--- jsr166/src/jsr166y/Phaser.java 2008/09/07 11:24:26 1.5 +++ jsr166/src/jsr166y/Phaser.java 2009/03/19 05:10:42 1.12 @@ -5,6 +5,7 @@ */ package jsr166y; + import java.util.concurrent.*; import java.util.concurrent.atomic.*; import java.util.concurrent.locks.LockSupport; @@ -13,9 +14,9 @@ import java.lang.reflect.*; /** * A reusable synchronization barrier, similar in functionality to a - * {@link java.util.concurrent.CyclicBarrier} and {@link - * java.util.concurrent.CountDownLatch} but supporting more flexible - * usage. + * {@link java.util.concurrent.CyclicBarrier CyclicBarrier} and + * {@link java.util.concurrent.CountDownLatch CountDownLatch} + * but supporting more flexible usage. * * * *

Sample usages: * - *

A Phaser may be used instead of a CountdownLatch to control + *

A Phaser may be used instead of a {@code CountDownLatch} to control * a one-shot action serving a variable number of parties. The typical * idiom is for the method setting this up to first register, then * start the actions, then deregister, as in: @@ -102,15 +106,18 @@ import java.lang.reflect.*; * } * }.start(); * } + * + * doSomethingOnBehalfOfWorkers(); * phaser.arrive(); // allow threads to start - * int p = phaser.arriveAndDeregister(); // deregister self + * int p = phaser.arriveAndDeregister(); // deregister self ... + * p = phaser.awaitAdvance(p); // ... and await arrival * otherActions(); // do other things while tasks execute - * phaser.awaitAdvance(p); // wait for all tasks to arrive + * phaser.awaitAdvance(p); // await final completion * } * * *

One way to cause a set of threads to repeatedly perform actions - * for a given number of iterations is to override onAdvance: + * for a given number of iterations is to override {@code onAdvance}: * *

  *  void startTasks(List<Runnable> list, final int iterations) {
@@ -160,7 +167,7 @@ import java.lang.reflect.*;
  *  build(new Task[n], 0, n, new Phaser());
  * 
* - * The best value of TASKS_PER_PHASER depends mainly on + * The best value of {@code TASKS_PER_PHASER} depends mainly on * expected barrier synchronization rates. A value as low as four may * be appropriate for extremely small per-barrier task bodies (thus * high rates), or up to hundreds for extremely large ones. @@ -192,24 +199,24 @@ public class Phaser { * However, to efficiently maintain atomicity, these values are * packed into a single (atomic) long. Termination uses the sign * bit of 32 bit representation of phase, so phase is set to -1 on - * termination. Good performace relies on keeping state decoding + * termination. Good performance relies on keeping state decoding * and encoding simple, and keeping race windows short. * * Note: there are some cheats in arrive() that rely on unarrived - * being lowest 16 bits. + * count being lowest 16 bits. */ private volatile long state; private static final int ushortBits = 16; - private static final int ushortMask = (1 << ushortBits) - 1; - private static final int phaseMask = 0x7fffffff; + private static final int ushortMask = 0xffff; + private static final int phaseMask = 0x7fffffff; private static int unarrivedOf(long s) { return (int)(s & ushortMask); } private static int partiesOf(long s) { - return (int)(s & (ushortMask << 16)) >>> 16; + return ((int)s) >>> 16; } private static int phaseOf(long s) { @@ -221,17 +228,21 @@ public class Phaser { } private static long stateFor(int phase, int parties, int unarrived) { - return (((long)phase) << 32) | ((parties << 16) | unarrived); + return ((((long)phase) << 32) | (((long)parties) << 16) | + (long)unarrived); } private static long trippedStateFor(int phase, int parties) { - return (((long)phase) << 32) | ((parties << 16) | parties); + long lp = (long)parties; + return (((long)phase) << 32) | (lp << 16) | lp; } - private static IllegalStateException badBounds(int parties, int unarrived) { - return new IllegalStateException - ("Attempt to set " + unarrived + - " unarrived of " + parties + " parties"); + /** + * Returns message string for bad bounds exceptions + */ + private static String badBounds(int parties, int unarrived) { + return ("Attempt to set " + unarrived + + " unarrived of " + parties + " parties"); } /** @@ -248,7 +259,7 @@ public class Phaser { // Wait queues /** - * Heads of Treiber stacks waiting for nonFJ threads. To eliminate + * Heads of Treiber stacks for waiting threads. To eliminate * contention while releasing some threads while adding others, we * use two of them, alternating across even and odd phases. */ @@ -292,7 +303,8 @@ public class Phaser { /** * Creates a new Phaser without any initially registered parties, - * initial phase number 0, and no parent. + * initial phase number 0, and no parent. Any thread using this + * Phaser will need to first register for it. */ public Phaser() { this(null); @@ -390,7 +402,7 @@ public class Phaser { if (phase < 0) break; if (parties > ushortMask || unarrived > ushortMask) - throw badBounds(parties, unarrived); + throw new IllegalStateException(badBounds(parties, unarrived)); if (phase == phaseOf(root.state) && casState(s, stateFor(phase, parties, unarrived))) break; @@ -412,6 +424,8 @@ public class Phaser { for (;;) { long s = state; phase = phaseOf(s); + if (phase < 0) + break; int parties = partiesOf(s); int unarrived = unarrivedOf(s) - 1; if (unarrived > 0) { // Not the last arrival @@ -437,12 +451,10 @@ public class Phaser { } } } - else if (phase < 0) // Don't throw exception if terminated - break; else if (phase != phaseOf(root.state)) // or if unreconciled reconcileState(); else - throw badBounds(parties, unarrived); + throw new IllegalStateException(badBounds(parties, unarrived)); } return phase; } @@ -466,6 +478,8 @@ public class Phaser { for (;;) { long s = state; phase = phaseOf(s); + if (phase < 0) + break; int parties = partiesOf(s) - 1; int unarrived = unarrivedOf(s) - 1; if (parties >= 0) { @@ -491,21 +505,19 @@ public class Phaser { } continue; } - if (phase < 0) - break; if (par != null && phase != phaseOf(root.state)) { reconcileState(); continue; } } - throw badBounds(parties, unarrived); + throw new IllegalStateException(badBounds(parties, unarrived)); } return phase; } /** * Arrives at the barrier and awaits others. Equivalent in effect - * to awaitAdvance(arrive()). If you instead need to + * to {@code awaitAdvance(arrive())}. If you instead need to * await with interruption of timeout, and/or deregister upon * arrival, you can arrange them using analogous constructions. * @return the phase on entry to this method @@ -530,7 +542,7 @@ public class Phaser { int p = phaseOf(s); if (p != phase) return p; - if (unarrivedOf(s) == 0) + if (unarrivedOf(s) == 0 && parent != null) parent.awaitAdvance(phase); // Fall here even if parent waited, to reconcile and help release return untimedWait(phase); @@ -538,21 +550,22 @@ public class Phaser { /** * Awaits the phase of the barrier to advance from the given - * value, or returns immediately if argumet is negative or this + * value, or returns immediately if argument is negative or this * barrier is terminated, or throws InterruptedException if * interrupted while waiting. * @param phase the phase on entry to this method * @return the phase on exit from this method * @throws InterruptedException if thread interrupted while waiting */ - public int awaitAdvanceInterruptibly(int phase) throws InterruptedException { + public int awaitAdvanceInterruptibly(int phase) + throws InterruptedException { if (phase < 0) return phase; long s = getReconciledState(); int p = phaseOf(s); if (p != phase) return p; - if (unarrivedOf(s) != 0) + if (unarrivedOf(s) == 0 && parent != null) parent.awaitAdvanceInterruptibly(phase); return interruptibleWait(phase); } @@ -574,7 +587,7 @@ public class Phaser { int p = phaseOf(s); if (p != phase) return p; - if (unarrivedOf(s) == 0) + if (unarrivedOf(s) == 0 && parent != null) parent.awaitAdvanceInterruptibly(phase, timeout, unit); return timedWait(phase, unit.toNanos(timeout)); } @@ -605,7 +618,7 @@ public class Phaser { /** * Returns the current phase number. The maximum phase number is - * Integer.MAX_VALUE, after which it restarts at + * {@code Integer.MAX_VALUE}, after which it restarts at * zero. Upon termination, the phase number is negative. * @return the phase number, or a negative value if terminated */ @@ -614,9 +627,9 @@ public class Phaser { } /** - * Returns true if the current phase number equals the given phase. + * Returns {@code true} if the current phase number equals the given phase. * @param phase the phase - * @return true if the current phase number equals the given phase. + * @return {@code true} if the current phase number equals the given phase */ public final boolean hasPhase(int phase) { return phaseOf(getReconciledState()) == phase; @@ -650,7 +663,7 @@ public class Phaser { /** * Returns the parent of this phaser, or null if none. - * @return the parent of this phaser, or null if none. + * @return the parent of this phaser, or null if none */ public Phaser getParent() { return parent; @@ -659,15 +672,15 @@ public class Phaser { /** * Returns the root ancestor of this phaser, which is the same as * this phaser if it has no parent. - * @return the root ancestor of this phaser. + * @return the root ancestor of this phaser */ public Phaser getRoot() { return root; } /** - * Returns true if this barrier has been terminated. - * @return true if this barrier has been terminated + * Returns {@code true} if this barrier has been terminated. + * @return {@code true} if this barrier has been terminated */ public boolean isTerminated() { return getPhase() < 0; @@ -679,7 +692,7 @@ public class Phaser { * barrier is tripped (and thus all other waiting parties are * dormant). If it returns true, then, rather than advance the * phase number, this barrier will be set to a final termination - * state, and subsequent calls to isTerminated will + * state, and subsequent calls to {@code isTerminated} will * return true. * *

The default version returns true when the number of @@ -690,7 +703,7 @@ public class Phaser { *

You may override this method to perform an action with side * effects visible to participating tasks, but it is in general * only sensible to do so in designs where all parties register - * before any arrive, and all awaitAdvance at each phase. + * before any arrive, and all {@code awaitAdvance} at each phase. * Otherwise, you cannot ensure lack of interference. In * particular, this method may be invoked more than once per * transition if other parties successfully register while the @@ -699,9 +712,8 @@ public class Phaser { * method. * * @param phase the phase number on entering the barrier - * @param registeredParties the current number of registered - * parties. - * @return true if this barrier should terminate + * @param registeredParties the current number of registered parties + * @return {@code true} if this barrier should terminate */ protected boolean onAdvance(int phase, int registeredParties) { return registeredParties <= 0; @@ -710,51 +722,67 @@ public class Phaser { /** * Returns a string identifying this phaser, as well as its * state. The state, in brackets, includes the String {@code - * "phase ="} followed by the phase number, {@code "parties ="} + * "phase = "} followed by the phase number, {@code "parties = "} * followed by the number of registered parties, and {@code - * "arrived ="} followed by the number of arrived parties + * "arrived = "} followed by the number of arrived parties. * * @return a string identifying this barrier, as well as its state */ public String toString() { long s = getReconciledState(); - return super.toString() + "[phase = " + phaseOf(s) + " parties = " + partiesOf(s) + " arrived = " + arrivedOf(s) + "]"; + return super.toString() + + "[phase = " + phaseOf(s) + + " parties = " + partiesOf(s) + + " arrived = " + arrivedOf(s) + "]"; } // methods for waiting - /** The number of CPUs, for spin control */ - static final int NCPUS = Runtime.getRuntime().availableProcessors(); - - /** - * The number of times to spin before blocking in timed waits. - * The value is empirically derived. - */ - static final int maxTimedSpins = (NCPUS < 2)? 0 : 32; - - /** - * The number of times to spin before blocking in untimed waits. - * This is greater than timed value because untimed waits spin - * faster since they don't need to check times on each spin. - */ - static final int maxUntimedSpins = maxTimedSpins * 32; - /** - * The number of nanoseconds for which it is faster to spin - * rather than to use timed park. A rough estimate suffices. + * Wait nodes for Treiber stack representing wait queue */ - static final long spinForTimeoutThreshold = 1000L; - - /** - * Wait nodes for Treiber stack representing wait queue for non-FJ - * tasks. - */ - static final class QNode { - QNode next; + static final class QNode implements ForkJoinPool.ManagedBlocker { + final Phaser phaser; + final int phase; + final long startTime; + final long nanos; + final boolean timed; + final boolean interruptible; + volatile boolean wasInterrupted = false; volatile Thread thread; // nulled to cancel wait - QNode() { + QNode next; + QNode(Phaser phaser, int phase, boolean interruptible, + boolean timed, long startTime, long nanos) { + this.phaser = phaser; + this.phase = phase; + this.timed = timed; + this.interruptible = interruptible; + this.startTime = startTime; + this.nanos = nanos; thread = Thread.currentThread(); } + public boolean isReleasable() { + return (thread == null || + phaser.getPhase() != phase || + (interruptible && wasInterrupted) || + (timed && (nanos - (System.nanoTime() - startTime)) <= 0)); + } + public boolean block() { + if (Thread.interrupted()) { + wasInterrupted = true; + if (interruptible) + return true; + } + if (!timed) + LockSupport.park(this); + else { + long waitTime = nanos - (System.nanoTime() - startTime); + if (waitTime <= 0) + return true; + LockSupport.parkNanos(this, waitTime); + } + return isReleasable(); + } void signal() { Thread t = thread; if (t != null) { @@ -762,6 +790,16 @@ public class Phaser { LockSupport.unpark(t); } } + boolean doWait() { + if (thread != null) { + try { + ForkJoinPool.managedBlock(this, false); + } catch (InterruptedException ie) { + } + } + return wasInterrupted; + } + } /** @@ -777,132 +815,141 @@ public class Phaser { } /** + * Tries to enqueue given node in the appropriate wait queue + * @return true if successful + */ + private boolean tryEnqueue(QNode node) { + AtomicReference head = queueFor(node.phase); + return head.compareAndSet(node.next = head.get(), node); + } + + /** * Enqueues node and waits unless aborted or signalled. + * @return current phase */ private int untimedWait(int phase) { - int spins = maxUntimedSpins; QNode node = null; - boolean interrupted = false; boolean queued = false; + boolean interrupted = false; int p; while ((p = getPhase()) == phase) { - interrupted = Thread.interrupted(); - if (node != null) { - if (!queued) { - AtomicReference head = queueFor(phase); - queued = head.compareAndSet(node.next = head.get(), node); - } - else if (node.thread != null) - LockSupport.park(this); - } - else if (spins <= 0) - node = new QNode(); + if (Thread.interrupted()) + interrupted = true; + else if (node == null) + node = new QNode(this, phase, false, false, 0, 0); + else if (!queued) + queued = tryEnqueue(node); else - --spins; + interrupted = node.doWait(); } if (node != null) node.thread = null; + releaseWaiters(phase); if (interrupted) Thread.currentThread().interrupt(); - releaseWaiters(phase); return p; } /** - * Messier interruptible version + * Interruptible version + * @return current phase */ private int interruptibleWait(int phase) throws InterruptedException { - int spins = maxUntimedSpins; QNode node = null; boolean queued = false; boolean interrupted = false; int p; - while ((p = getPhase()) == phase) { - if (interrupted = Thread.interrupted()) - break; - if (node != null) { - if (!queued) { - AtomicReference head = queueFor(phase); - queued = head.compareAndSet(node.next = head.get(), node); - } - else if (node.thread != null) - LockSupport.park(this); - } - else if (spins <= 0) - node = new QNode(); + while ((p = getPhase()) == phase && !interrupted) { + if (Thread.interrupted()) + interrupted = true; + else if (node == null) + node = new QNode(this, phase, true, false, 0, 0); + else if (!queued) + queued = tryEnqueue(node); else - --spins; + interrupted = node.doWait(); } if (node != null) node.thread = null; + if (p != phase || (p = getPhase()) != phase) + releaseWaiters(phase); if (interrupted) throw new InterruptedException(); - releaseWaiters(phase); return p; } /** - * Even messier timeout version. + * Timeout version. + * @return current phase */ private int timedWait(int phase, long nanos) throws InterruptedException, TimeoutException { + long startTime = System.nanoTime(); + QNode node = null; + boolean queued = false; + boolean interrupted = false; int p; - if ((p = getPhase()) == phase) { - long lastTime = System.nanoTime(); - int spins = maxTimedSpins; - QNode node = null; - boolean queued = false; - boolean interrupted = false; - while ((p = getPhase()) == phase) { - if (interrupted = Thread.interrupted()) - break; - long now = System.nanoTime(); - if ((nanos -= now - lastTime) <= 0) - break; - lastTime = now; - if (node != null) { - if (!queued) { - AtomicReference head = queueFor(phase); - queued = head.compareAndSet(node.next = head.get(), node); - } - else if (node.thread != null && - nanos > spinForTimeoutThreshold) { - LockSupport.parkNanos(this, nanos); - } - } - else if (spins <= 0) - node = new QNode(); - else - --spins; - } - if (node != null) - node.thread = null; - if (interrupted) - throw new InterruptedException(); - if (p == phase && (p = getPhase()) == phase) - throw new TimeoutException(); + while ((p = getPhase()) == phase && !interrupted) { + if (Thread.interrupted()) + interrupted = true; + else if (nanos - (System.nanoTime() - startTime) <= 0) + break; + else if (node == null) + node = new QNode(this, phase, true, true, startTime, nanos); + else if (!queued) + queued = tryEnqueue(node); + else + interrupted = node.doWait(); } - releaseWaiters(phase); + if (node != null) + node.thread = null; + if (p != phase || (p = getPhase()) != phase) + releaseWaiters(phase); + if (interrupted) + throw new InterruptedException(); + if (p == phase) + throw new TimeoutException(); return p; } // Temporary Unsafe mechanics for preliminary release + private static Unsafe getUnsafe() throws Throwable { + try { + return Unsafe.getUnsafe(); + } catch (SecurityException se) { + try { + return java.security.AccessController.doPrivileged + (new java.security.PrivilegedExceptionAction() { + public Unsafe run() throws Exception { + return getUnsafePrivileged(); + }}); + } catch (java.security.PrivilegedActionException e) { + throw e.getCause(); + } + } + } + + private static Unsafe getUnsafePrivileged() + throws NoSuchFieldException, IllegalAccessException { + Field f = Unsafe.class.getDeclaredField("theUnsafe"); + f.setAccessible(true); + return (Unsafe) f.get(null); + } + + private static long fieldOffset(String fieldName) + throws NoSuchFieldException { + return _unsafe.objectFieldOffset + (Phaser.class.getDeclaredField(fieldName)); + } static final Unsafe _unsafe; static final long stateOffset; static { try { - if (Phaser.class.getClassLoader() != null) { - Field f = Unsafe.class.getDeclaredField("theUnsafe"); - f.setAccessible(true); - _unsafe = (Unsafe)f.get(null); - } - else - _unsafe = Unsafe.getUnsafe(); - stateOffset = _unsafe.objectFieldOffset - (Phaser.class.getDeclaredField("state")); - } catch (Exception e) { + _unsafe = getUnsafe(); + stateOffset = fieldOffset("state"); + } catch (Throwable e) { throw new RuntimeException("Could not initialize intrinsics", e); } }