5 |
|
*/ |
6 |
|
|
7 |
|
package jsr166y; |
8 |
– |
|
8 |
|
import java.util.ArrayList; |
9 |
|
import java.util.Arrays; |
10 |
|
import java.util.Collection; |
176 |
|
* If an attempted steal fails, a thief always chooses a different |
177 |
|
* random victim target to try next. So, in order for one thief to |
178 |
|
* progress, it suffices for any in-progress poll or new push on |
179 |
< |
* any empty queue to complete. |
179 |
> |
* any empty queue to complete. (This is why we normally use |
180 |
> |
* method pollAt and its variants that try once at the apparent |
181 |
> |
* base index, else consider alternative actions, rather than |
182 |
> |
* method poll.) |
183 |
|
* |
184 |
|
* This approach also enables support of a user mode in which local |
185 |
|
* task processing is in FIFO, not LIFO order, simply by using |
209 |
|
* lock (mainly to protect in the case of resizing) but we use |
210 |
|
* only a simple spinlock (using bits in field runState), because |
211 |
|
* submitters encountering a busy queue move on to try or create |
212 |
< |
* other queues, so never block. |
212 |
> |
* other queues -- they block only when creating and registering |
213 |
> |
* new queues. |
214 |
|
* |
215 |
|
* Management |
216 |
|
* ========== |
236 |
|
* deregister WorkQueues, as well as to enable shutdown. It is |
237 |
|
* only modified under a lock (normally briefly held, but |
238 |
|
* occasionally protecting allocations and resizings) but even |
239 |
< |
* when locked remains available to check consistency. An |
237 |
< |
* auxiliary field "growHints", also only modified under lock, |
238 |
< |
* contains a candidate index for the next WorkQueue and |
239 |
< |
* a mask for submission queue indices. |
239 |
> |
* when locked remains available to check consistency. |
240 |
|
* |
241 |
|
* Recording WorkQueues. WorkQueues are recorded in the |
242 |
|
* "workQueues" array that is created upon pool construction and |
248 |
|
* readers must tolerate null slots. Shared (submission) queues |
249 |
|
* are at even indices, worker queues at odd indices. Grouping |
250 |
|
* them together in this way simplifies and speeds up task |
251 |
< |
* scanning. To avoid flailing during start-up, the array is |
252 |
< |
* presized to hold twice #parallelism workers (which is unlikely |
253 |
< |
* to need further resizing during execution). But to avoid |
254 |
< |
* dealing with so many null slots, variable runState includes a |
255 |
< |
* mask for the nearest power of two that contains all currently |
256 |
< |
* used indices. |
251 |
> |
* scanning. |
252 |
|
* |
253 |
|
* All worker thread creation is on-demand, triggered by task |
254 |
|
* submissions, replacement of terminated workers, and/or |
380 |
|
* (http://portal.acm.org/citation.cfm?id=155354). It differs in |
381 |
|
* that: (1) We only maintain dependency links across workers upon |
382 |
|
* steals, rather than use per-task bookkeeping. This sometimes |
383 |
< |
* requires a linear scan of workQueues array to locate stealers, but |
384 |
< |
* often doesn't because stealers leave hints (that may become |
383 |
> |
* requires a linear scan of workQueues array to locate stealers, |
384 |
> |
* but often doesn't because stealers leave hints (that may become |
385 |
|
* stale/wrong) of where to locate them. A stealHint is only a |
386 |
|
* hint because a worker might have had multiple steals and the |
387 |
|
* hint records only one of them (usually the most current). |
392 |
|
* which means that we miss links in the chain during long-lived |
393 |
|
* tasks, GC stalls etc (which is OK since blocking in such cases |
394 |
|
* is usually a good idea). (4) We bound the number of attempts |
395 |
< |
* to find work (see MAX_HELP_DEPTH) and fall back to suspending |
396 |
< |
* the worker and if necessary replacing it with another. |
395 |
> |
* to find work (see MAX_HELP) and fall back to suspending the |
396 |
> |
* worker and if necessary replacing it with another. |
397 |
|
* |
398 |
|
* It is impossible to keep exactly the target parallelism number |
399 |
|
* of threads running at any given time. Determining the |
400 |
|
* existence of conservatively safe helping targets, the |
401 |
|
* availability of already-created spares, and the apparent need |
402 |
|
* to create new spares are all racy, so we rely on multiple |
403 |
< |
* retries of each. Currently, in keeping with on-demand |
404 |
< |
* signalling policy, we compensate only if blocking would leave |
405 |
< |
* less than one active (non-waiting, non-blocked) worker. |
406 |
< |
* Additionally, to avoid some false alarms due to GC, lagging |
407 |
< |
* counters, system activity, etc, compensated blocking for joins |
408 |
< |
* is only attempted after rechecks stabilize in |
409 |
< |
* ForkJoinTask.awaitJoin. (Retries are interspersed with |
410 |
< |
* Thread.yield, for good citizenship.) |
403 |
> |
* retries of each. Compensation in the apparent absence of |
404 |
> |
* helping opportunities is challenging to control on JVMs, where |
405 |
> |
* GC and other activities can stall progress of tasks that in |
406 |
> |
* turn stall out many other dependent tasks, without us being |
407 |
> |
* able to determine whether they will ever require compensation. |
408 |
> |
* Even though work-stealing otherwise encounters little |
409 |
> |
* degradation in the presence of more threads than cores, |
410 |
> |
* aggressively adding new threads in such cases entails risk of |
411 |
> |
* unwanted positive feedback control loops in which more threads |
412 |
> |
* cause more dependent stalls (as well as delayed progress of |
413 |
> |
* unblocked threads to the point that we know they are available) |
414 |
> |
* leading to more situations requiring more threads, and so |
415 |
> |
* on. This aspect of control can be seen as an (analytically |
416 |
> |
* intractible) game with an opponent that may choose the worst |
417 |
> |
* (for us) active thread to stall at any time. We take several |
418 |
> |
* precautions to bound losses (and thus bound gains), mainly in |
419 |
> |
* methods tryCompensate and awaitJoin: (1) We only try |
420 |
> |
* compensation after attempting enough helping steps (measured |
421 |
> |
* via counting and timing) that we have already consumed the |
422 |
> |
* estimated cost of creating and activating a new thread. (2) We |
423 |
> |
* allow up to 50% of threads to be blocked before initially |
424 |
> |
* adding any others, and unless completely saturated, check that |
425 |
> |
* some work is available for a new worker before adding. Also, we |
426 |
> |
* create up to only 50% more threads until entering a mode that |
427 |
> |
* only adds a thread if all others are possibly blocked. All |
428 |
> |
* together, this means that we might be half as fast to react, |
429 |
> |
* and create half as many threads as possible in the ideal case, |
430 |
> |
* but present vastly fewer anomalies in all other cases compared |
431 |
> |
* to both more aggressive and more conservative alternatives. |
432 |
|
* |
433 |
|
* Style notes: There is a lot of representation-level coupling |
434 |
|
* among classes ForkJoinPool, ForkJoinWorkerThread, and |
465 |
|
// Static utilities |
466 |
|
|
467 |
|
/** |
452 |
– |
* Computes an initial hash code (also serving as a non-zero |
453 |
– |
* random seed) for a thread id. This method is expected to |
454 |
– |
* provide higher-quality hash codes than using method hashCode(). |
455 |
– |
*/ |
456 |
– |
static final int hashId(long id) { |
457 |
– |
int h = (int)id ^ (int)(id >>> 32); // Use MurmurHash of thread id |
458 |
– |
h ^= h >>> 16; h *= 0x85ebca6b; |
459 |
– |
h ^= h >>> 13; h *= 0xc2b2ae35; |
460 |
– |
h ^= h >>> 16; |
461 |
– |
return (h == 0) ? 1 : h; // ensure nonzero |
462 |
– |
} |
463 |
– |
|
464 |
– |
/** |
468 |
|
* If there is a security manager, makes sure caller has |
469 |
|
* permission to modify threads. |
470 |
|
*/ |
595 |
|
static final class WorkQueue { |
596 |
|
/** |
597 |
|
* Capacity of work-stealing queue array upon initialization. |
598 |
< |
* Must be a power of two; at least 4, but set larger to |
599 |
< |
* reduce cacheline sharing among queues. |
598 |
> |
* Must be a power of two; at least 4, but should be larger to |
599 |
> |
* reduce or eliminate cacheline sharing among queues. |
600 |
> |
* Currently, it is much larger, as a partial workaround for |
601 |
> |
* the fact that JVMs often place arrays in locations that |
602 |
> |
* share GC bookkeeping (especially cardmarks) such that |
603 |
> |
* per-write accesses encounter serious memory contention. |
604 |
|
*/ |
605 |
< |
static final int INITIAL_QUEUE_CAPACITY = 1 << 8; |
605 |
> |
static final int INITIAL_QUEUE_CAPACITY = 1 << 13; |
606 |
|
|
607 |
|
/** |
608 |
|
* Maximum size for queue arrays. Must be a power of two less |
626 |
|
volatile int base; // index of next slot for poll |
627 |
|
int top; // index of next slot for push |
628 |
|
ForkJoinTask<?>[] array; // the elements (initially unallocated) |
629 |
+ |
final ForkJoinPool pool; // the containing pool (may be null) |
630 |
|
final ForkJoinWorkerThread owner; // owning thread or null if shared |
631 |
|
volatile Thread parker; // == owner during call to park; else null |
632 |
|
ForkJoinTask<?> currentJoin; // task being joined in awaitJoin |
633 |
|
ForkJoinTask<?> currentSteal; // current non-local task being executed |
634 |
|
// Heuristic padding to ameliorate unfortunate memory placements |
635 |
< |
Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a; |
635 |
> |
Object p00, p01, p02, p03, p04, p05, p06, p07; |
636 |
> |
Object p08, p09, p0a, p0b, p0c, p0d, p0e; |
637 |
|
|
638 |
< |
WorkQueue(ForkJoinWorkerThread owner, int mode) { |
630 |
< |
this.owner = owner; |
638 |
> |
WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) { |
639 |
|
this.mode = mode; |
640 |
+ |
this.pool = pool; |
641 |
+ |
this.owner = owner; |
642 |
|
// Place indices in the center of array (that is not yet allocated) |
643 |
|
base = top = INITIAL_QUEUE_CAPACITY >>> 1; |
644 |
|
} |
645 |
|
|
646 |
|
/** |
647 |
< |
* Returns number of tasks in the queue. |
647 |
> |
* Returns the approximate number of tasks in the queue. |
648 |
|
*/ |
649 |
|
final int queueSize() { |
650 |
< |
int n = base - top; // non-owner callers must read base first |
651 |
< |
return (n >= 0) ? 0 : -n; |
650 |
> |
int n = base - top; // non-owner callers must read base first |
651 |
> |
return (n >= 0) ? 0 : -n; // ignore transient negative |
652 |
> |
} |
653 |
> |
|
654 |
> |
/** |
655 |
> |
* Provides a more accurate estimate of whether this queue has |
656 |
> |
* any tasks than does queueSize, by checking whether a |
657 |
> |
* near-empty queue has at least one unclaimed task. |
658 |
> |
*/ |
659 |
> |
final boolean isEmpty() { |
660 |
> |
ForkJoinTask<?>[] a; int m, s; |
661 |
> |
int n = base - (s = top); |
662 |
> |
return (n >= 0 || |
663 |
> |
(n == -1 && |
664 |
> |
((a = array) == null || |
665 |
> |
(m = a.length - 1) < 0 || |
666 |
> |
U.getObjectVolatile |
667 |
> |
(a, ((m & (s - 1)) << ASHIFT) + ABASE) == null))); |
668 |
|
} |
669 |
|
|
670 |
|
/** |
671 |
|
* Pushes a task. Call only by owner in unshared queues. |
672 |
|
* |
673 |
|
* @param task the task. Caller must ensure non-null. |
648 |
– |
* @param p if non-null, pool to signal if necessary |
674 |
|
* @throw RejectedExecutionException if array cannot be resized |
675 |
|
*/ |
676 |
< |
final void push(ForkJoinTask<?> task, ForkJoinPool p) { |
677 |
< |
ForkJoinTask<?>[] a; |
676 |
> |
final void push(ForkJoinTask<?> task) { |
677 |
> |
ForkJoinTask<?>[] a; ForkJoinPool p; |
678 |
|
int s = top, m, n; |
679 |
|
if ((a = array) != null) { // ignore if queue removed |
680 |
|
U.putOrderedObject |
681 |
|
(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task); |
682 |
|
if ((n = (top = s + 1) - base) <= 2) { |
683 |
< |
if (p != null) |
683 |
> |
if ((p = pool) != null) |
684 |
|
p.signalWork(); |
685 |
|
} |
686 |
|
else if (n >= m) |
716 |
|
} |
717 |
|
|
718 |
|
/** |
694 |
– |
* Takes next task, if one exists, in FIFO order. |
695 |
– |
*/ |
696 |
– |
final ForkJoinTask<?> poll() { |
697 |
– |
ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t; |
698 |
– |
while ((b = base) - top < 0 && (a = array) != null) { |
699 |
– |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
700 |
– |
if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null && |
701 |
– |
base == b && |
702 |
– |
U.compareAndSwapObject(a, j, t, null)) { |
703 |
– |
base = b + 1; |
704 |
– |
return t; |
705 |
– |
} |
706 |
– |
} |
707 |
– |
return null; |
708 |
– |
} |
709 |
– |
|
710 |
– |
/** |
719 |
|
* Takes next task, if one exists, in LIFO order. Call only |
720 |
|
* by owner in unshared queues. (We do not have a shared |
721 |
|
* version of this method because it is never needed.) |
738 |
|
} |
739 |
|
|
740 |
|
/** |
741 |
+ |
* Takes a task in FIFO order if b is base of queue and a task |
742 |
+ |
* can be claimed without contention. Specialized versions |
743 |
+ |
* appear in ForkJoinPool methods scan and tryHelpStealer. |
744 |
+ |
*/ |
745 |
+ |
final ForkJoinTask<?> pollAt(int b) { |
746 |
+ |
ForkJoinTask<?> t; ForkJoinTask<?>[] a; |
747 |
+ |
if ((a = array) != null) { |
748 |
+ |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
749 |
+ |
if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null && |
750 |
+ |
base == b && |
751 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
752 |
+ |
base = b + 1; |
753 |
+ |
return t; |
754 |
+ |
} |
755 |
+ |
} |
756 |
+ |
return null; |
757 |
+ |
} |
758 |
+ |
|
759 |
+ |
/** |
760 |
+ |
* Takes next task, if one exists, in FIFO order. |
761 |
+ |
*/ |
762 |
+ |
final ForkJoinTask<?> poll() { |
763 |
+ |
ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t; |
764 |
+ |
while ((b = base) - top < 0 && (a = array) != null) { |
765 |
+ |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
766 |
+ |
t = (ForkJoinTask<?>)U.getObjectVolatile(a, j); |
767 |
+ |
if (t != null) { |
768 |
+ |
if (base == b && |
769 |
+ |
U.compareAndSwapObject(a, j, t, null)) { |
770 |
+ |
base = b + 1; |
771 |
+ |
return t; |
772 |
+ |
} |
773 |
+ |
} |
774 |
+ |
else if (base == b) { |
775 |
+ |
if (b + 1 == top) |
776 |
+ |
break; |
777 |
+ |
Thread.yield(); // wait for lagging update |
778 |
+ |
} |
779 |
+ |
} |
780 |
+ |
return null; |
781 |
+ |
} |
782 |
+ |
|
783 |
+ |
/** |
784 |
|
* Takes next task, if one exists, in order specified by mode. |
785 |
|
*/ |
786 |
|
final ForkJoinTask<?> nextLocalTask() { |
800 |
|
} |
801 |
|
|
802 |
|
/** |
752 |
– |
* Returns task at index b if b is current base of queue. |
753 |
– |
*/ |
754 |
– |
final ForkJoinTask<?> pollAt(int b) { |
755 |
– |
ForkJoinTask<?> t; ForkJoinTask<?>[] a; |
756 |
– |
if ((a = array) != null) { |
757 |
– |
int j = (((a.length - 1) & b) << ASHIFT) + ABASE; |
758 |
– |
if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null && |
759 |
– |
base == b && |
760 |
– |
U.compareAndSwapObject(a, j, t, null)) { |
761 |
– |
base = b + 1; |
762 |
– |
return t; |
763 |
– |
} |
764 |
– |
} |
765 |
– |
return null; |
766 |
– |
} |
767 |
– |
|
768 |
– |
/** |
803 |
|
* Pops the given task only if it is at the current top. |
804 |
|
*/ |
805 |
|
final boolean tryUnpush(ForkJoinTask<?> t) { |
926 |
|
* Computes next value for random probes. Scans don't require |
927 |
|
* a very high quality generator, but also not a crummy one. |
928 |
|
* Marsaglia xor-shift is cheap and works well enough. Note: |
929 |
< |
* This is manually inlined in several usages in ForkJoinPool |
930 |
< |
* to avoid writes inside busy scan loops. |
929 |
> |
* This is manually inlined in its usages in ForkJoinPool to |
930 |
> |
* avoid writes inside busy scan loops. |
931 |
|
*/ |
932 |
|
final int nextSeed() { |
933 |
|
int r = seed; |
940 |
|
|
941 |
|
/** |
942 |
|
* Removes and runs tasks until empty, using local mode |
943 |
< |
* ordering. |
943 |
> |
* ordering. Normally called only after checking for apparent |
944 |
> |
* non-emptiness. |
945 |
|
*/ |
946 |
|
final void runLocalTasks() { |
947 |
< |
if (base - top < 0) { |
948 |
< |
for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; ) |
949 |
< |
t.doExec(); |
947 |
> |
// hoist checks from repeated pop/poll |
948 |
> |
ForkJoinTask<?>[] a; int m; |
949 |
> |
if ((a = array) != null && (m = a.length - 1) >= 0) { |
950 |
> |
if (mode == 0) { |
951 |
> |
for (int s; (s = top - 1) - base >= 0;) { |
952 |
> |
int j = ((m & s) << ASHIFT) + ABASE; |
953 |
> |
ForkJoinTask<?> t = |
954 |
> |
(ForkJoinTask<?>)U.getObjectVolatile(a, j); |
955 |
> |
if (t != null) { |
956 |
> |
if (U.compareAndSwapObject(a, j, t, null)) { |
957 |
> |
top = s; |
958 |
> |
t.doExec(); |
959 |
> |
} |
960 |
> |
} |
961 |
> |
else |
962 |
> |
break; |
963 |
> |
} |
964 |
> |
} |
965 |
> |
else { |
966 |
> |
for (int b; (b = base) - top < 0;) { |
967 |
> |
int j = ((m & b) << ASHIFT) + ABASE; |
968 |
> |
ForkJoinTask<?> t = |
969 |
> |
(ForkJoinTask<?>)U.getObjectVolatile(a, j); |
970 |
> |
if (t != null) { |
971 |
> |
if (base == b && |
972 |
> |
U.compareAndSwapObject(a, j, t, null)) { |
973 |
> |
base = b + 1; |
974 |
> |
t.doExec(); |
975 |
> |
} |
976 |
> |
} else if (base == b) { |
977 |
> |
if (b + 1 == top) |
978 |
> |
break; |
979 |
> |
Thread.yield(); // wait for lagging update |
980 |
> |
} |
981 |
> |
} |
982 |
> |
} |
983 |
|
} |
984 |
|
} |
985 |
|
|
994 |
|
if (t != null) { |
995 |
|
currentSteal = t; |
996 |
|
t.doExec(); |
997 |
< |
runLocalTasks(); |
997 |
> |
if (top != base) // conservative guard |
998 |
> |
runLocalTasks(); |
999 |
|
++nsteals; |
1000 |
|
currentSteal = null; |
1001 |
|
} |
1002 |
< |
else if (runState < 0) // terminating |
1002 |
> |
else if (runState < 0) // terminating |
1003 |
|
alive = false; |
1004 |
|
return alive; |
1005 |
|
} |
1074 |
|
* submission queues in method doSubmit. In the future, this may |
1075 |
|
* also incorporate a means to implement different task rejection |
1076 |
|
* and resubmission policies. |
1077 |
+ |
* |
1078 |
+ |
* Seeds for submitters and workers/workQueues work in basically |
1079 |
+ |
* the same way but are initialized and updated using slightly |
1080 |
+ |
* different mechanics. Both are initialized using the same |
1081 |
+ |
* approach as in class ThreadLocal, where successive values are |
1082 |
+ |
* unlikely to collide with previous values. This is done during |
1083 |
+ |
* registration for workers, but requires a separate AtomicInteger |
1084 |
+ |
* for submitters. Seeds are then randomly modified upon |
1085 |
+ |
* collisions using xorshifts, which requires a non-zero seed. |
1086 |
|
*/ |
1087 |
|
static final class Submitter { |
1088 |
|
int seed; |
1089 |
< |
Submitter() { seed = hashId(Thread.currentThread().getId()); } |
1089 |
> |
Submitter() { |
1090 |
> |
int s = nextSubmitterSeed.getAndAdd(SEED_INCREMENT); |
1091 |
> |
seed = (s == 0) ? 1 : s; // ensure non-zero |
1092 |
> |
} |
1093 |
|
} |
1094 |
|
|
1095 |
|
/** ThreadLocal class for Submitters */ |
1112 |
|
private static final AtomicInteger poolNumberGenerator; |
1113 |
|
|
1114 |
|
/** |
1115 |
+ |
* Generator for initial hashes/seeds for submitters. Accessed by |
1116 |
+ |
* Submitter class constructor. |
1117 |
+ |
*/ |
1118 |
+ |
static final AtomicInteger nextSubmitterSeed; |
1119 |
+ |
|
1120 |
+ |
/** |
1121 |
|
* Permission required for callers of methods that may start or |
1122 |
|
* kill threads. |
1123 |
|
*/ |
1150 |
|
private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10); |
1151 |
|
|
1152 |
|
/** |
1153 |
< |
* The maximum stolen->joining link depth allowed in tryHelpStealer. |
1154 |
< |
* Depths for legitimate chains are unbounded, but we use a fixed |
1155 |
< |
* constant to avoid (otherwise unchecked) cycles and to bound |
1156 |
< |
* staleness of traversal parameters at the expense of sometimes |
1157 |
< |
* blocking when we could be helping. |
1153 |
> |
* The maximum stolen->joining link depth allowed in method |
1154 |
> |
* tryHelpStealer. Must be a power of two. This value also |
1155 |
> |
* controls the maximum number of times to try to help join a task |
1156 |
> |
* without any apparent progress or change in pool state before |
1157 |
> |
* giving up and blocking (see awaitJoin). Depths for legitimate |
1158 |
> |
* chains are unbounded, but we use a fixed constant to avoid |
1159 |
> |
* (otherwise unchecked) cycles and to bound staleness of |
1160 |
> |
* traversal parameters at the expense of sometimes blocking when |
1161 |
> |
* we could be helping. |
1162 |
> |
*/ |
1163 |
> |
private static final int MAX_HELP = 32; |
1164 |
> |
|
1165 |
> |
/** |
1166 |
> |
* Secondary time-based bound (in nanosecs) for helping attempts |
1167 |
> |
* before trying compensated blocking in awaitJoin. Used in |
1168 |
> |
* conjunction with MAX_HELP to reduce variance due to different |
1169 |
> |
* polling rates associated with different helping options. The |
1170 |
> |
* value should roughly approximate the time required to create |
1171 |
> |
* and/or activate a worker thread. |
1172 |
> |
*/ |
1173 |
> |
private static final long COMPENSATION_DELAY = 100L * 1000L; // 0.1 millisec |
1174 |
> |
|
1175 |
> |
/** |
1176 |
> |
* Increment for seed generators. See class ThreadLocal for |
1177 |
> |
* explanation. |
1178 |
|
*/ |
1179 |
< |
private static final int MAX_HELP_DEPTH = 16; |
1179 |
> |
private static final int SEED_INCREMENT = 0x61c88647; |
1180 |
|
|
1181 |
|
/** |
1182 |
|
* Bits and masks for control variables |
1208 |
|
* |
1209 |
|
* Field runState is an int packed with: |
1210 |
|
* SHUTDOWN: true if shutdown is enabled (1 bit) |
1211 |
< |
* SEQ: a sequence number updated upon (de)registering workers (15 bits) |
1212 |
< |
* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits) |
1211 |
> |
* SEQ: a sequence number updated upon (de)registering workers (30 bits) |
1212 |
> |
* INIT: set true after workQueues array construction (1 bit) |
1213 |
|
* |
1214 |
< |
* The combination of mask and sequence number enables simple |
1215 |
< |
* consistency checks: Staleness of read-only operations on the |
1216 |
< |
* workQueues array can be checked by comparing runState before vs |
1110 |
< |
* after the reads. The low 16 bits (i.e, anding with SMASK) hold |
1111 |
< |
* the smallest power of two covering all indices, minus |
1112 |
< |
* one. |
1214 |
> |
* The sequence number enables simple consistency checks: |
1215 |
> |
* Staleness of read-only operations on the workQueues array can |
1216 |
> |
* be checked by comparing runState before vs after the reads. |
1217 |
|
*/ |
1218 |
|
|
1219 |
|
// bit positions/shifts for fields |
1223 |
|
private static final int EC_SHIFT = 16; |
1224 |
|
|
1225 |
|
// bounds |
1122 |
– |
private static final int POOL_MAX = 0x7fff; // max #workers - 1 |
1226 |
|
private static final int SMASK = 0xffff; // short bits |
1227 |
+ |
private static final int MAX_CAP = 0x7fff; // max #workers - 1 |
1228 |
|
private static final int SQMASK = 0xfffe; // even short bits |
1229 |
|
private static final int SHORT_SIGN = 1 << 15; |
1230 |
|
private static final int INT_SIGN = 1 << 31; |
1252 |
|
|
1253 |
|
// runState bits |
1254 |
|
private static final int SHUTDOWN = 1 << 31; |
1151 |
– |
private static final int RS_SEQ = 1 << 16; |
1152 |
– |
private static final int RS_SEQ_MASK = 0x7fff0000; |
1255 |
|
|
1256 |
|
// access mode for WorkQueue |
1257 |
|
static final int LIFO_QUEUE = 0; |
1270 |
|
volatile long ctl; // main pool control |
1271 |
|
final int parallelism; // parallelism level |
1272 |
|
final int localMode; // per-worker scheduling mode |
1273 |
< |
int growHints; // for expanding indices/ranges |
1274 |
< |
volatile int runState; // shutdown status, seq, and mask |
1273 |
> |
final int submitMask; // submit queue index bound |
1274 |
> |
int nextSeed; // for initializing worker seeds |
1275 |
> |
volatile int runState; // shutdown status and seq |
1276 |
|
WorkQueue[] workQueues; // main registry |
1277 |
|
final Mutex lock; // for registration |
1278 |
|
final Condition termination; // for awaitTermination |
1282 |
|
final AtomicInteger nextWorkerNumber; // to create worker name string |
1283 |
|
final String workerNamePrefix; // to create worker name string |
1284 |
|
|
1285 |
< |
// Creating, registering, deregistering and running workers |
1285 |
> |
// Creating, registering, and deregistering workers |
1286 |
|
|
1287 |
|
/** |
1288 |
|
* Tries to create and start a worker |
1313 |
|
} |
1314 |
|
|
1315 |
|
/** |
1316 |
< |
* Callback from ForkJoinWorkerThread constructor to establish and |
1317 |
< |
* record its WorkQueue. |
1316 |
> |
* Callback from ForkJoinWorkerThread constructor to establish its |
1317 |
> |
* poolIndex and record its WorkQueue. To avoid scanning bias due |
1318 |
> |
* to packing entries in front of the workQueues array, we treat |
1319 |
> |
* the array as a simple power-of-two hash table using per-thread |
1320 |
> |
* seed as hash, expanding as needed. |
1321 |
|
* |
1322 |
< |
* @param wt the worker thread |
1322 |
> |
* @param w the worker's queue |
1323 |
|
*/ |
1324 |
< |
final void registerWorker(ForkJoinWorkerThread wt) { |
1219 |
< |
WorkQueue w = wt.workQueue; |
1324 |
> |
final void registerWorker(WorkQueue w) { |
1325 |
|
Mutex lock = this.lock; |
1326 |
|
lock.lock(); |
1327 |
|
try { |
1223 |
– |
int g = growHints, k = g & SMASK; |
1328 |
|
WorkQueue[] ws = workQueues; |
1329 |
< |
if (ws != null) { // ignore on shutdown |
1330 |
< |
int n = ws.length; |
1331 |
< |
if ((k & 1) == 0 || k >= n || ws[k] != null) { |
1332 |
< |
for (k = 1; k < n && ws[k] != null; k += 2) |
1333 |
< |
; // workers are at odd indices |
1334 |
< |
if (k >= n) // resize |
1335 |
< |
workQueues = ws = Arrays.copyOf(ws, n << 1); |
1336 |
< |
} |
1337 |
< |
w.eventCount = w.poolIndex = k; // establish before recording |
1338 |
< |
ws[k] = w; |
1339 |
< |
growHints = (g & ~SMASK) | ((k + 2) & SMASK); |
1340 |
< |
int rs = runState; |
1341 |
< |
int m = rs & SMASK; // recalculate runState mask |
1342 |
< |
if (k > m) |
1239 |
< |
m = (m << 1) + 1; |
1240 |
< |
runState = (rs & SHUTDOWN) | ((rs + RS_SEQ) & RS_SEQ_MASK) | m; |
1329 |
> |
if (w != null && ws != null) { // skip on shutdown/failure |
1330 |
> |
int rs, n; |
1331 |
> |
while ((n = ws.length) < // ensure can hold total |
1332 |
> |
(parallelism + (short)(ctl >>> TC_SHIFT) << 1)) |
1333 |
> |
workQueues = ws = Arrays.copyOf(ws, n << 1); |
1334 |
> |
int m = n - 1; |
1335 |
> |
int s = nextSeed += SEED_INCREMENT; // rarely-colliding sequence |
1336 |
> |
w.seed = (s == 0) ? 1 : s; // ensure non-zero seed |
1337 |
> |
int r = (s << 1) | 1; // use odd-numbered indices |
1338 |
> |
while (ws[r &= m] != null) // step by approx half size |
1339 |
> |
r += ((n >>> 1) & SQMASK) + 2; |
1340 |
> |
w.eventCount = w.poolIndex = r; // establish before recording |
1341 |
> |
ws[r] = w; // also update seq |
1342 |
> |
runState = ((rs = runState) & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN); |
1343 |
|
} |
1344 |
|
} finally { |
1345 |
|
lock.unlock(); |
1356 |
|
* @param ex the exception causing failure, or null if none |
1357 |
|
*/ |
1358 |
|
final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) { |
1359 |
+ |
Mutex lock = this.lock; |
1360 |
|
WorkQueue w = null; |
1361 |
|
if (wt != null && (w = wt.workQueue) != null) { |
1362 |
|
w.runState = -1; // ensure runState is set |
1363 |
|
stealCount.getAndAdd(w.totalSteals + w.nsteals); |
1364 |
|
int idx = w.poolIndex; |
1262 |
– |
Mutex lock = this.lock; |
1365 |
|
lock.lock(); |
1366 |
|
try { // remove record from array |
1367 |
|
WorkQueue[] ws = workQueues; |
1368 |
< |
if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) { |
1368 |
> |
if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w) |
1369 |
|
ws[idx] = null; |
1268 |
– |
growHints = (growHints & ~SMASK) | idx; |
1269 |
– |
} |
1370 |
|
} finally { |
1371 |
|
lock.unlock(); |
1372 |
|
} |
1390 |
|
U.throwException(ex); |
1391 |
|
} |
1392 |
|
|
1293 |
– |
/** |
1294 |
– |
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1295 |
– |
*/ |
1296 |
– |
final void runWorker(ForkJoinWorkerThread wt) { |
1297 |
– |
// Initialize queue array and seed in this thread |
1298 |
– |
WorkQueue w = wt.workQueue; |
1299 |
– |
w.growArray(false); |
1300 |
– |
w.seed = hashId(Thread.currentThread().getId()); |
1301 |
– |
|
1302 |
– |
do {} while (w.runTask(scan(w))); |
1303 |
– |
} |
1393 |
|
|
1394 |
|
// Submissions |
1395 |
|
|
1396 |
|
/** |
1397 |
|
* Unless shutting down, adds the given task to a submission queue |
1398 |
|
* at submitter's current queue index (modulo submission |
1399 |
< |
* range). If no queue exists at the index, one is created unless |
1400 |
< |
* pool lock is busy. If the queue and/or lock are busy, another |
1401 |
< |
* index is randomly chosen. The mask in growHints controls the |
1402 |
< |
* effective index range of queues considered. The mask is |
1403 |
< |
* expanded, up to the current workerQueue mask, upon any detected |
1404 |
< |
* contention but otherwise remains small to avoid needlessly |
1316 |
< |
* creating queues when there is no contention. |
1399 |
> |
* range). If no queue exists at the index, one is created. If |
1400 |
> |
* the queue is busy, another index is randomly chosen. The |
1401 |
> |
* submitMask bounds the effective number of queues to the |
1402 |
> |
* (nearest poswer of two for) parallelism level. |
1403 |
> |
* |
1404 |
> |
* @param task the task. Caller must ensure non-null. |
1405 |
|
*/ |
1406 |
|
private void doSubmit(ForkJoinTask<?> task) { |
1319 |
– |
if (task == null) |
1320 |
– |
throw new NullPointerException(); |
1407 |
|
Submitter s = submitters.get(); |
1408 |
< |
for (int r = s.seed, m = growHints >>> 16;;) { |
1409 |
< |
WorkQueue[] ws; WorkQueue q; Mutex lk; |
1408 |
> |
for (int r = s.seed, m = submitMask;;) { |
1409 |
> |
WorkQueue[] ws; WorkQueue q; |
1410 |
|
int k = r & m & SQMASK; // use only even indices |
1411 |
|
if (runState < 0 || (ws = workQueues) == null || ws.length <= k) |
1412 |
|
throw new RejectedExecutionException(); // shutting down |
1413 |
< |
if ((q = ws[k]) == null && (lk = lock).tryAcquire(0)) { |
1414 |
< |
try { // try to create new queue |
1415 |
< |
if (ws == workQueues && (q = ws[k]) == null) { |
1416 |
< |
int rs; // update runState seq |
1417 |
< |
ws[k] = q = new WorkQueue(null, SHARED_QUEUE); |
1418 |
< |
runState = (((rs = runState) & SHUTDOWN) | |
1419 |
< |
((rs + RS_SEQ) & ~SHUTDOWN)); |
1413 |
> |
else if ((q = ws[k]) == null) { // create new queue |
1414 |
> |
WorkQueue nq = new WorkQueue(this, null, SHARED_QUEUE); |
1415 |
> |
Mutex lock = this.lock; // construct outside lock |
1416 |
> |
lock.lock(); |
1417 |
> |
try { // recheck under lock |
1418 |
> |
int rs = runState; // to update seq |
1419 |
> |
if (ws == workQueues && ws[k] == null) { |
1420 |
> |
ws[k] = nq; |
1421 |
> |
runState = ((rs & SHUTDOWN) | ((rs + 2) & ~SHUTDOWN)); |
1422 |
|
} |
1423 |
|
} finally { |
1424 |
< |
lk.unlock(); |
1424 |
> |
lock.unlock(); |
1425 |
|
} |
1426 |
|
} |
1427 |
< |
if (q != null) { |
1428 |
< |
if (q.trySharedPush(task)) { |
1429 |
< |
signalWork(); |
1430 |
< |
return; |
1431 |
< |
} |
1432 |
< |
else if (m < parallelism - 1 && m < (runState & SMASK)) { |
1433 |
< |
Mutex lock = this.lock; |
1346 |
< |
lock.lock(); // block until lock free |
1347 |
< |
int g = growHints; |
1348 |
< |
if (g >>> 16 == m) // expand range |
1349 |
< |
growHints = (((m << 1) + 1) << 16) | (g & SMASK); |
1350 |
< |
lock.unlock(); // no need for try/finally |
1351 |
< |
} |
1352 |
< |
else if ((r & m) == 0) |
1353 |
< |
Thread.yield(); // occasionally yield if busy |
1354 |
< |
} |
1355 |
< |
if (m == (m = growHints >>> 16)) { |
1356 |
< |
r ^= r << 13; // update seed unless new range |
1357 |
< |
r ^= r >>> 17; // same xorshift as WorkQueues |
1427 |
> |
else if (q.trySharedPush(task)) { |
1428 |
> |
signalWork(); |
1429 |
> |
return; |
1430 |
> |
} |
1431 |
> |
else if (m > 1) { // move to a different index |
1432 |
> |
r ^= r << 13; // same xorshift as WorkQueues |
1433 |
> |
r ^= r >>> 17; |
1434 |
|
s.seed = r ^= r << 5; |
1435 |
|
} |
1436 |
+ |
else |
1437 |
+ |
Thread.yield(); // yield if no alternatives |
1438 |
|
} |
1439 |
|
} |
1440 |
|
|
1483 |
|
} |
1484 |
|
} |
1485 |
|
|
1486 |
+ |
|
1487 |
+ |
// Scanning for tasks |
1488 |
+ |
|
1489 |
|
/** |
1490 |
< |
* Tries to decrement active count (sometimes implicitly) and |
1410 |
< |
* possibly release or create a compensating worker in preparation |
1411 |
< |
* for blocking. Fails on contention or termination. |
1412 |
< |
* |
1413 |
< |
* @return true if the caller can block, else should recheck and retry |
1490 |
> |
* Top-level runloop for workers, called by ForkJoinWorkerThread.run. |
1491 |
|
*/ |
1492 |
< |
final boolean tryCompensate() { |
1493 |
< |
WorkQueue w; Thread p; |
1494 |
< |
int pc = parallelism, e, u, ac, tc, i; |
1418 |
< |
long c = ctl; |
1419 |
< |
WorkQueue[] ws = workQueues; |
1420 |
< |
if ((e = (int)c) >= 0) { |
1421 |
< |
if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 && |
1422 |
< |
e != 0 && ws != null && (i = e & SMASK) < ws.length && |
1423 |
< |
(w = ws[i]) != null) { |
1424 |
< |
long nc = (long)(w.nextWait & E_MASK) | (c & (AC_MASK|TC_MASK)); |
1425 |
< |
if (w.eventCount == (e | INT_SIGN) && |
1426 |
< |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1427 |
< |
w.eventCount = (e + E_SEQ) & E_MASK; |
1428 |
< |
if ((p = w.parker) != null) |
1429 |
< |
U.unpark(p); |
1430 |
< |
return true; // release an idle worker |
1431 |
< |
} |
1432 |
< |
} |
1433 |
< |
else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) { |
1434 |
< |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1435 |
< |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1436 |
< |
return true; // no compensation needed |
1437 |
< |
} |
1438 |
< |
else if (tc + pc < POOL_MAX) { |
1439 |
< |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1440 |
< |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1441 |
< |
addWorker(); |
1442 |
< |
return true; // create replacement |
1443 |
< |
} |
1444 |
< |
} |
1445 |
< |
} |
1446 |
< |
return false; |
1492 |
> |
final void runWorker(WorkQueue w) { |
1493 |
> |
w.growArray(false); // initialize queue array in this thread |
1494 |
> |
do {} while (w.runTask(scan(w))); |
1495 |
|
} |
1496 |
|
|
1449 |
– |
// Scanning for tasks |
1450 |
– |
|
1497 |
|
/** |
1498 |
|
* Scans for and, if found, returns one task, else possibly |
1499 |
|
* inactivates the worker. This method operates on single reads of |
1500 |
< |
* volatile state and is designed to be re-invoked continuously in |
1501 |
< |
* part because it returns upon detecting inconsistencies, |
1500 |
> |
* volatile state and is designed to be re-invoked continuously, |
1501 |
> |
* in part because it returns upon detecting inconsistencies, |
1502 |
|
* contention, or state changes that indicate possible success on |
1503 |
|
* re-invocation. |
1504 |
|
* |
1505 |
< |
* The scan searches for tasks across queues, randomly selecting |
1506 |
< |
* the first #queues probes, favoring steals over submissions |
1507 |
< |
* (by exploiting even/odd indexing), and then performing a |
1508 |
< |
* circular sweep of all queues. The scan terminates upon either |
1509 |
< |
* finding a non-empty queue, or completing a full sweep. If the |
1510 |
< |
* worker is not inactivated, it takes and returns a task from |
1511 |
< |
* this queue. On failure to find a task, we take one of the |
1512 |
< |
* following actions, after which the caller will retry calling |
1467 |
< |
* this method unless terminated. |
1505 |
> |
* The scan searches for tasks across a random permutation of |
1506 |
> |
* queues (starting at a random index and stepping by a random |
1507 |
> |
* relative prime, checking each at least once). The scan |
1508 |
> |
* terminates upon either finding a non-empty queue, or completing |
1509 |
> |
* the sweep. If the worker is not inactivated, it takes and |
1510 |
> |
* returns a task from this queue. On failure to find a task, we |
1511 |
> |
* take one of the following actions, after which the caller will |
1512 |
> |
* retry calling this method unless terminated. |
1513 |
|
* |
1514 |
|
* * If pool is terminating, terminate the worker. |
1515 |
|
* |
1520 |
|
* another worker, but with same net effect. Releasing in other |
1521 |
|
* cases as well ensures that we have enough workers running. |
1522 |
|
* |
1478 |
– |
* * If the caller has run a task since the last empty scan, |
1479 |
– |
* return (to allow rescan) if other workers are not also yet |
1480 |
– |
* enqueued. Field WorkQueue.rescans counts down on each scan to |
1481 |
– |
* ensure eventual inactivation and blocking. |
1482 |
– |
* |
1523 |
|
* * If not already enqueued, try to inactivate and enqueue the |
1524 |
< |
* worker on wait queue. |
1524 |
> |
* worker on wait queue. Or, if inactivating has caused the pool |
1525 |
> |
* to be quiescent, relay to idleAwaitWork to check for |
1526 |
> |
* termination and possibly shrink pool. |
1527 |
> |
* |
1528 |
> |
* * If already inactive, and the caller has run a task since the |
1529 |
> |
* last empty scan, return (to allow rescan) unless others are |
1530 |
> |
* also inactivated. Field WorkQueue.rescans counts down on each |
1531 |
> |
* scan to ensure eventual inactivation and blocking. |
1532 |
|
* |
1533 |
< |
* * If already enqueued and none of the above apply, either park |
1534 |
< |
* awaiting signal, or if this is the most recent waiter and pool |
1488 |
< |
* is quiescent, relay to idleAwaitWork to check for termination |
1489 |
< |
* and possibly shrink pool. |
1533 |
> |
* * If already enqueued and none of the above apply, park |
1534 |
> |
* awaiting signal, |
1535 |
|
* |
1536 |
|
* @param w the worker (via its WorkQueue) |
1537 |
|
* @return a task or null of none found |
1538 |
|
*/ |
1539 |
|
private final ForkJoinTask<?> scan(WorkQueue w) { |
1540 |
< |
boolean swept = false; // true after full empty scan |
1541 |
< |
WorkQueue[] ws; // volatile read order matters |
1542 |
< |
int r = w.seed, ec = w.eventCount; // ec is negative if inactive |
1543 |
< |
int rs = runState, m = rs & SMASK; |
1544 |
< |
if ((ws = workQueues) != null && ws.length > m) { // consistency check |
1545 |
< |
for (int k = 0, j = -1 - m; ; ++j) { |
1546 |
< |
WorkQueue q; int b; |
1547 |
< |
if (j < 0) { // random probes while j negative |
1548 |
< |
r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) | (j & 1); |
1549 |
< |
} // worker (not submit) for odd j |
1550 |
< |
else // cyclic scan when j >= 0 |
1551 |
< |
k += 7; // step 7 reduces array packing bias |
1552 |
< |
if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) { |
1553 |
< |
ForkJoinTask<?> t = (ec >= 0) ? q.pollAt(b) : null; |
1554 |
< |
w.seed = r; // save seed for next scan |
1510 |
< |
if (t != null) |
1540 |
> |
WorkQueue[] ws; // first update random seed |
1541 |
> |
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
1542 |
> |
int rs = runState, m; // volatile read order matters |
1543 |
> |
if ((ws = workQueues) != null && (m = ws.length - 1) > 0) { |
1544 |
> |
int ec = w.eventCount; // ec is negative if inactive |
1545 |
> |
int step = (r >>> 16) | 1; // relative prime |
1546 |
> |
for (int j = (m + 1) << 2; ; r += step) { |
1547 |
> |
WorkQueue q; ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b; |
1548 |
> |
if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 && |
1549 |
> |
(a = q.array) != null) { // probably nonempty |
1550 |
> |
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1551 |
> |
t = (ForkJoinTask<?>)U.getObjectVolatile(a, i); |
1552 |
> |
if (q.base == b && ec >= 0 && t != null && |
1553 |
> |
U.compareAndSwapObject(a, i, t, null)) { |
1554 |
> |
q.base = b + 1; // specialization of pollAt |
1555 |
|
return t; |
1556 |
< |
break; |
1556 |
> |
} |
1557 |
> |
else if ((t != null || b + 1 != q.top) && |
1558 |
> |
(ec < 0 || j <= m)) { |
1559 |
> |
rs = 0; // mark scan as imcomplete |
1560 |
> |
break; // caller can retry after release |
1561 |
> |
} |
1562 |
|
} |
1563 |
< |
else if (j - m > m) { |
1515 |
< |
if (rs == runState) // staleness check |
1516 |
< |
swept = true; |
1563 |
> |
if (--j < 0) |
1564 |
|
break; |
1518 |
– |
} |
1565 |
|
} |
1520 |
– |
|
1521 |
– |
// Decode ctl on empty scan |
1566 |
|
long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns; |
1567 |
< |
if (e < 0) // pool is terminating |
1568 |
< |
w.runState = -1; |
1569 |
< |
else if (!swept) { // try to release a waiter |
1570 |
< |
WorkQueue v; Thread p; |
1571 |
< |
if (e > 0 && a < 0 && (v = ws[e & m]) != null && |
1572 |
< |
v.eventCount == (e | INT_SIGN)) { |
1567 |
> |
if (e < 0) // decode ctl on empty scan |
1568 |
> |
w.runState = -1; // pool is terminating |
1569 |
> |
else if (rs == 0 || rs != runState) { // incomplete scan |
1570 |
> |
WorkQueue v; Thread p; // try to release a waiter |
1571 |
> |
if (e > 0 && a < 0 && w.eventCount == ec && |
1572 |
> |
(v = ws[e & m]) != null && v.eventCount == (e | INT_SIGN)) { |
1573 |
|
long nc = ((long)(v.nextWait & E_MASK) | |
1574 |
|
((c + AC_UNIT) & (AC_MASK|TC_MASK))); |
1575 |
< |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1575 |
> |
if (ctl == c && U.compareAndSwapLong(this, CTL, c, nc)) { |
1576 |
|
v.eventCount = (e + E_SEQ) & E_MASK; |
1577 |
|
if ((p = v.parker) != null) |
1578 |
|
U.unpark(p); |
1579 |
|
} |
1580 |
|
} |
1581 |
|
} |
1582 |
< |
else if ((nr = w.rescans) > 0) { // continue rescanning |
1539 |
< |
int ac = a + parallelism; |
1540 |
< |
if (((w.rescans = (ac < nr) ? ac : nr - 1) & 3) == 0 && |
1541 |
< |
w.eventCount == ec) |
1542 |
< |
Thread.yield(); // occasionally yield |
1543 |
< |
} |
1544 |
< |
else if (ec >= 0) { // try to enqueue |
1582 |
> |
else if (ec >= 0) { // try to enqueue/inactivate |
1583 |
|
long nc = (long)ec | ((c - AC_UNIT) & (AC_MASK|TC_MASK)); |
1584 |
|
w.nextWait = e; |
1585 |
< |
w.eventCount = ec | INT_SIGN;// mark as inactive |
1586 |
< |
if (!U.compareAndSwapLong(this, CTL, c, nc)) |
1587 |
< |
w.eventCount = ec; // unmark on CAS failure |
1588 |
< |
else if ((ns = w.nsteals) != 0) { |
1589 |
< |
w.nsteals = 0; // set rescans if ran task |
1590 |
< |
w.rescans = a + parallelism; |
1591 |
< |
w.totalSteals += ns; |
1585 |
> |
w.eventCount = ec | INT_SIGN; // mark as inactive |
1586 |
> |
if (ctl != c || !U.compareAndSwapLong(this, CTL, c, nc)) |
1587 |
> |
w.eventCount = ec; // unmark on CAS failure |
1588 |
> |
else { |
1589 |
> |
if ((ns = w.nsteals) != 0) { |
1590 |
> |
w.nsteals = 0; // set rescans if ran task |
1591 |
> |
w.rescans = (a > 0)? 0 : a + parallelism; |
1592 |
> |
w.totalSteals += ns; |
1593 |
> |
} |
1594 |
> |
if (a == 1 - parallelism) // quiescent |
1595 |
> |
idleAwaitWork(w, nc, c); |
1596 |
|
} |
1597 |
|
} |
1598 |
< |
else { // already queued |
1599 |
< |
if (parallelism == -a) |
1600 |
< |
idleAwaitWork(w); // quiescent |
1601 |
< |
if (w.eventCount == ec) { |
1602 |
< |
Thread.interrupted(); // clear status |
1603 |
< |
ForkJoinWorkerThread wt = w.owner; |
1598 |
> |
else if (w.eventCount < 0) { // already queued |
1599 |
> |
if ((nr = w.rescans) > 0) { // continue rescanning |
1600 |
> |
int ac = a + parallelism; |
1601 |
> |
if (((w.rescans = (ac < nr) ? ac : nr - 1) & 3) == 0) |
1602 |
> |
Thread.yield(); // yield before block |
1603 |
> |
} |
1604 |
> |
else { |
1605 |
> |
Thread.interrupted(); // clear status |
1606 |
> |
Thread wt = Thread.currentThread(); |
1607 |
|
U.putObject(wt, PARKBLOCKER, this); |
1608 |
< |
w.parker = wt; // emulate LockSupport.park |
1609 |
< |
if (w.eventCount == ec) // recheck |
1610 |
< |
U.park(false, 0L); // block |
1608 |
> |
w.parker = wt; // emulate LockSupport.park |
1609 |
> |
if (w.eventCount < 0) // recheck |
1610 |
> |
U.park(false, 0L); |
1611 |
|
w.parker = null; |
1612 |
|
U.putObject(wt, PARKBLOCKER, null); |
1613 |
|
} |
1617 |
|
} |
1618 |
|
|
1619 |
|
/** |
1620 |
< |
* If inactivating worker w has caused pool to become quiescent, |
1621 |
< |
* checks for pool termination, and, so long as this is not the |
1622 |
< |
* only worker, waits for event for up to SHRINK_RATE nanosecs. |
1623 |
< |
* On timeout, if ctl has not changed, terminates the worker, |
1624 |
< |
* which will in turn wake up another worker to possibly repeat |
1625 |
< |
* this process. |
1620 |
> |
* If inactivating worker w has caused the pool to become |
1621 |
> |
* quiescent, checks for pool termination, and, so long as this is |
1622 |
> |
* not the only worker, waits for event for up to SHRINK_RATE |
1623 |
> |
* nanosecs. On timeout, if ctl has not changed, terminates the |
1624 |
> |
* worker, which will in turn wake up another worker to possibly |
1625 |
> |
* repeat this process. |
1626 |
|
* |
1627 |
|
* @param w the calling worker |
1628 |
+ |
* @param currentCtl the ctl value triggering possible quiescence |
1629 |
+ |
* @param prevCtl the ctl value to restore if thread is terminated |
1630 |
|
*/ |
1631 |
< |
private void idleAwaitWork(WorkQueue w) { |
1632 |
< |
long c; int nw, ec; |
1633 |
< |
if (!tryTerminate(false, false) && |
1634 |
< |
(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 && |
1635 |
< |
(ec = w.eventCount) == ((int)c | INT_SIGN) && |
1636 |
< |
(nw = w.nextWait) != 0) { |
1590 |
< |
long nc = ((long)(nw & E_MASK) | // ctl to restore on timeout |
1591 |
< |
((c + AC_UNIT) & AC_MASK) | (c & TC_MASK)); |
1592 |
< |
ForkJoinWorkerThread wt = w.owner; |
1593 |
< |
while (ctl == c) { |
1631 |
> |
private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) { |
1632 |
> |
if (w.eventCount < 0 && !tryTerminate(false, false) && |
1633 |
> |
(int)prevCtl != 0 && ctl == currentCtl) { |
1634 |
> |
Thread wt = Thread.currentThread(); |
1635 |
> |
Thread.yield(); // yield before block |
1636 |
> |
while (ctl == currentCtl) { |
1637 |
|
long startTime = System.nanoTime(); |
1638 |
|
Thread.interrupted(); // timed variant of version in scan() |
1639 |
|
U.putObject(wt, PARKBLOCKER, this); |
1640 |
|
w.parker = wt; |
1641 |
< |
if (ctl == c) |
1641 |
> |
if (ctl == currentCtl) |
1642 |
|
U.park(false, SHRINK_RATE); |
1643 |
|
w.parker = null; |
1644 |
|
U.putObject(wt, PARKBLOCKER, null); |
1645 |
< |
if (ctl != c) |
1645 |
> |
if (ctl != currentCtl) |
1646 |
|
break; |
1647 |
|
if (System.nanoTime() - startTime >= SHRINK_TIMEOUT && |
1648 |
< |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1649 |
< |
w.eventCount = (ec + E_SEQ) | E_MASK; |
1650 |
< |
w.runState = -1; // shrink |
1648 |
> |
U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) { |
1649 |
> |
w.eventCount = (w.eventCount + E_SEQ) | E_MASK; |
1650 |
> |
w.runState = -1; // shrink |
1651 |
|
break; |
1652 |
|
} |
1653 |
|
} |
1665 |
|
* leaves hints in workers to speed up subsequent calls. The |
1666 |
|
* implementation is very branchy to cope with potential |
1667 |
|
* inconsistencies or loops encountering chains that are stale, |
1668 |
< |
* unknown, or of length greater than MAX_HELP_DEPTH links. All |
1669 |
< |
* of these cases are dealt with by just retrying by caller. |
1668 |
> |
* unknown, or so long that they are likely cyclic. All of these |
1669 |
> |
* cases are dealt with by just retrying by caller. |
1670 |
|
* |
1671 |
|
* @param joiner the joining worker |
1672 |
|
* @param task the task to join |
1673 |
|
* @return true if found or ran a task (and so is immediately retryable) |
1674 |
|
*/ |
1675 |
< |
final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) { |
1676 |
< |
ForkJoinTask<?> subtask; // current target |
1675 |
> |
private boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) { |
1676 |
> |
WorkQueue[] ws; |
1677 |
> |
int m, depth = MAX_HELP; // remaining chain depth |
1678 |
|
boolean progress = false; |
1679 |
< |
int depth = 0; // current chain depth |
1680 |
< |
int m = runState & SMASK; |
1681 |
< |
WorkQueue[] ws = workQueues; |
1682 |
< |
|
1683 |
< |
if (ws != null && ws.length > m && (subtask = task).status >= 0) { |
1684 |
< |
outer:for (WorkQueue j = joiner;;) { |
1641 |
< |
// Try to find the stealer of subtask, by first using hint |
1642 |
< |
WorkQueue stealer = null; |
1643 |
< |
WorkQueue v = ws[j.stealHint & m]; |
1679 |
> |
if ((ws = workQueues) != null && (m = ws.length - 1) > 0 && |
1680 |
> |
task.status >= 0) { |
1681 |
> |
ForkJoinTask<?> subtask = task; // current target |
1682 |
> |
outer: for (WorkQueue j = joiner;;) { |
1683 |
> |
WorkQueue stealer = null; // find stealer of subtask |
1684 |
> |
WorkQueue v = ws[j.stealHint & m]; // try hint |
1685 |
|
if (v != null && v.currentSteal == subtask) |
1686 |
|
stealer = v; |
1687 |
< |
else { |
1687 |
> |
else { // scan |
1688 |
|
for (int i = 1; i <= m; i += 2) { |
1689 |
< |
if ((v = ws[i]) != null && v.currentSteal == subtask) { |
1689 |
> |
if ((v = ws[i]) != null && v.currentSteal == subtask && |
1690 |
> |
v != joiner) { |
1691 |
|
stealer = v; |
1692 |
< |
j.stealHint = i; // save hint |
1692 |
> |
j.stealHint = i; // save hint |
1693 |
|
break; |
1694 |
|
} |
1695 |
|
} |
1697 |
|
break; |
1698 |
|
} |
1699 |
|
|
1700 |
< |
for (WorkQueue q = stealer;;) { // Try to help stealer |
1701 |
< |
ForkJoinTask<?> t; int b; |
1700 |
> |
for (WorkQueue q = stealer;;) { // try to help stealer |
1701 |
> |
ForkJoinTask[] a; ForkJoinTask<?> t; int b; |
1702 |
|
if (task.status < 0) |
1703 |
|
break outer; |
1704 |
< |
if ((b = q.base) - q.top < 0) { |
1704 |
> |
if ((b = q.base) - q.top < 0 && (a = q.array) != null) { |
1705 |
|
progress = true; |
1706 |
< |
if (subtask.status < 0) |
1707 |
< |
break outer; // stale |
1708 |
< |
if ((t = q.pollAt(b)) != null) { |
1709 |
< |
stealer.stealHint = joiner.poolIndex; |
1706 |
> |
int i = (((a.length - 1) & b) << ASHIFT) + ABASE; |
1707 |
> |
t = (ForkJoinTask<?>)U.getObjectVolatile(a, i); |
1708 |
> |
if (subtask.status < 0) // must recheck before taking |
1709 |
> |
break outer; |
1710 |
> |
if (t != null && |
1711 |
> |
q.base == b && |
1712 |
> |
U.compareAndSwapObject(a, i, t, null)) { |
1713 |
> |
q.base = b + 1; |
1714 |
|
joiner.runSubtask(t); |
1715 |
|
} |
1716 |
+ |
else if (q.base == b) |
1717 |
+ |
break outer; // possibly stalled |
1718 |
|
} |
1719 |
< |
else { // empty - try to descend to find stealer's stealer |
1719 |
> |
else { // descend |
1720 |
|
ForkJoinTask<?> next = stealer.currentJoin; |
1721 |
< |
if (++depth == MAX_HELP_DEPTH || subtask.status < 0 || |
1721 |
> |
if (--depth <= 0 || subtask.status < 0 || |
1722 |
|
next == null || next == subtask) |
1723 |
< |
break outer; // max depth, stale, dead-end, cyclic |
1723 |
> |
break outer; // stale, dead-end, or cyclic |
1724 |
|
subtask = next; |
1725 |
|
j = stealer; |
1726 |
|
break; |
1737 |
|
* @param joiner the joining worker |
1738 |
|
* @param task the task |
1739 |
|
*/ |
1740 |
< |
final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) { |
1740 |
> |
private void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) { |
1741 |
|
WorkQueue[] ws; |
1742 |
< |
int m = runState & SMASK; |
1743 |
< |
if ((ws = workQueues) != null && ws.length > m) { |
1696 |
< |
for (int j = 1; j <= m && task.status >= 0; j += 2) { |
1742 |
> |
if ((ws = workQueues) != null) { |
1743 |
> |
for (int j = 1; j < ws.length && task.status >= 0; j += 2) { |
1744 |
|
WorkQueue q = ws[j]; |
1745 |
|
if (q != null && q.pollFor(task)) { |
1746 |
|
joiner.runSubtask(task); |
1751 |
|
} |
1752 |
|
|
1753 |
|
/** |
1754 |
< |
* Returns a non-empty steal queue, if one is found during a random, |
1755 |
< |
* then cyclic scan, else null. This method must be retried by |
1756 |
< |
* caller if, by the time it tries to use the queue, it is empty. |
1754 |
> |
* Tries to decrement active count (sometimes implicitly) and |
1755 |
> |
* possibly release or create a compensating worker in preparation |
1756 |
> |
* for blocking. Fails on contention or termination. Otherwise, |
1757 |
> |
* adds a new thread if no idle workers are available and either |
1758 |
> |
* pool would become completely starved or: (at least half |
1759 |
> |
* starved, and fewer than 50% spares exist, and there is at least |
1760 |
> |
* one task apparently available). Even though the availablity |
1761 |
> |
* check requires a full scan, it is worthwhile in reducing false |
1762 |
> |
* alarms. |
1763 |
> |
* |
1764 |
> |
* @param task if nonnull, a task being waited for |
1765 |
> |
* @param blocker if nonnull, a blocker being waited for |
1766 |
> |
* @return true if the caller can block, else should recheck and retry |
1767 |
> |
*/ |
1768 |
> |
final boolean tryCompensate(ForkJoinTask<?> task, ManagedBlocker blocker) { |
1769 |
> |
int pc = parallelism, e; |
1770 |
> |
long c = ctl; |
1771 |
> |
WorkQueue[] ws = workQueues; |
1772 |
> |
if ((e = (int)c) >= 0 && ws != null) { |
1773 |
> |
int u, a, ac, hc; |
1774 |
> |
int tc = (short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) + pc; |
1775 |
> |
boolean replace = false; |
1776 |
> |
if ((a = u >> UAC_SHIFT) <= 0) { |
1777 |
> |
if ((ac = a + pc) <= 1) |
1778 |
> |
replace = true; |
1779 |
> |
else if ((e > 0 || (task != null && |
1780 |
> |
ac <= (hc = pc >>> 1) && tc < pc + hc))) { |
1781 |
> |
WorkQueue w; |
1782 |
> |
for (int j = 0; j < ws.length; ++j) { |
1783 |
> |
if ((w = ws[j]) != null && !w.isEmpty()) { |
1784 |
> |
replace = true; |
1785 |
> |
break; // in compensation range and tasks available |
1786 |
> |
} |
1787 |
> |
} |
1788 |
> |
} |
1789 |
> |
} |
1790 |
> |
if ((task == null || task.status >= 0) && // recheck need to block |
1791 |
> |
(blocker == null || !blocker.isReleasable()) && ctl == c) { |
1792 |
> |
if (!replace) { // no compensation |
1793 |
> |
long nc = ((c - AC_UNIT) & AC_MASK) | (c & ~AC_MASK); |
1794 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) |
1795 |
> |
return true; |
1796 |
> |
} |
1797 |
> |
else if (e != 0) { // release an idle worker |
1798 |
> |
WorkQueue w; Thread p; int i; |
1799 |
> |
if ((i = e & SMASK) < ws.length && (w = ws[i]) != null) { |
1800 |
> |
long nc = ((long)(w.nextWait & E_MASK) | |
1801 |
> |
(c & (AC_MASK|TC_MASK))); |
1802 |
> |
if (w.eventCount == (e | INT_SIGN) && |
1803 |
> |
U.compareAndSwapLong(this, CTL, c, nc)) { |
1804 |
> |
w.eventCount = (e + E_SEQ) & E_MASK; |
1805 |
> |
if ((p = w.parker) != null) |
1806 |
> |
U.unpark(p); |
1807 |
> |
return true; |
1808 |
> |
} |
1809 |
> |
} |
1810 |
> |
} |
1811 |
> |
else if (tc < MAX_CAP) { // create replacement |
1812 |
> |
long nc = ((c + TC_UNIT) & TC_MASK) | (c & ~TC_MASK); |
1813 |
> |
if (U.compareAndSwapLong(this, CTL, c, nc)) { |
1814 |
> |
addWorker(); |
1815 |
> |
return true; |
1816 |
> |
} |
1817 |
> |
} |
1818 |
> |
} |
1819 |
> |
} |
1820 |
> |
return false; |
1821 |
> |
} |
1822 |
> |
|
1823 |
> |
/** |
1824 |
> |
* Helps and/or blocks until the given task is done |
1825 |
> |
* |
1826 |
> |
* @param joiner the joining worker |
1827 |
> |
* @param task the task |
1828 |
> |
* @return task status on exit |
1829 |
> |
*/ |
1830 |
> |
final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) { |
1831 |
> |
ForkJoinTask<?> prevJoin = joiner.currentJoin; |
1832 |
> |
joiner.currentJoin = task; |
1833 |
> |
long startTime = 0L; |
1834 |
> |
for (int k = 0, s; ; ++k) { |
1835 |
> |
if ((joiner.isEmpty() ? // try to help |
1836 |
> |
!tryHelpStealer(joiner, task) : |
1837 |
> |
!joiner.tryRemoveAndExec(task))) { |
1838 |
> |
if (k == 0) { |
1839 |
> |
startTime = System.nanoTime(); |
1840 |
> |
tryPollForAndExec(joiner, task); // check uncommon case |
1841 |
> |
} |
1842 |
> |
else if ((k & (MAX_HELP - 1)) == 0 && |
1843 |
> |
System.nanoTime() - startTime >= COMPENSATION_DELAY && |
1844 |
> |
tryCompensate(task, null)) { |
1845 |
> |
if (task.trySetSignal() && task.status >= 0) { |
1846 |
> |
synchronized (task) { |
1847 |
> |
if (task.status >= 0) { |
1848 |
> |
try { // see ForkJoinTask |
1849 |
> |
task.wait(); // for explanation |
1850 |
> |
} catch (InterruptedException ie) { |
1851 |
> |
} |
1852 |
> |
} |
1853 |
> |
else |
1854 |
> |
task.notifyAll(); |
1855 |
> |
} |
1856 |
> |
} |
1857 |
> |
long c; // re-activate |
1858 |
> |
do {} while (!U.compareAndSwapLong |
1859 |
> |
(this, CTL, c = ctl, c + AC_UNIT)); |
1860 |
> |
} |
1861 |
> |
} |
1862 |
> |
if ((s = task.status) < 0) { |
1863 |
> |
joiner.currentJoin = prevJoin; |
1864 |
> |
return s; |
1865 |
> |
} |
1866 |
> |
else if ((k & (MAX_HELP - 1)) == MAX_HELP >>> 1) |
1867 |
> |
Thread.yield(); // for politeness |
1868 |
> |
} |
1869 |
> |
} |
1870 |
> |
|
1871 |
> |
/** |
1872 |
> |
* Stripped-down variant of awaitJoin used by timed joins. Tries |
1873 |
> |
* to help join only while there is continuous progress. (Caller |
1874 |
> |
* will then enter a timed wait.) |
1875 |
> |
* |
1876 |
> |
* @param joiner the joining worker |
1877 |
> |
* @param task the task |
1878 |
> |
* @return task status on exit |
1879 |
> |
*/ |
1880 |
> |
final int helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) { |
1881 |
> |
int s; |
1882 |
> |
while ((s = task.status) >= 0 && |
1883 |
> |
(joiner.isEmpty() ? |
1884 |
> |
tryHelpStealer(joiner, task) : |
1885 |
> |
joiner.tryRemoveAndExec(task))) |
1886 |
> |
; |
1887 |
> |
return s; |
1888 |
> |
} |
1889 |
> |
|
1890 |
> |
/** |
1891 |
> |
* Returns a (probably) non-empty steal queue, if one is found |
1892 |
> |
* during a random, then cyclic scan, else null. This method must |
1893 |
> |
* be retried by caller if, by the time it tries to use the queue, |
1894 |
> |
* it is empty. |
1895 |
|
*/ |
1896 |
|
private WorkQueue findNonEmptyStealQueue(WorkQueue w) { |
1897 |
< |
int r = w.seed; // Same idea as scan(), but ignoring submissions |
1897 |
> |
// Similar to loop in scan(), but ignoring submissions |
1898 |
> |
int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5; |
1899 |
> |
int step = (r >>> 16) | 1; |
1900 |
|
for (WorkQueue[] ws;;) { |
1901 |
< |
int m = runState & SMASK; |
1902 |
< |
if ((ws = workQueues) == null) |
1901 |
> |
int rs = runState, m; |
1902 |
> |
if ((ws = workQueues) == null || (m = ws.length - 1) < 1) |
1903 |
|
return null; |
1904 |
< |
if (ws.length > m) { |
1905 |
< |
WorkQueue q; |
1906 |
< |
for (int k = 0, j = -1 - m;; ++j) { |
1907 |
< |
if (j < 0) { |
1908 |
< |
r ^= r << 13; r ^= r >>> 17; k = r ^= r << 5; |
1909 |
< |
} |
1723 |
< |
else |
1724 |
< |
k += 7; |
1725 |
< |
if ((q = ws[(k | 1) & m]) != null && q.base - q.top < 0) { |
1726 |
< |
w.seed = r; |
1727 |
< |
return q; |
1728 |
< |
} |
1729 |
< |
else if (j - m > m) |
1904 |
> |
for (int j = (m + 1) << 2; ; r += step) { |
1905 |
> |
WorkQueue q = ws[((r << 1) | 1) & m]; |
1906 |
> |
if (q != null && !q.isEmpty()) |
1907 |
> |
return q; |
1908 |
> |
else if (--j < 0) { |
1909 |
> |
if (runState == rs) |
1910 |
|
return null; |
1911 |
+ |
break; |
1912 |
|
} |
1913 |
|
} |
1914 |
|
} |
1922 |
|
*/ |
1923 |
|
final void helpQuiescePool(WorkQueue w) { |
1924 |
|
for (boolean active = true;;) { |
1925 |
< |
w.runLocalTasks(); // exhaust local queue |
1925 |
> |
if (w.base - w.top < 0) |
1926 |
> |
w.runLocalTasks(); // exhaust local queue |
1927 |
|
WorkQueue q = findNonEmptyStealQueue(w); |
1928 |
|
if (q != null) { |
1929 |
< |
ForkJoinTask<?> t; |
1929 |
> |
ForkJoinTask<?> t; int b; |
1930 |
|
if (!active) { // re-establish active count |
1931 |
|
long c; |
1932 |
|
active = true; |
1933 |
|
do {} while (!U.compareAndSwapLong |
1934 |
|
(this, CTL, c = ctl, c + AC_UNIT)); |
1935 |
|
} |
1936 |
< |
if ((t = q.poll()) != null) |
1936 |
> |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
1937 |
|
w.runSubtask(t); |
1938 |
|
} |
1939 |
|
else { |
1961 |
|
*/ |
1962 |
|
final ForkJoinTask<?> nextTaskFor(WorkQueue w) { |
1963 |
|
for (ForkJoinTask<?> t;;) { |
1964 |
< |
WorkQueue q; |
1964 |
> |
WorkQueue q; int b; |
1965 |
|
if ((t = w.nextLocalTask()) != null) |
1966 |
|
return t; |
1967 |
|
if ((q = findNonEmptyStealQueue(w)) == null) |
1968 |
|
return null; |
1969 |
< |
if ((t = q.poll()) != null) |
1969 |
> |
if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null) |
1970 |
|
return t; |
1971 |
|
} |
1972 |
|
} |
2141 |
|
checkPermission(); |
2142 |
|
if (factory == null) |
2143 |
|
throw new NullPointerException(); |
2144 |
< |
if (parallelism <= 0 || parallelism > POOL_MAX) |
2144 |
> |
if (parallelism <= 0 || parallelism > MAX_CAP) |
2145 |
|
throw new IllegalArgumentException(); |
2146 |
|
this.parallelism = parallelism; |
2147 |
|
this.factory = factory; |
2148 |
|
this.ueh = handler; |
2149 |
|
this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE; |
1968 |
– |
this.growHints = 1; |
2150 |
|
long np = (long)(-parallelism); // offset ctl counts |
2151 |
|
this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); |
2152 |
< |
// initialize workQueues array with room for 2*parallelism if possible |
2153 |
< |
int n = parallelism << 1; |
2154 |
< |
if (n >= POOL_MAX) |
2155 |
< |
n = POOL_MAX; |
2156 |
< |
else { // See Hackers Delight, sec 3.2, where n < (1 << 16) |
2157 |
< |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; |
1977 |
< |
} |
1978 |
< |
this.workQueues = new WorkQueue[(n + 1) << 1]; // #slots = 2 * #workers |
2152 |
> |
// Use nearest power 2 for workQueues size. See Hackers Delight sec 3.2. |
2153 |
> |
int n = parallelism - 1; |
2154 |
> |
n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16; |
2155 |
> |
int size = (n + 1) << 1; // #slots = 2*#workers |
2156 |
> |
this.submitMask = size - 1; // room for max # of submit queues |
2157 |
> |
this.workQueues = new WorkQueue[size]; |
2158 |
|
this.termination = (this.lock = new Mutex()).newCondition(); |
2159 |
|
this.stealCount = new AtomicLong(); |
2160 |
|
this.nextWorkerNumber = new AtomicInteger(); |
2161 |
+ |
int pn = poolNumberGenerator.incrementAndGet(); |
2162 |
|
StringBuilder sb = new StringBuilder("ForkJoinPool-"); |
2163 |
< |
sb.append(poolNumberGenerator.incrementAndGet()); |
2163 |
> |
sb.append(Integer.toString(pn)); |
2164 |
|
sb.append("-worker-"); |
2165 |
|
this.workerNamePrefix = sb.toString(); |
2166 |
+ |
lock.lock(); |
2167 |
+ |
this.runState = 1; // set init flag |
2168 |
+ |
lock.unlock(); |
2169 |
|
} |
2170 |
|
|
2171 |
|
// Execution methods |
2187 |
|
* scheduled for execution |
2188 |
|
*/ |
2189 |
|
public <T> T invoke(ForkJoinTask<T> task) { |
2190 |
+ |
if (task == null) |
2191 |
+ |
throw new NullPointerException(); |
2192 |
|
doSubmit(task); |
2193 |
|
return task.join(); |
2194 |
|
} |
2202 |
|
* scheduled for execution |
2203 |
|
*/ |
2204 |
|
public void execute(ForkJoinTask<?> task) { |
2205 |
+ |
if (task == null) |
2206 |
+ |
throw new NullPointerException(); |
2207 |
|
doSubmit(task); |
2208 |
|
} |
2209 |
|
|
2221 |
|
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2222 |
|
job = (ForkJoinTask<?>) task; |
2223 |
|
else |
2224 |
< |
job = ForkJoinTask.adapt(task, null); |
2224 |
> |
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2225 |
|
doSubmit(job); |
2226 |
|
} |
2227 |
|
|
2235 |
|
* scheduled for execution |
2236 |
|
*/ |
2237 |
|
public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { |
2238 |
+ |
if (task == null) |
2239 |
+ |
throw new NullPointerException(); |
2240 |
|
doSubmit(task); |
2241 |
|
return task; |
2242 |
|
} |
2247 |
|
* scheduled for execution |
2248 |
|
*/ |
2249 |
|
public <T> ForkJoinTask<T> submit(Callable<T> task) { |
2250 |
< |
if (task == null) |
2062 |
< |
throw new NullPointerException(); |
2063 |
< |
ForkJoinTask<T> job = ForkJoinTask.adapt(task); |
2250 |
> |
ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task); |
2251 |
|
doSubmit(job); |
2252 |
|
return job; |
2253 |
|
} |
2258 |
|
* scheduled for execution |
2259 |
|
*/ |
2260 |
|
public <T> ForkJoinTask<T> submit(Runnable task, T result) { |
2261 |
< |
if (task == null) |
2075 |
< |
throw new NullPointerException(); |
2076 |
< |
ForkJoinTask<T> job = ForkJoinTask.adapt(task, result); |
2261 |
> |
ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result); |
2262 |
|
doSubmit(job); |
2263 |
|
return job; |
2264 |
|
} |
2275 |
|
if (task instanceof ForkJoinTask<?>) // avoid re-wrap |
2276 |
|
job = (ForkJoinTask<?>) task; |
2277 |
|
else |
2278 |
< |
job = ForkJoinTask.adapt(task, null); |
2278 |
> |
job = new ForkJoinTask.AdaptedRunnableAction(task); |
2279 |
|
doSubmit(job); |
2280 |
|
return job; |
2281 |
|
} |
2297 |
|
boolean done = false; |
2298 |
|
try { |
2299 |
|
for (Callable<T> t : tasks) { |
2300 |
< |
ForkJoinTask<T> f = ForkJoinTask.adapt(t); |
2300 |
> |
ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t); |
2301 |
|
doSubmit(f); |
2302 |
|
fs.add(f); |
2303 |
|
} |
2483 |
|
WorkQueue[] ws; WorkQueue w; |
2484 |
|
if ((ws = workQueues) != null) { |
2485 |
|
for (int i = 0; i < ws.length; i += 2) { |
2486 |
< |
if ((w = ws[i]) != null && w.queueSize() != 0) |
2486 |
> |
if ((w = ws[i]) != null && !w.isEmpty()) |
2487 |
|
return true; |
2488 |
|
} |
2489 |
|
} |
2797 |
|
ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ? |
2798 |
|
((ForkJoinWorkerThread)t).pool : null); |
2799 |
|
while (!blocker.isReleasable()) { |
2800 |
< |
if (p == null || p.tryCompensate()) { |
2800 |
> |
if (p == null || p.tryCompensate(null, blocker)) { |
2801 |
|
try { |
2802 |
|
do {} while (!blocker.isReleasable() && !blocker.block()); |
2803 |
|
} finally { |
2814 |
|
// implement RunnableFuture. |
2815 |
|
|
2816 |
|
protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) { |
2817 |
< |
return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value); |
2817 |
> |
return new ForkJoinTask.AdaptedRunnable<T>(runnable, value); |
2818 |
|
} |
2819 |
|
|
2820 |
|
protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) { |
2821 |
< |
return (RunnableFuture<T>) ForkJoinTask.adapt(callable); |
2821 |
> |
return new ForkJoinTask.AdaptedCallable<T>(callable); |
2822 |
|
} |
2823 |
|
|
2824 |
|
// Unsafe mechanics |
2825 |
|
private static final sun.misc.Unsafe U; |
2826 |
|
private static final long CTL; |
2827 |
|
private static final long PARKBLOCKER; |
2828 |
+ |
private static final int ABASE; |
2829 |
+ |
private static final int ASHIFT; |
2830 |
|
|
2831 |
|
static { |
2832 |
|
poolNumberGenerator = new AtomicInteger(); |
2833 |
+ |
nextSubmitterSeed = new AtomicInteger(0x55555555); |
2834 |
|
modifyThreadPermission = new RuntimePermission("modifyThread"); |
2835 |
|
defaultForkJoinWorkerThreadFactory = |
2836 |
|
new DefaultForkJoinWorkerThreadFactory(); |
2837 |
|
submitters = new ThreadSubmitter(); |
2838 |
+ |
int s; |
2839 |
|
try { |
2840 |
|
U = getUnsafe(); |
2841 |
|
Class<?> k = ForkJoinPool.class; |
2842 |
+ |
Class<?> ak = ForkJoinTask[].class; |
2843 |
|
CTL = U.objectFieldOffset |
2844 |
|
(k.getDeclaredField("ctl")); |
2845 |
|
Class<?> tk = Thread.class; |
2846 |
|
PARKBLOCKER = U.objectFieldOffset |
2847 |
|
(tk.getDeclaredField("parkBlocker")); |
2848 |
+ |
ABASE = U.arrayBaseOffset(ak); |
2849 |
+ |
s = U.arrayIndexScale(ak); |
2850 |
|
} catch (Exception e) { |
2851 |
|
throw new Error(e); |
2852 |
|
} |
2853 |
+ |
if ((s & (s-1)) != 0) |
2854 |
+ |
throw new Error("data type scale not a power of two"); |
2855 |
+ |
ASHIFT = 31 - Integer.numberOfLeadingZeros(s); |
2856 |
|
} |
2857 |
|
|
2858 |
|
/** |