1 /* 2 * Written by Doug Lea, Bill Scherer, and Michael Scott with 3 * assistance from members of JCP JSR-166 Expert Group and released to 4 * the public domain, as explained at 5 * http://creativecommons.org/licenses/publicdomain 6 */ 7 8 package java.util.concurrent; 9 10 import java.util.concurrent.atomic.AtomicInteger; 11 import java.util.concurrent.atomic.AtomicReference; 12 import java.util.concurrent.locks.LockSupport; 13 14 /** 15 * A synchronization point at which threads can pair and swap elements 16 * within pairs. Each thread presents some object on entry to the 17 * {@link #exchange exchange} method, matches with a partner thread, 18 * and receives its partner's object on return. An Exchanger may be 19 * viewed as a bidirectional form of a {@link SynchronousQueue}. 20 * Exchangers may be useful in applications such as genetic algorithms 21 * and pipeline designs. 22 * 23 * <p><b>Sample Usage:</b> 24 * Here are the highlights of a class that uses an {@code Exchanger} 25 * to swap buffers between threads so that the thread filling the 26 * buffer gets a freshly emptied one when it needs it, handing off the 27 * filled one to the thread emptying the buffer. 28 * <pre>{@code 29 * class FillAndEmpty { 30 * Exchanger<DataBuffer> exchanger = new Exchanger<DataBuffer>(); 31 * DataBuffer initialEmptyBuffer = ... a made-up type 32 * DataBuffer initialFullBuffer = ... 33 * 34 * class FillingLoop implements Runnable { 35 * public void run() { 36 * DataBuffer currentBuffer = initialEmptyBuffer; 37 * try { 38 * while (currentBuffer != null) { 39 * addToBuffer(currentBuffer); 40 * if (currentBuffer.isFull()) 41 * currentBuffer = exchanger.exchange(currentBuffer); 42 * } 43 * } catch (InterruptedException ex) { ... handle ... } 44 * } 45 * } 46 * 47 * class EmptyingLoop implements Runnable { 48 * public void run() { 49 * DataBuffer currentBuffer = initialFullBuffer; 50 * try { 51 * while (currentBuffer != null) { 52 * takeFromBuffer(currentBuffer); 53 * if (currentBuffer.isEmpty()) 54 * currentBuffer = exchanger.exchange(currentBuffer); 55 * } 56 * } catch (InterruptedException ex) { ... handle ...} 57 * } 58 * } 59 * 60 * void start() { 61 * new Thread(new FillingLoop()).start(); 62 * new Thread(new EmptyingLoop()).start(); 63 * } 64 * } 65 * }</pre> 66 * 67 * <p>Memory consistency effects: For each pair of threads that 68 * successfully exchange objects via an {@code Exchanger}, actions 69 * prior to the {@code exchange()} in each thread 70 * <a href="package-summary.html#MemoryVisibility"><i>happen-before</i></a> 71 * those subsequent to a return from the corresponding {@code exchange()} 72 * in the other thread. 73 * 74 * @since 1.5 75 * @author Doug Lea and Bill Scherer and Michael Scott 76 * @param <V> The type of objects that may be exchanged 77 */ 78 public class Exchanger<V> { 79 /* 80 * Algorithm Description: 81 * 82 * The basic idea is to maintain a "slot", which is a reference to 83 * a Node containing both an Item to offer and a "hole" waiting to 84 * get filled in. If an incoming "occupying" thread sees that the 85 * slot is null, it CAS'es (compareAndSets) a Node there and waits 86 * for another to invoke exchange. That second "fulfilling" thread 87 * sees that the slot is non-null, and so CASes it back to null, 88 * also exchanging items by CASing the hole, plus waking up the 89 * occupying thread if it is blocked. In each case CAS'es may 90 * fail because a slot at first appears non-null but is null upon 91 * CAS, or vice-versa. So threads may need to retry these 92 * actions. 93 * 94 * This simple approach works great when there are only a few 95 * threads using an Exchanger, but performance rapidly 96 * deteriorates due to CAS contention on the single slot when 97 * there are lots of threads using an exchanger. So instead we use 98 * an "arena"; basically a kind of hash table with a dynamically 99 * varying number of slots, any one of which can be used by 100 * threads performing an exchange. Incoming threads pick slots 101 * based on a hash of their Thread ids. If an incoming thread 102 * fails to CAS in its chosen slot, it picks an alternative slot 103 * instead. And similarly from there. If a thread successfully 104 * CASes into a slot but no other thread arrives, it tries 105 * another, heading toward the zero slot, which always exists even 106 * if the table shrinks. The particular mechanics controlling this 107 * are as follows: 108 * 109 * Waiting: Slot zero is special in that it is the only slot that 110 * exists when there is no contention. A thread occupying slot 111 * zero will block if no thread fulfills it after a short spin. 112 * In other cases, occupying threads eventually give up and try 113 * another slot. Waiting threads spin for a while (a period that 114 * should be a little less than a typical context-switch time) 115 * before either blocking (if slot zero) or giving up (if other 116 * slots) and restarting. There is no reason for threads to block 117 * unless there are unlikely to be any other threads present. 118 * Occupants are mainly avoiding memory contention so sit there 119 * quietly polling for a shorter period than it would take to 120 * block and then unblock them. Non-slot-zero waits that elapse 121 * because of lack of other threads waste around one extra 122 * context-switch time per try, which is still on average much 123 * faster than alternative approaches. 124 * 125 * Sizing: Usually, using only a few slots suffices to reduce 126 * contention. Especially with small numbers of threads, using 127 * too many slots can lead to just as poor performance as using 128 * too few of them, and there's not much room for error. The 129 * variable "max" maintains the number of slots actually in 130 * use. It is increased when a thread sees too many CAS 131 * failures. (This is analogous to resizing a regular hash table 132 * based on a target load factor, except here, growth steps are 133 * just one-by-one rather than proportional.) Growth requires 134 * contention failures in each of three tried slots. Requiring 135 * multiple failures for expansion copes with the fact that some 136 * failed CASes are not due to contention but instead to simple 137 * races between two threads or thread pre-emptions occurring 138 * between reading and CASing. Also, very transient peak 139 * contention can be much higher than the average sustainable 140 * levels. The max limit is decreased on average 50% of the times 141 * that a non-slot-zero wait elapses without being fulfilled. 142 * Threads experiencing elapsed waits move closer to zero, so 143 * eventually find existing (or future) threads even if the table 144 * has been shrunk due to inactivity. The chosen mechanics and 145 * thresholds for growing and shrinking are intrinsically 146 * entangled with indexing and hashing inside the exchange code, 147 * and can't be nicely abstracted out. 148 * 149 * Hashing: Each thread picks its initial slot to use in accord 150 * with a simple hashcode. The sequence is the same on each 151 * encounter by any given thread, but effectively random across 152 * threads. Using arenas encounters the classic cost vs quality 153 * tradeoffs of all hash tables. Here, we use a one-step FNV-1a 154 * hash code based on the current thread's Thread.getId(), along 155 * with a cheap approximation to a mod operation to select an 156 * index. The downside of optimizing index selection in this way 157 * is that the code is hardwired to use a maximum table size of 158 * 32. But this value more than suffices for known platforms and 159 * applications. 160 * 161 * Probing: On sensed contention of a selected slot, we probe 162 * sequentially through the table, analogously to linear probing 163 * after collision in a hash table. (We move circularly, in 164 * reverse order, to mesh best with table growth and shrinkage 165 * rules.) Except that to minimize the effects of false-alarms 166 * and cache thrashing, we try the first selected slot twice 167 * before moving. 168 * 169 * Padding: Even with contention management, slots are heavily 170 * contended, so use cache-padding to avoid poor memory 171 * performance. Because of this, slots are lazily constructed 172 * only when used, to avoid wasting this space unnecessarily. 173 * While isolation of locations is not much of an issue at first 174 * in an application, as time goes on and garbage-collectors 175 * perform compaction, slots are very likely to be moved adjacent 176 * to each other, which can cause much thrashing of cache lines on 177 * MPs unless padding is employed. 178 * 179 * This is an improvement of the algorithm described in the paper 180 * "A Scalable Elimination-based Exchange Channel" by William 181 * Scherer, Doug Lea, and Michael Scott in Proceedings of SCOOL05 182 * workshop. Available at: http://hdl.handle.net/1802/2104 183 */ 184 185 /** The number of CPUs, for sizing and spin control */ 186 private static final int NCPU = Runtime.getRuntime().availableProcessors(); 187 188 /** 189 * The capacity of the arena. Set to a value that provides more 190 * than enough space to handle contention. On small machines 191 * most slots won't be used, but it is still not wasted because 192 * the extra space provides some machine-level address padding 193 * to minimize interference with heavily CAS'ed Slot locations. 194 * And on very large machines, performance eventually becomes 195 * bounded by memory bandwidth, not numbers of threads/CPUs. 196 * This constant cannot be changed without also modifying 197 * indexing and hashing algorithms. 198 */ 199 private static final int CAPACITY = 32; 200 201 /** 202 * The value of "max" that will hold all threads without 203 * contention. When this value is less than CAPACITY, some 204 * otherwise wasted expansion can be avoided. 205 */ 206 private static final int FULL = 207 Math.max(0, Math.min(CAPACITY, NCPU / 2) - 1); 208 209 /** 210 * The number of times to spin (doing nothing except polling a 211 * memory location) before blocking or giving up while waiting to 212 * be fulfilled. Should be zero on uniprocessors. On 213 * multiprocessors, this value should be large enough so that two 214 * threads exchanging items as fast as possible block only when 215 * one of them is stalled (due to GC or preemption), but not much 216 * longer, to avoid wasting CPU resources. Seen differently, this 217 * value is a little over half the number of cycles of an average 218 * context switch time on most systems. The value here is 219 * approximately the average of those across a range of tested 220 * systems. 221 */ 222 private static final int SPINS = (NCPU == 1) ? 0 : 2000; 223 224 /** 225 * The number of times to spin before blocking in timed waits. 226 * Timed waits spin more slowly because checking the time takes 227 * time. The best value relies mainly on the relative rate of 228 * System.nanoTime vs memory accesses. The value is empirically 229 * derived to work well across a variety of systems. 230 */ 231 private static final int TIMED_SPINS = SPINS / 20; 232 233 /** 234 * Sentinel item representing cancellation of a wait due to 235 * interruption, timeout, or elapsed spin-waits. This value is 236 * placed in holes on cancellation, and used as a return value 237 * from waiting methods to indicate failure to set or get hole. 238 */ 239 private static final Object CANCEL = new Object(); 240 241 /** 242 * Value representing null arguments/returns from public 243 * methods. This disambiguates from internal requirement that 244 * holes start out as null to mean they are not yet set. 245 */ 246 private static final Object NULL_ITEM = new Object(); 247 248 /** 249 * Nodes hold partially exchanged data. This class 250 * opportunistically subclasses AtomicReference to represent the 251 * hole. So get() returns hole, and compareAndSet CAS'es value 252 * into hole. This class cannot be parameterized as "V" because 253 * of the use of non-V CANCEL sentinels. 254 */ 255 private static final class Node extends AtomicReference<Object> { 256 /** The element offered by the Thread creating this node. */ 257 public final Object item; 258 259 /** The Thread waiting to be signalled; null until waiting. */ 260 public volatile Thread waiter; 261 262 /** 263 * Creates node with given item and empty hole. 264 * @param item the item 265 */ 266 public Node(Object item) { 267 this.item = item; 268 } 269 } 270 271 /** 272 * A Slot is an AtomicReference with heuristic padding to lessen 273 * cache effects of this heavily CAS'ed location. While the 274 * padding adds noticeable space, all slots are created only on 275 * demand, and there will be more than one of them only when it 276 * would improve throughput more than enough to outweigh using 277 * extra space. 278 */ 279 private static final class Slot extends AtomicReference<Object> { 280 // Improve likelihood of isolation on <= 64 byte cache lines 281 long q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, qa, qb, qc, qd, qe; 282 } 283 284 /** 285 * Slot array. Elements are lazily initialized when needed. 286 * Declared volatile to enable double-checked lazy construction. 287 */ 288 private volatile Slot[] arena = new Slot[CAPACITY]; 289 290 /** 291 * The maximum slot index being used. The value sometimes 292 * increases when a thread experiences too many CAS contentions, 293 * and sometimes decreases when a spin-wait elapses. Changes 294 * are performed only via compareAndSet, to avoid stale values 295 * when a thread happens to stall right before setting. 296 */ 297 private final AtomicInteger max = new AtomicInteger(); 298 299 /** 300 * Main exchange function, handling the different policy variants. 301 * Uses Object, not "V" as argument and return value to simplify 302 * handling of sentinel values. Callers from public methods decode 303 * and cast accordingly. 304 * 305 * @param item the (non-null) item to exchange 306 * @param timed true if the wait is timed 307 * @param nanos if timed, the maximum wait time 308 * @return the other thread's item, or CANCEL if interrupted or timed out 309 */ 310 private Object doExchange(Object item, boolean timed, long nanos) { 311 Node me = new Node(item); // Create in case occupying 312 int index = hashIndex(); // Index of current slot 313 int fails = 0; // Number of CAS failures 314 315 for (;;) { 316 Object y; // Contents of current slot 317 Slot slot = arena[index]; 318 if (slot == null) // Lazily initialize slots 319 createSlot(index); // Continue loop to reread 320 else if ((y = slot.get()) != null && // Try to fulfill 321 slot.compareAndSet(y, null)) { 322 Node you = (Node)y; // Transfer item 323 if (you.compareAndSet(null, item)) { 324 LockSupport.unpark(you.waiter); 325 return you.item; 326 } // Else cancelled; continue 327 } 328 else if (y == null && // Try to occupy 329 slot.compareAndSet(null, me)) { 330 if (index == 0) // Blocking wait for slot 0 331 return timed? awaitNanos(me, slot, nanos): await(me, slot); 332 Object v = spinWait(me, slot); // Spin wait for non-0 333 if (v != CANCEL) 334 return v; 335 me = new Node(item); // Throw away cancelled node 336 int m = max.get(); 337 if (m > (index >>>= 1)) // Decrease index 338 max.compareAndSet(m, m - 1); // Maybe shrink table 339 } 340 else if (++fails > 1) { // Allow 2 fails on 1st slot 341 int m = max.get(); 342 if (fails > 3 && m < FULL && max.compareAndSet(m, m + 1)) 343 index = m + 1; // Grow on 3rd failed slot 344 else if (--index < 0) 345 index = m; // Circularly traverse 346 } 347 } 348 } 349 350 /** 351 * Returns a hash index for the current thread. Uses a one-step 352 * FNV-1a hash code (http://www.isthe.com/chongo/tech/comp/fnv/) 353 * based on the current thread's Thread.getId(). These hash codes 354 * have more uniform distribution properties with respect to small 355 * moduli (here 1-31) than do other simple hashing functions. 356 * 357 * <p>To return an index between 0 and max, we use a cheap 358 * approximation to a mod operation, that also corrects for bias 359 * due to non-power-of-2 remaindering (see {@link 360 * java.util.Random#nextInt}). Bits of the hashcode are masked 361 * with "nbits", the ceiling power of two of table size (looked up 362 * in a table packed into three ints). If too large, this is 363 * retried after rotating the hash by nbits bits, while forcing new 364 * top bit to 0, which guarantees eventual termination (although 365 * with a non-random-bias). This requires an average of less than 366 * 2 tries for all table sizes, and has a maximum 2% difference 367 * from perfectly uniform slot probabilities when applied to all 368 * possible hash codes for sizes less than 32. 369 * 370 * @return a per-thread-random index, 0 <= index < max 371 */ 372 private final int hashIndex() { 373 long id = Thread.currentThread().getId(); 374 int hash = (((int)(id ^ (id >>> 32))) ^ 0x811c9dc5) * 0x01000193; 375 376 int m = max.get(); 377 int nbits = (((0xfffffc00 >> m) & 4) | // Compute ceil(log2(m+1)) 378 ((0x000001f8 >>> m) & 2) | // The constants hold 379 ((0xffff00f2 >>> m) & 1)); // a lookup table 380 int index; 381 while ((index = hash & ((1 << nbits) - 1)) > m) // May retry on 382 hash = (hash >>> nbits) | (hash << (33 - nbits)); // non-power-2 m 383 return index; 384 } 385 386 /** 387 * Creates a new slot at given index. Called only when the slot 388 * appears to be null. Relies on double-check using builtin 389 * locks, since they rarely contend. This in turn relies on the 390 * arena array being declared volatile. 391 * 392 * @param index the index to add slot at 393 */ 394 private void createSlot(int index) { 395 // Create slot outside of lock to narrow sync region 396 Slot newSlot = new Slot(); 397 Slot[] a = arena; 398 synchronized (a) { 399 if (a[index] == null) 400 a[index] = newSlot; 401 } 402 } 403 404 /** 405 * Tries to cancel a wait for the given node waiting in the given 406 * slot, if so, helping clear the node from its slot to avoid 407 * garbage retention. 408 * 409 * @param node the waiting node 410 * @param slot the slot it is waiting in 411 * @return true if successfully cancelled 412 */ 413 private static boolean tryCancel(Node node, Slot slot) { 414 if (!node.compareAndSet(null, CANCEL)) 415 return false; 416 if (slot.get() == node) // pre-check to minimize contention 417 slot.compareAndSet(node, null); 418 return true; 419 } 420 421 // Three forms of waiting. Each just different enough not to merge 422 // code with others. 423 424 /** 425 * Spin-waits for hole for a non-0 slot. Fails if spin elapses 426 * before hole filled. Does not check interrupt, relying on check 427 * in public exchange method to abort if interrupted on entry. 428 * 429 * @param node the waiting node 430 * @return on success, the hole; on failure, CANCEL 431 */ 432 private static Object spinWait(Node node, Slot slot) { 433 int spins = SPINS; 434 for (;;) { 435 Object v = node.get(); 436 if (v != null) 437 return v; 438 else if (spins > 0) 439 --spins; 440 else 441 tryCancel(node, slot); 442 } 443 } 444 445 /** 446 * Waits for (by spinning and/or blocking) and gets the hole 447 * filled in by another thread. Fails if interrupted before 448 * hole filled. 449 * 450 * When a node/thread is about to block, it sets its waiter field 451 * and then rechecks state at least one more time before actually 452 * parking, thus covering race vs fulfiller noticing that waiter 453 * is non-null so should be woken. 454 * 455 * Thread interruption status is checked only surrounding calls to 456 * park. The caller is assumed to have checked interrupt status 457 * on entry. 458 * 459 * @param node the waiting node 460 * @return on success, the hole; on failure, CANCEL 461 */ 462 private static Object await(Node node, Slot slot) { 463 Thread w = Thread.currentThread(); 464 int spins = SPINS; 465 for (;;) { 466 Object v = node.get(); 467 if (v != null) 468 return v; 469 else if (spins > 0) // Spin-wait phase 470 --spins; 471 else if (node.waiter == null) // Set up to block next 472 node.waiter = w; 473 else if (w.isInterrupted()) // Abort on interrupt 474 tryCancel(node, slot); 475 else // Block 476 LockSupport.park(node); 477 } 478 } 479 480 /** 481 * Waits for (at index 0) and gets the hole filled in by another 482 * thread. Fails if timed out or interrupted before hole filled. 483 * Same basic logic as untimed version, but a bit messier. 484 * 485 * @param node the waiting node 486 * @param nanos the wait time 487 * @return on success, the hole; on failure, CANCEL 488 */ 489 private Object awaitNanos(Node node, Slot slot, long nanos) { 490 int spins = TIMED_SPINS; 491 long lastTime = 0; 492 Thread w = null; 493 for (;;) { 494 Object v = node.get(); 495 if (v != null) 496 return v; 497 long now = System.nanoTime(); 498 if (w == null) 499 w = Thread.currentThread(); 500 else 501 nanos -= now - lastTime; 502 lastTime = now; 503 if (nanos > 0) { 504 if (spins > 0) 505 --spins; 506 else if (node.waiter == null) 507 node.waiter = w; 508 else if (w.isInterrupted()) 509 tryCancel(node, slot); 510 else 511 LockSupport.parkNanos(node, nanos); 512 } 513 else if (tryCancel(node, slot) && !w.isInterrupted()) 514 return scanOnTimeout(node); 515 } 516 } 517 518 /** 519 * Sweeps through arena checking for any waiting threads. Called 520 * only upon return from timeout while waiting in slot 0. When a 521 * thread gives up on a timed wait, it is possible that a 522 * previously-entered thread is still waiting in some other 523 * slot. So we scan to check for any. This is almost always 524 * overkill, but decreases the likelihood of timeouts when there 525 * are other threads present to far less than that in lock-based 526 * exchangers in which earlier-arriving threads may still be 527 * waiting on entry locks. 528 * 529 * @param node the waiting node 530 * @return another thread's item, or CANCEL 531 */ 532 private Object scanOnTimeout(Node node) { 533 Object y; 534 for (int j = arena.length - 1; j >= 0; --j) { 535 Slot slot = arena[j]; 536 if (slot != null) { 537 while ((y = slot.get()) != null) { 538 if (slot.compareAndSet(y, null)) { 539 Node you = (Node)y; 540 if (you.compareAndSet(null, node.item)) { 541 LockSupport.unpark(you.waiter); 542 return you.item; 543 } 544 } 545 } 546 } 547 } 548 return CANCEL; 549 } 550 551 /** 552 * Creates a new Exchanger. 553 */ 554 public Exchanger() { 555 } 556 557 /** 558 * Waits for another thread to arrive at this exchange point (unless 559 * the current thread is {@linkplain Thread#interrupt interrupted}), 560 * and then transfers the given object to it, receiving its object 561 * in return. 562 * 563 * <p>If another thread is already waiting at the exchange point then 564 * it is resumed for thread scheduling purposes and receives the object 565 * passed in by the current thread. The current thread returns immediately, 566 * receiving the object passed to the exchange by that other thread. 567 * 568 * <p>If no other thread is already waiting at the exchange then the 569 * current thread is disabled for thread scheduling purposes and lies 570 * dormant until one of two things happens: 571 * <ul> 572 * <li>Some other thread enters the exchange; or 573 * <li>Some other thread {@linkplain Thread#interrupt interrupts} the current 574 * thread. 575 * </ul> 576 * <p>If the current thread: 577 * <ul> 578 * <li>has its interrupted status set on entry to this method; or 579 * <li>is {@linkplain Thread#interrupt interrupted} while waiting 580 * for the exchange, 581 * </ul> 582 * then {@link InterruptedException} is thrown and the current thread's 583 * interrupted status is cleared. 584 * 585 * @param x the object to exchange 586 * @return the object provided by the other thread 587 * @throws InterruptedException if the current thread was 588 * interrupted while waiting 589 */ 590 public V exchange(V x) throws InterruptedException { 591 if (!Thread.interrupted()) { 592 Object v = doExchange(x == null? NULL_ITEM : x, false, 0); 593 if (v == NULL_ITEM) 594 return null; 595 if (v != CANCEL) 596 return (V)v; 597 Thread.interrupted(); // Clear interrupt status on IE throw 598 } 599 throw new InterruptedException(); 600 } 601 602 /** 603 * Waits for another thread to arrive at this exchange point (unless 604 * the current thread is {@linkplain Thread#interrupt interrupted} or 605 * the specified waiting time elapses), and then transfers the given 606 * object to it, receiving its object in return. 607 * 608 * <p>If another thread is already waiting at the exchange point then 609 * it is resumed for thread scheduling purposes and receives the object 610 * passed in by the current thread. The current thread returns immediately, 611 * receiving the object passed to the exchange by that other thread. 612 * 613 * <p>If no other thread is already waiting at the exchange then the 614 * current thread is disabled for thread scheduling purposes and lies 615 * dormant until one of three things happens: 616 * <ul> 617 * <li>Some other thread enters the exchange; or 618 * <li>Some other thread {@linkplain Thread#interrupt interrupts} 619 * the current thread; or 620 * <li>The specified waiting time elapses. 621 * </ul> 622 * <p>If the current thread: 623 * <ul> 624 * <li>has its interrupted status set on entry to this method; or 625 * <li>is {@linkplain Thread#interrupt interrupted} while waiting 626 * for the exchange, 627 * </ul> 628 * then {@link InterruptedException} is thrown and the current thread's 629 * interrupted status is cleared. 630 * 631 * <p>If the specified waiting time elapses then {@link 632 * TimeoutException} is thrown. If the time is less than or equal 633 * to zero, the method will not wait at all. 634 * 635 * @param x the object to exchange 636 * @param timeout the maximum time to wait 637 * @param unit the time unit of the <tt>timeout</tt> argument 638 * @return the object provided by the other thread 639 * @throws InterruptedException if the current thread was 640 * interrupted while waiting 641 * @throws TimeoutException if the specified waiting time elapses 642 * before another thread enters the exchange 643 */ 644 public V exchange(V x, long timeout, TimeUnit unit) 645 throws InterruptedException, TimeoutException { 646 if (!Thread.interrupted()) { 647 Object v = doExchange(x == null? NULL_ITEM : x, 648 true, unit.toNanos(timeout)); 649 if (v == NULL_ITEM) 650 return null; 651 if (v != CANCEL) 652 return (V)v; 653 if (!Thread.interrupted()) 654 throw new TimeoutException(); 655 } 656 throw new InterruptedException(); 657 } 658 } 659