ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/jsr166/jsr166/src/jsr166y/Phaser.java
(Generate patch)

Comparing jsr166/src/jsr166y/Phaser.java (file contents):
Revision 1.2 by jsr166, Fri Jul 25 18:10:41 2008 UTC vs.
Revision 1.68 by dl, Sat Dec 4 15:25:08 2010 UTC

# Line 5 | Line 5
5   */
6  
7   package jsr166y;
8 < import jsr166y.forkjoin.*;
9 < import java.util.concurrent.*;
10 < import java.util.concurrent.atomic.*;
8 >
9 > import java.util.concurrent.TimeUnit;
10 > import java.util.concurrent.TimeoutException;
11 > import java.util.concurrent.atomic.AtomicReference;
12   import java.util.concurrent.locks.LockSupport;
13  
14   /**
15 < * A reusable synchronization barrier, similar in functionality to a
16 < * {@link java.util.concurrent.CyclicBarrier}, but supporting more
17 < * flexible usage.
15 > * A reusable synchronization barrier, similar in functionality to
16 > * {@link java.util.concurrent.CyclicBarrier CyclicBarrier} and
17 > * {@link java.util.concurrent.CountDownLatch CountDownLatch}
18 > * but supporting more flexible usage.
19   *
20 < * <ul>
20 > * <p> <b>Registration.</b> Unlike the case for other barriers, the
21 > * number of parties <em>registered</em> to synchronize on a phaser
22 > * may vary over time.  Tasks may be registered at any time (using
23 > * methods {@link #register}, {@link #bulkRegister}, or forms of
24 > * constructors establishing initial numbers of parties), and
25 > * optionally deregistered upon any arrival (using {@link
26 > * #arriveAndDeregister}).  As is the case with most basic
27 > * synchronization constructs, registration and deregistration affect
28 > * only internal counts; they do not establish any further internal
29 > * bookkeeping, so tasks cannot query whether they are registered.
30 > * (However, you can introduce such bookkeeping by subclassing this
31 > * class.)
32   *
33 < * <li> The number of parties synchronizing on the barrier may vary
34 < * over time.  A task may register to be a party in a barrier at any
35 < * time, and may deregister upon arriving at the barrier.  As is the
36 < * case with most basic synchronization constructs, registration
37 < * and deregistration affect only internal counts; they do not
38 < * establish any further internal bookkeeping, so tasks cannot query
39 < * whether they are registered.
40 < *
41 < * <li> Each generation has an associated phase value, starting at
42 < * zero, and advancing when all parties reach the barrier (wrapping
43 < * around to zero after reaching <tt>Integer.MAX_VALUE</tt>).
31 < *
32 < * <li> Like a CyclicBarrier, a Phaser may be repeatedly awaited.
33 < * Method <tt>arriveAndAwaitAdvance</tt> has effect analogous to
34 < * <tt>CyclicBarrier.await</tt>.  However, Phasers separate two
35 < * aspects of coordination, that may be invoked independently:
33 > * <p> <b>Synchronization.</b> Like a {@code CyclicBarrier}, a {@code
34 > * Phaser} may be repeatedly awaited.  Method {@link
35 > * #arriveAndAwaitAdvance} has effect analogous to {@link
36 > * java.util.concurrent.CyclicBarrier#await CyclicBarrier.await}. Each
37 > * generation of a phaser has an associated phase number. The phase
38 > * number starts at zero, and advances when all parties arrive at the
39 > * phaser, wrapping around to zero after reaching {@code
40 > * Integer.MAX_VALUE}. The use of phase numbers enables independent
41 > * control of actions upon arrival at a phaser and upon awaiting
42 > * others, via two kinds of methods that may be invoked by any
43 > * registered party:
44   *
45   * <ul>
46   *
47 < *   <li> Arriving at a barrier. Methods <tt>arrive</tt> and
48 < *       <tt>arriveAndDeregister</tt> do not block, but return
49 < *       the phase value on entry to the method.
50 < *
51 < *   <li> Awaiting others. Method <tt>awaitAdvance</tt> requires an
52 < *       argument indicating the entry phase, and returns when the
53 < *       barrier advances to a new phase.
47 > *   <li> <b>Arrival.</b> Methods {@link #arrive} and
48 > *       {@link #arriveAndDeregister} record arrival.  These methods
49 > *       do not block, but return an associated <em>arrival phase
50 > *       number</em>; that is, the phase number of the phaser to which
51 > *       the arrival applied. When the final party for a given phase
52 > *       arrives, an optional action is performed and the phase
53 > *       advances.  These actions are performed by the party
54 > *       triggering a phase advance, and are arranged by overriding
55 > *       method {@link #onAdvance(int, int)}, which also controls
56 > *       termination. Overriding this method is similar to, but more
57 > *       flexible than, providing a barrier action to a {@code
58 > *       CyclicBarrier}.
59 > *
60 > *   <li> <b>Waiting.</b> Method {@link #awaitAdvance} requires an
61 > *       argument indicating an arrival phase number, and returns when
62 > *       the phaser advances to (or is already at) a different phase.
63 > *       Unlike similar constructions using {@code CyclicBarrier},
64 > *       method {@code awaitAdvance} continues to wait even if the
65 > *       waiting thread is interrupted. Interruptible and timeout
66 > *       versions are also available, but exceptions encountered while
67 > *       tasks wait interruptibly or with timeout do not change the
68 > *       state of the phaser. If necessary, you can perform any
69 > *       associated recovery within handlers of those exceptions,
70 > *       often after invoking {@code forceTermination}.  Phasers may
71 > *       also be used by tasks executing in a {@link ForkJoinPool},
72 > *       which will ensure sufficient parallelism to execute tasks
73 > *       when others are blocked waiting for a phase to advance.
74 > *
75   * </ul>
76   *
77 + * <p> <b>Termination.</b> A phaser may enter a <em>termination</em>
78 + * state, that may be checked using method {@link #isTerminated}. Upon
79 + * termination, all synchronization methods immediately return without
80 + * waiting for advance, as indicated by a negative return
81 + * value. Similarly, attempts to register upon termination have no
82 + * effect.  Termination is triggered when an invocation of {@code
83 + * onAdvance} returns {@code true}. The default implementation returns
84 + * {@code true} if a deregistration has caused the number of
85 + * registered parties to become zero.  As illustrated below, when
86 + * phasers control actions with a fixed number of iterations, it is
87 + * often convenient to override this method to cause termination when
88 + * the current phase number reaches a threshold. Method {@link
89 + * #forceTermination} is also available to abruptly release waiting
90 + * threads and allow them to terminate.
91   *
92 < * <li> Barrier actions, performed by the task triggering a phase
93 < * advance while others may be waiting, are arranged by overriding
94 < * method <tt>onAdvance</tt>, that also controls termination.
95 < *
96 < * <li> Phasers may enter a <em>termination</em> state in which all
97 < * await actions immediately return, indicating (via a negative phase
98 < * value) that execution is complete.  Termination is triggered by
56 < * executing the overridable <tt>onAdvance</tt> method that is invoked
57 < * each time the barrier is tripped. When a Phaser is controlling an
58 < * action with a fixed number of iterations, it is often convenient to
59 < * override this method to cause termination when the current phase
60 < * number reaches a threshold.  Method <tt>forceTermination</tt> is
61 < * also available to assist recovery actions upon failure.
62 < *
63 < * <li> Unlike most synchronizers, a Phaser may also be used with
64 < * ForkJoinTasks (as well as plain threads).
65 < *
66 < * <li> By default, <tt>awaitAdvance</tt> continues to wait even if
67 < * the current thread is interrupted. And unlike the case in
68 < * CyclicBarriers, exceptions encountered while tasks wait
69 < * interruptibly or with timeout do not change the state of the
70 < * barrier. If necessary, you can perform any associated recovery
71 < * within handlers of those exceptions.
92 > * <p> <b>Tiering.</b> Phasers may be <em>tiered</em> (i.e.,
93 > * constructed in tree structures) to reduce contention. Phasers with
94 > * large numbers of parties that would otherwise experience heavy
95 > * synchronization contention costs may instead be set up so that
96 > * groups of sub-phasers share a common parent.  This may greatly
97 > * increase throughput even though it incurs greater per-operation
98 > * overhead.
99   *
100 < * </ul>
100 > * <p>In a tree of tiered phasers, registration and deregistration of
101 > * child phasers with their parent are managed automatically.
102 > * Whenever the number of registered parties of a child phaser becomes
103 > * non-zero (as established in the {@link #Phaser(Phaser,int)}
104 > * constructor, {@link #register}, or {@link #bulkRegister}), the
105 > * child phaser is registered with its parent.  Whenever the number of
106 > * registered parties becomes zero as the result of an invocation of
107 > * {@link #arriveAndDeregister}, the child phaser is deregistered
108 > * from its parent.
109 > *
110 > * <p><b>Monitoring.</b> While synchronization methods may be invoked
111 > * only by registered parties, the current state of a phaser may be
112 > * monitored by any caller.  At any given moment there are {@link
113 > * #getRegisteredParties} parties in total, of which {@link
114 > * #getArrivedParties} have arrived at the current phase ({@link
115 > * #getPhase}).  When the remaining ({@link #getUnarrivedParties})
116 > * parties arrive, the phase advances.  The values returned by these
117 > * methods may reflect transient states and so are not in general
118 > * useful for synchronization control.  Method {@link #toString}
119 > * returns snapshots of these state queries in a form convenient for
120 > * informal monitoring.
121 > *
122 > * <p><b>Sample usages:</b>
123 > *
124 > * <p>A {@code Phaser} may be used instead of a {@code CountDownLatch}
125 > * to control a one-shot action serving a variable number of parties.
126 > * The typical idiom is for the method setting this up to first
127 > * register, then start the actions, then deregister, as in:
128 > *
129 > *  <pre> {@code
130 > * void runTasks(List<Runnable> tasks) {
131 > *   final Phaser phaser = new Phaser(1); // "1" to register self
132 > *   // create and start threads
133 > *   for (Runnable task : tasks) {
134 > *     phaser.register();
135 > *     new Thread() {
136 > *       public void run() {
137 > *         phaser.arriveAndAwaitAdvance(); // await all creation
138 > *         task.run();
139 > *       }
140 > *     }.start();
141 > *   }
142 > *
143 > *   // allow threads to start and deregister self
144 > *   phaser.arriveAndDeregister();
145 > * }}</pre>
146 > *
147 > * <p>One way to cause a set of threads to repeatedly perform actions
148 > * for a given number of iterations is to override {@code onAdvance}:
149 > *
150 > *  <pre> {@code
151 > * void startTasks(List<Runnable> tasks, final int iterations) {
152 > *   final Phaser phaser = new Phaser() {
153 > *     protected boolean onAdvance(int phase, int registeredParties) {
154 > *       return phase >= iterations || registeredParties == 0;
155 > *     }
156 > *   };
157 > *   phaser.register();
158 > *   for (final Runnable task : tasks) {
159 > *     phaser.register();
160 > *     new Thread() {
161 > *       public void run() {
162 > *         do {
163 > *           task.run();
164 > *           phaser.arriveAndAwaitAdvance();
165 > *         } while (!phaser.isTerminated());
166 > *       }
167 > *     }.start();
168 > *   }
169 > *   phaser.arriveAndDeregister(); // deregister self, don't wait
170 > * }}</pre>
171   *
172 < * <p><b>Sample usage:</b>
172 > * If the main task must later await termination, it
173 > * may re-register and then execute a similar loop:
174 > *  <pre> {@code
175 > *   // ...
176 > *   phaser.register();
177 > *   while (!phaser.isTerminated())
178 > *     phaser.arriveAndAwaitAdvance();}</pre>
179   *
180 < * <p>[todo: non-FJ example]
180 > * <p>Related constructions may be used to await particular phase numbers
181 > * in contexts where you are sure that the phase will never wrap around
182 > * {@code Integer.MAX_VALUE}. For example:
183   *
184 < * <p> A Phaser may be used to support a style of programming in
185 < * which a task waits for others to complete, without otherwise
186 < * needing to keep track of which tasks it is waiting for. This is
187 < * similar to the "sync" construct in Cilk and "clocks" in X10.
188 < * Special constructions based on such barriers are available using
189 < * the <tt>LinkedAsyncAction</tt> and <tt>CyclicAction</tt> classes,
190 < * but they can be useful in other contexts as well.  For a simple
191 < * (but not very useful) example, here is a variant of Fibonacci:
87 < *
88 < * <pre>
89 < * class BarrierFibonacci extends RecursiveAction {
90 < *   int argument, result;
91 < *   final Phaser parentBarrier;
92 < *   BarrierFibonacci(int n, Phaser parentBarrier) {
93 < *     this.argument = n;
94 < *     this.parentBarrier = parentBarrier;
95 < *     parentBarrier.register();
184 > *  <pre> {@code
185 > * void awaitPhase(Phaser phaser, int phase) {
186 > *   int p = phaser.register(); // assumes caller not already registered
187 > *   while (p < phase) {
188 > *     if (phaser.isTerminated())
189 > *       // ... deal with unexpected termination
190 > *     else
191 > *       p = phaser.arriveAndAwaitAdvance();
192   *   }
193 < *   protected void compute() {
194 < *     int n = argument;
195 < *     if (n &lt;= 1)
196 < *        result = n;
197 < *     else {
198 < *        Phaser childBarrier = new Phaser(1);
199 < *        BarrierFibonacci f1 = new BarrierFibonacci(n - 1, childBarrier);
200 < *        BarrierFibonacci f2 = new BarrierFibonacci(n - 2, childBarrier);
201 < *        f1.fork();
202 < *        f2.fork();
203 < *        childBarrier.arriveAndAwait();
204 < *        result = f1.result + f2.result;
193 > *   phaser.arriveAndDeregister();
194 > * }}</pre>
195 > *
196 > *
197 > * <p>To create a set of {@code n} tasks using a tree of phasers, you
198 > * could use code of the following form, assuming a Task class with a
199 > * constructor accepting a {@code Phaser} that it registers with upon
200 > * construction. After invocation of {@code build(new Task[n], 0, n,
201 > * new Phaser())}, these tasks could then be started, for example by
202 > * submitting to a pool:
203 > *
204 > *  <pre> {@code
205 > * void build(Task[] tasks, int lo, int hi, Phaser ph) {
206 > *   if (hi - lo > TASKS_PER_PHASER) {
207 > *     for (int i = lo; i < hi; i += TASKS_PER_PHASER) {
208 > *       int j = Math.min(i + TASKS_PER_PHASER, hi);
209 > *       build(tasks, i, j, new Phaser(ph));
210   *     }
211 < *     parentBarrier.arriveAndDeregister();
211 > *   } else {
212 > *     for (int i = lo; i < hi; ++i)
213 > *       tasks[i] = new Task(ph);
214 > *       // assumes new Task(ph) performs ph.register()
215   *   }
216 < * }
217 < * </pre>
216 > * }}</pre>
217 > *
218 > * The best value of {@code TASKS_PER_PHASER} depends mainly on
219 > * expected synchronization rates. A value as low as four may
220 > * be appropriate for extremely small per-phase task bodies (thus
221 > * high rates), or up to hundreds for extremely large ones.
222   *
223   * <p><b>Implementation notes</b>: This implementation restricts the
224 < * maximum number of parties to 65535. Attempts to register
225 < * additional parties result in IllegalStateExceptions.  
224 > * maximum number of parties to 65535. Attempts to register additional
225 > * parties result in {@code IllegalStateException}. However, you can and
226 > * should create tiered phasers to accommodate arbitrarily large sets
227 > * of participants.
228 > *
229 > * @since 1.7
230 > * @author Doug Lea
231   */
232   public class Phaser {
233      /*
234       * This class implements an extension of X10 "clocks".  Thanks to
235 <     * Vijay Saraswat for the idea of applying it to ForkJoinTasks,
236 <     * and to Vivek Sarkar for enhancements to extend functionality.
235 >     * Vijay Saraswat for the idea, and to Vivek Sarkar for
236 >     * enhancements to extend functionality.
237       */
238  
239      /**
240 <     * Barrier state representation. Conceptually, a barrier contains
128 <     * four values:
129 <     *
130 <     * * parties -- the number of parties to wait (16 bits)
131 <     * * unarrived -- the number of parties yet to hit barrier (16 bits)
132 <     * * phase -- the generation of the barrier (31 bits)
133 <     * * terminated -- set if barrier is terminated (1 bit)
240 >     * Primary state representation, holding four fields:
241       *
242 <     * However, to efficiently maintain atomicity, these values are
243 <     * packed into a single AtomicLong. Termination uses the sign bit
244 <     * of 32 bit representation of phase, so phase is set to -1 on
245 <     * termination.
246 <     */
247 <    private final AtomicLong state;
248 <
249 <    /**
250 <     * Head of Treiber stack for waiting nonFJ threads.
251 <     */
252 <    private final AtomicReference<QNode> head = new AtomicReference<QNode>();
242 >     * * unarrived -- the number of parties yet to hit barrier (bits  0-15)
243 >     * * parties -- the number of parties to wait              (bits 16-31)
244 >     * * phase -- the generation of the barrier                (bits 32-62)
245 >     * * terminated -- set if barrier is terminated            (bit  63 / sign)
246 >     *
247 >     * Except that a phaser with no registered parties is
248 >     * distinguished with the otherwise illegal state of having zero
249 >     * parties and one unarrived parties (encoded as EMPTY below).
250 >     *
251 >     * To efficiently maintain atomicity, these values are packed into
252 >     * a single (atomic) long. Good performance relies on keeping
253 >     * state decoding and encoding simple, and keeping race windows
254 >     * short.
255 >     *
256 >     * All state updates are performed via CAS except initial
257 >     * registration of a sub-phaser (i.e., one with a non-null
258 >     * parent).  In this (relatively rare) case, we use built-in
259 >     * synchronization to lock while first registering with its
260 >     * parent.
261 >     *
262 >     * The phase of a subphaser is allowed to lag that of its
263 >     * ancestors until it is actually accessed.  Method reconcileState
264 >     * is usually attempted only only when the number of unarrived
265 >     * parties appears to be zero, which indicates a potential lag in
266 >     * updating phase after the root advanced.
267 >     */
268 >    private volatile long state;
269 >
270 >    private static final int  MAX_PARTIES     = 0xffff;
271 >    private static final int  MAX_PHASE       = 0x7fffffff;
272 >    private static final int  PARTIES_SHIFT   = 16;
273 >    private static final int  PHASE_SHIFT     = 32;
274 >    private static final int  UNARRIVED_MASK  = 0xffff;      // to mask ints
275 >    private static final long PARTIES_MASK    = 0xffff0000L; // to mask longs
276 >    private static final long TERMINATION_BIT = 1L << 63;
277 >
278 >    // some special values
279 >    private static final int  ONE_ARRIVAL     = 1;
280 >    private static final int  ONE_PARTY       = 1 << PARTIES_SHIFT;
281 >    private static final int  EMPTY           = 1;
282  
283 <    private static final int ushortBits = 16;
148 <    private static final int ushortMask =  (1 << ushortBits) - 1;
149 <    private static final int phaseMask = 0x7fffffff;
283 >    // The following unpacking methods are usually manually inlined
284  
285      private static int unarrivedOf(long s) {
286 <        return (int)(s & ushortMask);
286 >        int counts = (int)s;
287 >        return (counts == EMPTY) ? 0 : counts & UNARRIVED_MASK;
288      }
289  
290      private static int partiesOf(long s) {
291 <        return (int)(s & (ushortMask << 16)) >>> 16;
291 >        return (int)s >>> PARTIES_SHIFT;
292      }
293  
294      private static int phaseOf(long s) {
295 <        return (int)(s >>> 32);
295 >        return (int) (s >>> PHASE_SHIFT);
296      }
297  
298      private static int arrivedOf(long s) {
299 <        return partiesOf(s) - unarrivedOf(s);
299 >        int counts = (int)s;
300 >        return (counts == EMPTY) ? 0 :
301 >            (counts >>> PARTIES_SHIFT) - (counts & UNARRIVED_MASK);
302      }
303  
304 <    private static long stateFor(int phase, int parties, int unarrived) {
305 <        return (((long)phase) << 32) | ((parties << 16) | unarrived);
306 <    }
304 >    /**
305 >     * The parent of this phaser, or null if none
306 >     */
307 >    private final Phaser parent;
308  
309 <    private static IllegalStateException badBounds(int parties, int unarrived) {
310 <        return new IllegalStateException("Attempt to set " + unarrived +
311 <                                         " unarrived of " + parties + " parties");
312 <    }
309 >    /**
310 >     * The root of phaser tree. Equals this if not in a tree.
311 >     */
312 >    private final Phaser root;
313  
314      /**
315 <     * Creates a new Phaser without any initially registered parties,
316 <     * and initial phase number 0.
315 >     * Heads of Treiber stacks for waiting threads. To eliminate
316 >     * contention when releasing some threads while adding others, we
317 >     * use two of them, alternating across even and odd phases.
318 >     * Subphasers share queues with root to speed up releases.
319       */
320 <    public Phaser() {
321 <        state = new AtomicLong(stateFor(0, 0, 0));
320 >    private final AtomicReference<QNode> evenQ;
321 >    private final AtomicReference<QNode> oddQ;
322 >
323 >    private AtomicReference<QNode> queueFor(int phase) {
324 >        return ((phase & 1) == 0) ? evenQ : oddQ;
325      }
326  
327      /**
328 <     * Creates a new Phaser with the given numbers of registered
186 <     * unarrived parties and initial phase number 0.
187 <     * @param parties the number of parties required to trip barrier.
188 <     * @throws IllegalArgumentException if parties less than zero
189 <     * or greater than the maximum number of parties supported.
328 >     * Returns message string for bounds exceptions on arrival.
329       */
330 <    public Phaser(int parties) {
331 <        if (parties < 0 || parties > ushortMask)
332 <            throw new IllegalArgumentException("Illegal number of parties");
194 <        state = new AtomicLong(stateFor(0, parties, parties));
330 >    private String badArrive(long s) {
331 >        return "Attempted arrival of unregistered party for " +
332 >            stateToString(s);
333      }
334  
335      /**
336 <     * Adds a new unarrived party to this phaser.
199 <     * @return the current barrier phase number upon registration
200 <     * @throws IllegalStateException if attempting to register more
201 <     * than the maximum supported number of parties.
336 >     * Returns message string for bounds exceptions on registration.
337       */
338 <    public int register() { // increment both parties and unarrived
339 <        final AtomicLong state = this.state;
340 <        for (;;) {
206 <            long s = state.get();
207 <            int phase = phaseOf(s);
208 <            int parties = partiesOf(s) + 1;
209 <            int unarrived = unarrivedOf(s) + 1;
210 <            if (parties > ushortMask || unarrived > ushortMask)
211 <                throw badBounds(parties, unarrived);
212 <            if (state.compareAndSet(s, stateFor(phase, parties, unarrived)))
213 <                return phase;
214 <        }
338 >    private String badRegister(long s) {
339 >        return "Attempt to register more than " +
340 >            MAX_PARTIES + " parties for " + stateToString(s);
341      }
342  
343      /**
344 <     * Arrives at the barrier, but does not wait for others.  (You can
345 <     * in turn wait for others via {@link #awaitAdvance}).
344 >     * Main implementation for methods arrive and arriveAndDeregister.
345 >     * Manually tuned to speed up and minimize race windows for the
346 >     * common case of just decrementing unarrived field.
347       *
348 <     * @return the current barrier phase number upon entry to
222 <     * this method, or a negative value if terminated;
223 <     * @throws IllegalStateException if the number of unarrived
224 <     * parties would become negative.
348 >     * @param deregister false for arrive, true for arriveAndDeregister
349       */
350 <    public int arrive() { // decrement unarrived. If zero, trip
351 <        final AtomicLong state = this.state;
350 >    private int doArrive(boolean deregister) {
351 >        int adj = deregister ? ONE_ARRIVAL|ONE_PARTY : ONE_ARRIVAL;
352 >        final Phaser root = this.root;
353          for (;;) {
354 <            long s = state.get();
355 <            int phase = phaseOf(s);
356 <            int parties = partiesOf(s);
357 <            int unarrived = unarrivedOf(s) - 1;
358 <            if (unarrived < 0)
234 <                throw badBounds(parties, unarrived);
235 <            if (unarrived == 0 && phase >= 0) {
236 <                trip(phase, parties);
354 >            long s = (root == this) ? state : reconcileState();
355 >            int phase = (int)(s >>> PHASE_SHIFT);
356 >            int counts = (int)s;
357 >            int unarrived = (counts & UNARRIVED_MASK) - 1;
358 >            if (phase < 0)
359                  return phase;
360 +            else if (counts == EMPTY || unarrived < 0) {
361 +                if (root == this || reconcileState() == s)
362 +                    throw new IllegalStateException(badArrive(s));
363              }
364 <            if (state.compareAndSet(s, stateFor(phase, parties, unarrived)))
364 >            else if (UNSAFE.compareAndSwapLong(this, stateOffset, s, s-=adj)) {
365 >                if (unarrived == 0) {
366 >                    long n = s & PARTIES_MASK;  // base of next state
367 >                    int nextUnarrived = ((int)n) >>> PARTIES_SHIFT;
368 >                    if (root != this)
369 >                        return parent.doArrive(nextUnarrived == 0);
370 >                    if (onAdvance(phase, nextUnarrived))
371 >                        n |= TERMINATION_BIT;
372 >                    else if (nextUnarrived == 0)
373 >                        n |= EMPTY;
374 >                    else
375 >                        n |= nextUnarrived;
376 >                    n |= ((long)((phase + 1) & MAX_PHASE)) << PHASE_SHIFT;
377 >                    UNSAFE.compareAndSwapLong(this, stateOffset, s, n);
378 >                    releaseWaiters(phase);
379 >                }
380                  return phase;
381 +            }
382          }
383      }
384  
385      /**
386 <     * Arrives at the barrier, and deregisters from it, without
246 <     * waiting for others.
386 >     * Implementation of register, bulkRegister
387       *
388 <     * @return the current barrier phase number upon entry to
389 <     * this method, or a negative value if terminated;
250 <     * @throws IllegalStateException if the number of registered or
251 <     * unarrived parties would become negative.
388 >     * @param registrations number to add to both parties and
389 >     * unarrived fields. Must be greater than zero.
390       */
391 <    public int arriveAndDeregister() { // Same as arrive, plus decrement parties
392 <        final AtomicLong state = this.state;
391 >    private int doRegister(int registrations) {
392 >        // adjustment to state
393 >        long adj = ((long)registrations << PARTIES_SHIFT) | registrations;
394 >        Phaser par = parent;
395 >        int phase;
396          for (;;) {
397 <            long s = state.get();
398 <            int phase = phaseOf(s);
399 <            int parties = partiesOf(s) - 1;
400 <            int unarrived = unarrivedOf(s) - 1;
401 <            if (parties < 0 || unarrived < 0)
402 <                throw badBounds(parties, unarrived);
403 <            if (unarrived == 0 && phase >= 0) {
404 <                trip(phase, parties);
405 <                return phase;
397 >            long s = state;
398 >            int counts = (int)s;
399 >            int parties = counts >>> PARTIES_SHIFT;
400 >            int unarrived = counts & UNARRIVED_MASK;
401 >            if (registrations > MAX_PARTIES - parties)
402 >                throw new IllegalStateException(badRegister(s));
403 >            else if ((phase = (int)(s >>> PHASE_SHIFT)) < 0)
404 >                break;
405 >            else if (counts != EMPTY) {             // not 1st registration
406 >                if (par == null || reconcileState() == s) {
407 >                    if (unarrived == 0)             // wait out advance
408 >                        root.internalAwaitAdvance(phase, null);
409 >                    else if (UNSAFE.compareAndSwapLong(this, stateOffset,
410 >                                                       s, s + adj))
411 >                        break;
412 >                }
413              }
414 <            if (state.compareAndSet(s, stateFor(phase, parties, unarrived)))
415 <                return phase;
414 >            else if (par == null) {                 // 1st root registration
415 >                long next = (((long) phase) << PHASE_SHIFT) | adj;
416 >                if (UNSAFE.compareAndSwapLong(this, stateOffset, s, next))
417 >                    break;
418 >            }
419 >            else {
420 >                synchronized (this) {               // 1st sub registration
421 >                    if (state == s) {               // recheck under lock
422 >                        par.doRegister(1);
423 >                        do {                        // force current phase
424 >                            phase = (int)(root.state >>> PHASE_SHIFT);
425 >                            // assert phase < 0 || (int)state == EMPTY;
426 >                        } while (!UNSAFE.compareAndSwapLong
427 >                                 (this, stateOffset, state,
428 >                                  (((long) phase) << PHASE_SHIFT) | adj));
429 >                        break;
430 >                    }
431 >                }
432 >            }
433 >        }
434 >        return phase;
435 >    }
436 >
437 >    /**
438 >     * Resolves lagged phase propagation from root if necessary.
439 >     */
440 >    private long reconcileState() {
441 >        Phaser rt = root;
442 >        long s = state;
443 >        if (rt != this) {
444 >            int phase;
445 >            while ((phase = (int)(rt.state >>> PHASE_SHIFT)) !=
446 >                   (int)(s >>> PHASE_SHIFT)) {
447 >                // assert phase < 0 || unarrivedOf(s) == 0
448 >                long t;                             // to reread s
449 >                long p = s & PARTIES_MASK;          // unshifted parties field
450 >                long n = (((long) phase) << PHASE_SHIFT) | p;
451 >                if (phase >= 0) {
452 >                    if (p == 0L)
453 >                        n |= EMPTY;                 // reset to empty
454 >                    else
455 >                        n |= p >>> PARTIES_SHIFT;   // set unarr to parties
456 >                }
457 >                if ((t = state) == s &&
458 >                    UNSAFE.compareAndSwapLong(this, stateOffset, s, s = n))
459 >                    break;
460 >                s = t;
461 >            }
462 >        }
463 >        return s;
464 >    }
465 >
466 >    /**
467 >     * Creates a new phaser with no initially registered parties, no
468 >     * parent, and initial phase number 0. Any thread using this
469 >     * phaser will need to first register for it.
470 >     */
471 >    public Phaser() {
472 >        this(null, 0);
473 >    }
474 >
475 >    /**
476 >     * Creates a new phaser with the given number of registered
477 >     * unarrived parties, no parent, and initial phase number 0.
478 >     *
479 >     * @param parties the number of parties required to advance to the
480 >     * next phase
481 >     * @throws IllegalArgumentException if parties less than zero
482 >     * or greater than the maximum number of parties supported
483 >     */
484 >    public Phaser(int parties) {
485 >        this(null, parties);
486 >    }
487 >
488 >    /**
489 >     * Equivalent to {@link #Phaser(Phaser, int) Phaser(parent, 0)}.
490 >     *
491 >     * @param parent the parent phaser
492 >     */
493 >    public Phaser(Phaser parent) {
494 >        this(parent, 0);
495 >    }
496 >
497 >    /**
498 >     * Creates a new phaser with the given parent and number of
499 >     * registered unarrived parties.  When the given parent is non-null
500 >     * and the given number of parties is greater than zero, this
501 >     * child phaser is registered with its parent.
502 >     *
503 >     * @param parent the parent phaser
504 >     * @param parties the number of parties required to advance to the
505 >     * next phase
506 >     * @throws IllegalArgumentException if parties less than zero
507 >     * or greater than the maximum number of parties supported
508 >     */
509 >    public Phaser(Phaser parent, int parties) {
510 >        if (parties >>> PARTIES_SHIFT != 0)
511 >            throw new IllegalArgumentException("Illegal number of parties");
512 >        int phase = 0;
513 >        this.parent = parent;
514 >        if (parent != null) {
515 >            final Phaser root = parent.root;
516 >            this.root = root;
517 >            this.evenQ = root.evenQ;
518 >            this.oddQ = root.oddQ;
519 >            if (parties != 0)
520 >                phase = parent.doRegister(1);
521          }
522 +        else {
523 +            this.root = this;
524 +            this.evenQ = new AtomicReference<QNode>();
525 +            this.oddQ = new AtomicReference<QNode>();
526 +        }
527 +        this.state = (parties == 0) ? (long) EMPTY :
528 +            ((((long) phase) << PHASE_SHIFT) |
529 +             (((long) parties) << PARTIES_SHIFT) |
530 +             ((long) parties));
531 +    }
532 +
533 +    /**
534 +     * Adds a new unarrived party to this phaser.  If an ongoing
535 +     * invocation of {@link #onAdvance} is in progress, this method
536 +     * may await its completion before returning.  If this phaser has
537 +     * a parent, and this phaser previously had no registered parties,
538 +     * this child phaser is also registered with its parent. If
539 +     * this phaser is terminated, the attempt to register has
540 +     * no effect, and a negative value is returned.
541 +     *
542 +     * @return the arrival phase number to which this registration
543 +     * applied.  If this value is negative, then this phaser has
544 +     * terminated, in which casem registration has no effect.
545 +     * @throws IllegalStateException if attempting to register more
546 +     * than the maximum supported number of parties
547 +     */
548 +    public int register() {
549 +        return doRegister(1);
550 +    }
551 +
552 +    /**
553 +     * Adds the given number of new unarrived parties to this phaser.
554 +     * If an ongoing invocation of {@link #onAdvance} is in progress,
555 +     * this method may await its completion before returning.  If this
556 +     * phaser has a parent, and the given number of parties is greater
557 +     * than zero, and this phaser previously had no registered
558 +     * parties, this child phaser is also registered with its parent.
559 +     * If this phaser is terminated, the attempt to register has no
560 +     * effect, and a negative value is returned.
561 +     *
562 +     * @param parties the number of additional parties required to
563 +     * advance to the next phase
564 +     * @return the arrival phase number to which this registration
565 +     * applied.  If this value is negative, then this phaser has
566 +     * terminated, in which casem registration has no effect.
567 +     * @throws IllegalStateException if attempting to register more
568 +     * than the maximum supported number of parties
569 +     * @throws IllegalArgumentException if {@code parties < 0}
570 +     */
571 +    public int bulkRegister(int parties) {
572 +        if (parties < 0)
573 +            throw new IllegalArgumentException();
574 +        if (parties == 0)
575 +            return getPhase();
576 +        return doRegister(parties);
577 +    }
578 +
579 +    /**
580 +     * Arrives at this phaser, without waiting for others to arrive.
581 +     *
582 +     * <p>It is a usage error for an unregistered party to invoke this
583 +     * method.  However, this error may result in an {@code
584 +     * IllegalStateException} only upon some subsequent operation on
585 +     * this phaser, if ever.
586 +     *
587 +     * @return the arrival phase number, or a negative value if terminated
588 +     * @throws IllegalStateException if not terminated and the number
589 +     * of unarrived parties would become negative
590 +     */
591 +    public int arrive() {
592 +        return doArrive(false);
593      }
594  
595      /**
596 <     * Arrives at the barrier and awaits others. Unlike other arrival
597 <     * methods, this method returns the arrival index of the
598 <     * caller. The caller tripping the barrier returns zero, the
599 <     * previous caller 1, and so on.
600 <     * @return the arrival index
601 <     * @throws IllegalStateException if the number of unarrived
602 <     * parties would become negative.
596 >     * Arrives at this phaser and deregisters from it without waiting
597 >     * for others to arrive. Deregistration reduces the number of
598 >     * parties required to advance in future phases.  If this phaser
599 >     * has a parent, and deregistration causes this phaser to have
600 >     * zero parties, this phaser is also deregistered from its parent.
601 >     *
602 >     * <p>It is a usage error for an unregistered party to invoke this
603 >     * method.  However, this error may result in an {@code
604 >     * IllegalStateException} only upon some subsequent operation on
605 >     * this phaser, if ever.
606 >     *
607 >     * @return the arrival phase number, or a negative value if terminated
608 >     * @throws IllegalStateException if not terminated and the number
609 >     * of registered or unarrived parties would become negative
610 >     */
611 >    public int arriveAndDeregister() {
612 >        return doArrive(true);
613 >    }
614 >
615 >    /**
616 >     * Arrives at this phaser and awaits others. Equivalent in effect
617 >     * to {@code awaitAdvance(arrive())}.  If you need to await with
618 >     * interruption or timeout, you can arrange this with an analogous
619 >     * construction using one of the other forms of the {@code
620 >     * awaitAdvance} method.  If instead you need to deregister upon
621 >     * arrival, use {@code awaitAdvance(arriveAndDeregister())}.
622 >     *
623 >     * <p>It is a usage error for an unregistered party to invoke this
624 >     * method.  However, this error may result in an {@code
625 >     * IllegalStateException} only upon some subsequent operation on
626 >     * this phaser, if ever.
627 >     *
628 >     * @return the arrival phase number, or the (negative)
629 >     * {@linkplain #getPhase() current phase} if terminated
630 >     * @throws IllegalStateException if not terminated and the number
631 >     * of unarrived parties would become negative
632       */
633      public int arriveAndAwaitAdvance() {
634 <        final AtomicLong state = this.state;
634 >        // Specialization of doArrive+awaitAdvance eliminating some reads/paths
635 >        final Phaser root = this.root;
636          for (;;) {
637 <            long s = state.get();
638 <            int phase = phaseOf(s);
639 <            int parties = partiesOf(s);
640 <            int unarrived = unarrivedOf(s) - 1;
641 <            if (unarrived < 0)
642 <                throw badBounds(parties, unarrived);
643 <            if (unarrived == 0 && phase >= 0) {
644 <                trip(phase, parties);
645 <                return 0;
637 >            long s = (root == this) ? state : reconcileState();
638 >            int phase = (int)(s >>> PHASE_SHIFT);
639 >            int counts = (int)s;
640 >            int unarrived = (counts & UNARRIVED_MASK) - 1;
641 >            if (phase < 0)
642 >                return phase;
643 >            else if (counts == EMPTY || unarrived < 0) {
644 >                if (reconcileState() == s)
645 >                    throw new IllegalStateException(badArrive(s));
646              }
647 <            if (state.compareAndSet(s, stateFor(phase, parties, unarrived))) {
648 <                awaitAdvance(phase);
649 <                return unarrived;
647 >            else if (UNSAFE.compareAndSwapLong(this, stateOffset, s,
648 >                                               s -= ONE_ARRIVAL)) {
649 >                if (unarrived != 0)
650 >                    return root.internalAwaitAdvance(phase, null);
651 >                if (root != this)
652 >                    return parent.arriveAndAwaitAdvance();
653 >                long n = s & PARTIES_MASK;  // base of next state
654 >                int nextUnarrived = ((int)n) >>> PARTIES_SHIFT;
655 >                if (onAdvance(phase, nextUnarrived))
656 >                    n |= TERMINATION_BIT;
657 >                else if (nextUnarrived == 0)
658 >                    n |= EMPTY;
659 >                else
660 >                    n |= nextUnarrived;
661 >                int nextPhase = (phase + 1) & MAX_PHASE;
662 >                n |= (long)nextPhase << PHASE_SHIFT;
663 >                if (!UNSAFE.compareAndSwapLong(this, stateOffset, s, n))
664 >                    return (int)(state >>> PHASE_SHIFT); // terminated
665 >                releaseWaiters(phase);
666 >                return nextPhase;
667              }
668          }
669      }
670  
671      /**
672 <     * Awaits the phase of the barrier to advance from the given
673 <     * value, or returns immediately if this barrier is terminated.
674 <     * @param phase the phase on entry to this method
675 <     * @return the phase on exit from this method
672 >     * Awaits the phase of this phaser to advance from the given phase
673 >     * value, returning immediately if the current phase is not equal
674 >     * to the given phase value or this phaser is terminated.
675 >     *
676 >     * @param phase an arrival phase number, or negative value if
677 >     * terminated; this argument is normally the value returned by a
678 >     * previous call to {@code arrive} or {@code arriveAndDeregister}.
679 >     * @return the next arrival phase number, or the argument if it is
680 >     * negative, or the (negative) {@linkplain #getPhase() current phase}
681 >     * if terminated
682       */
683      public int awaitAdvance(int phase) {
684 +        final Phaser root = this.root;
685 +        int p = (int)((root == this? state : reconcileState()) >>> PHASE_SHIFT);
686          if (phase < 0)
687              return phase;
688 <        Thread current = Thread.currentThread();
689 <        if (current instanceof ForkJoinWorkerThread)
690 <            return helpingWait(phase);
312 <        if (untimedWait(current, phase, false))
313 <            current.interrupt();
314 <        return phaseOf(state.get());
688 >        if (p == phase)
689 >            return root.internalAwaitAdvance(phase, null);
690 >        return p;
691      }
692  
693      /**
694 <     * Awaits the phase of the barrier to advance from the given
695 <     * value, or returns immediately if this barrier is terminated, or
696 <     * throws InterruptedException if interrupted while waiting.
697 <     * @param phase the phase on entry to this method
698 <     * @return the phase on exit from this method
694 >     * Awaits the phase of this phaser to advance from the given phase
695 >     * value, throwing {@code InterruptedException} if interrupted
696 >     * while waiting, or returning immediately if the current phase is
697 >     * not equal to the given phase value or this phaser is
698 >     * terminated.
699 >     *
700 >     * @param phase an arrival phase number, or negative value if
701 >     * terminated; this argument is normally the value returned by a
702 >     * previous call to {@code arrive} or {@code arriveAndDeregister}.
703 >     * @return the next arrival phase number, or the argument if it is
704 >     * negative, or the (negative) {@linkplain #getPhase() current phase}
705 >     * if terminated
706       * @throws InterruptedException if thread interrupted while waiting
707       */
708 <    public int awaitAdvanceInterruptibly(int phase) throws InterruptedException {
708 >    public int awaitAdvanceInterruptibly(int phase)
709 >        throws InterruptedException {
710 >        final Phaser root = this.root;
711 >        int p = (int)((root == this? state : reconcileState()) >>> PHASE_SHIFT);
712          if (phase < 0)
713              return phase;
714 <        Thread current = Thread.currentThread();
715 <        if (current instanceof ForkJoinWorkerThread)
716 <            return helpingWait(phase);
717 <        else if (Thread.interrupted() || untimedWait(current, phase, true))
718 <            throw new InterruptedException();
719 <        else
720 <            return phaseOf(state.get());
714 >        if (p == phase) {
715 >            QNode node = new QNode(this, phase, true, false, 0L);
716 >            p = root.internalAwaitAdvance(phase, node);
717 >            if (node.wasInterrupted)
718 >                throw new InterruptedException();
719 >        }
720 >        return p;
721      }
722  
723      /**
724 <     * Awaits the phase of the barrier to advance from the given value
725 <     * or the given timeout elapses, or returns immediately if this
726 <     * barrier is terminated.
727 <     * @param phase the phase on entry to this method
728 <     * @return the phase on exit from this method
724 >     * Awaits the phase of this phaser to advance from the given phase
725 >     * value or the given timeout to elapse, throwing {@code
726 >     * InterruptedException} if interrupted while waiting, or
727 >     * returning immediately if the current phase is not equal to the
728 >     * given phase value or this phaser is terminated.
729 >     *
730 >     * @param phase an arrival phase number, or negative value if
731 >     * terminated; this argument is normally the value returned by a
732 >     * previous call to {@code arrive} or {@code arriveAndDeregister}.
733 >     * @param timeout how long to wait before giving up, in units of
734 >     *        {@code unit}
735 >     * @param unit a {@code TimeUnit} determining how to interpret the
736 >     *        {@code timeout} parameter
737 >     * @return the next arrival phase number, or the argument if it is
738 >     * negative, or the (negative) {@linkplain #getPhase() current phase}
739 >     * if terminated
740       * @throws InterruptedException if thread interrupted while waiting
741       * @throws TimeoutException if timed out while waiting
742       */
743 <    public int awaitAdvanceInterruptibly(int phase, long timeout, TimeUnit unit)
743 >    public int awaitAdvanceInterruptibly(int phase,
744 >                                         long timeout, TimeUnit unit)
745          throws InterruptedException, TimeoutException {
746 +        long nanos = unit.toNanos(timeout);
747 +        final Phaser root = this.root;
748 +        int p = (int)((root == this? state : reconcileState()) >>> PHASE_SHIFT);
749          if (phase < 0)
750              return phase;
751 <        long nanos = unit.toNanos(timeout);
752 <        Thread current = Thread.currentThread();
753 <        if (current instanceof ForkJoinWorkerThread)
754 <            return timedHelpingWait(phase, nanos);
755 <        timedWait(current, phase, nanos);
756 <        return phaseOf(state.get());
751 >        if (p == phase) {
752 >            QNode node = new QNode(this, phase, true, true, nanos);
753 >            p = root.internalAwaitAdvance(phase, node);
754 >            if (node.wasInterrupted)
755 >                throw new InterruptedException();
756 >            else if (p == phase)
757 >                throw new TimeoutException();
758 >        }
759 >        return p;
760      }
761  
762      /**
763 <     * Forces this barrier to enter termination state. Counts of
764 <     * arrived and registered parties are unaffected. This method may
765 <     * be useful for coordinating recovery after one or more tasks
766 <     * encounter unexpected exceptions.
763 >     * Forces this phaser to enter termination state.  Counts of
764 >     * registered parties are unaffected.  If this phaser is a member
765 >     * of a tiered set of phasers, then all of the phasers in the set
766 >     * are terminated.  If this phaser is already terminated, this
767 >     * method has no effect.  This method may be useful for
768 >     * coordinating recovery after one or more tasks encounter
769 >     * unexpected exceptions.
770       */
771      public void forceTermination() {
772 <        final AtomicLong state = this.state;
773 <        for (;;) {
774 <            long s = state.get();
775 <            int phase = phaseOf(s);
776 <            int parties = partiesOf(s);
777 <            int unarrived = unarrivedOf(s);
778 <            if (phase < 0 ||
779 <                state.compareAndSet(s, stateFor(-1, parties, unarrived))) {
780 <                if (head.get() != null)
374 <                    releaseWaiters(-1);
772 >        // Only need to change root state
773 >        final Phaser root = this.root;
774 >        long s;
775 >        while ((s = root.state) >= 0) {
776 >            long next = (s & ~((long)UNARRIVED_MASK)) | TERMINATION_BIT;
777 >            if (UNSAFE.compareAndSwapLong(root, stateOffset, s, next)) {
778 >                // signal all threads
779 >                releaseWaiters(0);
780 >                releaseWaiters(1);
781                  return;
782              }
783          }
784      }
785  
786      /**
381     * Resets the barrier with the given numbers of registered unarrived
382     * parties and phase number 0. This method allows repeated reuse
383     * of this barrier, but only if it is somehow known not to be in
384     * use for other purposes.
385     * @param parties the number of parties required to trip barrier.
386     * @throws IllegalArgumentException if parties less than zero
387     * or greater than the maximum number of parties supported.
388     */
389    public void reset(int parties) {
390        if (parties < 0 || parties > ushortMask)
391            throw new IllegalArgumentException("Illegal number of parties");
392        state.set(stateFor(0, parties, parties));
393        if (head.get() != null)
394            releaseWaiters(0);
395    }
396
397    /**
787       * Returns the current phase number. The maximum phase number is
788 <     * <tt>Integer.MAX_VALUE</tt>, after which it restarts at
789 <     * zero. Upon termination, the phase number is negative.
788 >     * {@code Integer.MAX_VALUE}, after which it restarts at
789 >     * zero. Upon termination, the phase number is negative,
790 >     * in which case the prevailing phase prior to termination
791 >     * may be obtained via {@code getPhase() + Integer.MIN_VALUE}.
792 >     *
793       * @return the phase number, or a negative value if terminated
794       */
795 <    public int getPhase() {
796 <        return phaseOf(state.get());
795 >    public final int getPhase() {
796 >        return (int)(root.state >>> PHASE_SHIFT);
797      }
798  
799      /**
800 <     * Returns the number of parties registered at this barrier.
800 >     * Returns the number of parties registered at this phaser.
801 >     *
802       * @return the number of parties
803       */
804      public int getRegisteredParties() {
805 <        return partiesOf(state.get());
805 >        return partiesOf(state);
806      }
807  
808      /**
809 <     * Returns the number of parties that have arrived at the current
810 <     * phase of this barrier.
809 >     * Returns the number of registered parties that have arrived at
810 >     * the current phase of this phaser.
811 >     *
812       * @return the number of arrived parties
813       */
814      public int getArrivedParties() {
815 <        return arrivedOf(state.get());
815 >        return arrivedOf(reconcileState());
816      }
817  
818      /**
819       * Returns the number of registered parties that have not yet
820 <     * arrived at the current phase of this barrier.
820 >     * arrived at the current phase of this phaser.
821 >     *
822       * @return the number of unarrived parties
823       */
824      public int getUnarrivedParties() {
825 <        return unarrivedOf(state.get());
825 >        return unarrivedOf(reconcileState());
826 >    }
827 >
828 >    /**
829 >     * Returns the parent of this phaser, or {@code null} if none.
830 >     *
831 >     * @return the parent of this phaser, or {@code null} if none
832 >     */
833 >    public Phaser getParent() {
834 >        return parent;
835      }
836  
837      /**
838 <     * Returns true if this barrier has been terminated.
839 <     * @return true if this barrier has been terminated
838 >     * Returns the root ancestor of this phaser, which is the same as
839 >     * this phaser if it has no parent.
840 >     *
841 >     * @return the root ancestor of this phaser
842 >     */
843 >    public Phaser getRoot() {
844 >        return root;
845 >    }
846 >
847 >    /**
848 >     * Returns {@code true} if this phaser has been terminated.
849 >     *
850 >     * @return {@code true} if this phaser has been terminated
851       */
852      public boolean isTerminated() {
853 <        return phaseOf(state.get()) < 0;
853 >        return root.state < 0L;
854      }
855  
856      /**
857 <     * Overridable method to perform an action upon phase advance, and
858 <     * to control termination. This method is invoked whenever the
859 <     * barrier is tripped (and thus all other waiting parties are
860 <     * dormant). If it returns true, then, rather than advance the
861 <     * phase number, this barrier will be set to a final termination
862 <     * state, and subsequent calls to <tt>isTerminated</tt> will
863 <     * return true.
864 <     *
865 <     * <p> The default version returns true when the number of
866 <     * registered parties is zero. Normally, overrides that arrange
867 <     * termination for other reasons should also preserve this
868 <     * property.
869 <     *
870 <     * @param phase the phase number on entering the barrier
871 <     * @param registeredParties the current number of registered
872 <     * parties.
873 <     * @return true if this barrier should terminate
857 >     * Overridable method to perform an action upon impending phase
858 >     * advance, and to control termination. This method is invoked
859 >     * upon arrival of the party advancing this phaser (when all other
860 >     * waiting parties are dormant).  If this method returns {@code
861 >     * true}, this phaser will be set to a final termination state
862 >     * upon advance, and subsequent calls to {@link #isTerminated}
863 >     * will return true. Any (unchecked) Exception or Error thrown by
864 >     * an invocation of this method is propagated to the party
865 >     * attempting to advance this phaser, in which case no advance
866 >     * occurs.
867 >     *
868 >     * <p>The arguments to this method provide the state of the phaser
869 >     * prevailing for the current transition.  The effects of invoking
870 >     * arrival, registration, and waiting methods on this phaser from
871 >     * within {@code onAdvance} are unspecified and should not be
872 >     * relied on.
873 >     *
874 >     * <p>If this phaser is a member of a tiered set of phasers, then
875 >     * {@code onAdvance} is invoked only for its root phaser on each
876 >     * advance.
877 >     *
878 >     * <p>To support the most common use cases, the default
879 >     * implementation of this method returns {@code true} when the
880 >     * number of registered parties has become zero as the result of a
881 >     * party invoking {@code arriveAndDeregister}.  You can disable
882 >     * this behavior, thus enabling continuation upon future
883 >     * registrations, by overriding this method to always return
884 >     * {@code false}:
885 >     *
886 >     * <pre> {@code
887 >     * Phaser phaser = new Phaser() {
888 >     *   protected boolean onAdvance(int phase, int parties) { return false; }
889 >     * }}</pre>
890 >     *
891 >     * @param phase the current phase number on entry to this method,
892 >     * before this phaser is advanced
893 >     * @param registeredParties the current number of registered parties
894 >     * @return {@code true} if this phaser should terminate
895       */
896      protected boolean onAdvance(int phase, int registeredParties) {
897 <        return registeredParties <= 0;
897 >        return registeredParties == 0;
898      }
899  
900      /**
901 <     * Returns a string identifying this barrier, as well as its
901 >     * Returns a string identifying this phaser, as well as its
902       * state.  The state, in brackets, includes the String {@code
903 <     * "phase ="} followed by the phase number, {@code "parties ="}
903 >     * "phase = "} followed by the phase number, {@code "parties = "}
904       * followed by the number of registered parties, and {@code
905 <     * "arrived ="} followed by the number of arrived parties
905 >     * "arrived = "} followed by the number of arrived parties.
906       *
907 <     * @return a string identifying this barrier, as well as its state
907 >     * @return a string identifying this phaser, as well as its state
908       */
909      public String toString() {
910 <        long s = state.get();
475 <        return super.toString() + "[phase = " + phaseOf(s) + " parties = " + partiesOf(s) + " arrived = " + arrivedOf(s) + "]";
910 >        return stateToString(reconcileState());
911      }
912  
478    // methods for tripping and waiting
479
913      /**
914 <     * Advance the current phase (or terminate)
914 >     * Implementation of toString and string-based error messages
915       */
916 <    private void trip(int phase, int parties) {
917 <        int next = onAdvance(phase, parties)? -1 : ((phase + 1) & phaseMask);
918 <        state.set(stateFor(next, parties, parties));
919 <        if (head.get() != null)
920 <            releaseWaiters(next);
916 >    private String stateToString(long s) {
917 >        return super.toString() +
918 >            "[phase = " + phaseOf(s) +
919 >            " parties = " + partiesOf(s) +
920 >            " arrived = " + arrivedOf(s) + "]";
921      }
922  
923 <    private int helpingWait(int phase) {
491 <        final AtomicLong state = this.state;
492 <        int p;
493 <        while ((p = phaseOf(state.get())) == phase) {
494 <            ForkJoinTask<?> t = ForkJoinWorkerThread.pollTask();
495 <            if (t != null) {
496 <                if ((p = phaseOf(state.get())) == phase)
497 <                    t.exec();
498 <                else {   // push task and exit if barrier advanced
499 <                    t.fork();
500 <                    break;
501 <                }
502 <            }
503 <        }
504 <        return p;
505 <    }
923 >    // Waiting mechanics
924  
925 <    private int timedHelpingWait(int phase, long nanos) throws TimeoutException {
926 <        final AtomicLong state = this.state;
927 <        long lastTime = System.nanoTime();
928 <        int p;
929 <        while ((p = phaseOf(state.get())) == phase) {
930 <            long now = System.nanoTime();
931 <            nanos -= now - lastTime;
932 <            lastTime = now;
933 <            if (nanos <= 0) {
934 <                if ((p = phaseOf(state.get())) == phase)
935 <                    throw new TimeoutException();
936 <                else
937 <                    break;
520 <            }
521 <            ForkJoinTask<?> t = ForkJoinWorkerThread.pollTask();
522 <            if (t != null) {
523 <                if ((p = phaseOf(state.get())) == phase)
524 <                    t.exec();
525 <                else {   // push task and exit if barrier advanced
526 <                    t.fork();
527 <                    break;
528 <                }
925 >    /**
926 >     * Removes and signals threads from queue for phase.
927 >     */
928 >    private void releaseWaiters(int phase) {
929 >        QNode q;   // first element of queue
930 >        Thread t;  // its thread
931 >        AtomicReference<QNode> head = (phase & 1) == 0 ? evenQ : oddQ;
932 >        while ((q = head.get()) != null &&
933 >               q.phase != (int)(root.state >>> PHASE_SHIFT)) {
934 >            if (head.compareAndSet(q, q.next) &&
935 >                (t = q.thread) != null) {
936 >                q.thread = null;
937 >                LockSupport.unpark(t);
938              }
939          }
531        return p;
940      }
941  
942      /**
943 <     * Wait nodes for Treiber stack representing wait queue for non-FJ
944 <     * tasks. The waiting scheme is an adaptation of the one used in
945 <     * forkjoin.PoolBarrier.
943 >     * Variant of releaseWaiters that additionally tries to remove any
944 >     * nodes no longer waiting for advance due to timeout or
945 >     * interrupt. Currently, nodes are removed only if they are at
946 >     * head of queue, which suffices to reduce memory footprint in
947 >     * most usages.
948 >     *
949 >     * @return current phase on exit
950       */
951 <    static final class QNode {
952 <        QNode next;
953 <        volatile Thread thread; // nulled to cancel wait
954 <        final int phase;
955 <        QNode(Thread t, int c) {
956 <            thread = t;
957 <            phase = c;
958 <        }
959 <    }
960 <
961 <    private void releaseWaiters(int currentPhase) {
550 <        final AtomicReference<QNode> head = this.head;
551 <        QNode p;
552 <        while ((p = head.get()) != null && p.phase != currentPhase) {
553 <            if (head.compareAndSet(p, null)) {
554 <                do {
555 <                    Thread t = p.thread;
556 <                    if (t != null) {
557 <                        p.thread = null;
558 <                        LockSupport.unpark(t);
559 <                    }
560 <                } while ((p = p.next) != null);
951 >    private int abortWait(int phase) {
952 >        AtomicReference<QNode> head = (phase & 1) == 0 ? evenQ : oddQ;
953 >        for (;;) {
954 >            Thread t;
955 >            QNode q = head.get();
956 >            int p = (int)(root.state >>> PHASE_SHIFT);
957 >            if (q == null || ((t = q.thread) != null && q.phase == p))
958 >                return p;
959 >            if (head.compareAndSet(q, q.next) && t != null) {
960 >                q.thread = null;
961 >                LockSupport.unpark(t);
962              }
963          }
964      }
965  
966      /** The number of CPUs, for spin control */
967 <    static final int NCPUS = Runtime.getRuntime().availableProcessors();
567 <
568 <    /**
569 <     * The number of times to spin before blocking in timed waits.
570 <     * The value is empirically derived.
571 <     */
572 <    static final int maxTimedSpins = (NCPUS < 2)? 0 : 32;
573 <
574 <    /**
575 <     * The number of times to spin before blocking in untimed waits.
576 <     * This is greater than timed value because untimed waits spin
577 <     * faster since they don't need to check times on each spin.
578 <     */
579 <    static final int maxUntimedSpins = maxTimedSpins * 32;
967 >    private static final int NCPU = Runtime.getRuntime().availableProcessors();
968  
969      /**
970 <     * The number of nanoseconds for which it is faster to spin
971 <     * rather than to use timed park. A rough estimate suffices.
970 >     * The number of times to spin before blocking while waiting for
971 >     * advance, per arrival while waiting. On multiprocessors, fully
972 >     * blocking and waking up a large number of threads all at once is
973 >     * usually a very slow process, so we use rechargeable spins to
974 >     * avoid it when threads regularly arrive: When a thread in
975 >     * internalAwaitAdvance notices another arrival before blocking,
976 >     * and there appear to be enough CPUs available, it spins
977 >     * SPINS_PER_ARRIVAL more times before blocking. The value trades
978 >     * off good-citizenship vs big unnecessary slowdowns.
979       */
980 <    static final long spinForTimeoutThreshold = 1000L;
980 >    static final int SPINS_PER_ARRIVAL = (NCPU < 2) ? 1 : 1 << 8;
981  
982      /**
983 <     * Enqueues node and waits unless aborted or signalled.
984 <     */
985 <    private boolean untimedWait(Thread thread, int currentPhase,
986 <                               boolean abortOnInterrupt) {
987 <        final AtomicReference<QNode> head = this.head;
988 <        final AtomicLong state = this.state;
989 <        boolean wasInterrupted = false;
990 <        QNode node = null;
991 <        boolean queued = false;
992 <        int spins = maxUntimedSpins;
993 <        while (phaseOf(state.get()) == currentPhase) {
994 <            QNode h;
995 <            if (node != null && queued) {
996 <                if (node.thread != null) {
997 <                    LockSupport.park();
998 <                    if (Thread.interrupted()) {
999 <                        wasInterrupted = true;
1000 <                        if (abortOnInterrupt)
1001 <                            break;
1002 <                    }
983 >     * Possibly blocks and waits for phase to advance unless aborted.
984 >     * Call only from root node.
985 >     *
986 >     * @param phase current phase
987 >     * @param node if non-null, the wait node to track interrupt and timeout;
988 >     * if null, denotes noninterruptible wait
989 >     * @return current phase
990 >     */
991 >    private int internalAwaitAdvance(int phase, QNode node) {
992 >        releaseWaiters(phase-1);          // ensure old queue clean
993 >        boolean queued = false;           // true when node is enqueued
994 >        int lastUnarrived = 0;            // to increase spins upon change
995 >        int spins = SPINS_PER_ARRIVAL;
996 >        long s;
997 >        int p;
998 >        while ((p = (int)((s = state) >>> PHASE_SHIFT)) == phase) {
999 >            if (node == null) {           // spinning in noninterruptible mode
1000 >                int unarrived = (int)s & UNARRIVED_MASK;
1001 >                if (unarrived != lastUnarrived &&
1002 >                    (lastUnarrived = unarrived) < NCPU)
1003 >                    spins += SPINS_PER_ARRIVAL;
1004 >                boolean interrupted = Thread.interrupted();
1005 >                if (interrupted || --spins < 0) { // need node to record intr
1006 >                    node = new QNode(this, phase, false, false, 0L);
1007 >                    node.wasInterrupted = interrupted;
1008                  }
1009              }
1010 <            else if ((h = head.get()) != null && h.phase != currentPhase) {
1011 <                if (phaseOf(state.get()) == currentPhase) { // must recheck
1012 <                    if (head.compareAndSet(h, h.next)) {
1013 <                        Thread t = h.thread; // help clear out old waiters
1014 <                        if (t != null) {
1015 <                            h.thread = null;
1016 <                            LockSupport.unpark(t);
1017 <                        }
1018 <                    }
1010 >            else if (node.isReleasable()) // done or aborted
1011 >                break;
1012 >            else if (!queued) {           // push onto queue
1013 >                AtomicReference<QNode> head = (phase & 1) == 0 ? evenQ : oddQ;
1014 >                QNode q = node.next = head.get();
1015 >                if ((q == null || q.phase == phase) &&
1016 >                    (int)(state >>> PHASE_SHIFT) == phase) // avoid stale enq
1017 >                    queued = head.compareAndSet(q, node);
1018 >            }
1019 >            else {
1020 >                try {
1021 >                    ForkJoinPool.managedBlock(node);
1022 >                } catch (InterruptedException ie) {
1023 >                    node.wasInterrupted = true;
1024                  }
620                else
621                    break;
1025              }
623            else if (node != null)
624                queued = head.compareAndSet(node.next = h, node);
625            else if (spins <= 0)
626                node = new QNode(thread, currentPhase);
627            else
628                --spins;
1026          }
1027 <        if (node != null)
1028 <            node.thread = null;
1029 <        return wasInterrupted;
1027 >
1028 >        if (node != null) {
1029 >            if (node.thread != null)
1030 >                node.thread = null;       // avoid need for unpark()
1031 >            if (node.wasInterrupted && !node.interruptible)
1032 >                Thread.currentThread().interrupt();
1033 >            if (p == phase && (p = (int)(state >>> PHASE_SHIFT)) == phase)
1034 >                return abortWait(phase); // possibly clean up on abort
1035 >        }
1036 >        releaseWaiters(phase);
1037 >        return p;
1038      }
1039  
1040      /**
1041 <     * Messier timeout version
1041 >     * Wait nodes for Treiber stack representing wait queue
1042       */
1043 <    private void timedWait(Thread thread, int currentPhase, long nanos)
1044 <        throws InterruptedException, TimeoutException {
1045 <        final AtomicReference<QNode> head = this.head;
1046 <        final AtomicLong state = this.state;
1047 <        long lastTime = System.nanoTime();
1048 <        QNode node = null;
1049 <        boolean queued = false;
1050 <        int spins = maxTimedSpins;
1051 <        while (phaseOf(state.get()) == currentPhase) {
1052 <            QNode h;
1053 <            long now = System.nanoTime();
1054 <            nanos -= now - lastTime;
1055 <            lastTime = now;
1056 <            if (nanos <= 0) {
1057 <                if (node != null)
1058 <                    node.thread = null;
1059 <                if (phaseOf(state.get()) == currentPhase)
1060 <                    throw new TimeoutException();
1061 <                else
1062 <                    break;
1043 >    static final class QNode implements ForkJoinPool.ManagedBlocker {
1044 >        final Phaser phaser;
1045 >        final int phase;
1046 >        final boolean interruptible;
1047 >        final boolean timed;
1048 >        boolean wasInterrupted;
1049 >        long nanos;
1050 >        long lastTime;
1051 >        volatile Thread thread; // nulled to cancel wait
1052 >        QNode next;
1053 >
1054 >        QNode(Phaser phaser, int phase, boolean interruptible,
1055 >              boolean timed, long nanos) {
1056 >            this.phaser = phaser;
1057 >            this.phase = phase;
1058 >            this.interruptible = interruptible;
1059 >            this.nanos = nanos;
1060 >            this.timed = timed;
1061 >            this.lastTime = timed ? System.nanoTime() : 0L;
1062 >            thread = Thread.currentThread();
1063 >        }
1064 >
1065 >        public boolean isReleasable() {
1066 >            if (thread == null)
1067 >                return true;
1068 >            if (phaser.getPhase() != phase) {
1069 >                thread = null;
1070 >                return true;
1071              }
1072 <            else if (node != null && queued) {
1073 <                if (node.thread != null &&
1074 <                    nanos > spinForTimeoutThreshold) {
1075 <                    //                LockSupport.parkNanos(this, nanos);
1076 <                    LockSupport.parkNanos(nanos);
664 <                    if (Thread.interrupted()) {
665 <                        node.thread = null;
666 <                        throw new InterruptedException();
667 <                    }
668 <                }
1072 >            if (Thread.interrupted())
1073 >                wasInterrupted = true;
1074 >            if (wasInterrupted && interruptible) {
1075 >                thread = null;
1076 >                return true;
1077              }
1078 <            else if ((h = head.get()) != null && h.phase != currentPhase) {
1079 <                if (phaseOf(state.get()) == currentPhase) { // must recheck
1080 <                    if (head.compareAndSet(h, h.next)) {
1081 <                        Thread t = h.thread; // help clear out old waiters
1082 <                        if (t != null) {
1083 <                            h.thread = null;
1084 <                            LockSupport.unpark(t);
1085 <                        }
1086 <                    }
1078 >            if (timed) {
1079 >                if (nanos > 0L) {
1080 >                    long now = System.nanoTime();
1081 >                    nanos -= now - lastTime;
1082 >                    lastTime = now;
1083 >                }
1084 >                if (nanos <= 0L) {
1085 >                    thread = null;
1086 >                    return true;
1087                  }
680                else
681                    break;
1088              }
1089 <            else if (node != null)
1090 <                queued = head.compareAndSet(node.next = h, node);
1091 <            else if (spins <= 0)
1092 <                node = new QNode(thread, currentPhase);
1093 <            else
1094 <                --spins;
1089 >            return false;
1090 >        }
1091 >
1092 >        public boolean block() {
1093 >            if (isReleasable())
1094 >                return true;
1095 >            else if (!timed)
1096 >                LockSupport.park(this);
1097 >            else if (nanos > 0)
1098 >                LockSupport.parkNanos(this, nanos);
1099 >            return isReleasable();
1100          }
690        if (node != null)
691            node.thread = null;
1101      }
1102  
1103 < }
1103 >    // Unsafe mechanics
1104  
1105 +    private static final sun.misc.Unsafe UNSAFE = getUnsafe();
1106 +    private static final long stateOffset =
1107 +        objectFieldOffset("state", Phaser.class);
1108 +
1109 +    private static long objectFieldOffset(String field, Class<?> klazz) {
1110 +        try {
1111 +            return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
1112 +        } catch (NoSuchFieldException e) {
1113 +            // Convert Exception to corresponding Error
1114 +            NoSuchFieldError error = new NoSuchFieldError(field);
1115 +            error.initCause(e);
1116 +            throw error;
1117 +        }
1118 +    }
1119 +
1120 +    /**
1121 +     * Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
1122 +     * Replace with a simple call to Unsafe.getUnsafe when integrating
1123 +     * into a jdk.
1124 +     *
1125 +     * @return a sun.misc.Unsafe
1126 +     */
1127 +    private static sun.misc.Unsafe getUnsafe() {
1128 +        try {
1129 +            return sun.misc.Unsafe.getUnsafe();
1130 +        } catch (SecurityException se) {
1131 +            try {
1132 +                return java.security.AccessController.doPrivileged
1133 +                    (new java.security
1134 +                     .PrivilegedExceptionAction<sun.misc.Unsafe>() {
1135 +                        public sun.misc.Unsafe run() throws Exception {
1136 +                            java.lang.reflect.Field f = sun.misc
1137 +                                .Unsafe.class.getDeclaredField("theUnsafe");
1138 +                            f.setAccessible(true);
1139 +                            return (sun.misc.Unsafe) f.get(null);
1140 +                        }});
1141 +            } catch (java.security.PrivilegedActionException e) {
1142 +                throw new RuntimeException("Could not initialize intrinsics",
1143 +                                           e.getCause());
1144 +            }
1145 +        }
1146 +    }
1147 + }

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines