1 /*
   2  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   3  *
   4  * This code is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License version 2 only, as
   6  * published by the Free Software Foundation.  Oracle designates this
   7  * particular file as subject to the "Classpath" exception as provided
   8  * by Oracle in the LICENSE file that accompanied this code.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  */
  24 
  25 /*
  26  * This file is available under and governed by the GNU General Public
  27  * License version 2 only, as published by the Free Software Foundation.
  28  * However, the following notice accompanied the original version of this
  29  * file:
  30  *
  31  * Written by Doug Lea, Bill Scherer, and Michael Scott with
  32  * assistance from members of JCP JSR-166 Expert Group and released to
  33  * the public domain, as explained at
  34  * http://creativecommons.org/publicdomain/zero/1.0/
  35  */
  36 
  37 package java.util.concurrent;
  38 import java.util.concurrent.locks.*;
  39 import java.util.*;
  40 
  41 /**
  42  * A {@linkplain BlockingQueue blocking queue} in which each insert
  43  * operation must wait for a corresponding remove operation by another
  44  * thread, and vice versa.  A synchronous queue does not have any
  45  * internal capacity, not even a capacity of one.  You cannot
  46  * <tt>peek</tt> at a synchronous queue because an element is only
  47  * present when you try to remove it; you cannot insert an element
  48  * (using any method) unless another thread is trying to remove it;
  49  * you cannot iterate as there is nothing to iterate.  The
  50  * <em>head</em> of the queue is the element that the first queued
  51  * inserting thread is trying to add to the queue; if there is no such
  52  * queued thread then no element is available for removal and
  53  * <tt>poll()</tt> will return <tt>null</tt>.  For purposes of other
  54  * <tt>Collection</tt> methods (for example <tt>contains</tt>), a
  55  * <tt>SynchronousQueue</tt> acts as an empty collection.  This queue
  56  * does not permit <tt>null</tt> elements.
  57  *
  58  * <p>Synchronous queues are similar to rendezvous channels used in
  59  * CSP and Ada. They are well suited for handoff designs, in which an
  60  * object running in one thread must sync up with an object running
  61  * in another thread in order to hand it some information, event, or
  62  * task.
  63  *
  64  * <p> This class supports an optional fairness policy for ordering
  65  * waiting producer and consumer threads.  By default, this ordering
  66  * is not guaranteed. However, a queue constructed with fairness set
  67  * to <tt>true</tt> grants threads access in FIFO order.
  68  *
  69  * <p>This class and its iterator implement all of the
  70  * <em>optional</em> methods of the {@link Collection} and {@link
  71  * Iterator} interfaces.
  72  *
  73  * <p>This class is a member of the
  74  * <a href="{@docRoot}/../technotes/guides/collections/index.html">
  75  * Java Collections Framework</a>.
  76  *
  77  * @since 1.5
  78  * @author Doug Lea and Bill Scherer and Michael Scott
  79  * @param <E> the type of elements held in this collection
  80  */
  81 public class SynchronousQueue<E> extends AbstractQueue<E>
  82     implements BlockingQueue<E>, java.io.Serializable {
  83     private static final long serialVersionUID = -3223113410248163686L;
  84 
  85     /*
  86      * This class implements extensions of the dual stack and dual
  87      * queue algorithms described in "Nonblocking Concurrent Objects
  88      * with Condition Synchronization", by W. N. Scherer III and
  89      * M. L. Scott.  18th Annual Conf. on Distributed Computing,
  90      * Oct. 2004 (see also
  91      * http://www.cs.rochester.edu/u/scott/synchronization/pseudocode/duals.html).
  92      * The (Lifo) stack is used for non-fair mode, and the (Fifo)
  93      * queue for fair mode. The performance of the two is generally
  94      * similar. Fifo usually supports higher throughput under
  95      * contention but Lifo maintains higher thread locality in common
  96      * applications.
  97      *
  98      * A dual queue (and similarly stack) is one that at any given
  99      * time either holds "data" -- items provided by put operations,
 100      * or "requests" -- slots representing take operations, or is
 101      * empty. A call to "fulfill" (i.e., a call requesting an item
 102      * from a queue holding data or vice versa) dequeues a
 103      * complementary node.  The most interesting feature of these
 104      * queues is that any operation can figure out which mode the
 105      * queue is in, and act accordingly without needing locks.
 106      *
 107      * Both the queue and stack extend abstract class Transferer
 108      * defining the single method transfer that does a put or a
 109      * take. These are unified into a single method because in dual
 110      * data structures, the put and take operations are symmetrical,
 111      * so nearly all code can be combined. The resulting transfer
 112      * methods are on the long side, but are easier to follow than
 113      * they would be if broken up into nearly-duplicated parts.
 114      *
 115      * The queue and stack data structures share many conceptual
 116      * similarities but very few concrete details. For simplicity,
 117      * they are kept distinct so that they can later evolve
 118      * separately.
 119      *
 120      * The algorithms here differ from the versions in the above paper
 121      * in extending them for use in synchronous queues, as well as
 122      * dealing with cancellation. The main differences include:
 123      *
 124      *  1. The original algorithms used bit-marked pointers, but
 125      *     the ones here use mode bits in nodes, leading to a number
 126      *     of further adaptations.
 127      *  2. SynchronousQueues must block threads waiting to become
 128      *     fulfilled.
 129      *  3. Support for cancellation via timeout and interrupts,
 130      *     including cleaning out cancelled nodes/threads
 131      *     from lists to avoid garbage retention and memory depletion.
 132      *
 133      * Blocking is mainly accomplished using LockSupport park/unpark,
 134      * except that nodes that appear to be the next ones to become
 135      * fulfilled first spin a bit (on multiprocessors only). On very
 136      * busy synchronous queues, spinning can dramatically improve
 137      * throughput. And on less busy ones, the amount of spinning is
 138      * small enough not to be noticeable.
 139      *
 140      * Cleaning is done in different ways in queues vs stacks.  For
 141      * queues, we can almost always remove a node immediately in O(1)
 142      * time (modulo retries for consistency checks) when it is
 143      * cancelled. But if it may be pinned as the current tail, it must
 144      * wait until some subsequent cancellation. For stacks, we need a
 145      * potentially O(n) traversal to be sure that we can remove the
 146      * node, but this can run concurrently with other threads
 147      * accessing the stack.
 148      *
 149      * While garbage collection takes care of most node reclamation
 150      * issues that otherwise complicate nonblocking algorithms, care
 151      * is taken to "forget" references to data, other nodes, and
 152      * threads that might be held on to long-term by blocked
 153      * threads. In cases where setting to null would otherwise
 154      * conflict with main algorithms, this is done by changing a
 155      * node's link to now point to the node itself. This doesn't arise
 156      * much for Stack nodes (because blocked threads do not hang on to
 157      * old head pointers), but references in Queue nodes must be
 158      * aggressively forgotten to avoid reachability of everything any
 159      * node has ever referred to since arrival.
 160      */
 161 
 162     /**
 163      * Shared internal API for dual stacks and queues.
 164      */
 165     abstract static class Transferer<E> {
 166         /**
 167          * Performs a put or take.
 168          *
 169          * @param e if non-null, the item to be handed to a consumer;
 170          *          if null, requests that transfer return an item
 171          *          offered by producer.
 172          * @param timed if this operation should timeout
 173          * @param nanos the timeout, in nanoseconds
 174          * @return if non-null, the item provided or received; if null,
 175          *         the operation failed due to timeout or interrupt --
 176          *         the caller can distinguish which of these occurred
 177          *         by checking Thread.interrupted.
 178          */
 179         abstract E transfer(E e, boolean timed, long nanos);
 180     }
 181 
 182     /** The number of CPUs, for spin control */
 183     static final int NCPUS = Runtime.getRuntime().availableProcessors();
 184 
 185     /**
 186      * The number of times to spin before blocking in timed waits.
 187      * The value is empirically derived -- it works well across a
 188      * variety of processors and OSes. Empirically, the best value
 189      * seems not to vary with number of CPUs (beyond 2) so is just
 190      * a constant.
 191      */
 192     static final int maxTimedSpins = (NCPUS < 2) ? 0 : 32;
 193 
 194     /**
 195      * The number of times to spin before blocking in untimed waits.
 196      * This is greater than timed value because untimed waits spin
 197      * faster since they don't need to check times on each spin.
 198      */
 199     static final int maxUntimedSpins = maxTimedSpins * 16;
 200 
 201     /**
 202      * The number of nanoseconds for which it is faster to spin
 203      * rather than to use timed park. A rough estimate suffices.
 204      */
 205     static final long spinForTimeoutThreshold = 1000L;
 206 
 207     /** Dual stack */
 208     static final class TransferStack<E> extends Transferer<E> {
 209         /*
 210          * This extends Scherer-Scott dual stack algorithm, differing,
 211          * among other ways, by using "covering" nodes rather than
 212          * bit-marked pointers: Fulfilling operations push on marker
 213          * nodes (with FULFILLING bit set in mode) to reserve a spot
 214          * to match a waiting node.
 215          */
 216 
 217         /* Modes for SNodes, ORed together in node fields */
 218         /** Node represents an unfulfilled consumer */
 219         static final int REQUEST    = 0;
 220         /** Node represents an unfulfilled producer */
 221         static final int DATA       = 1;
 222         /** Node is fulfilling another unfulfilled DATA or REQUEST */
 223         static final int FULFILLING = 2;
 224 
 225         /** Return true if m has fulfilling bit set */
 226         static boolean isFulfilling(int m) { return (m & FULFILLING) != 0; }
 227 
 228         /** Node class for TransferStacks. */
 229         static final class SNode {
 230             volatile SNode next;        // next node in stack
 231             volatile SNode match;       // the node matched to this
 232             volatile Thread waiter;     // to control park/unpark
 233             Object item;                // data; or null for REQUESTs
 234             int mode;
 235             // Note: item and mode fields don't need to be volatile
 236             // since they are always written before, and read after,
 237             // other volatile/atomic operations.
 238 
 239             SNode(Object item) {
 240                 this.item = item;
 241             }
 242 
 243             boolean casNext(SNode cmp, SNode val) {
 244                 return cmp == next &&
 245                     UNSAFE.compareAndSwapObject(this, nextOffset, cmp, val);
 246             }
 247 
 248             /**
 249              * Tries to match node s to this node, if so, waking up thread.
 250              * Fulfillers call tryMatch to identify their waiters.
 251              * Waiters block until they have been matched.
 252              *
 253              * @param s the node to match
 254              * @return true if successfully matched to s
 255              */
 256             boolean tryMatch(SNode s) {
 257                 if (match == null &&
 258                     UNSAFE.compareAndSwapObject(this, matchOffset, null, s)) {
 259                     Thread w = waiter;
 260                     if (w != null) {    // waiters need at most one unpark
 261                         waiter = null;
 262                         LockSupport.unpark(w);
 263                     }
 264                     return true;
 265                 }
 266                 return match == s;
 267             }
 268 
 269             /**
 270              * Tries to cancel a wait by matching node to itself.
 271              */
 272             void tryCancel() {
 273                 UNSAFE.compareAndSwapObject(this, matchOffset, null, this);
 274             }
 275 
 276             boolean isCancelled() {
 277                 return match == this;
 278             }
 279 
 280             // Unsafe mechanics
 281             private static final sun.misc.Unsafe UNSAFE;
 282             private static final long matchOffset;
 283             private static final long nextOffset;
 284 
 285             static {
 286                 try {
 287                     UNSAFE = sun.misc.Unsafe.getUnsafe();
 288                     Class<?> k = SNode.class;
 289                     matchOffset = UNSAFE.objectFieldOffset
 290                         (k.getDeclaredField("match"));
 291                     nextOffset = UNSAFE.objectFieldOffset
 292                         (k.getDeclaredField("next"));
 293                 } catch (Exception e) {
 294                     throw new Error(e);
 295                 }
 296             }
 297         }
 298 
 299         /** The head (top) of the stack */
 300         volatile SNode head;
 301 
 302         boolean casHead(SNode h, SNode nh) {
 303             return h == head &&
 304                 UNSAFE.compareAndSwapObject(this, headOffset, h, nh);
 305         }
 306 
 307         /**
 308          * Creates or resets fields of a node. Called only from transfer
 309          * where the node to push on stack is lazily created and
 310          * reused when possible to help reduce intervals between reads
 311          * and CASes of head and to avoid surges of garbage when CASes
 312          * to push nodes fail due to contention.
 313          */
 314         static SNode snode(SNode s, Object e, SNode next, int mode) {
 315             if (s == null) s = new SNode(e);
 316             s.mode = mode;
 317             s.next = next;
 318             return s;
 319         }
 320 
 321         /**
 322          * Puts or takes an item.
 323          */
 324         @SuppressWarnings("unchecked")
 325         E transfer(E e, boolean timed, long nanos) {
 326             /*
 327              * Basic algorithm is to loop trying one of three actions:
 328              *
 329              * 1. If apparently empty or already containing nodes of same
 330              *    mode, try to push node on stack and wait for a match,
 331              *    returning it, or null if cancelled.
 332              *
 333              * 2. If apparently containing node of complementary mode,
 334              *    try to push a fulfilling node on to stack, match
 335              *    with corresponding waiting node, pop both from
 336              *    stack, and return matched item. The matching or
 337              *    unlinking might not actually be necessary because of
 338              *    other threads performing action 3:
 339              *
 340              * 3. If top of stack already holds another fulfilling node,
 341              *    help it out by doing its match and/or pop
 342              *    operations, and then continue. The code for helping
 343              *    is essentially the same as for fulfilling, except
 344              *    that it doesn't return the item.
 345              */
 346 
 347             SNode s = null; // constructed/reused as needed
 348             int mode = (e == null) ? REQUEST : DATA;
 349 
 350             for (;;) {
 351                 SNode h = head;
 352                 if (h == null || h.mode == mode) {  // empty or same-mode
 353                     if (timed && nanos <= 0) {      // can't wait
 354                         if (h != null && h.isCancelled())
 355                             casHead(h, h.next);     // pop cancelled node
 356                         else
 357                             return null;
 358                     } else if (casHead(h, s = snode(s, e, h, mode))) {
 359                         SNode m = awaitFulfill(s, timed, nanos);
 360                         if (m == s) {               // wait was cancelled
 361                             clean(s);
 362                             return null;
 363                         }
 364                         if ((h = head) != null && h.next == s)
 365                             casHead(h, s.next);     // help s's fulfiller
 366                         return (E) ((mode == REQUEST) ? m.item : s.item);
 367                     }
 368                 } else if (!isFulfilling(h.mode)) { // try to fulfill
 369                     if (h.isCancelled())            // already cancelled
 370                         casHead(h, h.next);         // pop and retry
 371                     else if (casHead(h, s=snode(s, e, h, FULFILLING|mode))) {
 372                         for (;;) { // loop until matched or waiters disappear
 373                             SNode m = s.next;       // m is s's match
 374                             if (m == null) {        // all waiters are gone
 375                                 casHead(s, null);   // pop fulfill node
 376                                 s = null;           // use new node next time
 377                                 break;              // restart main loop
 378                             }
 379                             SNode mn = m.next;
 380                             if (m.tryMatch(s)) {
 381                                 casHead(s, mn);     // pop both s and m
 382                                 return (E) ((mode == REQUEST) ? m.item : s.item);
 383                             } else                  // lost match
 384                                 s.casNext(m, mn);   // help unlink
 385                         }
 386                     }
 387                 } else {                            // help a fulfiller
 388                     SNode m = h.next;               // m is h's match
 389                     if (m == null)                  // waiter is gone
 390                         casHead(h, null);           // pop fulfilling node
 391                     else {
 392                         SNode mn = m.next;
 393                         if (m.tryMatch(h))          // help match
 394                             casHead(h, mn);         // pop both h and m
 395                         else                        // lost match
 396                             h.casNext(m, mn);       // help unlink
 397                     }
 398                 }
 399             }
 400         }
 401 
 402         /**
 403          * Spins/blocks until node s is matched by a fulfill operation.
 404          *
 405          * @param s the waiting node
 406          * @param timed true if timed wait
 407          * @param nanos timeout value
 408          * @return matched node, or s if cancelled
 409          */
 410         SNode awaitFulfill(SNode s, boolean timed, long nanos) {
 411             /*
 412              * When a node/thread is about to block, it sets its waiter
 413              * field and then rechecks state at least one more time
 414              * before actually parking, thus covering race vs
 415              * fulfiller noticing that waiter is non-null so should be
 416              * woken.
 417              *
 418              * When invoked by nodes that appear at the point of call
 419              * to be at the head of the stack, calls to park are
 420              * preceded by spins to avoid blocking when producers and
 421              * consumers are arriving very close in time.  This can
 422              * happen enough to bother only on multiprocessors.
 423              *
 424              * The order of checks for returning out of main loop
 425              * reflects fact that interrupts have precedence over
 426              * normal returns, which have precedence over
 427              * timeouts. (So, on timeout, one last check for match is
 428              * done before giving up.) Except that calls from untimed
 429              * SynchronousQueue.{poll/offer} don't check interrupts
 430              * and don't wait at all, so are trapped in transfer
 431              * method rather than calling awaitFulfill.
 432              */
 433             long lastTime = timed ? System.nanoTime() : 0;
 434             Thread w = Thread.currentThread();
 435             SNode h = head;
 436             int spins = (shouldSpin(s) ?
 437                          (timed ? maxTimedSpins : maxUntimedSpins) : 0);
 438             for (;;) {
 439                 if (w.isInterrupted())
 440                     s.tryCancel();
 441                 SNode m = s.match;
 442                 if (m != null)
 443                     return m;
 444                 if (timed) {
 445                     long now = System.nanoTime();
 446                     nanos -= now - lastTime;
 447                     lastTime = now;
 448                     if (nanos <= 0) {
 449                         s.tryCancel();
 450                         continue;
 451                     }
 452                 }
 453                 if (spins > 0)
 454                     spins = shouldSpin(s) ? (spins-1) : 0;
 455                 else if (s.waiter == null)
 456                     s.waiter = w; // establish waiter so can park next iter
 457                 else if (!timed)
 458                     LockSupport.park(this);
 459                 else if (nanos > spinForTimeoutThreshold)
 460                     LockSupport.parkNanos(this, nanos);
 461             }
 462         }
 463 
 464         /**
 465          * Returns true if node s is at head or there is an active
 466          * fulfiller.
 467          */
 468         boolean shouldSpin(SNode s) {
 469             SNode h = head;
 470             return (h == s || h == null || isFulfilling(h.mode));
 471         }
 472 
 473         /**
 474          * Unlinks s from the stack.
 475          */
 476         void clean(SNode s) {
 477             s.item = null;   // forget item
 478             s.waiter = null; // forget thread
 479 
 480             /*
 481              * At worst we may need to traverse entire stack to unlink
 482              * s. If there are multiple concurrent calls to clean, we
 483              * might not see s if another thread has already removed
 484              * it. But we can stop when we see any node known to
 485              * follow s. We use s.next unless it too is cancelled, in
 486              * which case we try the node one past. We don't check any
 487              * further because we don't want to doubly traverse just to
 488              * find sentinel.
 489              */
 490 
 491             SNode past = s.next;
 492             if (past != null && past.isCancelled())
 493                 past = past.next;
 494 
 495             // Absorb cancelled nodes at head
 496             SNode p;
 497             while ((p = head) != null && p != past && p.isCancelled())
 498                 casHead(p, p.next);
 499 
 500             // Unsplice embedded nodes
 501             while (p != null && p != past) {
 502                 SNode n = p.next;
 503                 if (n != null && n.isCancelled())
 504                     p.casNext(n, n.next);
 505                 else
 506                     p = n;
 507             }
 508         }
 509 
 510         // Unsafe mechanics
 511         private static final sun.misc.Unsafe UNSAFE;
 512         private static final long headOffset;
 513         static {
 514             try {
 515                 UNSAFE = sun.misc.Unsafe.getUnsafe();
 516                 Class<?> k = TransferStack.class;
 517                 headOffset = UNSAFE.objectFieldOffset
 518                     (k.getDeclaredField("head"));
 519             } catch (Exception e) {
 520                 throw new Error(e);
 521             }
 522         }
 523     }
 524 
 525     /** Dual Queue */
 526     static final class TransferQueue<E> extends Transferer<E> {
 527         /*
 528          * This extends Scherer-Scott dual queue algorithm, differing,
 529          * among other ways, by using modes within nodes rather than
 530          * marked pointers. The algorithm is a little simpler than
 531          * that for stacks because fulfillers do not need explicit
 532          * nodes, and matching is done by CAS'ing QNode.item field
 533          * from non-null to null (for put) or vice versa (for take).
 534          */
 535 
 536         /** Node class for TransferQueue. */
 537         static final class QNode {
 538             volatile QNode next;          // next node in queue
 539             volatile Object item;         // CAS'ed to or from null
 540             volatile Thread waiter;       // to control park/unpark
 541             final boolean isData;
 542 
 543             QNode(Object item, boolean isData) {
 544                 this.item = item;
 545                 this.isData = isData;
 546             }
 547 
 548             boolean casNext(QNode cmp, QNode val) {
 549                 return next == cmp &&
 550                     UNSAFE.compareAndSwapObject(this, nextOffset, cmp, val);
 551             }
 552 
 553             boolean casItem(Object cmp, Object val) {
 554                 return item == cmp &&
 555                     UNSAFE.compareAndSwapObject(this, itemOffset, cmp, val);
 556             }
 557 
 558             /**
 559              * Tries to cancel by CAS'ing ref to this as item.
 560              */
 561             void tryCancel(Object cmp) {
 562                 UNSAFE.compareAndSwapObject(this, itemOffset, cmp, this);
 563             }
 564 
 565             boolean isCancelled() {
 566                 return item == this;
 567             }
 568 
 569             /**
 570              * Returns true if this node is known to be off the queue
 571              * because its next pointer has been forgotten due to
 572              * an advanceHead operation.
 573              */
 574             boolean isOffList() {
 575                 return next == this;
 576             }
 577 
 578             // Unsafe mechanics
 579             private static final sun.misc.Unsafe UNSAFE;
 580             private static final long itemOffset;
 581             private static final long nextOffset;
 582 
 583             static {
 584                 try {
 585                     UNSAFE = sun.misc.Unsafe.getUnsafe();
 586                     Class<?> k = QNode.class;
 587                     itemOffset = UNSAFE.objectFieldOffset
 588                         (k.getDeclaredField("item"));
 589                     nextOffset = UNSAFE.objectFieldOffset
 590                         (k.getDeclaredField("next"));
 591                 } catch (Exception e) {
 592                     throw new Error(e);
 593                 }
 594             }
 595         }
 596 
 597         /** Head of queue */
 598         transient volatile QNode head;
 599         /** Tail of queue */
 600         transient volatile QNode tail;
 601         /**
 602          * Reference to a cancelled node that might not yet have been
 603          * unlinked from queue because it was the last inserted node
 604          * when it cancelled.
 605          */
 606         transient volatile QNode cleanMe;
 607 
 608         TransferQueue() {
 609             QNode h = new QNode(null, false); // initialize to dummy node.
 610             head = h;
 611             tail = h;
 612         }
 613 
 614         /**
 615          * Tries to cas nh as new head; if successful, unlink
 616          * old head's next node to avoid garbage retention.
 617          */
 618         void advanceHead(QNode h, QNode nh) {
 619             if (h == head &&
 620                 UNSAFE.compareAndSwapObject(this, headOffset, h, nh))
 621                 h.next = h; // forget old next
 622         }
 623 
 624         /**
 625          * Tries to cas nt as new tail.
 626          */
 627         void advanceTail(QNode t, QNode nt) {
 628             if (tail == t)
 629                 UNSAFE.compareAndSwapObject(this, tailOffset, t, nt);
 630         }
 631 
 632         /**
 633          * Tries to CAS cleanMe slot.
 634          */
 635         boolean casCleanMe(QNode cmp, QNode val) {
 636             return cleanMe == cmp &&
 637                 UNSAFE.compareAndSwapObject(this, cleanMeOffset, cmp, val);
 638         }
 639 
 640         /**
 641          * Puts or takes an item.
 642          */
 643         @SuppressWarnings("unchecked")
 644         E transfer(E e, boolean timed, long nanos) {
 645             /* Basic algorithm is to loop trying to take either of
 646              * two actions:
 647              *
 648              * 1. If queue apparently empty or holding same-mode nodes,
 649              *    try to add node to queue of waiters, wait to be
 650              *    fulfilled (or cancelled) and return matching item.
 651              *
 652              * 2. If queue apparently contains waiting items, and this
 653              *    call is of complementary mode, try to fulfill by CAS'ing
 654              *    item field of waiting node and dequeuing it, and then
 655              *    returning matching item.
 656              *
 657              * In each case, along the way, check for and try to help
 658              * advance head and tail on behalf of other stalled/slow
 659              * threads.
 660              *
 661              * The loop starts off with a null check guarding against
 662              * seeing uninitialized head or tail values. This never
 663              * happens in current SynchronousQueue, but could if
 664              * callers held non-volatile/final ref to the
 665              * transferer. The check is here anyway because it places
 666              * null checks at top of loop, which is usually faster
 667              * than having them implicitly interspersed.
 668              */
 669 
 670             QNode s = null; // constructed/reused as needed
 671             boolean isData = (e != null);
 672 
 673             for (;;) {
 674                 QNode t = tail;
 675                 QNode h = head;
 676                 if (t == null || h == null)         // saw uninitialized value
 677                     continue;                       // spin
 678 
 679                 if (h == t || t.isData == isData) { // empty or same-mode
 680                     QNode tn = t.next;
 681                     if (t != tail)                  // inconsistent read
 682                         continue;
 683                     if (tn != null) {               // lagging tail
 684                         advanceTail(t, tn);
 685                         continue;
 686                     }
 687                     if (timed && nanos <= 0)        // can't wait
 688                         return null;
 689                     if (s == null)
 690                         s = new QNode(e, isData);
 691                     if (!t.casNext(null, s))        // failed to link in
 692                         continue;
 693 
 694                     advanceTail(t, s);              // swing tail and wait
 695                     Object x = awaitFulfill(s, e, timed, nanos);
 696                     if (x == s) {                   // wait was cancelled
 697                         clean(t, s);
 698                         return null;
 699                     }
 700 
 701                     if (!s.isOffList()) {           // not already unlinked
 702                         advanceHead(t, s);          // unlink if head
 703                         if (x != null)              // and forget fields
 704                             s.item = s;
 705                         s.waiter = null;
 706                     }
 707                     return (x != null) ? (E)x : e;
 708 
 709                 } else {                            // complementary-mode
 710                     QNode m = h.next;               // node to fulfill
 711                     if (t != tail || m == null || h != head)
 712                         continue;                   // inconsistent read
 713 
 714                     Object x = m.item;
 715                     if (isData == (x != null) ||    // m already fulfilled
 716                         x == m ||                   // m cancelled
 717                         !m.casItem(x, e)) {         // lost CAS
 718                         advanceHead(h, m);          // dequeue and retry
 719                         continue;
 720                     }
 721 
 722                     advanceHead(h, m);              // successfully fulfilled
 723                     LockSupport.unpark(m.waiter);
 724                     return (x != null) ? (E)x : e;
 725                 }
 726             }
 727         }
 728 
 729         /**
 730          * Spins/blocks until node s is fulfilled.
 731          *
 732          * @param s the waiting node
 733          * @param e the comparison value for checking match
 734          * @param timed true if timed wait
 735          * @param nanos timeout value
 736          * @return matched item, or s if cancelled
 737          */
 738         Object awaitFulfill(QNode s, E e, boolean timed, long nanos) {
 739             /* Same idea as TransferStack.awaitFulfill */
 740             long lastTime = timed ? System.nanoTime() : 0;
 741             Thread w = Thread.currentThread();
 742             int spins = ((head.next == s) ?
 743                          (timed ? maxTimedSpins : maxUntimedSpins) : 0);
 744             for (;;) {
 745                 if (w.isInterrupted())
 746                     s.tryCancel(e);
 747                 Object x = s.item;
 748                 if (x != e)
 749                     return x;
 750                 if (timed) {
 751                     long now = System.nanoTime();
 752                     nanos -= now - lastTime;
 753                     lastTime = now;
 754                     if (nanos <= 0) {
 755                         s.tryCancel(e);
 756                         continue;
 757                     }
 758                 }
 759                 if (spins > 0)
 760                     --spins;
 761                 else if (s.waiter == null)
 762                     s.waiter = w;
 763                 else if (!timed)
 764                     LockSupport.park(this);
 765                 else if (nanos > spinForTimeoutThreshold)
 766                     LockSupport.parkNanos(this, nanos);
 767             }
 768         }
 769 
 770         /**
 771          * Gets rid of cancelled node s with original predecessor pred.
 772          */
 773         void clean(QNode pred, QNode s) {
 774             s.waiter = null; // forget thread
 775             /*
 776              * At any given time, exactly one node on list cannot be
 777              * deleted -- the last inserted node. To accommodate this,
 778              * if we cannot delete s, we save its predecessor as
 779              * "cleanMe", deleting the previously saved version
 780              * first. At least one of node s or the node previously
 781              * saved can always be deleted, so this always terminates.
 782              */
 783             while (pred.next == s) { // Return early if already unlinked
 784                 QNode h = head;
 785                 QNode hn = h.next;   // Absorb cancelled first node as head
 786                 if (hn != null && hn.isCancelled()) {
 787                     advanceHead(h, hn);
 788                     continue;
 789                 }
 790                 QNode t = tail;      // Ensure consistent read for tail
 791                 if (t == h)
 792                     return;
 793                 QNode tn = t.next;
 794                 if (t != tail)
 795                     continue;
 796                 if (tn != null) {
 797                     advanceTail(t, tn);
 798                     continue;
 799                 }
 800                 if (s != t) {        // If not tail, try to unsplice
 801                     QNode sn = s.next;
 802                     if (sn == s || pred.casNext(s, sn))
 803                         return;
 804                 }
 805                 QNode dp = cleanMe;
 806                 if (dp != null) {    // Try unlinking previous cancelled node
 807                     QNode d = dp.next;
 808                     QNode dn;
 809                     if (d == null ||               // d is gone or
 810                         d == dp ||                 // d is off list or
 811                         !d.isCancelled() ||        // d not cancelled or
 812                         (d != t &&                 // d not tail and
 813                          (dn = d.next) != null &&  //   has successor
 814                          dn != d &&                //   that is on list
 815                          dp.casNext(d, dn)))       // d unspliced
 816                         casCleanMe(dp, null);
 817                     if (dp == pred)
 818                         return;      // s is already saved node
 819                 } else if (casCleanMe(null, pred))
 820                     return;          // Postpone cleaning s
 821             }
 822         }
 823 
 824         private static final sun.misc.Unsafe UNSAFE;
 825         private static final long headOffset;
 826         private static final long tailOffset;
 827         private static final long cleanMeOffset;
 828         static {
 829             try {
 830                 UNSAFE = sun.misc.Unsafe.getUnsafe();
 831                 Class<?> k = TransferQueue.class;
 832                 headOffset = UNSAFE.objectFieldOffset
 833                     (k.getDeclaredField("head"));
 834                 tailOffset = UNSAFE.objectFieldOffset
 835                     (k.getDeclaredField("tail"));
 836                 cleanMeOffset = UNSAFE.objectFieldOffset
 837                     (k.getDeclaredField("cleanMe"));
 838             } catch (Exception e) {
 839                 throw new Error(e);
 840             }
 841         }
 842     }
 843 
 844     /**
 845      * The transferer. Set only in constructor, but cannot be declared
 846      * as final without further complicating serialization.  Since
 847      * this is accessed only at most once per public method, there
 848      * isn't a noticeable performance penalty for using volatile
 849      * instead of final here.
 850      */
 851     private transient volatile Transferer<E> transferer;
 852 
 853     /**
 854      * Creates a <tt>SynchronousQueue</tt> with nonfair access policy.
 855      */
 856     public SynchronousQueue() {
 857         this(false);
 858     }
 859 
 860     /**
 861      * Creates a <tt>SynchronousQueue</tt> with the specified fairness policy.
 862      *
 863      * @param fair if true, waiting threads contend in FIFO order for
 864      *        access; otherwise the order is unspecified.
 865      */
 866     public SynchronousQueue(boolean fair) {
 867         transferer = fair ? new TransferQueue<E>() : new TransferStack<E>();
 868     }
 869 
 870     /**
 871      * Adds the specified element to this queue, waiting if necessary for
 872      * another thread to receive it.
 873      *
 874      * @throws InterruptedException {@inheritDoc}
 875      * @throws NullPointerException {@inheritDoc}
 876      */
 877     public void put(E o) throws InterruptedException {
 878         if (o == null) throw new NullPointerException();
 879         if (transferer.transfer(o, false, 0) == null) {
 880             Thread.interrupted();
 881             throw new InterruptedException();
 882         }
 883     }
 884 
 885     /**
 886      * Inserts the specified element into this queue, waiting if necessary
 887      * up to the specified wait time for another thread to receive it.
 888      *
 889      * @return <tt>true</tt> if successful, or <tt>false</tt> if the
 890      *         specified waiting time elapses before a consumer appears.
 891      * @throws InterruptedException {@inheritDoc}
 892      * @throws NullPointerException {@inheritDoc}
 893      */
 894     public boolean offer(E o, long timeout, TimeUnit unit)
 895         throws InterruptedException {
 896         if (o == null) throw new NullPointerException();
 897         if (transferer.transfer(o, true, unit.toNanos(timeout)) != null)
 898             return true;
 899         if (!Thread.interrupted())
 900             return false;
 901         throw new InterruptedException();
 902     }
 903 
 904     /**
 905      * Inserts the specified element into this queue, if another thread is
 906      * waiting to receive it.
 907      *
 908      * @param e the element to add
 909      * @return <tt>true</tt> if the element was added to this queue, else
 910      *         <tt>false</tt>
 911      * @throws NullPointerException if the specified element is null
 912      */
 913     public boolean offer(E e) {
 914         if (e == null) throw new NullPointerException();
 915         return transferer.transfer(e, true, 0) != null;
 916     }
 917 
 918     /**
 919      * Retrieves and removes the head of this queue, waiting if necessary
 920      * for another thread to insert it.
 921      *
 922      * @return the head of this queue
 923      * @throws InterruptedException {@inheritDoc}
 924      */
 925     public E take() throws InterruptedException {
 926         E e = transferer.transfer(null, false, 0);
 927         if (e != null)
 928             return e;
 929         Thread.interrupted();
 930         throw new InterruptedException();
 931     }
 932 
 933     /**
 934      * Retrieves and removes the head of this queue, waiting
 935      * if necessary up to the specified wait time, for another thread
 936      * to insert it.
 937      *
 938      * @return the head of this queue, or <tt>null</tt> if the
 939      *         specified waiting time elapses before an element is present.
 940      * @throws InterruptedException {@inheritDoc}
 941      */
 942     public E poll(long timeout, TimeUnit unit) throws InterruptedException {
 943         E e = transferer.transfer(null, true, unit.toNanos(timeout));
 944         if (e != null || !Thread.interrupted())
 945             return e;
 946         throw new InterruptedException();
 947     }
 948 
 949     /**
 950      * Retrieves and removes the head of this queue, if another thread
 951      * is currently making an element available.
 952      *
 953      * @return the head of this queue, or <tt>null</tt> if no
 954      *         element is available.
 955      */
 956     public E poll() {
 957         return transferer.transfer(null, true, 0);
 958     }
 959 
 960     /**
 961      * Always returns <tt>true</tt>.
 962      * A <tt>SynchronousQueue</tt> has no internal capacity.
 963      *
 964      * @return <tt>true</tt>
 965      */
 966     public boolean isEmpty() {
 967         return true;
 968     }
 969 
 970     /**
 971      * Always returns zero.
 972      * A <tt>SynchronousQueue</tt> has no internal capacity.
 973      *
 974      * @return zero.
 975      */
 976     public int size() {
 977         return 0;
 978     }
 979 
 980     /**
 981      * Always returns zero.
 982      * A <tt>SynchronousQueue</tt> has no internal capacity.
 983      *
 984      * @return zero.
 985      */
 986     public int remainingCapacity() {
 987         return 0;
 988     }
 989 
 990     /**
 991      * Does nothing.
 992      * A <tt>SynchronousQueue</tt> has no internal capacity.
 993      */
 994     public void clear() {
 995     }
 996 
 997     /**
 998      * Always returns <tt>false</tt>.
 999      * A <tt>SynchronousQueue</tt> has no internal capacity.
1000      *
1001      * @param o the element
1002      * @return <tt>false</tt>
1003      */
1004     public boolean contains(Object o) {
1005         return false;
1006     }
1007 
1008     /**
1009      * Always returns <tt>false</tt>.
1010      * A <tt>SynchronousQueue</tt> has no internal capacity.
1011      *
1012      * @param o the element to remove
1013      * @return <tt>false</tt>
1014      */
1015     public boolean remove(Object o) {
1016         return false;
1017     }
1018 
1019     /**
1020      * Returns <tt>false</tt> unless the given collection is empty.
1021      * A <tt>SynchronousQueue</tt> has no internal capacity.
1022      *
1023      * @param c the collection
1024      * @return <tt>false</tt> unless given collection is empty
1025      */
1026     public boolean containsAll(Collection<?> c) {
1027         return c.isEmpty();
1028     }
1029 
1030     /**
1031      * Always returns <tt>false</tt>.
1032      * A <tt>SynchronousQueue</tt> has no internal capacity.
1033      *
1034      * @param c the collection
1035      * @return <tt>false</tt>
1036      */
1037     public boolean removeAll(Collection<?> c) {
1038         return false;
1039     }
1040 
1041     /**
1042      * Always returns <tt>false</tt>.
1043      * A <tt>SynchronousQueue</tt> has no internal capacity.
1044      *
1045      * @param c the collection
1046      * @return <tt>false</tt>
1047      */
1048     public boolean retainAll(Collection<?> c) {
1049         return false;
1050     }
1051 
1052     /**
1053      * Always returns <tt>null</tt>.
1054      * A <tt>SynchronousQueue</tt> does not return elements
1055      * unless actively waited on.
1056      *
1057      * @return <tt>null</tt>
1058      */
1059     public E peek() {
1060         return null;
1061     }
1062 
1063     /**
1064      * Returns an empty iterator in which <tt>hasNext</tt> always returns
1065      * <tt>false</tt>.
1066      *
1067      * @return an empty iterator
1068      */
1069     @SuppressWarnings("unchecked")
1070     public Iterator<E> iterator() {
1071         return (Iterator<E>) EmptyIterator.EMPTY_ITERATOR;
1072     }
1073 
1074     // Replicated from a previous version of Collections
1075     private static class EmptyIterator<E> implements Iterator<E> {
1076         static final EmptyIterator<Object> EMPTY_ITERATOR
1077             = new EmptyIterator<Object>();
1078 
1079         public boolean hasNext() { return false; }
1080         public E next() { throw new NoSuchElementException(); }
1081         public void remove() { throw new IllegalStateException(); }
1082     }
1083 
1084     /**
1085      * Returns a zero-length array.
1086      * @return a zero-length array
1087      */
1088     public Object[] toArray() {
1089         return new Object[0];
1090     }
1091 
1092     /**
1093      * Sets the zeroeth element of the specified array to <tt>null</tt>
1094      * (if the array has non-zero length) and returns it.
1095      *
1096      * @param a the array
1097      * @return the specified array
1098      * @throws NullPointerException if the specified array is null
1099      */
1100     public <T> T[] toArray(T[] a) {
1101         if (a.length > 0)
1102             a[0] = null;
1103         return a;
1104     }
1105 
1106     /**
1107      * @throws UnsupportedOperationException {@inheritDoc}
1108      * @throws ClassCastException            {@inheritDoc}
1109      * @throws NullPointerException          {@inheritDoc}
1110      * @throws IllegalArgumentException      {@inheritDoc}
1111      */
1112     public int drainTo(Collection<? super E> c) {
1113         if (c == null)
1114             throw new NullPointerException();
1115         if (c == this)
1116             throw new IllegalArgumentException();
1117         int n = 0;
1118         for (E e; (e = poll()) != null;) {
1119             c.add(e);
1120             ++n;
1121         }
1122         return n;
1123     }
1124 
1125     /**
1126      * @throws UnsupportedOperationException {@inheritDoc}
1127      * @throws ClassCastException            {@inheritDoc}
1128      * @throws NullPointerException          {@inheritDoc}
1129      * @throws IllegalArgumentException      {@inheritDoc}
1130      */
1131     public int drainTo(Collection<? super E> c, int maxElements) {
1132         if (c == null)
1133             throw new NullPointerException();
1134         if (c == this)
1135             throw new IllegalArgumentException();
1136         int n = 0;
1137         for (E e; n < maxElements && (e = poll()) != null;) {
1138             c.add(e);
1139             ++n;
1140         }
1141         return n;
1142     }
1143 
1144     /*
1145      * To cope with serialization strategy in the 1.5 version of
1146      * SynchronousQueue, we declare some unused classes and fields
1147      * that exist solely to enable serializability across versions.
1148      * These fields are never used, so are initialized only if this
1149      * object is ever serialized or deserialized.
1150      */
1151 
1152     @SuppressWarnings("serial")
1153     static class WaitQueue implements java.io.Serializable { }
1154     static class LifoWaitQueue extends WaitQueue {
1155         private static final long serialVersionUID = -3633113410248163686L;
1156     }
1157     static class FifoWaitQueue extends WaitQueue {
1158         private static final long serialVersionUID = -3623113410248163686L;
1159     }
1160     private ReentrantLock qlock;
1161     private WaitQueue waitingProducers;
1162     private WaitQueue waitingConsumers;
1163 
1164     /**
1165      * Saves the state to a stream (that is, serializes it).
1166      *
1167      * @param s the stream
1168      */
1169     private void writeObject(java.io.ObjectOutputStream s)
1170         throws java.io.IOException {
1171         boolean fair = transferer instanceof TransferQueue;
1172         if (fair) {
1173             qlock = new ReentrantLock(true);
1174             waitingProducers = new FifoWaitQueue();
1175             waitingConsumers = new FifoWaitQueue();
1176         }
1177         else {
1178             qlock = new ReentrantLock();
1179             waitingProducers = new LifoWaitQueue();
1180             waitingConsumers = new LifoWaitQueue();
1181         }
1182         s.defaultWriteObject();
1183     }
1184 
1185     private void readObject(final java.io.ObjectInputStream s)
1186         throws java.io.IOException, ClassNotFoundException {
1187         s.defaultReadObject();
1188         if (waitingProducers instanceof FifoWaitQueue)
1189             transferer = new TransferQueue<E>();
1190         else
1191             transferer = new TransferStack<E>();
1192     }
1193 
1194     // Unsafe mechanics
1195     static long objectFieldOffset(sun.misc.Unsafe UNSAFE,
1196                                   String field, Class<?> klazz) {
1197         try {
1198             return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
1199         } catch (NoSuchFieldException e) {
1200             // Convert Exception to corresponding Error
1201             NoSuchFieldError error = new NoSuchFieldError(field);
1202             error.initCause(e);
1203             throw error;
1204         }
1205     }
1206 
1207 }