1161 |
|
* eventCount updates lag releases so usage requires care. |
1162 |
|
* |
1163 |
< |
* Field runState is an int packed with: |
1163 |
> |
* Field plock is an int packed with: |
1164 |
|
* SHUTDOWN: true if shutdown is enabled (1 bit) |
1165 |
< |
* SEQ: a sequence number updated upon (de)registering workers (30 bits) |
1166 |
< |
* INIT: set true after workQueues array construction (1 bit) |
1165 |
> |
* SEQ: a sequence lock, with PL_LOCK bit set if locked (30 bits) |
1166 |
> |
* SIGNAL: set when threads may be waiting on the lock (1 bit) |
1167 |
|
* |
1168 |
|
* The sequence number enables simple consistency checks: |
1169 |
|
* Staleness of read-only operations on the workQueues array can |
1170 |
< |
* be checked by comparing runState before vs after the reads. |
1170 |
> |
* be checked by comparing plock before vs after the reads. |
1171 |
|
*/ |
1172 |
|
|
1173 |
|
// bit positions/shifts for fields |
1179 |
|
// bounds |
1180 |
|
private static final int SMASK = 0xffff; // short bits |
1181 |
|
private static final int MAX_CAP = 0x7fff; // max #workers - 1 |
1182 |
< |
private static final int SQMASK = 0xfffe; // even short bits |
1182 |
> |
private static final int EVENMASK = 0xfffe; // even short bits |
1183 |
> |
private static final int SQMASK = 0x007e; // max 64 (even) slots |
1184 |
|
private static final int SHORT_SIGN = 1 << 15; |
1185 |
|
private static final int INT_SIGN = 1 << 31; |
1186 |
|
|
1205 |
|
private static final int E_MASK = 0x7fffffff; // no STOP_BIT |
1206 |
|
private static final int E_SEQ = 1 << EC_SHIFT; |
1207 |
|
|
1208 |
< |
// runState bits |
1208 |
> |
// plock bits |
1209 |
|
private static final int SHUTDOWN = 1 << 31; |
1210 |
+ |
private static final int PL_LOCK = 2; |
1211 |
+ |
private static final int PL_SIGNAL = 1; |
1212 |
+ |
private static final int PL_SPINS = 1 << 8; |
1213 |
|
|
1214 |
|
// access mode for WorkQueue |
1215 |
|
static final int LIFO_QUEUE = 0; |
1216 |
|
static final int FIFO_QUEUE = 1; |
1217 |
|
static final int SHARED_QUEUE = -1; |
1218 |
|
|
1219 |
+ |
// bounds for #steps in scan loop -- must be power 2 minus 1 |
1220 |
+ |
private static final int MIN_SCAN = 0x1ff; // cover estimation slop |
1221 |
+ |
private static final int MAX_SCAN = 0x1ffff; // 4 * max workers |
1222 |
+ |
|
1223 |
|
// Instance fields |
1224 |
|
|
1225 |
|
/* |
1226 |
< |
* Field layout order in this class tends to matter more than one |
1227 |
< |
* would like. Runtime layout order is only loosely related to |
1226 |
> |
* Field layout of this class tends to matter more than one would |
1227 |
> |
* like. Runtime layout order is only loosely related to |
1228 |
|
* declaration order and may differ across JVMs, but the following |
1229 |
|
* empirically works OK on current JVMs. |
1230 |
|
*/ |
1231 |
|
|
1232 |
+ |
// Heuristic padding to ameliorate unfortunate memory placements |
1233 |
+ |
volatile long pad00, pad01, pad02, pad03, pad04, pad05, pad06; |
1234 |
+ |
|
1235 |
|
volatile long stealCount; // collects worker counts |
1236 |
|
volatile long ctl; // main pool control |
1237 |
< |
final int parallelism; // parallelism level |
1238 |
< |
final int localMode; // per-worker scheduling mode |
1239 |
< |
volatile int nextWorkerNumber; // to create worker name string |
1347 |
< |
final int submitMask; // submit queue index bound |
1348 |
< |
int nextSeed; // for initializing worker seeds |
1349 |
< |
volatile int mainLock; // spinlock for array updates |
1350 |
< |
volatile int runState; // shutdown status and seq |
1237 |
> |
volatile int plock; // shutdown status and seqLock |
1238 |
> |
volatile int indexSeed; // worker/submitter index seed |
1239 |
> |
final int config; // mode and parallelism level |
1240 |
|
WorkQueue[] workQueues; // main registry |
1241 |
< |
final ForkJoinWorkerThreadFactory factory; // factory for new workers |
1241 |
> |
final ForkJoinWorkerThreadFactory factory; |
1242 |
|
final Thread.UncaughtExceptionHandler ueh; // per-worker UEH |
1243 |
|
final String workerNamePrefix; // to create worker name string |
1244 |
|
|
1245 |
< |
/* |
1246 |
< |
* Mechanics for main lock protecting worker array updates. Uses |
1358 |
< |
* the same strategy as ConcurrentHashMap bins -- a spinLock for |
1359 |
< |
* normal cases, but falling back to builtin lock when (rarely) |
1360 |
< |
* needed. See internal ConcurrentHashMap documentation for |
1361 |
< |
* explanation. |
1362 |
< |
*/ |
1363 |
< |
|
1364 |
< |
static final int LOCK_WAITING = 2; // bit to indicate need for signal |
1365 |
< |
static final int MAX_LOCK_SPINS = 1 << 8; |
1245 |
> |
volatile Object pad10, pad11, pad12, pad13, pad14, pad15, pad16, pad17; |
1246 |
> |
volatile Object pad18, pad19, pad1a, pad1b; |
1247 |
|
|
1248 |
< |
private void tryAwaitMainLock() { |
1249 |
< |
int spins = MAX_LOCK_SPINS, r = 0, h; |
1250 |
< |
while (((h = mainLock) & 1) != 0) { |
1251 |
< |
if (r == 0) |
1252 |
< |
r = ThreadLocalRandom.current().nextInt(); // randomize spins |
1248 |
> |
/* |
1249 |
> |
* Acquires the plock lock to protect worker array and related |
1250 |
> |
* updates. This method is called only if an initial CAS on plock |
1251 |
> |
* fails. This acts as a spinLock for normal cases, but falls back |
1252 |
> |
* to builtin monitor to block when (rarely) needed. This would be |
1253 |
> |
* a terrible idea for a highly contended lock, but works fine as |
1254 |
> |
* a more conservative alternative to a pure spinlock. |
1255 |
> |
*/ |
1256 |
> |
private int acquirePlock() { |
1257 |
> |
int spins = PL_SPINS, r = 0, ps, nps; |
1258 |
> |
for (;;) { |
1259 |
> |
if (((ps = plock) & PL_LOCK) == 0 && |
1260 |
> |
U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK)) |
1261 |
> |
return nps; |
1262 |
> |
else if (r == 0) { // randomize spins if possible |
1263 |
> |
Thread t = Thread.currentThread(); WorkQueue w; Submitter z; |
1264 |
> |
if ((t instanceof ForkJoinWorkerThread) && |
1265 |
> |
(w = ((ForkJoinWorkerThread)t).workQueue) != null) |
1266 |
> |
r = w.seed; |
1267 |
> |
else if ((z = submitters.get()) != null) |
1268 |
> |
r = z.seed; |
1269 |
> |
else |
1270 |
> |
r = 1; |
1271 |
> |
} |
1272 |
|
else if (spins >= 0) { |
1273 |
|
r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift |
1274 |
|
if (r >= 0) |
1275 |
|
--spins; |
1276 |
|
} |
1277 |
< |
else if (U.compareAndSwapInt(this, MAINLOCK, h, h | LOCK_WAITING)) { |
1277 |
> |
else if (U.compareAndSwapInt(this, PLOCK, ps, ps | PL_SIGNAL)) { |
1278 |
|
synchronized (this) { |
1279 |
< |
if ((mainLock & LOCK_WAITING) != 0) { |
1279 |
> |
if ((plock & PL_SIGNAL) != 0) { |
1280 |
|
try { |
1281 |
|
wait(); |
1282 |
|
} catch (InterruptedException ie) { |
1283 |
< |
Thread.currentThread().interrupt(); |
1283 |
> |
try { |
1284 |
> |
Thread.currentThread().interrupt(); |
1285 |
> |
} catch (SecurityException ignore) { |
1286 |
> |
} |
1287 |
|
} |
1288 |
|
} |
1289 |
|
else |
1290 |
< |
notifyAll(); // possibly won race vs signaller |
1290 |
> |
notifyAll(); |
1291 |
|
} |
1389 |
– |
break; |
1292 |
|
} |
1293 |
|
} |
1294 |
|
} |
1295 |
|
|
1394 |
– |
// Creating, registering, and deregistering workers |
1395 |
– |
|
1296 |
|
/** |
1297 |
< |
* Tries to create and start a worker |
1297 |
> |
* Unlocks and signals any thread waiting for plock. Called only |
1298 |
> |
* when CAS of seq value for unlock fails. |
1299 |
|
*/ |
1300 |
< |
private void addWorker() { |
1301 |
< |
Throwable ex = null; |
1302 |
< |
ForkJoinWorkerThread wt = null; |
1402 |
< |
try { |
1403 |
< |
if ((wt = factory.newThread(this)) != null) { |
1404 |
< |
wt.start(); |
1405 |
< |
return; |
1406 |
< |
} |
1407 |
< |
} catch (Throwable e) { |
1408 |
< |
ex = e; |
1409 |
< |
} |
1410 |
< |
deregisterWorker(wt, ex); // adjust counts etc on failure |
1300 |
> |
private void releasePlock(int ps) { |
1301 |
> |
plock = ps; |
1302 |
> |
synchronized (this) { notifyAll(); } |
1303 |
|
} |
1304 |
|
|
1305 |
|
/** |
1306 |
< |
* Callback from ForkJoinWorkerThread constructor to assign a |
1307 |
< |
* public name. This must be separate from registerWorker because |
1416 |
< |
* it is called during the "super" constructor call in |
1417 |
< |
* ForkJoinWorkerThread. |
1306 |
> |
* Tries to create and start one worker if fewer than target |
1307 |
> |
* parallelism level exist. Adjusts counts etc on failure. |
1308 |
|
*/ |
1309 |
< |
final String nextWorkerName() { |
1310 |
< |
int n; |
1311 |
< |
do {} while (!U.compareAndSwapInt(this, NEXTWORKERNUMBER, |
1312 |
< |
n = nextWorkerNumber, ++n)); |
1313 |
< |
return workerNamePrefix.concat(Integer.toString(n)); |
1309 |
> |
private void tryAddWorker() { |
1310 |
> |
long c; int u; |
1311 |
> |
while ((u = (int)((c = ctl) >>> 32)) < 0 && |
1312 |
> |
(u & SHORT_SIGN) != 0 && (int)c == 0) { |
1313 |
> |
long nc = (long)(((u + UTC_UNIT) & UTC_MASK) | |
1314 |
> |
((u + UAC_UNIT) & UAC_MASK)) << 32; |
1315 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1316 |
> |
ForkJoinWorkerThreadFactory fac; |
1317 |
> |
Throwable ex = null; |
1318 |
> |
ForkJoinWorkerThread wt = null; |
1319 |
> |
try { |
1320 |
> |
if ((fac = factory) != null && |
1321 |
> |
(wt = fac.newThread(this)) != null) { |
1322 |
> |
wt.start(); |
1323 |
> |
break; |
1324 |
> |
} |
1325 |
> |
} catch (Throwable e) { |
1326 |
> |
ex = e; |
1327 |
> |
} |
1328 |
> |
deregisterWorker(wt, ex); |
1329 |
> |
break; |
1330 |
> |
} |
1331 |
> |
} |
1332 |
|
} |
1333 |
|
|
1334 |
+ |
// Registering and deregistering workers |
1335 |
+ |
|
1336 |
|
/** |
1337 |
< |
* Callback from ForkJoinWorkerThread constructor to establish its |
1338 |
< |
* poolIndex and record its WorkQueue. To avoid scanning bias due |
1339 |
< |
* to packing entries in front of the workQueues array, we treat |
1340 |
< |
* the array as a simple power-of-two hash table using per-thread |
1341 |
< |
* seed as hash, expanding as needed. |
1342 |
< |
* |
1343 |
< |
* @param w the worker's queue |
1344 |
< |
*/ |
1345 |
< |
final void registerWorker(WorkQueue w) { |
1346 |
< |
while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) |
1347 |
< |
tryAwaitMainLock(); |
1337 |
> |
* Callback from ForkJoinWorkerThread to establish and record its |
1338 |
> |
* WorkQueue. To avoid scanning bias due to packing entries in |
1339 |
> |
* front of the workQueues array, we treat the array as a simple |
1340 |
> |
* power-of-two hash table using per-thread seed as hash, |
1341 |
> |
* expanding as needed. |
1342 |
> |
* |
1343 |
> |
* @param wt the worker thread |
1344 |
> |
* @return the worker's queue |
1345 |
> |
*/ |
1346 |
> |
final WorkQueue registerWorker(ForkJoinWorkerThread wt) { |
1347 |
> |
Thread.UncaughtExceptionHandler handler; WorkQueue[] ws; int s, ps; |
1348 |
> |
wt.setDaemon(true); |
1349 |
> |
if ((handler = ueh) != null) |
1350 |
> |
wt.setUncaughtExceptionHandler(handler); |
1351 |
> |
do {} while (!U.compareAndSwapInt(this, INDEXSEED, s = indexSeed, |
1352 |
> |
s += SEED_INCREMENT) || |
1353 |
> |
s == 0); // skip 0 |
1354 |
> |
WorkQueue w = new WorkQueue(this, wt, config >>> 16, s); |
1355 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
1356 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1357 |
> |
ps = acquirePlock(); |
1358 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1359 |
|
try { |
1360 |
< |
WorkQueue[] ws; |
1361 |
< |
if ((ws = workQueues) == null) |
1362 |
< |
ws = workQueues = new WorkQueue[submitMask + 1]; |
1363 |
< |
if (w != null) { |
1364 |
< |
int rs, n = ws.length, m = n - 1; |
1365 |
< |
int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence |
1445 |
< |
w.seed = (s == 0) ? 1 : s; // ensure non-zero seed |
1446 |
< |
int r = (s << 1) | 1; // use odd-numbered indices |
1447 |
< |
if (ws[r &= m] != null) { // collision |
1448 |
< |
int probes = 0; // step by approx half size |
1449 |
< |
int step = (n <= 4) ? 2 : ((n >>> 1) & SQMASK) + 2; |
1360 |
> |
if ((ws = workQueues) != null) { // skip if shutting down |
1361 |
> |
int n = ws.length, m = n - 1; |
1362 |
> |
int r = (s << 1) | 1; // use odd-numbered indices |
1363 |
> |
if (ws[r &= m] != null) { // collision |
1364 |
> |
int probes = 0; // step by approx half size |
1365 |
> |
int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2; |
1366 |
|
while (ws[r = (r + step) & m] != null) { |
1367 |
|
if (++probes >= n) { |
1368 |
|
workQueues = ws = Arrays.copyOf(ws, n <<= 1); |
1371 |
|
} |
1372 |
|
} |
1373 |
|
} |
1374 |
< |
w.eventCount = w.poolIndex = r; // establish before recording |
1375 |
< |
ws[r] = w; // also update seq |
1460 |
< |
runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN); |
1374 |
> |
w.eventCount = w.poolIndex = r; // volatile write orders |
1375 |
> |
ws[r] = w; |
1376 |
|
} |
1377 |
|
} finally { |
1378 |
< |
if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { |
1379 |
< |
mainLock = 0; |
1465 |
< |
synchronized (this) { notifyAll(); }; |
1466 |
< |
} |
1378 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1379 |
> |
releasePlock(nps); |
1380 |
|
} |
1381 |
< |
|
1381 |
> |
wt.setName(workerNamePrefix.concat(Integer.toString(w.poolIndex))); |
1382 |
> |
return w; |
1383 |
|
} |
1384 |
|
|
1385 |
|
/** |
1386 |
|
* Final callback from terminating worker, as well as upon failure |
1387 |
< |
* to construct or start a worker in addWorker. Removes record of |
1388 |
< |
* worker from array, and adjusts counts. If pool is shutting |
1389 |
< |
* down, tries to complete termination. |
1387 |
> |
* to construct or start a worker. Removes record of worker from |
1388 |
> |
* array, and adjusts counts. If pool is shutting down, tries to |
1389 |
> |
* complete termination. |
1390 |
|
* |
1391 |
< |
* @param wt the worker thread or null if addWorker failed |
1391 |
> |
* @param wt the worker thread or null if construction failed |
1392 |
|
* @param ex the exception causing failure, or null if none |
1393 |
|
*/ |
1394 |
|
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1395 |
|
WorkQueue w = null; |
1396 |
|
if (wt != null && (w = wt.workQueue) != null) { |
1397 |
< |
w.runState = -1; // ensure runState is set |
1398 |
< |
long steals = w.totalSteals + w.nsteals, sc; |
1397 |
> |
int ps; |
1398 |
> |
w.qlock = -1; // ensure set |
1399 |
> |
long ns = w.nsteals, sc; // collect steal count |
1400 |
|
do {} while (!U.compareAndSwapLong(this, STEALCOUNT, |
1401 |
< |
sc = stealCount, sc + steals)); |
1402 |
< |
int idx = w.poolIndex; |
1403 |
< |
while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) |
1404 |
< |
tryAwaitMainLock(); |
1401 |
> |
sc = stealCount, sc + ns)); |
1402 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
1403 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1404 |
> |
ps = acquirePlock(); |
1405 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1406 |
|
try { |
1407 |
+ |
int idx = w.poolIndex; |
1408 |
|
WorkQueue[] ws = workQueues; |
1409 |
|
if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) |
1410 |
|
ws[idx] = null; |
1411 |
|
} finally { |
1412 |
< |
if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { |
1413 |
< |
mainLock = 0; |
1497 |
< |
synchronized (this) { notifyAll(); }; |
1498 |
< |
} |
1412 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1413 |
> |
releasePlock(nps); |
1414 |
|
} |
1415 |
|
} |
1416 |
|
|
1417 |
< |
long c; // adjust ctl counts |
1417 |
> |
long c; // adjust ctl counts |
1418 |
|
do {} while (!U.compareAndSwapLong |
1419 |
|
(this, CTL, c = ctl, (((c - AC_UNIT) & AC_MASK) | |
1420 |
|
((c - TC_UNIT) & TC_MASK) | |
1421 |
|
(c & ~(AC_MASK|TC_MASK))))); |
1422 |
|
|
1423 |
< |
if (!tryTerminate(false, false) && w != null) { |
1424 |
< |
w.cancelAll(); // cancel remaining tasks |
1425 |
< |
if (w.array != null) // suppress signal if never ran |
1426 |
< |
signalWork(); // wake up or create replacement |
1427 |
< |
if (ex == null) // help clean refs on way out |
1428 |
< |
ForkJoinTask.helpExpungeStaleExceptions(); |
1423 |
> |
if (!tryTerminate(false, false) && w != null && w.array != null) { |
1424 |
> |
w.cancelAll(); // cancel remaining tasks |
1425 |
> |
WorkQueue[] ws; WorkQueue v; Thread p; int u, i, e; |
1426 |
> |
while ((u = (int)((c = ctl) >>> 32)) < 0 && (e = (int)c) >= 0) { |
1427 |
> |
if (e > 0) { // activate or create replacement |
1428 |
> |
if ((ws = workQueues) == null || |
1429 |
> |
(i = e & SMASK) >= ws.length || |
1430 |
> |
(v = ws[i]) == null) |
1431 |
> |
break; |
1432 |
> |
long nc = (((long)(v.nextWait & E_MASK)) | |
1433 |
> |
((long)(u + UAC_UNIT) << 32)); |
1434 |
> |
if (v.eventCount != (e | INT_SIGN)) |
1435 |
> |
break; |
1436 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1437 |
> |
v.eventCount = (e + E_SEQ) & E_MASK; |
1438 |
> |
if ((p = v.parker) != null) |
1439 |
> |
U.unpark(p); |
1440 |
> |
break; |
1441 |
> |
} |
1442 |
> |
} |
1443 |
> |
else { |
1444 |
> |
if ((short)u < 0) |
1445 |
> |
tryAddWorker(); |
1446 |
> |
break; |
1447 |
> |
} |
1448 |
> |
} |
1449 |
|
} |
1450 |
< |
|
1451 |
< |
if (ex != null) // rethrow |
1452 |
< |
U.throwException(ex); |
1450 |
> |
if (ex == null) // help clean refs on way out |
1451 |
> |
ForkJoinTask.helpExpungeStaleExceptions(); |
1452 |
> |
else // rethrow |
1453 |
> |
ForkJoinTask.rethrow(ex); |
1454 |
|
} |
1455 |
|
|
1456 |
|
// Submissions |
1458 |
|
/** |
1459 |
|
* Unless shutting down, adds the given task to a submission queue |
1460 |
|
* at submitter's current queue index (modulo submission |
1461 |
< |
* range). If no queue exists at the index, one is created. If |
1462 |
< |
* the queue is busy, another index is randomly chosen. The |
1527 |
< |
* submitMask bounds the effective number of queues to the |
1528 |
< |
* (nearest power of two for) parallelism level. |
1461 |
> |
* range). Only the most common path is directly handled in this |
1462 |
> |
* method. All others are relayed to fullExternalPush. |
1463 |
|
* |
1464 |
|
* @param task the task. Caller must ensure non-null. |
1465 |
|
*/ |
1466 |
< |
private void doSubmit(ForkJoinTask<?> task) { |
1467 |
< |
Submitter s = submitters.get(); |
1468 |
< |
for (int r = s.seed, m = submitMask;;) { |
1469 |
< |
WorkQueue[] ws; WorkQueue q; |
1470 |
< |
int k = r & m & SQMASK; // use only even indices |
1471 |
< |
if (runState < 0) |
1472 |
< |
throw new RejectedExecutionException(); // shutting down |
1473 |
< |
else if ((ws = workQueues) == null || ws.length <= k) { |
1474 |
< |
while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) |
1475 |
< |
tryAwaitMainLock(); |
1476 |
< |
try { |
1477 |
< |
if (workQueues == null) |
1478 |
< |
workQueues = new WorkQueue[submitMask + 1]; |
1479 |
< |
} finally { |
1546 |
< |
if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { |
1547 |
< |
mainLock = 0; |
1548 |
< |
synchronized (this) { notifyAll(); }; |
1549 |
< |
} |
1550 |
< |
} |
1551 |
< |
} |
1552 |
< |
else if ((q = ws[k]) == null) { // create new queue |
1553 |
< |
WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE); |
1554 |
< |
while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) |
1555 |
< |
tryAwaitMainLock(); |
1556 |
< |
try { |
1557 |
< |
int rs = runState; // to update seq |
1558 |
< |
if (ws == workQueues && ws[k] == null) { |
1559 |
< |
ws[k] = nq; |
1560 |
< |
runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN)); |
1561 |
< |
} |
1562 |
< |
} finally { |
1563 |
< |
if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { |
1564 |
< |
mainLock = 0; |
1565 |
< |
synchronized (this) { notifyAll(); }; |
1566 |
< |
} |
1567 |
< |
} |
1568 |
< |
} |
1569 |
< |
else if (q.trySharedPush(task)) { |
1570 |
< |
signalWork(); |
1466 |
> |
final void externalPush(ForkJoinTask<?> task) { |
1467 |
> |
WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a; |
1468 |
> |
if ((z = submitters.get()) != null && plock > 0 && |
1469 |
> |
(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 && |
1470 |
> |
(q = ws[m & z.seed & SQMASK]) != null && |
1471 |
> |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock |
1472 |
> |
int b = q.base, s = q.top, n, an; |
1473 |
> |
if ((a = q.array) != null && (an = a.length) > (n = s + 1 - b)) { |
1474 |
> |
int j = (((an - 1) & s) << ASHIFT) + ABASE; |
1475 |
> |
U.putOrderedObject(a, j, task); |
1476 |
> |
q.top = s + 1; // push on to deque |
1477 |
> |
q.qlock = 0; |
1478 |
> |
if (n <= 2) |
1479 |
> |
signalWork(q); |
1480 |
|
return; |
1481 |
|
} |
1482 |
< |
else if (m > 1) { // move to a different index |
1574 |
< |
r ^= r << 13; // same xorshift as WorkQueues |
1575 |
< |
r ^= r >>> 17; |
1576 |
< |
s.seed = r ^= r << 5; |
1577 |
< |
} |
1578 |
< |
else |
1579 |
< |
Thread.yield(); // yield if no alternatives |
1482 |
> |
q.qlock = 0; |
1483 |
|
} |
1484 |
+ |
fullExternalPush(task); |
1485 |
|
} |
1486 |
|
|
1487 |
|
/** |
1488 |
< |
* Submits the given (non-null) task to the common pool, if possible. |
1489 |
< |
*/ |
1490 |
< |
static void submitToCommonPool(ForkJoinTask<?> task) { |
1491 |
< |
ForkJoinPool p; |
1492 |
< |
if ((p = commonPool) == null) |
1493 |
< |
throw new RejectedExecutionException("Common Pool Unavailable"); |
1494 |
< |
p.doSubmit(task); |
1495 |
< |
} |
1496 |
< |
|
1497 |
< |
/** |
1498 |
< |
* Returns true if caller is (or may be) submitter to the common |
1499 |
< |
* pool, and not all workers are active, and there appear to be |
1500 |
< |
* tasks in the associated submission queue. |
1501 |
< |
*/ |
1502 |
< |
static boolean canHelpCommonPool() { |
1503 |
< |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; |
1504 |
< |
int k = submitters.get().seed & SQMASK; |
1505 |
< |
return ((p = commonPool) != null && |
1506 |
< |
(int)(p.ctl >> AC_SHIFT) < 0 && |
1507 |
< |
(ws = p.workQueues) != null && |
1508 |
< |
ws.length > (k &= p.submitMask) && |
1509 |
< |
(q = ws[k]) != null && |
1510 |
< |
q.top - q.base > 0); |
1511 |
< |
} |
1512 |
< |
|
1513 |
< |
/** |
1514 |
< |
* Returns true if the given task was submitted to common pool |
1515 |
< |
* and has not yet commenced execution, and is available for |
1516 |
< |
* removal according to execution policies; if so removing the |
1517 |
< |
* submission from the pool. |
1518 |
< |
* |
1519 |
< |
* @param task the task |
1520 |
< |
* @return true if successful |
1521 |
< |
*/ |
1522 |
< |
static boolean tryUnsubmitFromCommonPool(ForkJoinTask<?> task) { |
1523 |
< |
// Peek, looking for task and eligibility before |
1524 |
< |
// using trySharedUnpush to actually take it under lock |
1525 |
< |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; |
1526 |
< |
ForkJoinTask<?>[] a; int s; |
1527 |
< |
int k = submitters.get().seed & SQMASK; |
1528 |
< |
return ((p = commonPool) != null && |
1529 |
< |
(int)(p.ctl >> AC_SHIFT) < 0 && |
1530 |
< |
(ws = p.workQueues) != null && |
1531 |
< |
ws.length > (k &= p.submitMask) && |
1532 |
< |
(q = ws[k]) != null && |
1533 |
< |
(a = q.array) != null && |
1534 |
< |
(s = q.top - 1) - q.base >= 0 && |
1535 |
< |
s >= 0 && s < a.length && |
1536 |
< |
a[s] == task && |
1537 |
< |
q.trySharedUnpush(task)); |
1538 |
< |
} |
1539 |
< |
|
1540 |
< |
/** |
1541 |
< |
* Tries to pop a task from common pool with given root |
1542 |
< |
*/ |
1543 |
< |
static ForkJoinTask<?> popCCFromCommonPool(CountedCompleter<?> root) { |
1544 |
< |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; |
1545 |
< |
ForkJoinTask<?> t; |
1546 |
< |
int k = submitters.get().seed & SQMASK; |
1547 |
< |
if (root != null && |
1548 |
< |
(p = commonPool) != null && |
1549 |
< |
(int)(p.ctl >> AC_SHIFT) < 0 && |
1550 |
< |
(ws = p.workQueues) != null && |
1551 |
< |
ws.length > (k &= p.submitMask) && |
1552 |
< |
(q = ws[k]) != null && q.top - q.base > 0 && |
1553 |
< |
root.status < 0 && |
1554 |
< |
(t = q.sharedPopCC(root)) != null) |
1555 |
< |
return t; |
1556 |
< |
return null; |
1488 |
> |
* Full version of externalPush. This method is called, among |
1489 |
> |
* other times, upon the first submission of the first task to the |
1490 |
> |
* pool, so must perform secondary initialization. It also |
1491 |
> |
* detects first submission by an external thread by looking up |
1492 |
> |
* its ThreadLocal, and creates a new shared queue if the one at |
1493 |
> |
* index if empty or contended. The plock lock body must be |
1494 |
> |
* exception-free (so no try/finally) so we optimistically |
1495 |
> |
* allocate new queues outside the lock and throw them away if |
1496 |
> |
* (very rarely) not needed. |
1497 |
> |
* |
1498 |
> |
* Secondary initialization occurs when plock is zero, to create |
1499 |
> |
* workQueue array and set plock to a valid value. This lock body |
1500 |
> |
* must also be exception-free. Because the plock seq value can |
1501 |
> |
* eventually wrap around zero, this method harmlessly fails to |
1502 |
> |
* reinitialize if workQueues exists, while still advancing plock. |
1503 |
> |
*/ |
1504 |
> |
private void fullExternalPush(ForkJoinTask<?> task) { |
1505 |
> |
int r = 0; // random index seed |
1506 |
> |
for (Submitter z = submitters.get();;) { |
1507 |
> |
WorkQueue[] ws; WorkQueue q; int ps, m, k; |
1508 |
> |
if (z == null) { |
1509 |
> |
if (U.compareAndSwapInt(this, INDEXSEED, r = indexSeed, |
1510 |
> |
r += SEED_INCREMENT) && r != 0) |
1511 |
> |
submitters.set(z = new Submitter(r)); |
1512 |
> |
} |
1513 |
> |
else if (r == 0) { // move to a different index |
1514 |
> |
r = z.seed; |
1515 |
> |
r ^= r << 13; // same xorshift as WorkQueues |
1516 |
> |
r ^= r >>> 17; |
1517 |
> |
z.seed = r ^ (r << 5); |
1518 |
> |
} |
1519 |
> |
else if ((ps = plock) < 0) |
1520 |
> |
throw new RejectedExecutionException(); |
1521 |
> |
else if (ps == 0 || (ws = workQueues) == null || |
1522 |
> |
(m = ws.length - 1) < 0) { // initialize workQueues |
1523 |
> |
int p = config & SMASK; // find power of two table size |
1524 |
> |
int n = (p > 1) ? p - 1 : 1; // ensure at least 2 slots |
1525 |
> |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; |
1526 |
> |
n |= n >>> 8; n |= n >>> 16; n = (n + 1) << 1; |
1527 |
> |
WorkQueue[] nws = ((ws = workQueues) == null || ws.length == 0 ? |
1528 |
> |
new WorkQueue[n] : null); |
1529 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
1530 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1531 |
> |
ps = acquirePlock(); |
1532 |
> |
if (((ws = workQueues) == null || ws.length == 0) && nws != null) |
1533 |
> |
workQueues = nws; |
1534 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1535 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1536 |
> |
releasePlock(nps); |
1537 |
> |
} |
1538 |
> |
else if ((q = ws[k = r & m & SQMASK]) != null) { |
1539 |
> |
if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
1540 |
> |
ForkJoinTask<?>[] a = q.array; |
1541 |
> |
int s = q.top; |
1542 |
> |
boolean submitted = false; |
1543 |
> |
try { // locked version of push |
1544 |
> |
if ((a != null && a.length > s + 1 - q.base) || |
1545 |
> |
(a = q.growArray()) != null) { // must presize |
1546 |
> |
int j = (((a.length - 1) & s) << ASHIFT) + ABASE; |
1547 |
> |
U.putOrderedObject(a, j, task); |
1548 |
> |
q.top = s + 1; |
1549 |
> |
submitted = true; |
1550 |
> |
} |
1551 |
> |
} finally { |
1552 |
> |
q.qlock = 0; // unlock |
1553 |
> |
} |
1554 |
> |
if (submitted) { |
1555 |
> |
signalWork(q); |
1556 |
> |
return; |
1557 |
> |
} |
1558 |
> |
} |
1559 |
> |
r = 0; // move on failure |
1560 |
> |
} |
1561 |
> |
else if (((ps = plock) & PL_LOCK) == 0) { // create new queue |
1562 |
> |
q = new WorkQueue(this, null, SHARED_QUEUE, r); |
1563 |
> |
if (((ps = plock) & PL_LOCK) != 0 || |
1564 |
> |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
1565 |
> |
ps = acquirePlock(); |
1566 |
> |
if ((ws = workQueues) != null && k < ws.length && ws[k] == null) |
1567 |
> |
ws[k] = q; |
1568 |
> |
int nps = (ps & SHUTDOWN) | ((ps + PL_LOCK) & ~SHUTDOWN); |
1569 |
> |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
1570 |
> |
releasePlock(nps); |
1571 |
> |
} |
1572 |
> |
else |
1573 |
> |
r = 0; // try elsewhere while lock held |
1574 |
> |
} |
1575 |
|
} |
1576 |
|
|
1655 |
– |
|
1577 |
|
// Maintaining ctl counts |
1578 |
|
|
1579 |
|
/** |
1585 |
|
} |
1586 |
|
|
1587 |
|
/** |
1588 |
< |
* Tries to create one or activate one or more workers if too few are active. |
1588 |
> |
* Tries to create or activate a worker if too few are active. |
1589 |
> |
* |
1590 |
> |
* @param q the (non-null) queue holding tasks to be signalled |
1591 |
|
*/ |
1592 |
< |
final void signalWork() { |
1593 |
< |
long c; int u; |
1594 |
< |
while ((u = (int)((c = ctl) >>> 32)) < 0) { // too few active |
1595 |
< |
WorkQueue[] ws = workQueues; int e, i; WorkQueue w; Thread p; |
1596 |
< |
if ((e = (int)c) > 0) { // at least one waiting |
1597 |
< |
if (ws != null && (i = e & SMASK) < ws.length && |
1592 |
> |
final void signalWork(WorkQueue q) { |
1593 |
> |
int hint = q.poolIndex; |
1594 |
> |
long c; int e, u, i, n; WorkQueue[] ws; WorkQueue w; Thread p; |
1595 |
> |
while ((u = (int)((c = ctl) >>> 32)) < 0) { |
1596 |
> |
if ((e = (int)c) > 0) { |
1597 |
> |
if ((ws = workQueues) != null && ws.length > (i = e & SMASK) && |
1598 |
|
(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { |
1599 |
|
long nc = (((long)(w.nextWait & E_MASK)) | |
1600 |
|
((long)(u + UAC_UNIT) << 32)); |
1601 |
|
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1602 |
+ |
w.hint = hint; |
1603 |
|
w.eventCount = (e + E_SEQ) & E_MASK; |
1604 |
|
if ((p = w.parker) != null) |
1605 |
< |
U.unpark(p); // activate and release |
1605 |
> |
U.unpark(p); |
1606 |
|
break; |
1607 |
|
} |
1608 |
+ |
if (q.top - q.base <= 0) |
1609 |
+ |
break; |
1610 |
|
} |
1611 |
|
else |
1612 |
|
break; |
1613 |
|
} |
1614 |
< |
else if (e == 0 && (u & SHORT_SIGN) != 0) { // too few total |
1615 |
< |
long nc = (long)(((u + UTC_UNIT) & UTC_MASK) | |
1616 |
< |
((u + UAC_UNIT) & UAC_MASK)) << 32; |
1691 |
< |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1692 |
< |
addWorker(); |
1693 |
< |
break; |
1694 |
< |
} |
1695 |
< |
} |
1696 |
< |
else |
1614 |
> |
else { |
1615 |
> |
if ((short)u < 0) |
1616 |
> |
tryAddWorker(); |
1617 |
|
break; |
1618 |
+ |
} |
1619 |
|
} |
1620 |
|
} |
1621 |
|
|
1625 |
|
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1626 |
|
*/ |
1627 |
|
final void runWorker(WorkQueue w) { |
1628 |
< |
w.growArray(false); // initialize queue array in this thread |
1629 |
< |
do { w.runTask(scan(w)); } while (w.runState >= 0); |
1628 |
> |
w.growArray(); // allocate queue |
1629 |
> |
do { w.runTask(scan(w)); } while (w.qlock >= 0); |
1630 |
|
} |
1631 |
|
|
1632 |
|
/** |
1637 |
|
* contention, or state changes that indicate possible success on |
1638 |
|
* re-invocation. |
1639 |
|
* |
1640 |
< |
* The scan searches for tasks across a random permutation of |
1641 |
< |
* queues (starting at a random index and stepping by a random |
1642 |
< |
* relative prime, checking each at least once). The scan |
1643 |
< |
* terminates upon either finding a non-empty queue, or completing |
1644 |
< |
* the sweep. If the worker is not inactivated, it takes and |
1645 |
< |
* returns a task from this queue. On failure to find a task, we |
1646 |
< |
* take one of the following actions, after which the caller will |
1647 |
< |
* retry calling this method unless terminated. |
1640 |
> |
* The scan searches for tasks across queues (starting at a random |
1641 |
> |
* index, and relying on registerWorker to irregularly scatter |
1642 |
> |
* them within array to avoid bias), checking each at least twice. |
1643 |
> |
* The scan terminates upon either finding a non-empty queue, or |
1644 |
> |
* completing the sweep. If the worker is not inactivated, it |
1645 |
> |
* takes and returns a task from this queue. Otherwise, if not |
1646 |
> |
* activated, it signals workers (that may include itself) and |
1647 |
> |
* returns so caller can retry. Also returns for true if the |
1648 |
> |
* worker array may have changed during an empty scan. On failure |
1649 |
> |
* to find a task, we take one of the following actions, after |
1650 |
> |
* which the caller will retry calling this method unless |
1651 |
> |
* terminated. |
1652 |
|
* |
1653 |
|
* * If pool is terminating, terminate the worker. |
1654 |
|
* |
1730 |
– |
* * If not a complete sweep, try to release a waiting worker. If |
1731 |
– |
* the scan terminated because the worker is inactivated, then the |
1732 |
– |
* released worker will often be the calling worker, and it can |
1733 |
– |
* succeed obtaining a task on the next call. Or maybe it is |
1734 |
– |
* another worker, but with same net effect. Releasing in other |
1735 |
– |
* cases as well ensures that we have enough workers running. |
1736 |
– |
* |
1655 |
|
* * If not already enqueued, try to inactivate and enqueue the |
1656 |
|
* worker on wait queue. Or, if inactivating has caused the pool |
1657 |
< |
* to be quiescent, relay to idleAwaitWork to check for |
1658 |
< |
* termination and possibly shrink pool. |
1657 |
> |
* to be quiescent, relay to idleAwaitWork to possibly shrink |
1658 |
> |
* pool. |
1659 |
|
* |
1660 |
< |
* * If already inactive, and the caller has run a task since the |
1661 |
< |
* last empty scan, return (to allow rescan) unless others are |
1744 |
< |
* also inactivated. Field WorkQueue.rescans counts down on each |
1745 |
< |
* scan to ensure eventual inactivation and blocking. |
1660 |
> |
* * If already enqueued and none of the above apply, possibly |
1661 |
> |
* park awaiting signal, else lingering to help scan and signal. |
1662 |
|
* |
1663 |
< |
* * If already enqueued and none of the above apply, park |
1664 |
< |
* awaiting signal, |
1663 |
> |
* * If a non-empty queue discovered or left as a hint, |
1664 |
> |
* help wake up other workers before return. |
1665 |
|
* |
1666 |
|
* @param w the worker (via its WorkQueue) |
1667 |
|
* @return a task or null if none found |
1668 |
|
*/ |
1669 |
|
private final ForkJoinTask<?> scan(WorkQueue w) { |
1670 |
< |
WorkQueue[] ws; // first update random seed |
1671 |
< |
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
1672 |
< |
int rs = runState, m; // volatile read order matters |
1673 |
< |
if ((ws = workQueues) != null && (m = ws.length - 1) > 0) { |
1674 |
< |
int ec = w.eventCount; // ec is negative if inactive |
1675 |
< |
int step = (r >>> 16) | 1; // relative prime |
1676 |
< |
for (int j = (m + 1) << 2; ; r += step) { |
1677 |
< |
WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b; |
1678 |
< |
if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 && |
1679 |
< |
(a = q.array) != null) { // probably nonempty |
1670 |
> |
WorkQueue[] ws; int m; |
1671 |
> |
int ps = plock; // read plock before ws |
1672 |
> |
if (w != null && (ws = workQueues) != null && (m = ws.length - 1) >= 0) { |
1673 |
> |
int ec = w.eventCount; // ec is negative if inactive |
1674 |
> |
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
1675 |
> |
w.hint = -1; // update seed and clear hint |
1676 |
> |
int j = ((m + m + 1) | MIN_SCAN) & MAX_SCAN; |
1677 |
> |
do { |
1678 |
> |
WorkQueue q; ForkJoinTask<?>[] a; int b; |
1679 |
> |
if ((q = ws[(r + j) & m]) != null && (b = q.base) - q.top < 0 && |
1680 |
> |
(a = q.array) != null) { // probably nonempty |
1681 |
|
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1682 |
< |
t = (ForkJoinTask<?>)U.getObjectVolatile(a, i); |
1682 |
> |
ForkJoinTask<?> t = (ForkJoinTask<?>) |
1683 |
> |
U.getObjectVolatile(a, i); |
1684 |
|
if (q.base == b && ec >= 0 && t != null && |
1685 |
|
U.compareAndSwapObject(a, i, t, null)) { |
1686 |
< |
if (q.top - (q.base = b + 1) > 0) |
1687 |
< |
signalWork(); // help pushes signal |
1688 |
< |
return t; |
1689 |
< |
} |
1690 |
< |
else if (ec < 0 || j <= m) { |
1691 |
< |
rs = 0; // mark scan as imcomplete |
1692 |
< |
break; // caller can retry after release |
1693 |
< |
} |
1694 |
< |
} |
1695 |
< |
if (--j < 0) |
1696 |
< |
break; |
1697 |
< |
} |
1698 |
< |
|
1699 |
< |
long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns; |
1700 |
< |
if (e < 0) // decode ctl on empty scan |
1701 |
< |
w.runState = -1; // pool is terminating |
1702 |
< |
else if (rs == 0 || rs != runState) { // incomplete scan |
1703 |
< |
WorkQueue v; Thread p; // try to release a waiter |
1704 |
< |
if (e > 0 && a < 0 && w.eventCount == ec && |
1705 |
< |
(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) { |
1706 |
< |
long nc = ((long)(v.nextWait & E_MASK) | |
1707 |
< |
((c + AC_UNIT) & (AC_MASK|TC_MASK))); |
1708 |
< |
if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) { |
1686 |
> |
if ((q.base = b + 1) - q.top < 0) |
1687 |
> |
signalWork(q); |
1688 |
> |
return t; // taken |
1689 |
> |
} |
1690 |
> |
else if ((ec < 0 || j < m) && (int)(ctl >> AC_SHIFT) <= 0) { |
1691 |
> |
w.hint = (r + j) & m; // help signal below |
1692 |
> |
break; // cannot take |
1693 |
> |
} |
1694 |
> |
} |
1695 |
> |
} while (--j >= 0); |
1696 |
> |
|
1697 |
> |
int h, e, ns; long c, sc; WorkQueue q; |
1698 |
> |
if ((ns = w.nsteals) != 0) { |
1699 |
> |
if (U.compareAndSwapLong(this, STEALCOUNT, |
1700 |
> |
sc = stealCount, sc + ns)) |
1701 |
> |
w.nsteals = 0; // collect steals and rescan |
1702 |
> |
} |
1703 |
> |
else if (plock != ps) // consistency check |
1704 |
> |
; // skip |
1705 |
> |
else if ((e = (int)(c = ctl)) < 0) |
1706 |
> |
w.qlock = -1; // pool is terminating |
1707 |
> |
else { |
1708 |
> |
if ((h = w.hint) < 0) { |
1709 |
> |
if (ec >= 0) { // try to enqueue/inactivate |
1710 |
> |
long nc = (((long)ec | |
1711 |
> |
((c - AC_UNIT) & (AC_MASK|TC_MASK)))); |
1712 |
> |
w.nextWait = e; // link and mark inactive |
1713 |
> |
w.eventCount = ec | INT_SIGN; |
1714 |
> |
if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc)) |
1715 |
> |
w.eventCount = ec; // unmark on CAS failure |
1716 |
> |
else if ((int)(c >> AC_SHIFT) == 1 - (config & SMASK)) |
1717 |
> |
idleAwaitWork(w, nc, c); |
1718 |
> |
} |
1719 |
> |
else if (w.eventCount < 0 && ctl == c) { |
1720 |
> |
Thread wt = Thread.currentThread(); |
1721 |
> |
Thread.interrupted(); // clear status |
1722 |
> |
U.putObject(wt, PARKBLOCKER, this); |
1723 |
> |
w.parker = wt; // emulate LockSupport.park |
1724 |
> |
if (w.eventCount < 0) // recheck |
1725 |
> |
U.park(false, 0L); // block |
1726 |
> |
w.parker = null; |
1727 |
> |
U.putObject(wt, PARKBLOCKER, null); |
1728 |
> |
} |
1729 |
> |
} |
1730 |
> |
if ((h >= 0 || (h = w.hint) >= 0) && |
1731 |
> |
(ws = workQueues) != null && h < ws.length && |
1732 |
> |
(q = ws[h]) != null) { // signal others before retry |
1733 |
> |
WorkQueue v; Thread p; int u, i, s; |
1734 |
> |
for (int n = (config & SMASK) - 1;;) { |
1735 |
> |
int idleCount = (w.eventCount < 0) ? 0 : -1; |
1736 |
> |
if (((s = idleCount - q.base + q.top) <= n && |
1737 |
> |
(n = s) <= 0) || |
1738 |
> |
(u = (int)((c = ctl) >>> 32)) >= 0 || |
1739 |
> |
(e = (int)c) <= 0 || m < (i = e & SMASK) || |
1740 |
> |
(v = ws[i]) == null) |
1741 |
> |
break; |
1742 |
> |
long nc = (((long)(v.nextWait & E_MASK)) | |
1743 |
> |
((long)(u + UAC_UNIT) << 32)); |
1744 |
> |
if (v.eventCount != (e | INT_SIGN) || |
1745 |
> |
!U.compareAndSwapLong(this, CTL, c, nc)) |
1746 |
> |
break; |
1747 |
> |
v.hint = h; |
1748 |
|
v.eventCount = (e + E_SEQ) & E_MASK; |
1749 |
|
if ((p = v.parker) != null) |
1750 |
|
U.unpark(p); |
1751 |
+ |
if (--n <= 0) |
1752 |
+ |
break; |
1753 |
|
} |
1754 |
|
} |
1755 |
|
} |
1797 |
– |
else if (ec >= 0) { // try to enqueue/inactivate |
1798 |
– |
long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); |
1799 |
– |
w.nextWait = e; |
1800 |
– |
w.eventCount = ec | INT_SIGN; // mark as inactive |
1801 |
– |
if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc)) |
1802 |
– |
w.eventCount = ec; // unmark on CAS failure |
1803 |
– |
else { |
1804 |
– |
if ((ns = w.nsteals) != 0) { |
1805 |
– |
w.nsteals = 0; // set rescans if ran task |
1806 |
– |
w.rescans = (a > 0) ? 0 : a + parallelism; |
1807 |
– |
w.totalSteals += ns; |
1808 |
– |
} |
1809 |
– |
if (a == 1 - parallelism) // quiescent |
1810 |
– |
idleAwaitWork(w, nc, c); |
1811 |
– |
} |
1812 |
– |
} |
1813 |
– |
else if (w.eventCount < 0) { // already queued |
1814 |
– |
int ac = a + parallelism; |
1815 |
– |
if ((nr = w.rescans) > 0) // continue rescanning |
1816 |
– |
w.rescans = (ac < nr) ? ac : nr - 1; |
1817 |
– |
else if (((w.seed >>> 16) & ac) == 0) { // randomize park |
1818 |
– |
Thread.interrupted(); // clear status |
1819 |
– |
Thread wt = Thread.currentThread(); |
1820 |
– |
U.putObject(wt, PARKBLOCKER, this); |
1821 |
– |
w.parker = wt; // emulate LockSupport.park |
1822 |
– |
if (w.eventCount < 0) // recheck |
1823 |
– |
U.park(false, 0L); |
1824 |
– |
w.parker = null; |
1825 |
– |
U.putObject(wt, PARKBLOCKER, null); |
1826 |
– |
} |
1827 |
– |
} |
1756 |
|
} |
1757 |
|
return null; |
1758 |
|
} |
1770 |
|
* @param prevCtl the ctl value to restore if thread is terminated |
1771 |
|
*/ |
1772 |
|
private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) { |
1773 |
< |
if (w.eventCount < 0 && !tryTerminate(false, false) && |
1774 |
< |
(int)prevCtl != 0 && !hasQueuedSubmissions() && ctl == currentCtl) { |
1773 |
> |
if (w != null && w.eventCount < 0 && |
1774 |
> |
!tryTerminate(false, false) && (int)prevCtl != 0 && |
1775 |
> |
ctl == currentCtl) { |
1776 |
|
int dc = -(short)(currentCtl >>> TC_SHIFT); |
1777 |
|
long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT; |
1778 |
< |
long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop |
1778 |
> |
long deadline = System.nanoTime() + parkTime - TIMEOUT_SLOP; |
1779 |
|
Thread wt = Thread.currentThread(); |
1780 |
|
while (ctl == currentCtl) { |
1781 |
|
Thread.interrupted(); // timed variant of version in scan() |
1790 |
|
if (deadline - System.nanoTime() <= 0L && |
1791 |
|
U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) { |
1792 |
|
w.eventCount = (w.eventCount + E_SEQ) | E_MASK; |
1793 |
< |
w.runState = -1; // shrink |
1793 |
> |
w.hint = -1; |
1794 |
> |
w.qlock = -1; // shrink |
1795 |
|
break; |
1796 |
|
} |
1797 |
|
} |
1799 |
|
} |
1800 |
|
|
1801 |
|
/** |
1802 |
+ |
* Scans through queues looking for work while joining a task; if |
1803 |
+ |
* any present, signals. May return early if more signalling is |
1804 |
+ |
* detectably unneeded. |
1805 |
+ |
* |
1806 |
+ |
* @param task return early if done |
1807 |
+ |
* @param origin an index to start scan |
1808 |
+ |
*/ |
1809 |
+ |
private void helpSignal(ForkJoinTask<?> task, int origin) { |
1810 |
+ |
WorkQueue[] ws; WorkQueue w; Thread p; long c; int m, u, e, i, s; |
1811 |
+ |
if (task != null && task.status >= 0 && |
1812 |
+ |
(u = (int)(ctl >>> 32)) < 0 && (u >> UAC_SHIFT) < 0 && |
1813 |
+ |
(ws = workQueues) != null && (m = ws.length - 1) >= 0) { |
1814 |
+ |
outer: for (int k = origin, j = m; j >= 0; --j) { |
1815 |
+ |
WorkQueue q = ws[k++ & m]; |
1816 |
+ |
for (int n = m;;) { // limit to at most m signals |
1817 |
+ |
if (task.status < 0) |
1818 |
+ |
break outer; |
1819 |
+ |
if (q == null || |
1820 |
+ |
((s = -q.base + q.top) <= n && (n = s) <= 0)) |
1821 |
+ |
break; |
1822 |
+ |
if ((u = (int)((c = ctl) >>> 32)) >= 0 || |
1823 |
+ |
(e = (int)c) <= 0 || m < (i = e & SMASK) || |
1824 |
+ |
(w = ws[i]) == null) |
1825 |
+ |
break outer; |
1826 |
+ |
long nc = (((long)(w.nextWait & E_MASK)) | |
1827 |
+ |
((long)(u + UAC_UNIT) << 32)); |
1828 |
+ |
if (w.eventCount != (e | INT_SIGN)) |
1829 |
+ |
break outer; |
1830 |
+ |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1831 |
+ |
w.eventCount = (e + E_SEQ) & E_MASK; |
1832 |
+ |
if ((p = w.parker) != null) |
1833 |
+ |
U.unpark(p); |
1834 |
+ |
if (--n <= 0) |
1835 |
+ |
break; |
1836 |
+ |
} |
1837 |
+ |
} |
1838 |
+ |
} |
1839 |
+ |
} |
1840 |
+ |
} |
1841 |
+ |
|
1842 |
+ |
/** |
1843 |
|
* Tries to locate and execute tasks for a stealer of the given |
1844 |
|
* task, or in turn one of its stealers, Traces currentSteal -> |
1845 |
|
* currentJoin links looking for a thread working on a descendant |
1870 |
|
} |
1871 |
|
if ((ws = workQueues) == null || (m = ws.length - 1) <= 0) |
1872 |
|
break restart; // shutting down |
1873 |
< |
if ((v = ws[h = (j.stealHint | 1) & m]) == null || |
1873 |
> |
if ((v = ws[h = (j.hint | 1) & m]) == null || |
1874 |
|
v.currentSteal != subtask) { |
1875 |
|
for (int origin = h;;) { // find stealer |
1876 |
|
if (((h = (h + 2) & m) & 15) == 1 && |
1878 |
|
continue restart; // occasional staleness check |
1879 |
|
if ((v = ws[h]) != null && |
1880 |
|
v.currentSteal == subtask) { |
1881 |
< |
j.stealHint = h; // save hint |
1881 |
> |
j.hint = h; // save hint |
1882 |
|
break; |
1883 |
|
} |
1884 |
|
if (h == origin) |
1926 |
|
} |
1927 |
|
|
1928 |
|
/** |
1929 |
< |
* If task is at base of some steal queue, steals and executes it. |
1929 |
> |
* Analog of tryHelpStealer for CountedCompleters. Tries to steal |
1930 |
> |
* and run tasks within the target's computation. |
1931 |
|
* |
1932 |
< |
* @param joiner the joining worker |
1933 |
< |
* @param task the task |
1932 |
> |
* @param task the task to join |
1933 |
> |
* @param mode if shared, exit upon completing any task |
1934 |
> |
* if all workers are active |
1935 |
|
*/ |
1936 |
< |
private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) { |
1937 |
< |
WorkQueue[] ws; |
1938 |
< |
if ((ws = workQueues) != null) { |
1939 |
< |
for (int j = 1; j < ws.length && task.status >= 0; j += 2) { |
1940 |
< |
WorkQueue q = ws[j]; |
1941 |
< |
if (q != null && q.pollFor(task)) { |
1942 |
< |
joiner.runSubtask(task); |
1943 |
< |
break; |
1936 |
> |
private int helpComplete(ForkJoinTask<?> task, int mode) { |
1937 |
> |
WorkQueue[] ws; WorkQueue q; int m, n, s, u; |
1938 |
> |
if (task != null && (ws = workQueues) != null && |
1939 |
> |
(m = ws.length - 1) >= 0) { |
1940 |
> |
for (int j = 1, origin = j;;) { |
1941 |
> |
if ((s = task.status) < 0) |
1942 |
> |
return s; |
1943 |
> |
if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) { |
1944 |
> |
origin = j; |
1945 |
> |
if (mode == SHARED_QUEUE && |
1946 |
> |
((u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0)) |
1947 |
> |
break; |
1948 |
|
} |
1949 |
+ |
else if ((j = (j + 2) & m) == origin) |
1950 |
+ |
break; |
1951 |
|
} |
1952 |
|
} |
1953 |
+ |
return 0; |
1954 |
|
} |
1955 |
|
|
1956 |
|
/** |
1957 |
|
* Tries to decrement active count (sometimes implicitly) and |
1958 |
|
* possibly release or create a compensating worker in preparation |
1959 |
|
* for blocking. Fails on contention or termination. Otherwise, |
1960 |
< |
* adds a new thread if no idle workers are available and either |
1961 |
< |
* pool would become completely starved or: (at least half |
1982 |
< |
* starved, and fewer than 50% spares exist, and there is at least |
1983 |
< |
* one task apparently available). Even though the availability |
1984 |
< |
* check requires a full scan, it is worthwhile in reducing false |
1985 |
< |
* alarms. |
1986 |
< |
* |
1987 |
< |
* @param task if non-null, a task being waited for |
1988 |
< |
* @param blocker if non-null, a blocker being waited for |
1989 |
< |
* @return true if the caller can block, else should recheck and retry |
1960 |
> |
* adds a new thread if no idle workers are available and pool |
1961 |
> |
* may become starved. |
1962 |
|
*/ |
1963 |
< |
final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) { |
1964 |
< |
int pc = parallelism, e; |
1965 |
< |
long c = ctl; |
1966 |
< |
WorkQueue[] ws = workQueues; |
1967 |
< |
if ((e = (int)c) >= 0 && ws != null) { |
1968 |
< |
int u, a, ac, hc; |
1969 |
< |
int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc; |
1970 |
< |
boolean replace = false; |
1971 |
< |
if ((a = u >> UAC_SHIFT) <= 0) { |
1972 |
< |
if ((ac = a + pc) <= 1) |
1973 |
< |
replace = true; |
1974 |
< |
else if ((e > 0 || (task != null && |
1975 |
< |
ac <= (hc = pc >>> 1) && tc < pc + hc))) { |
2004 |
< |
WorkQueue w; |
2005 |
< |
for (int j = 0; j < ws.length; ++j) { |
2006 |
< |
if ((w = ws[j]) != null && !w.isEmpty()) { |
2007 |
< |
replace = true; |
2008 |
< |
break; // in compensation range and tasks available |
2009 |
< |
} |
2010 |
< |
} |
1963 |
> |
final boolean tryCompensate() { |
1964 |
> |
int pc = config & SMASK, e, i, tc; long c; |
1965 |
> |
WorkQueue[] ws; WorkQueue w; Thread p; |
1966 |
> |
if ((ws = workQueues) != null && (e = (int)(c = ctl)) >= 0) { |
1967 |
> |
if (e != 0 && (i = e & SMASK) < ws.length && |
1968 |
> |
(w = ws[i]) != null && w.eventCount == (e | INT_SIGN)) { |
1969 |
> |
long nc = ((long)(w.nextWait & E_MASK) | |
1970 |
> |
(c & (AC_MASK|TC_MASK))); |
1971 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1972 |
> |
w.eventCount = (e + E_SEQ) & E_MASK; |
1973 |
> |
if ((p = w.parker) != null) |
1974 |
> |
U.unpark(p); |
1975 |
> |
return true; // replace with idle worker |
1976 |
|
} |
1977 |
|
} |
1978 |
< |
if ((task == null || task.status >= 0) && // recheck need to block |
1979 |
< |
(blocker == null || !blocker.isReleasable()) && ctl == c) { |
1980 |
< |
if (!replace) { // no compensation |
1981 |
< |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1982 |
< |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1983 |
< |
return true; |
1984 |
< |
} |
1985 |
< |
else if (e != 0) { // release an idle worker |
1986 |
< |
WorkQueue w; Thread p; int i; |
1987 |
< |
if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) { |
1988 |
< |
long nc = ((long)(w.nextWait & E_MASK) | |
1989 |
< |
(c & (AC_MASK|TC_MASK))); |
1990 |
< |
if (w.eventCount == (e | INT_SIGN) && |
1991 |
< |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1992 |
< |
w.eventCount = (e + E_SEQ) & E_MASK; |
1993 |
< |
if ((p = w.parker) != null) |
2029 |
< |
U.unpark(p); |
1978 |
> |
else if ((tc = (short)(c >>> TC_SHIFT)) >= 0 && |
1979 |
> |
(int)(c >> AC_SHIFT) + pc > 1) { |
1980 |
> |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1981 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1982 |
> |
return true; // no compensation |
1983 |
> |
} |
1984 |
> |
else if (tc + pc < MAX_CAP) { |
1985 |
> |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1986 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1987 |
> |
ForkJoinWorkerThreadFactory fac; |
1988 |
> |
Throwable ex = null; |
1989 |
> |
ForkJoinWorkerThread wt = null; |
1990 |
> |
try { |
1991 |
> |
if ((fac = factory) != null && |
1992 |
> |
(wt = fac.newThread(this)) != null) { |
1993 |
> |
wt.start(); |
1994 |
|
return true; |
1995 |
|
} |
1996 |
+ |
} catch (Throwable rex) { |
1997 |
+ |
ex = rex; |
1998 |
|
} |
1999 |
< |
} |
2034 |
< |
else if (tc < MAX_CAP) { // create replacement |
2035 |
< |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
2036 |
< |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
2037 |
< |
addWorker(); |
2038 |
< |
return true; |
2039 |
< |
} |
1999 |
> |
deregisterWorker(wt, ex); // clean up and return false |
2000 |
|
} |
2001 |
|
} |
2002 |
|
} |
2011 |
|
* @return task status on exit |
2012 |
|
*/ |
2013 |
|
final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) { |
2014 |
< |
int s; |
2015 |
< |
if ((s = task.status) >= 0) { |
2014 |
> |
int s = 0; |
2015 |
> |
if (joiner != null && task != null && (s = task.status) >= 0) { |
2016 |
|
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
2017 |
|
joiner.currentJoin = task; |
2018 |
< |
long startTime = 0L; |
2019 |
< |
for (int k = 0;;) { |
2020 |
< |
if ((s = (joiner.isEmpty() ? // try to help |
2021 |
< |
tryHelpStealer(joiner, task) : |
2022 |
< |
joiner.tryRemoveAndExec(task))) == 0 && |
2018 |
> |
do {} while ((s = task.status) >= 0 && !joiner.isEmpty() && |
2019 |
> |
joiner.tryRemoveAndExec(task)); // process local tasks |
2020 |
> |
if (s >= 0 && (s = task.status) >= 0) { |
2021 |
> |
helpSignal(task, joiner.poolIndex); |
2022 |
> |
if ((s = task.status) >= 0 && |
2023 |
> |
(task instanceof CountedCompleter)) |
2024 |
> |
s = helpComplete(task, LIFO_QUEUE); |
2025 |
> |
} |
2026 |
> |
while (s >= 0 && (s = task.status) >= 0) { |
2027 |
> |
if ((!joiner.isEmpty() || // try helping |
2028 |
> |
(s = tryHelpStealer(joiner, task)) == 0) && |
2029 |
|
(s = task.status) >= 0) { |
2030 |
< |
if (k == 0) { |
2031 |
< |
startTime = System.nanoTime(); |
2032 |
< |
tryPollForAndExec(joiner, task); // check uncommon case |
2067 |
< |
} |
2068 |
< |
else if ((k & (MAX_HELP - 1)) == 0 && |
2069 |
< |
System.nanoTime() - startTime >= |
2070 |
< |
COMPENSATION_DELAY && |
2071 |
< |
tryCompensate(task, null)) { |
2072 |
< |
if (task.trySetSignal()) { |
2030 |
> |
helpSignal(task, joiner.poolIndex); |
2031 |
> |
if ((s = task.status) >= 0 && tryCompensate()) { |
2032 |
> |
if (task.trySetSignal() && (s = task.status) >= 0) { |
2033 |
|
synchronized (task) { |
2034 |
|
if (task.status >= 0) { |
2035 |
|
try { // see ForkJoinTask |
2046 |
|
(this, CTL, c = ctl, c + AC_UNIT)); |
2047 |
|
} |
2048 |
|
} |
2089 |
– |
if (s < 0 || (s = task.status) < 0) { |
2090 |
– |
joiner.currentJoin = prevJoin; |
2091 |
– |
break; |
2092 |
– |
} |
2093 |
– |
else if ((k++ & (MAX_HELP - 1)) == MAX_HELP >>> 1) |
2094 |
– |
Thread.yield(); // for politeness |
2049 |
|
} |
2050 |
+ |
joiner.currentJoin = prevJoin; |
2051 |
|
} |
2052 |
|
return s; |
2053 |
|
} |
2059 |
|
* |
2060 |
|
* @param joiner the joining worker |
2061 |
|
* @param task the task |
2107 |
– |
* @return task status on exit |
2062 |
|
*/ |
2063 |
< |
final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) { |
2063 |
> |
final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) { |
2064 |
|
int s; |
2065 |
< |
while ((s = task.status) >= 0 && |
2066 |
< |
(joiner.isEmpty() ? |
2067 |
< |
tryHelpStealer(joiner, task) : |
2068 |
< |
joiner.tryRemoveAndExec(task)) != 0) |
2069 |
< |
; |
2070 |
< |
return s; |
2065 |
> |
if (joiner != null && task != null && (s = task.status) >= 0) { |
2066 |
> |
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
2067 |
> |
joiner.currentJoin = task; |
2068 |
> |
do {} while ((s = task.status) >= 0 && !joiner.isEmpty() && |
2069 |
> |
joiner.tryRemoveAndExec(task)); |
2070 |
> |
if (s >= 0 && (s = task.status) >= 0) { |
2071 |
> |
helpSignal(task, joiner.poolIndex); |
2072 |
> |
if ((s = task.status) >= 0 && |
2073 |
> |
(task instanceof CountedCompleter)) |
2074 |
> |
s = helpComplete(task, LIFO_QUEUE); |
2075 |
> |
} |
2076 |
> |
if (s >= 0 && joiner.isEmpty()) { |
2077 |
> |
do {} while (task.status >= 0 && |
2078 |
> |
tryHelpStealer(joiner, task) > 0); |
2079 |
> |
} |
2080 |
> |
joiner.currentJoin = prevJoin; |
2081 |
> |
} |
2082 |
|
} |
2083 |
|
|
2084 |
|
/** |
2085 |
|
* Returns a (probably) non-empty steal queue, if one is found |
2086 |
< |
* during a random, then cyclic scan, else null. This method must |
2087 |
< |
* be retried by caller if, by the time it tries to use the queue, |
2088 |
< |
* it is empty. |
2089 |
< |
*/ |
2090 |
< |
private WorkQueue findNonEmptyStealQueue(WorkQueue w) { |
2091 |
< |
// Similar to loop in scan(), but ignoring submissions |
2092 |
< |
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
2093 |
< |
int step = (r >>> 16) | 1; |
2094 |
< |
for (WorkQueue[] ws;;) { |
2095 |
< |
int rs = runState, m; |
2096 |
< |
if ((ws = workQueues) == null || (m = ws.length - 1) < 1) |
2097 |
< |
return null; |
2133 |
< |
for (int j = (m + 1) << 2; ; r += step) { |
2134 |
< |
WorkQueue q = ws[((r << 1) | 1) & m]; |
2135 |
< |
if (q != null && !q.isEmpty()) |
2136 |
< |
return q; |
2137 |
< |
else if (--j < 0) { |
2138 |
< |
if (runState == rs) |
2139 |
< |
return null; |
2140 |
< |
break; |
2086 |
> |
* during a scan, else null. This method must be retried by |
2087 |
> |
* caller if, by the time it tries to use the queue, it is empty. |
2088 |
> |
* @param r a (random) seed for scanning |
2089 |
> |
*/ |
2090 |
> |
private WorkQueue findNonEmptyStealQueue(int r) { |
2091 |
> |
for (;;) { |
2092 |
> |
int ps = plock, m; WorkQueue[] ws; WorkQueue q; |
2093 |
> |
if ((ws = workQueues) != null && (m = ws.length - 1) >= 0) { |
2094 |
> |
for (int j = (m + 1) << 2; j >= 0; --j) { |
2095 |
> |
if ((q = ws[(((r + j) << 1) | 1) & m]) != null && |
2096 |
> |
q.base - q.top < 0) |
2097 |
> |
return q; |
2098 |
|
} |
2099 |
|
} |
2100 |
+ |
if (plock == ps) |
2101 |
+ |
return null; |
2102 |
|
} |
2103 |
|
} |
2104 |
|
|
2110 |
|
*/ |
2111 |
|
final void helpQuiescePool(WorkQueue w) { |
2112 |
|
for (boolean active = true;;) { |
2113 |
< |
ForkJoinTask<?> localTask; // exhaust local queue |
2114 |
< |
while ((localTask = w.nextLocalTask()) != null) |
2115 |
< |
localTask.doExec(); |
2116 |
< |
WorkQueue q = findNonEmptyStealQueue(w); |
2117 |
< |
if (q != null) { |
2118 |
< |
ForkJoinTask<?> t; int b; |
2113 |
> |
long c; WorkQueue q; ForkJoinTask<?> t; int b; |
2114 |
> |
while ((t = w.nextLocalTask()) != null) { |
2115 |
> |
if (w.base - w.top < 0) |
2116 |
> |
signalWork(w); |
2117 |
> |
t.doExec(); |
2118 |
> |
} |
2119 |
> |
if ((q = findNonEmptyStealQueue(w.nextSeed())) != null) { |
2120 |
|
if (!active) { // re-establish active count |
2161 |
– |
long c; |
2121 |
|
active = true; |
2122 |
|
do {} while (!U.compareAndSwapLong |
2123 |
|
(this, CTL, c = ctl, c + AC_UNIT)); |
2124 |
|
} |
2125 |
< |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
2125 |
> |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) { |
2126 |
> |
if (q.base - q.top < 0) |
2127 |
> |
signalWork(q); |
2128 |
|
w.runSubtask(t); |
2168 |
– |
} |
2169 |
– |
else { |
2170 |
– |
long c; |
2171 |
– |
if (active) { // decrement active count without queuing |
2172 |
– |
active = false; |
2173 |
– |
do {} while (!U.compareAndSwapLong |
2174 |
– |
(this, CTL, c = ctl, c -= AC_UNIT)); |
2175 |
– |
} |
2176 |
– |
else |
2177 |
– |
c = ctl; // re-increment on exit |
2178 |
– |
if ((int)(c >> AC_SHIFT) + parallelism == 0) { |
2179 |
– |
do {} while (!U.compareAndSwapLong |
2180 |
– |
(this, CTL, c = ctl, c + AC_UNIT)); |
2181 |
– |
break; |
2129 |
|
} |
2130 |
|
} |
2131 |
< |
} |
2132 |
< |
} |
2133 |
< |
|
2134 |
< |
/** |
2135 |
< |
* Restricted version of helpQuiescePool for non-FJ callers |
2136 |
< |
*/ |
2190 |
< |
static void externalHelpQuiescePool() { |
2191 |
< |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q, sq; |
2192 |
< |
ForkJoinTask<?>[] a; int b; |
2193 |
< |
ForkJoinTask<?> t = null; |
2194 |
< |
int k = submitters.get().seed & SQMASK; |
2195 |
< |
if ((p = commonPool) != null && |
2196 |
< |
(int)(p.ctl >> AC_SHIFT) < 0 && |
2197 |
< |
(ws = p.workQueues) != null && |
2198 |
< |
ws.length > (k &= p.submitMask) && |
2199 |
< |
(q = ws[k]) != null) { |
2200 |
< |
while (q.top - q.base > 0) { |
2201 |
< |
if ((t = q.sharedPop()) != null) |
2202 |
< |
break; |
2131 |
> |
else if (active) { // decrement active count without queuing |
2132 |
> |
long nc = (c = ctl) - AC_UNIT; |
2133 |
> |
if ((int)(nc >> AC_SHIFT) + (config & SMASK) == 0) |
2134 |
> |
return; // bypass decrement-then-increment |
2135 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
2136 |
> |
active = false; |
2137 |
|
} |
2138 |
< |
if (t == null && (sq = p.findNonEmptyStealQueue(q)) != null && |
2139 |
< |
(b = sq.base) - sq.top < 0) |
2140 |
< |
t = sq.pollAt(b); |
2207 |
< |
if (t != null) |
2208 |
< |
t.doExec(); |
2138 |
> |
else if ((int)((c = ctl) >> AC_SHIFT) + (config & SMASK) == 0 && |
2139 |
> |
U.compareAndSwapLong(this, CTL, c, c + AC_UNIT)) |
2140 |
> |
return; |
2141 |
|
} |
2142 |
|
} |
2143 |
|
|
2151 |
|
WorkQueue q; int b; |
2152 |
|
if ((t = w.nextLocalTask()) != null) |
2153 |
|
return t; |
2154 |
< |
if ((q = findNonEmptyStealQueue(w)) == null) |
2154 |
> |
if ((q = findNonEmptyStealQueue(w.nextSeed())) == null) |
2155 |
|
return null; |
2156 |
< |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
2156 |
> |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) { |
2157 |
> |
if (q.base - q.top < 0) |
2158 |
> |
signalWork(q); |
2159 |
|
return t; |
2160 |
+ |
} |
2161 |
|
} |
2162 |
|
} |
2163 |
|
|
2164 |
|
/** |
2165 |
< |
* Returns the approximate (non-atomic) number of idle threads per |
2166 |
< |
* active thread to offset steal queue size for method |
2167 |
< |
* ForkJoinTask.getSurplusQueuedTaskCount(). |
2168 |
< |
*/ |
2169 |
< |
final int idlePerActive() { |
2170 |
< |
// Approximate at powers of two for small values, saturate past 4 |
2171 |
< |
int p = parallelism; |
2172 |
< |
int a = p + (int)(ctl >> AC_SHIFT); |
2173 |
< |
return (a > (p >>>= 1) ? 0 : |
2174 |
< |
a > (p >>>= 1) ? 1 : |
2175 |
< |
a > (p >>>= 1) ? 2 : |
2176 |
< |
a > (p >>>= 1) ? 4 : |
2177 |
< |
8); |
2178 |
< |
} |
2179 |
< |
|
2180 |
< |
/** |
2181 |
< |
* Returns approximate submission queue length for the given caller |
2182 |
< |
*/ |
2183 |
< |
static int getEstimatedSubmitterQueueLength() { |
2184 |
< |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; |
2185 |
< |
int k = submitters.get().seed & SQMASK; |
2186 |
< |
return ((p = commonPool) != null && |
2187 |
< |
p.runState >= 0 && |
2188 |
< |
(ws = p.workQueues) != null && |
2189 |
< |
ws.length > (k &= p.submitMask) && |
2190 |
< |
(q = ws[k]) != null) ? |
2191 |
< |
q.queueSize() : 0; |
2165 |
> |
* Returns a cheap heuristic guide for task partitioning when |
2166 |
> |
* programmers, frameworks, tools, or languages have little or no |
2167 |
> |
* idea about task granularity. In essence by offering this |
2168 |
> |
* method, we ask users only about tradeoffs in overhead vs |
2169 |
> |
* expected throughput and its variance, rather than how finely to |
2170 |
> |
* partition tasks. |
2171 |
> |
* |
2172 |
> |
* In a steady state strict (tree-structured) computation, each |
2173 |
> |
* thread makes available for stealing enough tasks for other |
2174 |
> |
* threads to remain active. Inductively, if all threads play by |
2175 |
> |
* the same rules, each thread should make available only a |
2176 |
> |
* constant number of tasks. |
2177 |
> |
* |
2178 |
> |
* The minimum useful constant is just 1. But using a value of 1 |
2179 |
> |
* would require immediate replenishment upon each steal to |
2180 |
> |
* maintain enough tasks, which is infeasible. Further, |
2181 |
> |
* partitionings/granularities of offered tasks should minimize |
2182 |
> |
* steal rates, which in general means that threads nearer the top |
2183 |
> |
* of computation tree should generate more than those nearer the |
2184 |
> |
* bottom. In perfect steady state, each thread is at |
2185 |
> |
* approximately the same level of computation tree. However, |
2186 |
> |
* producing extra tasks amortizes the uncertainty of progress and |
2187 |
> |
* diffusion assumptions. |
2188 |
> |
* |
2189 |
> |
* So, users will want to use values larger, but not much larger |
2190 |
> |
* than 1 to both smooth over transient shortages and hedge |
2191 |
> |
* against uneven progress; as traded off against the cost of |
2192 |
> |
* extra task overhead. We leave the user to pick a threshold |
2193 |
> |
* value to compare with the results of this call to guide |
2194 |
> |
* decisions, but recommend values such as 3. |
2195 |
> |
* |
2196 |
> |
* When all threads are active, it is on average OK to estimate |
2197 |
> |
* surplus strictly locally. In steady-state, if one thread is |
2198 |
> |
* maintaining say 2 surplus tasks, then so are others. So we can |
2199 |
> |
* just use estimated queue length. However, this strategy alone |
2200 |
> |
* leads to serious mis-estimates in some non-steady-state |
2201 |
> |
* conditions (ramp-up, ramp-down, other stalls). We can detect |
2202 |
> |
* many of these by further considering the number of "idle" |
2203 |
> |
* threads, that are known to have zero queued tasks, so |
2204 |
> |
* compensate by a factor of (#idle/#active) threads. |
2205 |
> |
* |
2206 |
> |
* Note: The approximation of #busy workers as #active workers is |
2207 |
> |
* not very good under current signalling scheme, and should be |
2208 |
> |
* improved. |
2209 |
> |
*/ |
2210 |
> |
static int getSurplusQueuedTaskCount() { |
2211 |
> |
Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q; |
2212 |
> |
if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) { |
2213 |
> |
int p = (pool = (wt = (ForkJoinWorkerThread)t).pool).config & SMASK; |
2214 |
> |
int n = (q = wt.workQueue).top - q.base; |
2215 |
> |
int a = (int)(pool.ctl >> AC_SHIFT) + p; |
2216 |
> |
return n - (a > (p >>>= 1) ? 0 : |
2217 |
> |
a > (p >>>= 1) ? 1 : |
2218 |
> |
a > (p >>>= 1) ? 2 : |
2219 |
> |
a > (p >>>= 1) ? 4 : |
2220 |
> |
8); |
2221 |
> |
} |
2222 |
> |
return 0; |
2223 |
|
} |
2224 |
|
|
2225 |
|
// Termination |
2239 |
|
* @return true if now terminating or terminated |
2240 |
|
*/ |
2241 |
|
private boolean tryTerminate(boolean now, boolean enable) { |
2242 |
+ |
int ps; |
2243 |
+ |
if (this == common) // cannot shut down |
2244 |
+ |
return false; |
2245 |
+ |
if ((ps = plock) >= 0) { // enable by setting plock |
2246 |
+ |
if (!enable) |
2247 |
+ |
return false; |
2248 |
+ |
if ((ps & PL_LOCK) != 0 || |
2249 |
+ |
!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK)) |
2250 |
+ |
ps = acquirePlock(); |
2251 |
+ |
int nps = ((ps + PL_LOCK) & ~SHUTDOWN) | SHUTDOWN; |
2252 |
+ |
if (!U.compareAndSwapInt(this, PLOCK, ps, nps)) |
2253 |
+ |
releasePlock(nps); |
2254 |
+ |
} |
2255 |
|
for (long c;;) { |
2256 |
< |
if (((c = ctl) & STOP_BIT) != 0) { // already terminating |
2257 |
< |
if ((short)(c >>> TC_SHIFT) == -parallelism) { |
2256 |
> |
if (((c = ctl) & STOP_BIT) != 0) { // already terminating |
2257 |
> |
if ((short)(c >>> TC_SHIFT) == -(config & SMASK)) { |
2258 |
|
synchronized (this) { |
2259 |
< |
notifyAll(); // signal when 0 workers |
2259 |
> |
notifyAll(); // signal when 0 workers |
2260 |
|
} |
2261 |
|
} |
2262 |
|
return true; |
2263 |
|
} |
2264 |
< |
if (runState >= 0) { // not yet enabled |
2265 |
< |
if (!enable) |
2266 |
< |
return false; |
2288 |
< |
while (!U.compareAndSwapInt(this, MAINLOCK, 0, 1)) |
2289 |
< |
tryAwaitMainLock(); |
2290 |
< |
try { |
2291 |
< |
runState |= SHUTDOWN; |
2292 |
< |
} finally { |
2293 |
< |
if (!U.compareAndSwapInt(this, MAINLOCK, 1, 0)) { |
2294 |
< |
mainLock = 0; |
2295 |
< |
synchronized (this) { notifyAll(); }; |
2296 |
< |
} |
2297 |
< |
} |
2298 |
< |
} |
2299 |
< |
if (!now) { // check if idle & no tasks |
2300 |
< |
if ((int)(c >> AC_SHIFT) != -parallelism || |
2301 |
< |
hasQueuedSubmissions()) |
2264 |
> |
if (!now) { // check if idle & no tasks |
2265 |
> |
WorkQueue[] ws; WorkQueue w; |
2266 |
> |
if ((int)(c >> AC_SHIFT) != -(config & SMASK)) |
2267 |
|
return false; |
2268 |
< |
// Check for unqueued inactive workers. One pass suffices. |
2269 |
< |
WorkQueue[] ws = workQueues; WorkQueue w; |
2270 |
< |
if (ws != null) { |
2271 |
< |
for (int i = 1; i < ws.length; i += 2) { |
2272 |
< |
if ((w = ws[i]) != null && w.eventCount >= 0) |
2273 |
< |
return false; |
2268 |
> |
if ((ws = workQueues) != null) { |
2269 |
> |
for (int i = 0; i < ws.length; ++i) { |
2270 |
> |
if ((w = ws[i]) != null) { |
2271 |
> |
if (!w.isEmpty()) { // signal unprocessed tasks |
2272 |
> |
signalWork(w); |
2273 |
> |
return false; |
2274 |
> |
} |
2275 |
> |
if ((i & 1) != 0 && w.eventCount >= 0) |
2276 |
> |
return false; // unqueued inactive worker |
2277 |
> |
} |
2278 |
|
} |
2279 |
|
} |
2280 |
|
} |
2281 |
|
if (U.compareAndSwapLong(this, CTL, c, c | STOP_BIT)) { |
2282 |
|
for (int pass = 0; pass < 3; ++pass) { |
2283 |
< |
WorkQueue[] ws = workQueues; |
2284 |
< |
if (ws != null) { |
2316 |
< |
WorkQueue w; |
2283 |
> |
WorkQueue[] ws; WorkQueue w; Thread wt; |
2284 |
> |
if ((ws = workQueues) != null) { |
2285 |
|
int n = ws.length; |
2286 |
|
for (int i = 0; i < n; ++i) { |
2287 |
|
if ((w = ws[i]) != null) { |
2288 |
< |
w.runState = -1; |
2288 |
> |
w.qlock = -1; |
2289 |
|
if (pass > 0) { |
2290 |
|
w.cancelAll(); |
2291 |
< |
if (pass > 1) |
2292 |
< |
w.interruptOwner(); |
2291 |
> |
if (pass > 1 && (wt = w.owner) != null) { |
2292 |
> |
if (!wt.isInterrupted()) { |
2293 |
> |
try { |
2294 |
> |
wt.interrupt(); |
2295 |
> |
} catch (Throwable ignore) { |
2296 |
> |
} |
2297 |
> |
} |
2298 |
> |
U.unpark(wt); |
2299 |
> |
} |
2300 |
|
} |
2301 |
|
} |
2302 |
|
} |
2303 |
|
// Wake up workers parked on event queue |
2304 |
|
int i, e; long cc; Thread p; |
2305 |
|
while ((e = (int)(cc = ctl) & E_MASK) != 0 && |
2306 |
< |
(i = e & SMASK) < n && |
2306 |
> |
(i = e & SMASK) < n && i >= 0 && |
2307 |
|
(w = ws[i]) != null) { |
2308 |
|
long nc = ((long)(w.nextWait & E_MASK) | |
2309 |
|
((cc + AC_UNIT) & AC_MASK) | |
2311 |
|
if (w.eventCount == (e | INT_SIGN) && |
2312 |
|
U.compareAndSwapLong(this, CTL, cc, nc)) { |
2313 |
|
w.eventCount = (e + E_SEQ) & E_MASK; |
2314 |
< |
w.runState = -1; |
2314 |
> |
w.qlock = -1; |
2315 |
|
if ((p = w.parker) != null) |
2316 |
|
U.unpark(p); |
2317 |
|
} |
2322 |
|
} |
2323 |
|
} |
2324 |
|
|
2325 |
+ |
// external operations on common pool |
2326 |
+ |
|
2327 |
+ |
/** |
2328 |
+ |
* Returns common pool queue for a thread that has submitted at |
2329 |
+ |
* least one task. |
2330 |
+ |
*/ |
2331 |
+ |
static WorkQueue commonSubmitterQueue() { |
2332 |
+ |
ForkJoinPool p; WorkQueue[] ws; int m; Submitter z; |
2333 |
+ |
return ((z = submitters.get()) != null && |
2334 |
+ |
(p = common) != null && |
2335 |
+ |
(ws = p.workQueues) != null && |
2336 |
+ |
(m = ws.length - 1) >= 0) ? |
2337 |
+ |
ws[m & z.seed & SQMASK] : null; |
2338 |
+ |
} |
2339 |
+ |
|
2340 |
+ |
/** |
2341 |
+ |
* Tries to pop the given task from submitter's queue in common pool. |
2342 |
+ |
*/ |
2343 |
+ |
static boolean tryExternalUnpush(ForkJoinTask<?> t) { |
2344 |
+ |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z; |
2345 |
+ |
ForkJoinTask<?>[] a; int m, s; |
2346 |
+ |
if (t != null && |
2347 |
+ |
(z = submitters.get()) != null && |
2348 |
+ |
(p = common) != null && |
2349 |
+ |
(ws = p.workQueues) != null && |
2350 |
+ |
(m = ws.length - 1) >= 0 && |
2351 |
+ |
(q = ws[m & z.seed & SQMASK]) != null && |
2352 |
+ |
(s = q.top) != q.base && |
2353 |
+ |
(a = q.array) != null) { |
2354 |
+ |
long j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE; |
2355 |
+ |
if (U.getObject(a, j) == t && |
2356 |
+ |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2357 |
+ |
if (q.array == a && q.top == s && // recheck |
2358 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
2359 |
+ |
q.top = s - 1; |
2360 |
+ |
q.qlock = 0; |
2361 |
+ |
return true; |
2362 |
+ |
} |
2363 |
+ |
q.qlock = 0; |
2364 |
+ |
} |
2365 |
+ |
} |
2366 |
+ |
return false; |
2367 |
+ |
} |
2368 |
+ |
|
2369 |
+ |
/** |
2370 |
+ |
* Tries to pop and run local tasks within the same computation |
2371 |
+ |
* as the given root. On failure, tries to help complete from |
2372 |
+ |
* other queues via helpComplete. |
2373 |
+ |
*/ |
2374 |
+ |
private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) { |
2375 |
+ |
ForkJoinTask<?>[] a; int m; |
2376 |
+ |
if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 && |
2377 |
+ |
root != null && root.status >= 0) { |
2378 |
+ |
for (;;) { |
2379 |
+ |
int s, u; Object o; CountedCompleter<?> task = null; |
2380 |
+ |
if ((s = q.top) - q.base > 0) { |
2381 |
+ |
long j = ((m & (s - 1)) << ASHIFT) + ABASE; |
2382 |
+ |
if ((o = U.getObject(a, j)) != null && |
2383 |
+ |
(o instanceof CountedCompleter)) { |
2384 |
+ |
CountedCompleter<?> t = (CountedCompleter<?>)o, r = t; |
2385 |
+ |
do { |
2386 |
+ |
if (r == root) { |
2387 |
+ |
if (U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2388 |
+ |
if (q.array == a && q.top == s && |
2389 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
2390 |
+ |
q.top = s - 1; |
2391 |
+ |
task = t; |
2392 |
+ |
} |
2393 |
+ |
q.qlock = 0; |
2394 |
+ |
} |
2395 |
+ |
break; |
2396 |
+ |
} |
2397 |
+ |
} while ((r = r.completer) != null); |
2398 |
+ |
} |
2399 |
+ |
} |
2400 |
+ |
if (task != null) |
2401 |
+ |
task.doExec(); |
2402 |
+ |
if (root.status < 0 || |
2403 |
+ |
(u = (int)(ctl >>> 32)) >= 0 || (u >> UAC_SHIFT) >= 0) |
2404 |
+ |
break; |
2405 |
+ |
if (task == null) { |
2406 |
+ |
helpSignal(root, q.poolIndex); |
2407 |
+ |
if (root.status >= 0) |
2408 |
+ |
helpComplete(root, SHARED_QUEUE); |
2409 |
+ |
break; |
2410 |
+ |
} |
2411 |
+ |
} |
2412 |
+ |
} |
2413 |
+ |
} |
2414 |
+ |
|
2415 |
+ |
/** |
2416 |
+ |
* Tries to help execute or signal availability of the given task |
2417 |
+ |
* from submitter's queue in common pool. |
2418 |
+ |
*/ |
2419 |
+ |
static void externalHelpJoin(ForkJoinTask<?> t) { |
2420 |
+ |
// Some hard-to-avoid overlap with tryExternalUnpush |
2421 |
+ |
ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z; |
2422 |
+ |
ForkJoinTask<?>[] a; int m, s, n; |
2423 |
+ |
if (t != null && |
2424 |
+ |
(z = submitters.get()) != null && |
2425 |
+ |
(p = common) != null && |
2426 |
+ |
(ws = p.workQueues) != null && |
2427 |
+ |
(m = ws.length - 1) >= 0 && |
2428 |
+ |
(q = ws[m & z.seed & SQMASK]) != null && |
2429 |
+ |
(a = q.array) != null) { |
2430 |
+ |
int am = a.length - 1; |
2431 |
+ |
if ((s = q.top) != q.base) { |
2432 |
+ |
long j = ((am & (s - 1)) << ASHIFT) + ABASE; |
2433 |
+ |
if (U.getObject(a, j) == t && |
2434 |
+ |
U.compareAndSwapInt(q, QLOCK, 0, 1)) { |
2435 |
+ |
if (q.array == a && q.top == s && |
2436 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
2437 |
+ |
q.top = s - 1; |
2438 |
+ |
q.qlock = 0; |
2439 |
+ |
t.doExec(); |
2440 |
+ |
} |
2441 |
+ |
else |
2442 |
+ |
q.qlock = 0; |
2443 |
+ |
} |
2444 |
+ |
} |
2445 |
+ |
if (t.status >= 0) { |
2446 |
+ |
if (t instanceof CountedCompleter) |
2447 |
+ |
p.externalHelpComplete(q, t); |
2448 |
+ |
else |
2449 |
+ |
p.helpSignal(t, q.poolIndex); |
2450 |
+ |
} |
2451 |
+ |
} |
2452 |
+ |
} |
2453 |
+ |
|
2454 |
|
// Exported methods |
2455 |
|
|
2456 |
|
// Constructors |
2522 |
|
throw new NullPointerException(); |
2523 |
|
if (parallelism <= 0 || parallelism > MAX_CAP) |
2524 |
|
throw new IllegalArgumentException(); |
2421 |
– |
this.parallelism = parallelism; |
2525 |
|
this.factory = factory; |
2526 |
|
this.ueh = handler; |
2527 |
< |
this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE; |
2527 |
> |
this.config = parallelism | (asyncMode ? (FIFO_QUEUE << 16) : 0); |
2528 |
|
long np = (long)(-parallelism); // offset ctl counts |
2529 |
|
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2530 |
< |
// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2. |
2428 |
< |
int n = parallelism - 1; |
2429 |
< |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16; |
2430 |
< |
this.submitMask = ((n + 1) << 1) - 1; |
2431 |
< |
int pn = poolNumberGenerator.incrementAndGet(); |
2530 |
> |
int pn = nextPoolId(); |
2531 |
|
StringBuilder sb = new StringBuilder("ForkJoinPool-"); |
2532 |
|
sb.append(Integer.toString(pn)); |
2533 |
|
sb.append("-worker-"); |
2534 |
|
this.workerNamePrefix = sb.toString(); |
2436 |
– |
this.runState = 1; // set init flag |
2535 |
|
} |
2536 |
|
|
2537 |
|
/** |
2538 |
|
* Constructor for common pool, suitable only for static initialization. |
2539 |
|
* Basically the same as above, but uses smallest possible initial footprint. |
2540 |
|
*/ |
2541 |
< |
ForkJoinPool(int parallelism, int submitMask, |
2541 |
> |
ForkJoinPool(int parallelism, long ctl, |
2542 |
|
ForkJoinWorkerThreadFactory factory, |
2543 |
|
Thread.UncaughtExceptionHandler handler) { |
2544 |
+ |
this.config = parallelism; |
2545 |
+ |
this.ctl = ctl; |
2546 |
|
this.factory = factory; |
2547 |
|
this.ueh = handler; |
2448 |
– |
this.submitMask = submitMask; |
2449 |
– |
this.parallelism = parallelism; |
2450 |
– |
long np = (long)(-parallelism); |
2451 |
– |
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2452 |
– |
this.localMode = LIFO_QUEUE; |
2548 |
|
this.workerNamePrefix = "ForkJoinPool.commonPool-worker-"; |
2454 |
– |
this.runState = 1; |
2549 |
|
} |
2550 |
|
|
2551 |
|
/** |
2552 |
< |
* Returns the common pool instance. |
2552 |
> |
* Returns the common pool instance. This pool is statically |
2553 |
> |
* constructed; its run state is unaffected by attempts to {@link |
2554 |
> |
* #shutdown} or {@link #shutdownNow}. However this pool and any |
2555 |
> |
* ongoing processing are automatically terminated upon program |
2556 |
> |
* {@link System#exit}. Any program that relies on asynchronous |
2557 |
> |
* task processing to complete before program termination should |
2558 |
> |
* invoke {@code commonPool().}{@link #awaitQuiescence}, before |
2559 |
> |
* exit. |
2560 |
|
* |
2561 |
|
* @return the common pool instance |
2562 |
+ |
* @since 1.8 |
2563 |
|
*/ |
2564 |
|
public static ForkJoinPool commonPool() { |
2565 |
< |
ForkJoinPool p; |
2566 |
< |
if ((p = commonPool) == null) |
2465 |
< |
throw new Error("Common Pool Unavailable"); |
2466 |
< |
return p; |
2565 |
> |
// assert common != null : "static init error"; |
2566 |
> |
return common; |
2567 |
|
} |
2568 |
|
|
2569 |
|
// Execution methods |
2587 |
|
public <T> T invoke(ForkJoinTask<T> task) { |
2588 |
|
if (task == null) |
2589 |
|
throw new NullPointerException(); |
2590 |
< |
doSubmit(task); |
2590 |
> |
externalPush(task); |
2591 |
|
return task.join(); |
2592 |
|
} |
2593 |
|
|
2602 |
|
public void execute(ForkJoinTask<?> task) { |
2603 |
|
if (task == null) |
2604 |
|
throw new NullPointerException(); |
2605 |
< |
doSubmit(task); |
2605 |
> |
externalPush(task); |
2606 |
|
} |
2607 |
|
|
2608 |
|
// AbstractExecutorService methods |
2620 |
|
job = (ForkJoinTask<?>) task; |
2621 |
|
else |
2622 |
|
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2623 |
< |
doSubmit(job); |
2623 |
> |
externalPush(job); |
2624 |
|
} |
2625 |
|
|
2626 |
|
/** |
2635 |
|
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2636 |
|
if (task == null) |
2637 |
|
throw new NullPointerException(); |
2638 |
< |
doSubmit(task); |
2638 |
> |
externalPush(task); |
2639 |
|
return task; |
2640 |
|
} |
2641 |
|
|
2646 |
|
*/ |
2647 |
|
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2648 |
|
ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task); |
2649 |
< |
doSubmit(job); |
2649 |
> |
externalPush(job); |
2650 |
|
return job; |
2651 |
|
} |
2652 |
|
|
2657 |
|
*/ |
2658 |
|
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2659 |
|
ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result); |
2660 |
< |
doSubmit(job); |
2660 |
> |
externalPush(job); |
2661 |
|
return job; |
2662 |
|
} |
2663 |
|
|
2674 |
|
job = (ForkJoinTask<?>) task; |
2675 |
|
else |
2676 |
|
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2677 |
< |
doSubmit(job); |
2677 |
> |
externalPush(job); |
2678 |
|
return job; |
2679 |
|
} |
2680 |
|
|
2696 |
|
try { |
2697 |
|
for (Callable<T> t : tasks) { |
2698 |
|
ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t); |
2699 |
< |
doSubmit(f); |
2699 |
> |
externalPush(f); |
2700 |
|
fs.add(f); |
2701 |
|
} |
2702 |
|
for (ForkJoinTask<T> f : fs) |
2735 |
|
* @return the targeted parallelism level of this pool |
2736 |
|
*/ |
2737 |
|
public int getParallelism() { |
2738 |
< |
return parallelism; |
2738 |
> |
return config & SMASK; |
2739 |
|
} |
2740 |
|
|
2741 |
|
/** |
2742 |
|
* Returns the targeted parallelism level of the common pool. |
2743 |
|
* |
2744 |
|
* @return the targeted parallelism level of the common pool |
2745 |
+ |
* @since 1.8 |
2746 |
|
*/ |
2747 |
|
public static int getCommonPoolParallelism() { |
2748 |
< |
return commonPoolParallelism; |
2748 |
> |
return commonParallelism; |
2749 |
|
} |
2750 |
|
|
2751 |
|
/** |
2757 |
|
* @return the number of worker threads |
2758 |
|
*/ |
2759 |
|
public int getPoolSize() { |
2760 |
< |
return parallelism + (short)(ctl >>> TC_SHIFT); |
2760 |
> |
return (config & SMASK) + (short)(ctl >>> TC_SHIFT); |
2761 |
|
} |
2762 |
|
|
2763 |
|
/** |
2767 |
|
* @return {@code true} if this pool uses async mode |
2768 |
|
*/ |
2769 |
|
public boolean getAsyncMode() { |
2770 |
< |
return localMode != 0; |
2770 |
> |
return (config >>> 16) == FIFO_QUEUE; |
2771 |
|
} |
2772 |
|
|
2773 |
|
/** |
2798 |
|
* @return the number of active threads |
2799 |
|
*/ |
2800 |
|
public int getActiveThreadCount() { |
2801 |
< |
int r = parallelism + (int)(ctl >> AC_SHIFT); |
2801 |
> |
int r = (config & SMASK) + (int)(ctl >> AC_SHIFT); |
2802 |
|
return (r <= 0) ? 0 : r; // suppress momentarily negative values |
2803 |
|
} |
2804 |
|
|
2814 |
|
* @return {@code true} if all threads are currently idle |
2815 |
|
*/ |
2816 |
|
public boolean isQuiescent() { |
2817 |
< |
return (int)(ctl >> AC_SHIFT) + parallelism == 0; |
2817 |
> |
return (int)(ctl >> AC_SHIFT) + (config & SMASK) == 0; |
2818 |
|
} |
2819 |
|
|
2820 |
|
/** |
2834 |
|
if ((ws = workQueues) != null) { |
2835 |
|
for (int i = 1; i < ws.length; i += 2) { |
2836 |
|
if ((w = ws[i]) != null) |
2837 |
< |
count += w.totalSteals; |
2837 |
> |
count += w.nsteals; |
2838 |
|
} |
2839 |
|
} |
2840 |
|
return count; |
2970 |
|
qs += size; |
2971 |
|
else { |
2972 |
|
qt += size; |
2973 |
< |
st += w.totalSteals; |
2973 |
> |
st += w.nsteals; |
2974 |
|
if (w.isApparentlyUnblocked()) |
2975 |
|
++rc; |
2976 |
|
} |
2977 |
|
} |
2978 |
|
} |
2979 |
|
} |
2980 |
< |
int pc = parallelism; |
2980 |
> |
int pc = (config & SMASK); |
2981 |
|
int tc = pc + (short)(c >>> TC_SHIFT); |
2982 |
|
int ac = pc + (int)(c >> AC_SHIFT); |
2983 |
|
if (ac < 0) // ignore transient negative |
2986 |
|
if ((c & STOP_BIT) != 0) |
2987 |
|
level = (tc == 0) ? "Terminated" : "Terminating"; |
2988 |
|
else |
2989 |
< |
level = runState < 0 ? "Shutting down" : "Running"; |
2989 |
> |
level = plock < 0 ? "Shutting down" : "Running"; |
2990 |
|
return super.toString() + |
2991 |
|
"[" + level + |
2992 |
|
", parallelism = " + pc + |
3003 |
|
* Possibly initiates an orderly shutdown in which previously |
3004 |
|
* submitted tasks are executed, but no new tasks will be |
3005 |
|
* accepted. Invocation has no effect on execution state if this |
3006 |
< |
* is the {@link #commonPool}, and no additional effect if |
3006 |
> |
* is the {@link #commonPool()}, and no additional effect if |
3007 |
|
* already shut down. Tasks that are in the process of being |
3008 |
|
* submitted concurrently during the course of this method may or |
3009 |
|
* may not be rejected. |
3015 |
|
*/ |
3016 |
|
public void shutdown() { |
3017 |
|
checkPermission(); |
3018 |
< |
if (this != commonPool) |
2918 |
< |
tryTerminate(false, true); |
3018 |
> |
tryTerminate(false, true); |
3019 |
|
} |
3020 |
|
|
3021 |
|
/** |
3022 |
|
* Possibly attempts to cancel and/or stop all tasks, and reject |
3023 |
|
* all subsequently submitted tasks. Invocation has no effect on |
3024 |
< |
* execution state if this is the {@link #commonPool}, and no |
3024 |
> |
* execution state if this is the {@link #commonPool()}, and no |
3025 |
|
* additional effect if already shut down. Otherwise, tasks that |
3026 |
|
* are in the process of being submitted or executed concurrently |
3027 |
|
* during the course of this method may or may not be |
3038 |
|
*/ |
3039 |
|
public List<Runnable> shutdownNow() { |
3040 |
|
checkPermission(); |
3041 |
< |
if (this != commonPool) |
2942 |
< |
tryTerminate(true, true); |
3041 |
> |
tryTerminate(true, true); |
3042 |
|
return Collections.emptyList(); |
3043 |
|
} |
3044 |
|
|
3050 |
|
public boolean isTerminated() { |
3051 |
|
long c = ctl; |
3052 |
|
return ((c & STOP_BIT) != 0L && |
3053 |
< |
(short)(c >>> TC_SHIFT) == -parallelism); |
3053 |
> |
(short)(c >>> TC_SHIFT) == -(config & SMASK)); |
3054 |
|
} |
3055 |
|
|
3056 |
|
/** |
3058 |
|
* commenced but not yet completed. This method may be useful for |
3059 |
|
* debugging. A return of {@code true} reported a sufficient |
3060 |
|
* period after shutdown may indicate that submitted tasks have |
3061 |
< |
* ignored or suppressed interruption, or are waiting for IO, |
3061 |
> |
* ignored or suppressed interruption, or are waiting for I/O, |
3062 |
|
* causing this executor not to properly terminate. (See the |
3063 |
|
* advisory notes for class {@link ForkJoinTask} stating that |
3064 |
|
* tasks should not normally entail blocking operations. But if |
3069 |
|
public boolean isTerminating() { |
3070 |
|
long c = ctl; |
3071 |
|
return ((c & STOP_BIT) != 0L && |
3072 |
< |
(short)(c >>> TC_SHIFT) != -parallelism); |
3072 |
> |
(short)(c >>> TC_SHIFT) != -(config & SMASK)); |
3073 |
|
} |
3074 |
|
|
3075 |
|
/** |
3078 |
|
* @return {@code true} if this pool has been shut down |
3079 |
|
*/ |
3080 |
|
public boolean isShutdown() { |
3081 |
< |
return runState < 0; |
3081 |
> |
return plock < 0; |
3082 |
|
} |
3083 |
|
|
3084 |
|
/** |
3085 |
< |
* Blocks until all tasks have completed execution after a shutdown |
3086 |
< |
* request, or the timeout occurs, or the current thread is |
3087 |
< |
* interrupted, whichever happens first. |
3085 |
> |
* Blocks until all tasks have completed execution after a |
3086 |
> |
* shutdown request, or the timeout occurs, or the current thread |
3087 |
> |
* is interrupted, whichever happens first. Because the {@link |
3088 |
> |
* #commonPool()} never terminates until program shutdown, when |
3089 |
> |
* applied to the common pool, this method is equivalent to {@link |
3090 |
> |
* #awaitQuiescence} but always returns {@code false}. |
3091 |
|
* |
3092 |
|
* @param timeout the maximum time to wait |
3093 |
|
* @param unit the time unit of the timeout argument |
3097 |
|
*/ |
3098 |
|
public boolean awaitTermination(long timeout, TimeUnit unit) |
3099 |
|
throws InterruptedException { |
3100 |
+ |
if (Thread.interrupted()) |
3101 |
+ |
throw new InterruptedException(); |
3102 |
+ |
if (this == common) { |
3103 |
+ |
awaitQuiescence(timeout, unit); |
3104 |
+ |
return false; |
3105 |
+ |
} |
3106 |
|
long nanos = unit.toNanos(timeout); |
3107 |
|
if (isTerminated()) |
3108 |
|
return true; |
3122 |
|
} |
3123 |
|
|
3124 |
|
/** |
3125 |
+ |
* If called by a ForkJoinTask operating in this pool, equivalent |
3126 |
+ |
* in effect to {@link ForkJoinTask#helpQuiesce}. Otherwise, |
3127 |
+ |
* waits and/or attempts to assist performing tasks until this |
3128 |
+ |
* pool {@link #isQuiescent} or the indicated timeout elapses. |
3129 |
+ |
* |
3130 |
+ |
* @param timeout the maximum time to wait |
3131 |
+ |
* @param unit the time unit of the timeout argument |
3132 |
+ |
* @return {@code true} if quiescent; {@code false} if the |
3133 |
+ |
* timeout elapsed. |
3134 |
+ |
*/ |
3135 |
+ |
public boolean awaitQuiescence(long timeout, TimeUnit unit) { |
3136 |
+ |
long nanos = unit.toNanos(timeout); |
3137 |
+ |
ForkJoinWorkerThread wt; |
3138 |
+ |
Thread thread = Thread.currentThread(); |
3139 |
+ |
if ((thread instanceof ForkJoinWorkerThread) && |
3140 |
+ |
(wt = (ForkJoinWorkerThread)thread).pool == this) { |
3141 |
+ |
helpQuiescePool(wt.workQueue); |
3142 |
+ |
return true; |
3143 |
+ |
} |
3144 |
+ |
long startTime = System.nanoTime(); |
3145 |
+ |
WorkQueue[] ws; |
3146 |
+ |
int r = 0, m; |
3147 |
+ |
boolean found = true; |
3148 |
+ |
while (!isQuiescent() && (ws = workQueues) != null && |
3149 |
+ |
(m = ws.length - 1) >= 0) { |
3150 |
+ |
if (!found) { |
3151 |
+ |
if ((System.nanoTime() - startTime) > nanos) |
3152 |
+ |
return false; |
3153 |
+ |
Thread.yield(); // cannot block |
3154 |
+ |
} |
3155 |
+ |
found = false; |
3156 |
+ |
for (int j = (m + 1) << 2; j >= 0; --j) { |
3157 |
+ |
ForkJoinTask<?> t; WorkQueue q; int b; |
3158 |
+ |
if ((q = ws[r++ & m]) != null && (b = q.base) - q.top < 0) { |
3159 |
+ |
found = true; |
3160 |
+ |
if ((t = q.pollAt(b)) != null) { |
3161 |
+ |
if (q.base - q.top < 0) |
3162 |
+ |
signalWork(q); |
3163 |
+ |
t.doExec(); |
3164 |
+ |
} |
3165 |
+ |
break; |
3166 |
+ |
} |
3167 |
+ |
} |
3168 |
+ |
} |
3169 |
+ |
return true; |
3170 |
+ |
} |
3171 |
+ |
|
3172 |
+ |
/** |
3173 |
+ |
* Waits and/or attempts to assist performing tasks indefinitely |
3174 |
+ |
* until the {@link #commonPool()} {@link #isQuiescent} |
3175 |
+ |
*/ |
3176 |
+ |
static void quiesceCommonPool() { |
3177 |
+ |
common.awaitQuiescence(Long.MAX_VALUE, TimeUnit.NANOSECONDS); |
3178 |
+ |
} |
3179 |
+ |
|
3180 |
+ |
/** |
3181 |
|
* Interface for extending managed parallelism for tasks running |
3182 |
|
* in {@link ForkJoinPool}s. |
3183 |
|
* |
3273 |
|
public static void managedBlock(ManagedBlocker blocker) |
3274 |
|
throws InterruptedException { |
3275 |
|
Thread t = Thread.currentThread(); |
3276 |
< |
ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ? |
3277 |
< |
((ForkJoinWorkerThread)t).pool : null); |
3278 |
< |
while (!blocker.isReleasable()) { |
3279 |
< |
if (p == null || p.tryCompensate(null, blocker)) { |
3280 |
< |
try { |
3281 |
< |
do {} while (!blocker.isReleasable() && !blocker.block()); |
3282 |
< |
} finally { |
3283 |
< |
if (p != null) |
3276 |
> |
if (t instanceof ForkJoinWorkerThread) { |
3277 |
> |
ForkJoinPool p = ((ForkJoinWorkerThread)t).pool; |
3278 |
> |
while (!blocker.isReleasable()) { // variant of helpSignal |
3279 |
> |
WorkQueue[] ws; WorkQueue q; int m, u; |
3280 |
> |
if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) { |
3281 |
> |
for (int i = 0; i <= m; ++i) { |
3282 |
> |
if (blocker.isReleasable()) |
3283 |
> |
return; |
3284 |
> |
if ((q = ws[i]) != null && q.base - q.top < 0) { |
3285 |
> |
p.signalWork(q); |
3286 |
> |
if ((u = (int)(p.ctl >>> 32)) >= 0 || |
3287 |
> |
(u >> UAC_SHIFT) >= 0) |
3288 |
> |
break; |
3289 |
> |
} |
3290 |
> |
} |
3291 |
> |
} |
3292 |
> |
if (p.tryCompensate()) { |
3293 |
> |
try { |
3294 |
> |
do {} while (!blocker.isReleasable() && |
3295 |
> |
!blocker.block()); |
3296 |
> |
} finally { |
3297 |
|
p.incrementActiveCount(); |
3298 |
+ |
} |
3299 |
+ |
break; |
3300 |
|
} |
3122 |
– |
break; |
3301 |
|
} |
3302 |
|
} |
3303 |
+ |
else { |
3304 |
+ |
do {} while (!blocker.isReleasable() && |
3305 |
+ |
!blocker.block()); |
3306 |
+ |
} |
3307 |
|
} |
3308 |
|
|
3309 |
|
// AbstractExecutorService overrides. These rely on undocumented |
3324 |
|
private static final long PARKBLOCKER; |
3325 |
|
private static final int ABASE; |
3326 |
|
private static final int ASHIFT; |
3145 |
– |
private static final long NEXTWORKERNUMBER; |
3327 |
|
private static final long STEALCOUNT; |
3328 |
< |
private static final long MAINLOCK; |
3328 |
> |
private static final long PLOCK; |
3329 |
> |
private static final long INDEXSEED; |
3330 |
> |
private static final long QLOCK; |
3331 |
|
|
3332 |
|
static { |
3333 |
< |
poolNumberGenerator = new AtomicInteger(); |
3151 |
< |
nextSubmitterSeed = new AtomicInteger(0x55555555); |
3152 |
< |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3153 |
< |
defaultForkJoinWorkerThreadFactory = |
3154 |
< |
new DefaultForkJoinWorkerThreadFactory(); |
3155 |
< |
submitters = new ThreadSubmitter(); |
3156 |
< |
int s; |
3333 |
> |
int s; // initialize field offsets for CAS etc |
3334 |
|
try { |
3335 |
|
U = getUnsafe(); |
3336 |
|
Class<?> k = ForkJoinPool.class; |
3160 |
– |
Class<?> ak = ForkJoinTask[].class; |
3337 |
|
CTL = U.objectFieldOffset |
3338 |
|
(k.getDeclaredField("ctl")); |
3163 |
– |
NEXTWORKERNUMBER = U.objectFieldOffset |
3164 |
– |
(k.getDeclaredField("nextWorkerNumber")); |
3339 |
|
STEALCOUNT = U.objectFieldOffset |
3340 |
|
(k.getDeclaredField("stealCount")); |
3341 |
< |
MAINLOCK = U.objectFieldOffset |
3342 |
< |
(k.getDeclaredField("mainLock")); |
3341 |
> |
PLOCK = U.objectFieldOffset |
3342 |
> |
(k.getDeclaredField("plock")); |
3343 |
> |
INDEXSEED = U.objectFieldOffset |
3344 |
> |
(k.getDeclaredField("indexSeed")); |
3345 |
|
Class<?> tk = Thread.class; |
3346 |
|
PARKBLOCKER = U.objectFieldOffset |
3347 |
|
(tk.getDeclaredField("parkBlocker")); |
3348 |
+ |
Class<?> wk = WorkQueue.class; |
3349 |
+ |
QLOCK = U.objectFieldOffset |
3350 |
+ |
(wk.getDeclaredField("qlock")); |
3351 |
+ |
Class<?> ak = ForkJoinTask[].class; |
3352 |
|
ABASE = U.arrayBaseOffset(ak); |
3353 |
|
s = U.arrayIndexScale(ak); |
3354 |
|
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
3357 |
|
} |
3358 |
|
if ((s & (s-1)) != 0) |
3359 |
|
throw new Error("data type scale not a power of two"); |
3360 |
< |
try { // Establish common pool |
3361 |
< |
String pp = System.getProperty(propPrefix + "parallelism"); |
3362 |
< |
String fp = System.getProperty(propPrefix + "threadFactory"); |
3363 |
< |
String up = System.getProperty(propPrefix + "exceptionHandler"); |
3364 |
< |
ForkJoinWorkerThreadFactory fac = (fp == null) ? |
3365 |
< |
defaultForkJoinWorkerThreadFactory : |
3366 |
< |
((ForkJoinWorkerThreadFactory)ClassLoader. |
3367 |
< |
getSystemClassLoader().loadClass(fp).newInstance()); |
3368 |
< |
Thread.UncaughtExceptionHandler ueh = (up == null) ? null : |
3369 |
< |
((Thread.UncaughtExceptionHandler)ClassLoader. |
3370 |
< |
getSystemClassLoader().loadClass(up).newInstance()); |
3371 |
< |
int par; |
3372 |
< |
if ((pp == null || (par = Integer.parseInt(pp)) <= 0)) |
3373 |
< |
par = Runtime.getRuntime().availableProcessors(); |
3374 |
< |
if (par > MAX_CAP) |
3375 |
< |
par = MAX_CAP; |
3376 |
< |
commonPoolParallelism = par; |
3377 |
< |
int n = par - 1; // precompute submit mask |
3378 |
< |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; |
3379 |
< |
n |= n >>> 8; n |= n >>> 16; |
3380 |
< |
int mask = ((n + 1) << 1) - 1; |
3381 |
< |
commonPool = new ForkJoinPool(par, mask, fac, ueh); |
3382 |
< |
} catch (Exception e) { |
3383 |
< |
throw new Error(e); |
3384 |
< |
} |
3360 |
> |
|
3361 |
> |
submitters = new ThreadLocal<Submitter>(); |
3362 |
> |
ForkJoinWorkerThreadFactory fac = defaultForkJoinWorkerThreadFactory = |
3363 |
> |
new DefaultForkJoinWorkerThreadFactory(); |
3364 |
> |
modifyThreadPermission = new RuntimePermission("modifyThread"); |
3365 |
> |
|
3366 |
> |
/* |
3367 |
> |
* Establish common pool parameters. For extra caution, |
3368 |
> |
* computations to set up common pool state are here; the |
3369 |
> |
* constructor just assigns these values to fields. |
3370 |
> |
*/ |
3371 |
> |
|
3372 |
> |
int par = 0; |
3373 |
> |
Thread.UncaughtExceptionHandler handler = null; |
3374 |
> |
try { // TBD: limit or report ignored exceptions? |
3375 |
> |
String pp = System.getProperty |
3376 |
> |
("java.util.concurrent.ForkJoinPool.common.parallelism"); |
3377 |
> |
String hp = System.getProperty |
3378 |
> |
("java.util.concurrent.ForkJoinPool.common.exceptionHandler"); |
3379 |
> |
String fp = System.getProperty |
3380 |
> |
("java.util.concurrent.ForkJoinPool.common.threadFactory"); |
3381 |
> |
if (fp != null) |
3382 |
> |
fac = ((ForkJoinWorkerThreadFactory)ClassLoader. |
3383 |
> |
getSystemClassLoader().loadClass(fp).newInstance()); |
3384 |
> |
if (hp != null) |
3385 |
> |
handler = ((Thread.UncaughtExceptionHandler)ClassLoader. |
3386 |
> |
getSystemClassLoader().loadClass(hp).newInstance()); |
3387 |
> |
if (pp != null) |
3388 |
> |
par = Integer.parseInt(pp); |
3389 |
> |
} catch (Exception ignore) { |
3390 |
> |
} |
3391 |
> |
|
3392 |
> |
if (par <= 0) |
3393 |
> |
par = Runtime.getRuntime().availableProcessors(); |
3394 |
> |
if (par > MAX_CAP) |
3395 |
> |
par = MAX_CAP; |
3396 |
> |
commonParallelism = par; |
3397 |
> |
long np = (long)(-par); // precompute initial ctl value |
3398 |
> |
long ct = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
3399 |
> |
|
3400 |
> |
common = new ForkJoinPool(par, ct, fac, handler); |
3401 |
|
} |
3402 |
|
|
3403 |
|
/** |
3410 |
|
private static sun.misc.Unsafe getUnsafe() { |
3411 |
|
try { |
3412 |
|
return sun.misc.Unsafe.getUnsafe(); |
3413 |
< |
} catch (SecurityException se) { |
3414 |
< |
try { |
3415 |
< |
return java.security.AccessController.doPrivileged |
3416 |
< |
(new java.security |
3417 |
< |
.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
3418 |
< |
public sun.misc.Unsafe run() throws Exception { |
3419 |
< |
java.lang.reflect.Field f = sun.misc |
3420 |
< |
.Unsafe.class.getDeclaredField("theUnsafe"); |
3421 |
< |
f.setAccessible(true); |
3422 |
< |
return (sun.misc.Unsafe) f.get(null); |
3423 |
< |
}}); |
3424 |
< |
} catch (java.security.PrivilegedActionException e) { |
3425 |
< |
throw new RuntimeException("Could not initialize intrinsics", |
3426 |
< |
e.getCause()); |
3427 |
< |
} |
3413 |
> |
} catch (SecurityException tryReflectionInstead) {} |
3414 |
> |
try { |
3415 |
> |
return java.security.AccessController.doPrivileged |
3416 |
> |
(new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
3417 |
> |
public sun.misc.Unsafe run() throws Exception { |
3418 |
> |
Class<sun.misc.Unsafe> k = sun.misc.Unsafe.class; |
3419 |
> |
for (java.lang.reflect.Field f : k.getDeclaredFields()) { |
3420 |
> |
f.setAccessible(true); |
3421 |
> |
Object x = f.get(null); |
3422 |
> |
if (k.isInstance(x)) |
3423 |
> |
return k.cast(x); |
3424 |
> |
} |
3425 |
> |
throw new NoSuchFieldError("the Unsafe"); |
3426 |
> |
}}); |
3427 |
> |
} catch (java.security.PrivilegedActionException e) { |
3428 |
> |
throw new RuntimeException("Could not initialize intrinsics", |
3429 |
> |
e.getCause()); |
3430 |
|
} |
3431 |
|
} |
3234 |
– |
|
3432 |
|
} |