| 13 |
|
/** |
| 14 |
|
* A thread managed by a {@link ForkJoinPool}. This class is |
| 15 |
|
* subclassable solely for the sake of adding functionality -- there |
| 16 |
< |
* are no overridable methods dealing with scheduling or |
| 17 |
< |
* execution. However, you can override initialization and termination |
| 18 |
< |
* methods surrounding the main task processing loop. If you do |
| 19 |
< |
* create such a subclass, you will also need to supply a custom |
| 20 |
< |
* ForkJoinWorkerThreadFactory to use it in a ForkJoinPool. |
| 16 |
> |
* are no overridable methods dealing with scheduling or execution. |
| 17 |
> |
* However, you can override initialization and termination methods |
| 18 |
> |
* surrounding the main task processing loop. If you do create such a |
| 19 |
> |
* subclass, you will also need to supply a custom {@link |
| 20 |
> |
* ForkJoinPool.ForkJoinWorkerThreadFactory} to use it in a {@code |
| 21 |
> |
* ForkJoinPool}. |
| 22 |
|
* |
| 23 |
|
* @since 1.7 |
| 24 |
|
* @author Doug Lea |
| 57 |
|
* considered individually, is not wait-free. One thief cannot |
| 58 |
|
* successfully continue until another in-progress one (or, if |
| 59 |
|
* previously empty, a push) completes. However, in the |
| 60 |
< |
* aggregate, we ensure at least probabilistic non-blockingness. If |
| 61 |
< |
* an attempted steal fails, a thief always chooses a different |
| 62 |
< |
* random victim target to try next. So, in order for one thief to |
| 63 |
< |
* progress, it suffices for any in-progress deq or new push on |
| 64 |
< |
* any empty queue to complete. One reason this works well here is |
| 65 |
< |
* that apparently-nonempty often means soon-to-be-stealable, |
| 66 |
< |
* which gives threads a chance to activate if necessary before |
| 67 |
< |
* stealing (see below). |
| 60 |
> |
* aggregate, we ensure at least probabilistic |
| 61 |
> |
* non-blockingness. If an attempted steal fails, a thief always |
| 62 |
> |
* chooses a different random victim target to try next. So, in |
| 63 |
> |
* order for one thief to progress, it suffices for any |
| 64 |
> |
* in-progress deq or new push on any empty queue to complete. One |
| 65 |
> |
* reason this works well here is that apparently-nonempty often |
| 66 |
> |
* means soon-to-be-stealable, which gives threads a chance to |
| 67 |
> |
* activate if necessary before stealing (see below). |
| 68 |
> |
* |
| 69 |
> |
* This approach also enables support for "async mode" where local |
| 70 |
> |
* task processing is in FIFO, not LIFO order; simply by using a |
| 71 |
> |
* version of deq rather than pop when locallyFifo is true (as set |
| 72 |
> |
* by the ForkJoinPool). This allows use in message-passing |
| 73 |
> |
* frameworks in which tasks are never joined. |
| 74 |
|
* |
| 75 |
|
* Efficient implementation of this approach currently relies on |
| 76 |
|
* an uncomfortable amount of "Unsafe" mechanics. To maintain |
| 80 |
|
* protected by volatile base reads, reads of the queue array and |
| 81 |
|
* its slots do not need volatile load semantics, but writes (in |
| 82 |
|
* push) require store order and CASes (in pop and deq) require |
| 83 |
< |
* (volatile) CAS semantics. Since these combinations aren't |
| 84 |
< |
* supported using ordinary volatiles, the only way to accomplish |
| 85 |
< |
* these efficiently is to use direct Unsafe calls. (Using external |
| 83 |
> |
* (volatile) CAS semantics. (See "Idempotent work stealing" by |
| 84 |
> |
* Michael, Saraswat, and Vechev, PPoPP 2009 |
| 85 |
> |
* http://portal.acm.org/citation.cfm?id=1504186 for an algorithm |
| 86 |
> |
* with similar properties, but without support for nulling |
| 87 |
> |
* slots.) Since these combinations aren't supported using |
| 88 |
> |
* ordinary volatiles, the only way to accomplish these |
| 89 |
> |
* efficiently is to use direct Unsafe calls. (Using external |
| 90 |
|
* AtomicIntegers and AtomicReferenceArrays for the indices and |
| 91 |
|
* array is significantly slower because of memory locality and |
| 92 |
< |
* indirection effects.) Further, performance on most platforms is |
| 93 |
< |
* very sensitive to placement and sizing of the (resizable) queue |
| 94 |
< |
* array. Even though these queues don't usually become all that |
| 95 |
< |
* big, the initial size must be large enough to counteract cache |
| 92 |
> |
* indirection effects.) |
| 93 |
> |
* |
| 94 |
> |
* Further, performance on most platforms is very sensitive to |
| 95 |
> |
* placement and sizing of the (resizable) queue array. Even |
| 96 |
> |
* though these queues don't usually become all that big, the |
| 97 |
> |
* initial size must be large enough to counteract cache |
| 98 |
|
* contention effects across multiple queues (especially in the |
| 99 |
|
* presence of GC cardmarking). Also, to improve thread-locality, |
| 100 |
|
* queues are currently initialized immediately after the thread |
| 111 |
|
* counter (activeCount) held by the pool. It uses an algorithm |
| 112 |
|
* similar to that in Herlihy and Shavit section 17.6 to cause |
| 113 |
|
* threads to eventually block when all threads declare they are |
| 114 |
< |
* inactive. (See variable "scans".) For this to work, threads |
| 115 |
< |
* must be declared active when executing tasks, and before |
| 116 |
< |
* stealing a task. They must be inactive before blocking on the |
| 117 |
< |
* Pool Barrier (awaiting a new submission or other Pool |
| 118 |
< |
* event). In between, there is some free play which we take |
| 119 |
< |
* advantage of to avoid contention and rapid flickering of the |
| 120 |
< |
* global activeCount: If inactive, we activate only if a victim |
| 121 |
< |
* queue appears to be nonempty (see above). Similarly, a thread |
| 122 |
< |
* tries to inactivate only after a full scan of other threads. |
| 123 |
< |
* The net effect is that contention on activeCount is rarely a |
| 124 |
< |
* measurable performance issue. (There are also a few other cases |
| 125 |
< |
* where we scan for work rather than retry/block upon |
| 113 |
< |
* contention.) |
| 114 |
> |
* inactive. For this to work, threads must be declared active |
| 115 |
> |
* when executing tasks, and before stealing a task. They must be |
| 116 |
> |
* inactive before blocking on the Pool Barrier (awaiting a new |
| 117 |
> |
* submission or other Pool event). In between, there is some free |
| 118 |
> |
* play which we take advantage of to avoid contention and rapid |
| 119 |
> |
* flickering of the global activeCount: If inactive, we activate |
| 120 |
> |
* only if a victim queue appears to be nonempty (see above). |
| 121 |
> |
* Similarly, a thread tries to inactivate only after a full scan |
| 122 |
> |
* of other threads. The net effect is that contention on |
| 123 |
> |
* activeCount is rarely a measurable performance issue. (There |
| 124 |
> |
* are also a few other cases where we scan for work rather than |
| 125 |
> |
* retry/block upon contention.) |
| 126 |
|
* |
| 127 |
|
* 3. Selection control. We maintain policy of always choosing to |
| 128 |
|
* run local tasks rather than stealing, and always trying to |
| 277 |
|
final boolean shutdownNow() { return transitionRunStateTo(TERMINATING); } |
| 278 |
|
|
| 279 |
|
/** |
| 280 |
< |
* Transitions to at least the given state. Returns true if not |
| 281 |
< |
* already at least at given state. |
| 280 |
> |
* Transitions to at least the given state. |
| 281 |
> |
* |
| 282 |
> |
* @return {@code true} if not already at least at given state |
| 283 |
|
*/ |
| 284 |
|
private boolean transitionRunStateTo(int state) { |
| 285 |
|
for (;;) { |
| 321 |
|
* one. Marsaglia xor-shift is cheap and works well. |
| 322 |
|
*/ |
| 323 |
|
private static int xorShift(int r) { |
| 324 |
< |
r ^= r << 1; |
| 325 |
< |
r ^= r >>> 3; |
| 326 |
< |
r ^= r << 10; |
| 314 |
< |
return r; |
| 324 |
> |
r ^= (r << 13); |
| 325 |
> |
r ^= (r >>> 17); |
| 326 |
> |
return r ^ (r << 5); |
| 327 |
|
} |
| 328 |
|
|
| 329 |
|
// Lifecycle methods |
| 384 |
|
* {@code super.onTermination} at the end of the overridden method. |
| 385 |
|
* |
| 386 |
|
* @param exception the exception causing this thread to abort due |
| 387 |
< |
* to an unrecoverable error, or null if completed normally |
| 387 |
> |
* to an unrecoverable error, or {@code null} if completed normally |
| 388 |
|
*/ |
| 389 |
|
protected void onTermination(Throwable exception) { |
| 390 |
|
// Execute remaining local tasks unless aborting or terminating |
| 391 |
< |
while (exception == null && !pool.isTerminating() && base != sp) { |
| 391 |
> |
while (exception == null && pool.isProcessingTasks() && base != sp) { |
| 392 |
|
try { |
| 393 |
|
ForkJoinTask<?> t = popTask(); |
| 394 |
|
if (t != null) |
| 415 |
|
|
| 416 |
|
// Intrinsics-based support for queue operations. |
| 417 |
|
|
| 418 |
+ |
private static long slotOffset(int i) { |
| 419 |
+ |
return ((long) i << qShift) + qBase; |
| 420 |
+ |
} |
| 421 |
+ |
|
| 422 |
|
/** |
| 423 |
|
* Adds in store-order the given task at given slot of q to null. |
| 424 |
|
* Caller must ensure q is non-null and index is in range. |
| 425 |
|
*/ |
| 426 |
|
private static void setSlot(ForkJoinTask<?>[] q, int i, |
| 427 |
|
ForkJoinTask<?> t) { |
| 428 |
< |
UNSAFE.putOrderedObject(q, (i << qShift) + qBase, t); |
| 428 |
> |
UNSAFE.putOrderedObject(q, slotOffset(i), t); |
| 429 |
|
} |
| 430 |
|
|
| 431 |
|
/** |
| 434 |
|
*/ |
| 435 |
|
private static boolean casSlotNull(ForkJoinTask<?>[] q, int i, |
| 436 |
|
ForkJoinTask<?> t) { |
| 437 |
< |
return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null); |
| 437 |
> |
return UNSAFE.compareAndSwapObject(q, slotOffset(i), t, null); |
| 438 |
|
} |
| 439 |
|
|
| 440 |
|
/** |
| 485 |
|
} |
| 486 |
|
|
| 487 |
|
/** |
| 488 |
+ |
* Tries to take a task from the base of own queue, activating if |
| 489 |
+ |
* necessary, failing only if empty. Called only by current thread. |
| 490 |
+ |
* |
| 491 |
+ |
* @return a task, or null if none |
| 492 |
+ |
*/ |
| 493 |
+ |
final ForkJoinTask<?> locallyDeqTask() { |
| 494 |
+ |
int b; |
| 495 |
+ |
while (sp != (b = base)) { |
| 496 |
+ |
if (tryActivate()) { |
| 497 |
+ |
ForkJoinTask<?>[] q = queue; |
| 498 |
+ |
int i = (q.length - 1) & b; |
| 499 |
+ |
ForkJoinTask<?> t = q[i]; |
| 500 |
+ |
if (t != null && casSlotNull(q, i, t)) { |
| 501 |
+ |
base = b + 1; |
| 502 |
+ |
return t; |
| 503 |
+ |
} |
| 504 |
+ |
} |
| 505 |
+ |
} |
| 506 |
+ |
return null; |
| 507 |
+ |
} |
| 508 |
+ |
|
| 509 |
+ |
/** |
| 510 |
|
* Returns a popped task, or null if empty. Ensures active status |
| 511 |
|
* if non-null. Called only by current thread. |
| 512 |
|
*/ |
| 546 |
|
} |
| 547 |
|
|
| 548 |
|
/** |
| 549 |
< |
* Returns next task. |
| 549 |
> |
* Returns next task or null if empty or contended |
| 550 |
|
*/ |
| 551 |
|
final ForkJoinTask<?> peekTask() { |
| 552 |
|
ForkJoinTask<?>[] q = queue; |
| 637 |
|
* @return a task, if available |
| 638 |
|
*/ |
| 639 |
|
final ForkJoinTask<?> pollTask() { |
| 640 |
< |
ForkJoinTask<?> t = locallyFifo ? deqTask() : popTask(); |
| 640 |
> |
ForkJoinTask<?> t = locallyFifo ? locallyDeqTask() : popTask(); |
| 641 |
|
if (t == null && (t = scan()) != null) |
| 642 |
|
++stealCount; |
| 643 |
|
return t; |
| 649 |
|
* @return a task, if available |
| 650 |
|
*/ |
| 651 |
|
final ForkJoinTask<?> pollLocalTask() { |
| 652 |
< |
return locallyFifo ? deqTask() : popTask(); |
| 652 |
> |
return locallyFifo ? locallyDeqTask() : popTask(); |
| 653 |
|
} |
| 654 |
|
|
| 655 |
|
/** |
| 684 |
|
* |
| 685 |
|
* @return the number of tasks drained |
| 686 |
|
*/ |
| 687 |
< |
final int drainTasksTo(Collection<ForkJoinTask<?>> c) { |
| 687 |
> |
final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) { |
| 688 |
|
int n = 0; |
| 689 |
|
ForkJoinTask<?> t; |
| 690 |
|
while (base != sp && (t = deqTask()) != null) { |
| 705 |
|
} |
| 706 |
|
|
| 707 |
|
/** |
| 708 |
< |
* Returns true if at least one worker in the given array appears |
| 709 |
< |
* to have at least one queued task. |
| 708 |
> |
* Returns {@code true} if at least one worker in the given array |
| 709 |
> |
* appears to have at least one queued task. |
| 710 |
|
* |
| 711 |
|
* @param ws array of workers |
| 712 |
|
*/ |
| 769 |
|
do {} while (!tryActivate()); // re-activate on exit |
| 770 |
|
} |
| 771 |
|
|
| 772 |
< |
// Unsafe mechanics for jsr166y 3rd party package. |
| 772 |
> |
// Unsafe mechanics |
| 773 |
> |
|
| 774 |
> |
private static final sun.misc.Unsafe UNSAFE = getUnsafe(); |
| 775 |
> |
private static final long spOffset = |
| 776 |
> |
objectFieldOffset("sp", ForkJoinWorkerThread.class); |
| 777 |
> |
private static final long runStateOffset = |
| 778 |
> |
objectFieldOffset("runState", ForkJoinWorkerThread.class); |
| 779 |
> |
private static final long qBase; |
| 780 |
> |
private static final int qShift; |
| 781 |
> |
|
| 782 |
> |
static { |
| 783 |
> |
qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class); |
| 784 |
> |
int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class); |
| 785 |
> |
if ((s & (s-1)) != 0) |
| 786 |
> |
throw new Error("data type scale not a power of two"); |
| 787 |
> |
qShift = 31 - Integer.numberOfLeadingZeros(s); |
| 788 |
> |
} |
| 789 |
> |
|
| 790 |
> |
private static long objectFieldOffset(String field, Class<?> klazz) { |
| 791 |
> |
try { |
| 792 |
> |
return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field)); |
| 793 |
> |
} catch (NoSuchFieldException e) { |
| 794 |
> |
// Convert Exception to corresponding Error |
| 795 |
> |
NoSuchFieldError error = new NoSuchFieldError(field); |
| 796 |
> |
error.initCause(e); |
| 797 |
> |
throw error; |
| 798 |
> |
} |
| 799 |
> |
} |
| 800 |
> |
|
| 801 |
> |
/** |
| 802 |
> |
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
| 803 |
> |
* Replace with a simple call to Unsafe.getUnsafe when integrating |
| 804 |
> |
* into a jdk. |
| 805 |
> |
* |
| 806 |
> |
* @return a sun.misc.Unsafe |
| 807 |
> |
*/ |
| 808 |
|
private static sun.misc.Unsafe getUnsafe() { |
| 809 |
|
try { |
| 810 |
|
return sun.misc.Unsafe.getUnsafe(); |
| 811 |
|
} catch (SecurityException se) { |
| 812 |
|
try { |
| 813 |
|
return java.security.AccessController.doPrivileged |
| 814 |
< |
(new java.security.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
| 814 |
> |
(new java.security |
| 815 |
> |
.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
| 816 |
|
public sun.misc.Unsafe run() throws Exception { |
| 817 |
< |
return getUnsafeByReflection(); |
| 817 |
> |
java.lang.reflect.Field f = sun.misc |
| 818 |
> |
.Unsafe.class.getDeclaredField("theUnsafe"); |
| 819 |
> |
f.setAccessible(true); |
| 820 |
> |
return (sun.misc.Unsafe) f.get(null); |
| 821 |
|
}}); |
| 822 |
|
} catch (java.security.PrivilegedActionException e) { |
| 823 |
|
throw new RuntimeException("Could not initialize intrinsics", |
| 825 |
|
} |
| 826 |
|
} |
| 827 |
|
} |
| 751 |
– |
|
| 752 |
– |
private static sun.misc.Unsafe getUnsafeByReflection() |
| 753 |
– |
throws NoSuchFieldException, IllegalAccessException { |
| 754 |
– |
java.lang.reflect.Field f = |
| 755 |
– |
sun.misc.Unsafe.class.getDeclaredField("theUnsafe"); |
| 756 |
– |
f.setAccessible(true); |
| 757 |
– |
return (sun.misc.Unsafe) f.get(null); |
| 758 |
– |
} |
| 759 |
– |
|
| 760 |
– |
private static long fieldOffset(String fieldName, Class<?> klazz) { |
| 761 |
– |
try { |
| 762 |
– |
return UNSAFE.objectFieldOffset(klazz.getDeclaredField(fieldName)); |
| 763 |
– |
} catch (NoSuchFieldException e) { |
| 764 |
– |
// Convert Exception to Error |
| 765 |
– |
NoSuchFieldError error = new NoSuchFieldError(fieldName); |
| 766 |
– |
error.initCause(e); |
| 767 |
– |
throw error; |
| 768 |
– |
} |
| 769 |
– |
} |
| 770 |
– |
|
| 771 |
– |
private static final sun.misc.Unsafe UNSAFE = getUnsafe(); |
| 772 |
– |
static final long baseOffset = |
| 773 |
– |
fieldOffset("base", ForkJoinWorkerThread.class); |
| 774 |
– |
static final long spOffset = |
| 775 |
– |
fieldOffset("sp", ForkJoinWorkerThread.class); |
| 776 |
– |
static final long runStateOffset = |
| 777 |
– |
fieldOffset("runState", ForkJoinWorkerThread.class); |
| 778 |
– |
static final long qBase; |
| 779 |
– |
static final int qShift; |
| 780 |
– |
|
| 781 |
– |
static { |
| 782 |
– |
qBase = UNSAFE.arrayBaseOffset(ForkJoinTask[].class); |
| 783 |
– |
int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class); |
| 784 |
– |
if ((s & (s-1)) != 0) |
| 785 |
– |
throw new Error("data type scale not a power of two"); |
| 786 |
– |
qShift = 31 - Integer.numberOfLeadingZeros(s); |
| 787 |
– |
} |
| 828 |
|
} |
