1/* 2 * Written by Doug Lea with assistance from members of JCP JSR-166 3 * Expert Group and released to the public domain, as explained at 4 * http://creativecommons.org/publicdomain/zero/1.0/ 5 */ 6 7package java.util.concurrent; 8 9import java.util.AbstractQueue; 10import java.util.Collection; 11import java.util.Iterator; 12import java.util.NoSuchElementException; 13import java.util.Queue; 14import java.util.concurrent.locks.LockSupport; 15 16// BEGIN android-note 17// removed link to collections framework docs 18// END android-note 19 20/** 21 * An unbounded {@link TransferQueue} based on linked nodes. 22 * This queue orders elements FIFO (first-in-first-out) with respect 23 * to any given producer. The <em>head</em> of the queue is that 24 * element that has been on the queue the longest time for some 25 * producer. The <em>tail</em> of the queue is that element that has 26 * been on the queue the shortest time for some producer. 27 * 28 * <p>Beware that, unlike in most collections, the {@code size} method 29 * is <em>NOT</em> a constant-time operation. Because of the 30 * asynchronous nature of these queues, determining the current number 31 * of elements requires a traversal of the elements, and so may report 32 * inaccurate results if this collection is modified during traversal. 33 * Additionally, the bulk operations {@code addAll}, 34 * {@code removeAll}, {@code retainAll}, {@code containsAll}, 35 * {@code equals}, and {@code toArray} are <em>not</em> guaranteed 36 * to be performed atomically. For example, an iterator operating 37 * concurrently with an {@code addAll} operation might view only some 38 * of the added elements. 39 * 40 * <p>This class and its iterator implement all of the 41 * <em>optional</em> methods of the {@link Collection} and {@link 42 * Iterator} interfaces. 43 * 44 * <p>Memory consistency effects: As with other concurrent 45 * collections, actions in a thread prior to placing an object into a 46 * {@code LinkedTransferQueue} 47 * <a href="package-summary.html#MemoryVisibility"><i>happen-before</i></a> 48 * actions subsequent to the access or removal of that element from 49 * the {@code LinkedTransferQueue} in another thread. 50 * 51 * @since 1.7 52 * @author Doug Lea 53 * @param <E> the type of elements held in this collection 54 */ 55public class LinkedTransferQueue<E> extends AbstractQueue<E> 56 implements TransferQueue<E>, java.io.Serializable { 57 private static final long serialVersionUID = -3223113410248163686L; 58 59 /* 60 * *** Overview of Dual Queues with Slack *** 61 * 62 * Dual Queues, introduced by Scherer and Scott 63 * (http://www.cs.rice.edu/~wns1/papers/2004-DISC-DDS.pdf) are 64 * (linked) queues in which nodes may represent either data or 65 * requests. When a thread tries to enqueue a data node, but 66 * encounters a request node, it instead "matches" and removes it; 67 * and vice versa for enqueuing requests. Blocking Dual Queues 68 * arrange that threads enqueuing unmatched requests block until 69 * other threads provide the match. Dual Synchronous Queues (see 70 * Scherer, Lea, & Scott 71 * http://www.cs.rochester.edu/u/scott/papers/2009_Scherer_CACM_SSQ.pdf) 72 * additionally arrange that threads enqueuing unmatched data also 73 * block. Dual Transfer Queues support all of these modes, as 74 * dictated by callers. 75 * 76 * A FIFO dual queue may be implemented using a variation of the 77 * Michael & Scott (M&S) lock-free queue algorithm 78 * (http://www.cs.rochester.edu/~scott/papers/1996_PODC_queues.pdf). 79 * It maintains two pointer fields, "head", pointing to a 80 * (matched) node that in turn points to the first actual 81 * (unmatched) queue node (or null if empty); and "tail" that 82 * points to the last node on the queue (or again null if 83 * empty). For example, here is a possible queue with four data 84 * elements: 85 * 86 * head tail 87 * | | 88 * v v 89 * M -> U -> U -> U -> U 90 * 91 * The M&S queue algorithm is known to be prone to scalability and 92 * overhead limitations when maintaining (via CAS) these head and 93 * tail pointers. This has led to the development of 94 * contention-reducing variants such as elimination arrays (see 95 * Moir et al http://portal.acm.org/citation.cfm?id=1074013) and 96 * optimistic back pointers (see Ladan-Mozes & Shavit 97 * http://people.csail.mit.edu/edya/publications/OptimisticFIFOQueue-journal.pdf). 98 * However, the nature of dual queues enables a simpler tactic for 99 * improving M&S-style implementations when dual-ness is needed. 100 * 101 * In a dual queue, each node must atomically maintain its match 102 * status. While there are other possible variants, we implement 103 * this here as: for a data-mode node, matching entails CASing an 104 * "item" field from a non-null data value to null upon match, and 105 * vice-versa for request nodes, CASing from null to a data 106 * value. (Note that the linearization properties of this style of 107 * queue are easy to verify -- elements are made available by 108 * linking, and unavailable by matching.) Compared to plain M&S 109 * queues, this property of dual queues requires one additional 110 * successful atomic operation per enq/deq pair. But it also 111 * enables lower cost variants of queue maintenance mechanics. (A 112 * variation of this idea applies even for non-dual queues that 113 * support deletion of interior elements, such as 114 * j.u.c.ConcurrentLinkedQueue.) 115 * 116 * Once a node is matched, its match status can never again 117 * change. We may thus arrange that the linked list of them 118 * contain a prefix of zero or more matched nodes, followed by a 119 * suffix of zero or more unmatched nodes. (Note that we allow 120 * both the prefix and suffix to be zero length, which in turn 121 * means that we do not use a dummy header.) If we were not 122 * concerned with either time or space efficiency, we could 123 * correctly perform enqueue and dequeue operations by traversing 124 * from a pointer to the initial node; CASing the item of the 125 * first unmatched node on match and CASing the next field of the 126 * trailing node on appends. (Plus some special-casing when 127 * initially empty). While this would be a terrible idea in 128 * itself, it does have the benefit of not requiring ANY atomic 129 * updates on head/tail fields. 130 * 131 * We introduce here an approach that lies between the extremes of 132 * never versus always updating queue (head and tail) pointers. 133 * This offers a tradeoff between sometimes requiring extra 134 * traversal steps to locate the first and/or last unmatched 135 * nodes, versus the reduced overhead and contention of fewer 136 * updates to queue pointers. For example, a possible snapshot of 137 * a queue is: 138 * 139 * head tail 140 * | | 141 * v v 142 * M -> M -> U -> U -> U -> U 143 * 144 * The best value for this "slack" (the targeted maximum distance 145 * between the value of "head" and the first unmatched node, and 146 * similarly for "tail") is an empirical matter. We have found 147 * that using very small constants in the range of 1-3 work best 148 * over a range of platforms. Larger values introduce increasing 149 * costs of cache misses and risks of long traversal chains, while 150 * smaller values increase CAS contention and overhead. 151 * 152 * Dual queues with slack differ from plain M&S dual queues by 153 * virtue of only sometimes updating head or tail pointers when 154 * matching, appending, or even traversing nodes; in order to 155 * maintain a targeted slack. The idea of "sometimes" may be 156 * operationalized in several ways. The simplest is to use a 157 * per-operation counter incremented on each traversal step, and 158 * to try (via CAS) to update the associated queue pointer 159 * whenever the count exceeds a threshold. Another, that requires 160 * more overhead, is to use random number generators to update 161 * with a given probability per traversal step. 162 * 163 * In any strategy along these lines, because CASes updating 164 * fields may fail, the actual slack may exceed targeted 165 * slack. However, they may be retried at any time to maintain 166 * targets. Even when using very small slack values, this 167 * approach works well for dual queues because it allows all 168 * operations up to the point of matching or appending an item 169 * (hence potentially allowing progress by another thread) to be 170 * read-only, thus not introducing any further contention. As 171 * described below, we implement this by performing slack 172 * maintenance retries only after these points. 173 * 174 * As an accompaniment to such techniques, traversal overhead can 175 * be further reduced without increasing contention of head 176 * pointer updates: Threads may sometimes shortcut the "next" link 177 * path from the current "head" node to be closer to the currently 178 * known first unmatched node, and similarly for tail. Again, this 179 * may be triggered with using thresholds or randomization. 180 * 181 * These ideas must be further extended to avoid unbounded amounts 182 * of costly-to-reclaim garbage caused by the sequential "next" 183 * links of nodes starting at old forgotten head nodes: As first 184 * described in detail by Boehm 185 * (http://portal.acm.org/citation.cfm?doid=503272.503282) if a GC 186 * delays noticing that any arbitrarily old node has become 187 * garbage, all newer dead nodes will also be unreclaimed. 188 * (Similar issues arise in non-GC environments.) To cope with 189 * this in our implementation, upon CASing to advance the head 190 * pointer, we set the "next" link of the previous head to point 191 * only to itself; thus limiting the length of connected dead lists. 192 * (We also take similar care to wipe out possibly garbage 193 * retaining values held in other Node fields.) However, doing so 194 * adds some further complexity to traversal: If any "next" 195 * pointer links to itself, it indicates that the current thread 196 * has lagged behind a head-update, and so the traversal must 197 * continue from the "head". Traversals trying to find the 198 * current tail starting from "tail" may also encounter 199 * self-links, in which case they also continue at "head". 200 * 201 * It is tempting in slack-based scheme to not even use CAS for 202 * updates (similarly to Ladan-Mozes & Shavit). However, this 203 * cannot be done for head updates under the above link-forgetting 204 * mechanics because an update may leave head at a detached node. 205 * And while direct writes are possible for tail updates, they 206 * increase the risk of long retraversals, and hence long garbage 207 * chains, which can be much more costly than is worthwhile 208 * considering that the cost difference of performing a CAS vs 209 * write is smaller when they are not triggered on each operation 210 * (especially considering that writes and CASes equally require 211 * additional GC bookkeeping ("write barriers") that are sometimes 212 * more costly than the writes themselves because of contention). 213 * 214 * *** Overview of implementation *** 215 * 216 * We use a threshold-based approach to updates, with a slack 217 * threshold of two -- that is, we update head/tail when the 218 * current pointer appears to be two or more steps away from the 219 * first/last node. The slack value is hard-wired: a path greater 220 * than one is naturally implemented by checking equality of 221 * traversal pointers except when the list has only one element, 222 * in which case we keep slack threshold at one. Avoiding tracking 223 * explicit counts across method calls slightly simplifies an 224 * already-messy implementation. Using randomization would 225 * probably work better if there were a low-quality dirt-cheap 226 * per-thread one available, but even ThreadLocalRandom is too 227 * heavy for these purposes. 228 * 229 * With such a small slack threshold value, it is not worthwhile 230 * to augment this with path short-circuiting (i.e., unsplicing 231 * interior nodes) except in the case of cancellation/removal (see 232 * below). 233 * 234 * We allow both the head and tail fields to be null before any 235 * nodes are enqueued; initializing upon first append. This 236 * simplifies some other logic, as well as providing more 237 * efficient explicit control paths instead of letting JVMs insert 238 * implicit NullPointerExceptions when they are null. While not 239 * currently fully implemented, we also leave open the possibility 240 * of re-nulling these fields when empty (which is complicated to 241 * arrange, for little benefit.) 242 * 243 * All enqueue/dequeue operations are handled by the single method 244 * "xfer" with parameters indicating whether to act as some form 245 * of offer, put, poll, take, or transfer (each possibly with 246 * timeout). The relative complexity of using one monolithic 247 * method outweighs the code bulk and maintenance problems of 248 * using separate methods for each case. 249 * 250 * Operation consists of up to three phases. The first is 251 * implemented within method xfer, the second in tryAppend, and 252 * the third in method awaitMatch. 253 * 254 * 1. Try to match an existing node 255 * 256 * Starting at head, skip already-matched nodes until finding 257 * an unmatched node of opposite mode, if one exists, in which 258 * case matching it and returning, also if necessary updating 259 * head to one past the matched node (or the node itself if the 260 * list has no other unmatched nodes). If the CAS misses, then 261 * a loop retries advancing head by two steps until either 262 * success or the slack is at most two. By requiring that each 263 * attempt advances head by two (if applicable), we ensure that 264 * the slack does not grow without bound. Traversals also check 265 * if the initial head is now off-list, in which case they 266 * start at the new head. 267 * 268 * If no candidates are found and the call was untimed 269 * poll/offer, (argument "how" is NOW) return. 270 * 271 * 2. Try to append a new node (method tryAppend) 272 * 273 * Starting at current tail pointer, find the actual last node 274 * and try to append a new node (or if head was null, establish 275 * the first node). Nodes can be appended only if their 276 * predecessors are either already matched or are of the same 277 * mode. If we detect otherwise, then a new node with opposite 278 * mode must have been appended during traversal, so we must 279 * restart at phase 1. The traversal and update steps are 280 * otherwise similar to phase 1: Retrying upon CAS misses and 281 * checking for staleness. In particular, if a self-link is 282 * encountered, then we can safely jump to a node on the list 283 * by continuing the traversal at current head. 284 * 285 * On successful append, if the call was ASYNC, return. 286 * 287 * 3. Await match or cancellation (method awaitMatch) 288 * 289 * Wait for another thread to match node; instead cancelling if 290 * the current thread was interrupted or the wait timed out. On 291 * multiprocessors, we use front-of-queue spinning: If a node 292 * appears to be the first unmatched node in the queue, it 293 * spins a bit before blocking. In either case, before blocking 294 * it tries to unsplice any nodes between the current "head" 295 * and the first unmatched node. 296 * 297 * Front-of-queue spinning vastly improves performance of 298 * heavily contended queues. And so long as it is relatively 299 * brief and "quiet", spinning does not much impact performance 300 * of less-contended queues. During spins threads check their 301 * interrupt status and generate a thread-local random number 302 * to decide to occasionally perform a Thread.yield. While 303 * yield has underdefined specs, we assume that it might help, 304 * and will not hurt, in limiting impact of spinning on busy 305 * systems. We also use smaller (1/2) spins for nodes that are 306 * not known to be front but whose predecessors have not 307 * blocked -- these "chained" spins avoid artifacts of 308 * front-of-queue rules which otherwise lead to alternating 309 * nodes spinning vs blocking. Further, front threads that 310 * represent phase changes (from data to request node or vice 311 * versa) compared to their predecessors receive additional 312 * chained spins, reflecting longer paths typically required to 313 * unblock threads during phase changes. 314 * 315 * 316 * ** Unlinking removed interior nodes ** 317 * 318 * In addition to minimizing garbage retention via self-linking 319 * described above, we also unlink removed interior nodes. These 320 * may arise due to timed out or interrupted waits, or calls to 321 * remove(x) or Iterator.remove. Normally, given a node that was 322 * at one time known to be the predecessor of some node s that is 323 * to be removed, we can unsplice s by CASing the next field of 324 * its predecessor if it still points to s (otherwise s must 325 * already have been removed or is now offlist). But there are two 326 * situations in which we cannot guarantee to make node s 327 * unreachable in this way: (1) If s is the trailing node of list 328 * (i.e., with null next), then it is pinned as the target node 329 * for appends, so can only be removed later after other nodes are 330 * appended. (2) We cannot necessarily unlink s given a 331 * predecessor node that is matched (including the case of being 332 * cancelled): the predecessor may already be unspliced, in which 333 * case some previous reachable node may still point to s. 334 * (For further explanation see Herlihy & Shavit "The Art of 335 * Multiprocessor Programming" chapter 9). Although, in both 336 * cases, we can rule out the need for further action if either s 337 * or its predecessor are (or can be made to be) at, or fall off 338 * from, the head of list. 339 * 340 * Without taking these into account, it would be possible for an 341 * unbounded number of supposedly removed nodes to remain 342 * reachable. Situations leading to such buildup are uncommon but 343 * can occur in practice; for example when a series of short timed 344 * calls to poll repeatedly time out but never otherwise fall off 345 * the list because of an untimed call to take at the front of the 346 * queue. 347 * 348 * When these cases arise, rather than always retraversing the 349 * entire list to find an actual predecessor to unlink (which 350 * won't help for case (1) anyway), we record a conservative 351 * estimate of possible unsplice failures (in "sweepVotes"). 352 * We trigger a full sweep when the estimate exceeds a threshold 353 * ("SWEEP_THRESHOLD") indicating the maximum number of estimated 354 * removal failures to tolerate before sweeping through, unlinking 355 * cancelled nodes that were not unlinked upon initial removal. 356 * We perform sweeps by the thread hitting threshold (rather than 357 * background threads or by spreading work to other threads) 358 * because in the main contexts in which removal occurs, the 359 * caller is already timed-out, cancelled, or performing a 360 * potentially O(n) operation (e.g. remove(x)), none of which are 361 * time-critical enough to warrant the overhead that alternatives 362 * would impose on other threads. 363 * 364 * Because the sweepVotes estimate is conservative, and because 365 * nodes become unlinked "naturally" as they fall off the head of 366 * the queue, and because we allow votes to accumulate even while 367 * sweeps are in progress, there are typically significantly fewer 368 * such nodes than estimated. Choice of a threshold value 369 * balances the likelihood of wasted effort and contention, versus 370 * providing a worst-case bound on retention of interior nodes in 371 * quiescent queues. The value defined below was chosen 372 * empirically to balance these under various timeout scenarios. 373 * 374 * Note that we cannot self-link unlinked interior nodes during 375 * sweeps. However, the associated garbage chains terminate when 376 * some successor ultimately falls off the head of the list and is 377 * self-linked. 378 */ 379 380 /** True if on multiprocessor */ 381 private static final boolean MP = 382 Runtime.getRuntime().availableProcessors() > 1; 383 384 /** 385 * The number of times to spin (with randomly interspersed calls 386 * to Thread.yield) on multiprocessor before blocking when a node 387 * is apparently the first waiter in the queue. See above for 388 * explanation. Must be a power of two. The value is empirically 389 * derived -- it works pretty well across a variety of processors, 390 * numbers of CPUs, and OSes. 391 */ 392 private static final int FRONT_SPINS = 1 << 7; 393 394 /** 395 * The number of times to spin before blocking when a node is 396 * preceded by another node that is apparently spinning. Also 397 * serves as an increment to FRONT_SPINS on phase changes, and as 398 * base average frequency for yielding during spins. Must be a 399 * power of two. 400 */ 401 private static final int CHAINED_SPINS = FRONT_SPINS >>> 1; 402 403 /** 404 * The maximum number of estimated removal failures (sweepVotes) 405 * to tolerate before sweeping through the queue unlinking 406 * cancelled nodes that were not unlinked upon initial 407 * removal. See above for explanation. The value must be at least 408 * two to avoid useless sweeps when removing trailing nodes. 409 */ 410 static final int SWEEP_THRESHOLD = 32; 411 412 /** 413 * Queue nodes. Uses Object, not E, for items to allow forgetting 414 * them after use. Relies heavily on Unsafe mechanics to minimize 415 * unnecessary ordering constraints: Writes that are intrinsically 416 * ordered wrt other accesses or CASes use simple relaxed forms. 417 */ 418 static final class Node { 419 final boolean isData; // false if this is a request node 420 volatile Object item; // initially non-null if isData; CASed to match 421 volatile Node next; 422 volatile Thread waiter; // null until waiting 423 424 // CAS methods for fields 425 final boolean casNext(Node cmp, Node val) { 426 return UNSAFE.compareAndSwapObject(this, nextOffset, cmp, val); 427 } 428 429 final boolean casItem(Object cmp, Object val) { 430 // assert cmp == null || cmp.getClass() != Node.class; 431 return UNSAFE.compareAndSwapObject(this, itemOffset, cmp, val); 432 } 433 434 /** 435 * Constructs a new node. Uses relaxed write because item can 436 * only be seen after publication via casNext. 437 */ 438 Node(Object item, boolean isData) { 439 UNSAFE.putObject(this, itemOffset, item); // relaxed write 440 this.isData = isData; 441 } 442 443 /** 444 * Links node to itself to avoid garbage retention. Called 445 * only after CASing head field, so uses relaxed write. 446 */ 447 final void forgetNext() { 448 UNSAFE.putObject(this, nextOffset, this); 449 } 450 451 /** 452 * Sets item to self and waiter to null, to avoid garbage 453 * retention after matching or cancelling. Uses relaxed writes 454 * because order is already constrained in the only calling 455 * contexts: item is forgotten only after volatile/atomic 456 * mechanics that extract items. Similarly, clearing waiter 457 * follows either CAS or return from park (if ever parked; 458 * else we don't care). 459 */ 460 final void forgetContents() { 461 UNSAFE.putObject(this, itemOffset, this); 462 UNSAFE.putObject(this, waiterOffset, null); 463 } 464 465 /** 466 * Returns true if this node has been matched, including the 467 * case of artificial matches due to cancellation. 468 */ 469 final boolean isMatched() { 470 Object x = item; 471 return (x == this) || ((x == null) == isData); 472 } 473 474 /** 475 * Returns true if this is an unmatched request node. 476 */ 477 final boolean isUnmatchedRequest() { 478 return !isData && item == null; 479 } 480 481 /** 482 * Returns true if a node with the given mode cannot be 483 * appended to this node because this node is unmatched and 484 * has opposite data mode. 485 */ 486 final boolean cannotPrecede(boolean haveData) { 487 boolean d = isData; 488 Object x; 489 return d != haveData && (x = item) != this && (x != null) == d; 490 } 491 492 /** 493 * Tries to artificially match a data node -- used by remove. 494 */ 495 final boolean tryMatchData() { 496 // assert isData; 497 Object x = item; 498 if (x != null && x != this && casItem(x, null)) { 499 LockSupport.unpark(waiter); 500 return true; 501 } 502 return false; 503 } 504 505 private static final long serialVersionUID = -3375979862319811754L; 506 507 // Unsafe mechanics 508 private static final sun.misc.Unsafe UNSAFE; 509 private static final long itemOffset; 510 private static final long nextOffset; 511 private static final long waiterOffset; 512 static { 513 try { 514 UNSAFE = sun.misc.Unsafe.getUnsafe(); 515 Class<?> k = Node.class; 516 itemOffset = UNSAFE.objectFieldOffset 517 (k.getDeclaredField("item")); 518 nextOffset = UNSAFE.objectFieldOffset 519 (k.getDeclaredField("next")); 520 waiterOffset = UNSAFE.objectFieldOffset 521 (k.getDeclaredField("waiter")); 522 } catch (Exception e) { 523 throw new Error(e); 524 } 525 } 526 } 527 528 /** head of the queue; null until first enqueue */ 529 transient volatile Node head; 530 531 /** tail of the queue; null until first append */ 532 private transient volatile Node tail; 533 534 /** The number of apparent failures to unsplice removed nodes */ 535 private transient volatile int sweepVotes; 536 537 // CAS methods for fields 538 private boolean casTail(Node cmp, Node val) { 539 return UNSAFE.compareAndSwapObject(this, tailOffset, cmp, val); 540 } 541 542 private boolean casHead(Node cmp, Node val) { 543 return UNSAFE.compareAndSwapObject(this, headOffset, cmp, val); 544 } 545 546 private boolean casSweepVotes(int cmp, int val) { 547 return UNSAFE.compareAndSwapInt(this, sweepVotesOffset, cmp, val); 548 } 549 550 /* 551 * Possible values for "how" argument in xfer method. 552 */ 553 private static final int NOW = 0; // for untimed poll, tryTransfer 554 private static final int ASYNC = 1; // for offer, put, add 555 private static final int SYNC = 2; // for transfer, take 556 private static final int TIMED = 3; // for timed poll, tryTransfer 557 558 @SuppressWarnings("unchecked") 559 static <E> E cast(Object item) { 560 // assert item == null || item.getClass() != Node.class; 561 return (E) item; 562 } 563 564 /** 565 * Implements all queuing methods. See above for explanation. 566 * 567 * @param e the item or null for take 568 * @param haveData true if this is a put, else a take 569 * @param how NOW, ASYNC, SYNC, or TIMED 570 * @param nanos timeout in nanosecs, used only if mode is TIMED 571 * @return an item if matched, else e 572 * @throws NullPointerException if haveData mode but e is null 573 */ 574 private E xfer(E e, boolean haveData, int how, long nanos) { 575 if (haveData && (e == null)) 576 throw new NullPointerException(); 577 Node s = null; // the node to append, if needed 578 579 retry: 580 for (;;) { // restart on append race 581 582 for (Node h = head, p = h; p != null;) { // find & match first node 583 boolean isData = p.isData; 584 Object item = p.item; 585 if (item != p && (item != null) == isData) { // unmatched 586 if (isData == haveData) // can't match 587 break; 588 if (p.casItem(item, e)) { // match 589 for (Node q = p; q != h;) { 590 Node n = q.next; // update by 2 unless singleton 591 if (head == h && casHead(h, n == null ? q : n)) { 592 h.forgetNext(); 593 break; 594 } // advance and retry 595 if ((h = head) == null || 596 (q = h.next) == null || !q.isMatched()) 597 break; // unless slack < 2 598 } 599 LockSupport.unpark(p.waiter); 600 return LinkedTransferQueue.<E>cast(item); 601 } 602 } 603 Node n = p.next; 604 p = (p != n) ? n : (h = head); // Use head if p offlist 605 } 606 607 if (how != NOW) { // No matches available 608 if (s == null) 609 s = new Node(e, haveData); 610 Node pred = tryAppend(s, haveData); 611 if (pred == null) 612 continue retry; // lost race vs opposite mode 613 if (how != ASYNC) 614 return awaitMatch(s, pred, e, (how == TIMED), nanos); 615 } 616 return e; // not waiting 617 } 618 } 619 620 /** 621 * Tries to append node s as tail. 622 * 623 * @param s the node to append 624 * @param haveData true if appending in data mode 625 * @return null on failure due to losing race with append in 626 * different mode, else s's predecessor, or s itself if no 627 * predecessor 628 */ 629 private Node tryAppend(Node s, boolean haveData) { 630 for (Node t = tail, p = t;;) { // move p to last node and append 631 Node n, u; // temps for reads of next & tail 632 if (p == null && (p = head) == null) { 633 if (casHead(null, s)) 634 return s; // initialize 635 } 636 else if (p.cannotPrecede(haveData)) 637 return null; // lost race vs opposite mode 638 else if ((n = p.next) != null) // not last; keep traversing 639 p = p != t && t != (u = tail) ? (t = u) : // stale tail 640 (p != n) ? n : null; // restart if off list 641 else if (!p.casNext(null, s)) 642 p = p.next; // re-read on CAS failure 643 else { 644 if (p != t) { // update if slack now >= 2 645 while ((tail != t || !casTail(t, s)) && 646 (t = tail) != null && 647 (s = t.next) != null && // advance and retry 648 (s = s.next) != null && s != t); 649 } 650 return p; 651 } 652 } 653 } 654 655 /** 656 * Spins/yields/blocks until node s is matched or caller gives up. 657 * 658 * @param s the waiting node 659 * @param pred the predecessor of s, or s itself if it has no 660 * predecessor, or null if unknown (the null case does not occur 661 * in any current calls but may in possible future extensions) 662 * @param e the comparison value for checking match 663 * @param timed if true, wait only until timeout elapses 664 * @param nanos timeout in nanosecs, used only if timed is true 665 * @return matched item, or e if unmatched on interrupt or timeout 666 */ 667 private E awaitMatch(Node s, Node pred, E e, boolean timed, long nanos) { 668 final long deadline = timed ? System.nanoTime() + nanos : 0L; 669 Thread w = Thread.currentThread(); 670 int spins = -1; // initialized after first item and cancel checks 671 ThreadLocalRandom randomYields = null; // bound if needed 672 673 for (;;) { 674 Object item = s.item; 675 if (item != e) { // matched 676 // assert item != s; 677 s.forgetContents(); // avoid garbage 678 return LinkedTransferQueue.<E>cast(item); 679 } 680 if ((w.isInterrupted() || (timed && nanos <= 0)) && 681 s.casItem(e, s)) { // cancel 682 unsplice(pred, s); 683 return e; 684 } 685 686 if (spins < 0) { // establish spins at/near front 687 if ((spins = spinsFor(pred, s.isData)) > 0) 688 randomYields = ThreadLocalRandom.current(); 689 } 690 else if (spins > 0) { // spin 691 --spins; 692 if (randomYields.nextInt(CHAINED_SPINS) == 0) 693 Thread.yield(); // occasionally yield 694 } 695 else if (s.waiter == null) { 696 s.waiter = w; // request unpark then recheck 697 } 698 else if (timed) { 699 nanos = deadline - System.nanoTime(); 700 if (nanos > 0L) 701 LockSupport.parkNanos(this, nanos); 702 } 703 else { 704 LockSupport.park(this); 705 } 706 } 707 } 708 709 /** 710 * Returns spin/yield value for a node with given predecessor and 711 * data mode. See above for explanation. 712 */ 713 private static int spinsFor(Node pred, boolean haveData) { 714 if (MP && pred != null) { 715 if (pred.isData != haveData) // phase change 716 return FRONT_SPINS + CHAINED_SPINS; 717 if (pred.isMatched()) // probably at front 718 return FRONT_SPINS; 719 if (pred.waiter == null) // pred apparently spinning 720 return CHAINED_SPINS; 721 } 722 return 0; 723 } 724 725 /* -------------- Traversal methods -------------- */ 726 727 /** 728 * Returns the successor of p, or the head node if p.next has been 729 * linked to self, which will only be true if traversing with a 730 * stale pointer that is now off the list. 731 */ 732 final Node succ(Node p) { 733 Node next = p.next; 734 return (p == next) ? head : next; 735 } 736 737 /** 738 * Returns the first unmatched node of the given mode, or null if 739 * none. Used by methods isEmpty, hasWaitingConsumer. 740 */ 741 private Node firstOfMode(boolean isData) { 742 for (Node p = head; p != null; p = succ(p)) { 743 if (!p.isMatched()) 744 return (p.isData == isData) ? p : null; 745 } 746 return null; 747 } 748 749 /** 750 * Returns the item in the first unmatched node with isData; or 751 * null if none. Used by peek. 752 */ 753 private E firstDataItem() { 754 for (Node p = head; p != null; p = succ(p)) { 755 Object item = p.item; 756 if (p.isData) { 757 if (item != null && item != p) 758 return LinkedTransferQueue.<E>cast(item); 759 } 760 else if (item == null) 761 return null; 762 } 763 return null; 764 } 765 766 /** 767 * Traverses and counts unmatched nodes of the given mode. 768 * Used by methods size and getWaitingConsumerCount. 769 */ 770 private int countOfMode(boolean data) { 771 int count = 0; 772 for (Node p = head; p != null; ) { 773 if (!p.isMatched()) { 774 if (p.isData != data) 775 return 0; 776 if (++count == Integer.MAX_VALUE) // saturated 777 break; 778 } 779 Node n = p.next; 780 if (n != p) 781 p = n; 782 else { 783 count = 0; 784 p = head; 785 } 786 } 787 return count; 788 } 789 790 final class Itr implements Iterator<E> { 791 private Node nextNode; // next node to return item for 792 private E nextItem; // the corresponding item 793 private Node lastRet; // last returned node, to support remove 794 private Node lastPred; // predecessor to unlink lastRet 795 796 /** 797 * Moves to next node after prev, or first node if prev null. 798 */ 799 private void advance(Node prev) { 800 /* 801 * To track and avoid buildup of deleted nodes in the face 802 * of calls to both Queue.remove and Itr.remove, we must 803 * include variants of unsplice and sweep upon each 804 * advance: Upon Itr.remove, we may need to catch up links 805 * from lastPred, and upon other removes, we might need to 806 * skip ahead from stale nodes and unsplice deleted ones 807 * found while advancing. 808 */ 809 810 Node r, b; // reset lastPred upon possible deletion of lastRet 811 if ((r = lastRet) != null && !r.isMatched()) 812 lastPred = r; // next lastPred is old lastRet 813 else if ((b = lastPred) == null || b.isMatched()) 814 lastPred = null; // at start of list 815 else { 816 Node s, n; // help with removal of lastPred.next 817 while ((s = b.next) != null && 818 s != b && s.isMatched() && 819 (n = s.next) != null && n != s) 820 b.casNext(s, n); 821 } 822 823 this.lastRet = prev; 824 825 for (Node p = prev, s, n;;) { 826 s = (p == null) ? head : p.next; 827 if (s == null) 828 break; 829 else if (s == p) { 830 p = null; 831 continue; 832 } 833 Object item = s.item; 834 if (s.isData) { 835 if (item != null && item != s) { 836 nextItem = LinkedTransferQueue.<E>cast(item); 837 nextNode = s; 838 return; 839 } 840 } 841 else if (item == null) 842 break; 843 // assert s.isMatched(); 844 if (p == null) 845 p = s; 846 else if ((n = s.next) == null) 847 break; 848 else if (s == n) 849 p = null; 850 else 851 p.casNext(s, n); 852 } 853 nextNode = null; 854 nextItem = null; 855 } 856 857 Itr() { 858 advance(null); 859 } 860 861 public final boolean hasNext() { 862 return nextNode != null; 863 } 864 865 public final E next() { 866 Node p = nextNode; 867 if (p == null) throw new NoSuchElementException(); 868 E e = nextItem; 869 advance(p); 870 return e; 871 } 872 873 public final void remove() { 874 final Node lastRet = this.lastRet; 875 if (lastRet == null) 876 throw new IllegalStateException(); 877 this.lastRet = null; 878 if (lastRet.tryMatchData()) 879 unsplice(lastPred, lastRet); 880 } 881 } 882 883 /* -------------- Removal methods -------------- */ 884 885 /** 886 * Unsplices (now or later) the given deleted/cancelled node with 887 * the given predecessor. 888 * 889 * @param pred a node that was at one time known to be the 890 * predecessor of s, or null or s itself if s is/was at head 891 * @param s the node to be unspliced 892 */ 893 final void unsplice(Node pred, Node s) { 894 s.forgetContents(); // forget unneeded fields 895 /* 896 * See above for rationale. Briefly: if pred still points to 897 * s, try to unlink s. If s cannot be unlinked, because it is 898 * trailing node or pred might be unlinked, and neither pred 899 * nor s are head or offlist, add to sweepVotes, and if enough 900 * votes have accumulated, sweep. 901 */ 902 if (pred != null && pred != s && pred.next == s) { 903 Node n = s.next; 904 if (n == null || 905 (n != s && pred.casNext(s, n) && pred.isMatched())) { 906 for (;;) { // check if at, or could be, head 907 Node h = head; 908 if (h == pred || h == s || h == null) 909 return; // at head or list empty 910 if (!h.isMatched()) 911 break; 912 Node hn = h.next; 913 if (hn == null) 914 return; // now empty 915 if (hn != h && casHead(h, hn)) 916 h.forgetNext(); // advance head 917 } 918 if (pred.next != pred && s.next != s) { // recheck if offlist 919 for (;;) { // sweep now if enough votes 920 int v = sweepVotes; 921 if (v < SWEEP_THRESHOLD) { 922 if (casSweepVotes(v, v + 1)) 923 break; 924 } 925 else if (casSweepVotes(v, 0)) { 926 sweep(); 927 break; 928 } 929 } 930 } 931 } 932 } 933 } 934 935 /** 936 * Unlinks matched (typically cancelled) nodes encountered in a 937 * traversal from head. 938 */ 939 private void sweep() { 940 for (Node p = head, s, n; p != null && (s = p.next) != null; ) { 941 if (!s.isMatched()) 942 // Unmatched nodes are never self-linked 943 p = s; 944 else if ((n = s.next) == null) // trailing node is pinned 945 break; 946 else if (s == n) // stale 947 // No need to also check for p == s, since that implies s == n 948 p = head; 949 else 950 p.casNext(s, n); 951 } 952 } 953 954 /** 955 * Main implementation of remove(Object) 956 */ 957 private boolean findAndRemove(Object e) { 958 if (e != null) { 959 for (Node pred = null, p = head; p != null; ) { 960 Object item = p.item; 961 if (p.isData) { 962 if (item != null && item != p && e.equals(item) && 963 p.tryMatchData()) { 964 unsplice(pred, p); 965 return true; 966 } 967 } 968 else if (item == null) 969 break; 970 pred = p; 971 if ((p = p.next) == pred) { // stale 972 pred = null; 973 p = head; 974 } 975 } 976 } 977 return false; 978 } 979 980 /** 981 * Creates an initially empty {@code LinkedTransferQueue}. 982 */ 983 public LinkedTransferQueue() { 984 } 985 986 /** 987 * Creates a {@code LinkedTransferQueue} 988 * initially containing the elements of the given collection, 989 * added in traversal order of the collection's iterator. 990 * 991 * @param c the collection of elements to initially contain 992 * @throws NullPointerException if the specified collection or any 993 * of its elements are null 994 */ 995 public LinkedTransferQueue(Collection<? extends E> c) { 996 this(); 997 addAll(c); 998 } 999 1000 /** 1001 * Inserts the specified element at the tail of this queue. 1002 * As the queue is unbounded, this method will never block. 1003 * 1004 * @throws NullPointerException if the specified element is null 1005 */ 1006 public void put(E e) { 1007 xfer(e, true, ASYNC, 0); 1008 } 1009 1010 /** 1011 * Inserts the specified element at the tail of this queue. 1012 * As the queue is unbounded, this method will never block or 1013 * return {@code false}. 1014 * 1015 * @return {@code true} (as specified by 1016 * {@link java.util.concurrent.BlockingQueue#offer(Object,long,TimeUnit) 1017 * BlockingQueue.offer}) 1018 * @throws NullPointerException if the specified element is null 1019 */ 1020 public boolean offer(E e, long timeout, TimeUnit unit) { 1021 xfer(e, true, ASYNC, 0); 1022 return true; 1023 } 1024 1025 /** 1026 * Inserts the specified element at the tail of this queue. 1027 * As the queue is unbounded, this method will never return {@code false}. 1028 * 1029 * @return {@code true} (as specified by {@link Queue#offer}) 1030 * @throws NullPointerException if the specified element is null 1031 */ 1032 public boolean offer(E e) { 1033 xfer(e, true, ASYNC, 0); 1034 return true; 1035 } 1036 1037 /** 1038 * Inserts the specified element at the tail of this queue. 1039 * As the queue is unbounded, this method will never throw 1040 * {@link IllegalStateException} or return {@code false}. 1041 * 1042 * @return {@code true} (as specified by {@link Collection#add}) 1043 * @throws NullPointerException if the specified element is null 1044 */ 1045 public boolean add(E e) { 1046 xfer(e, true, ASYNC, 0); 1047 return true; 1048 } 1049 1050 /** 1051 * Transfers the element to a waiting consumer immediately, if possible. 1052 * 1053 * <p>More precisely, transfers the specified element immediately 1054 * if there exists a consumer already waiting to receive it (in 1055 * {@link #take} or timed {@link #poll(long,TimeUnit) poll}), 1056 * otherwise returning {@code false} without enqueuing the element. 1057 * 1058 * @throws NullPointerException if the specified element is null 1059 */ 1060 public boolean tryTransfer(E e) { 1061 return xfer(e, true, NOW, 0) == null; 1062 } 1063 1064 /** 1065 * Transfers the element to a consumer, waiting if necessary to do so. 1066 * 1067 * <p>More precisely, transfers the specified element immediately 1068 * if there exists a consumer already waiting to receive it (in 1069 * {@link #take} or timed {@link #poll(long,TimeUnit) poll}), 1070 * else inserts the specified element at the tail of this queue 1071 * and waits until the element is received by a consumer. 1072 * 1073 * @throws NullPointerException if the specified element is null 1074 */ 1075 public void transfer(E e) throws InterruptedException { 1076 if (xfer(e, true, SYNC, 0) != null) { 1077 Thread.interrupted(); // failure possible only due to interrupt 1078 throw new InterruptedException(); 1079 } 1080 } 1081 1082 /** 1083 * Transfers the element to a consumer if it is possible to do so 1084 * before the timeout elapses. 1085 * 1086 * <p>More precisely, transfers the specified element immediately 1087 * if there exists a consumer already waiting to receive it (in 1088 * {@link #take} or timed {@link #poll(long,TimeUnit) poll}), 1089 * else inserts the specified element at the tail of this queue 1090 * and waits until the element is received by a consumer, 1091 * returning {@code false} if the specified wait time elapses 1092 * before the element can be transferred. 1093 * 1094 * @throws NullPointerException if the specified element is null 1095 */ 1096 public boolean tryTransfer(E e, long timeout, TimeUnit unit) 1097 throws InterruptedException { 1098 if (xfer(e, true, TIMED, unit.toNanos(timeout)) == null) 1099 return true; 1100 if (!Thread.interrupted()) 1101 return false; 1102 throw new InterruptedException(); 1103 } 1104 1105 public E take() throws InterruptedException { 1106 E e = xfer(null, false, SYNC, 0); 1107 if (e != null) 1108 return e; 1109 Thread.interrupted(); 1110 throw new InterruptedException(); 1111 } 1112 1113 public E poll(long timeout, TimeUnit unit) throws InterruptedException { 1114 E e = xfer(null, false, TIMED, unit.toNanos(timeout)); 1115 if (e != null || !Thread.interrupted()) 1116 return e; 1117 throw new InterruptedException(); 1118 } 1119 1120 public E poll() { 1121 return xfer(null, false, NOW, 0); 1122 } 1123 1124 /** 1125 * @throws NullPointerException {@inheritDoc} 1126 * @throws IllegalArgumentException {@inheritDoc} 1127 */ 1128 public int drainTo(Collection<? super E> c) { 1129 if (c == null) 1130 throw new NullPointerException(); 1131 if (c == this) 1132 throw new IllegalArgumentException(); 1133 int n = 0; 1134 for (E e; (e = poll()) != null;) { 1135 c.add(e); 1136 ++n; 1137 } 1138 return n; 1139 } 1140 1141 /** 1142 * @throws NullPointerException {@inheritDoc} 1143 * @throws IllegalArgumentException {@inheritDoc} 1144 */ 1145 public int drainTo(Collection<? super E> c, int maxElements) { 1146 if (c == null) 1147 throw new NullPointerException(); 1148 if (c == this) 1149 throw new IllegalArgumentException(); 1150 int n = 0; 1151 for (E e; n < maxElements && (e = poll()) != null;) { 1152 c.add(e); 1153 ++n; 1154 } 1155 return n; 1156 } 1157 1158 /** 1159 * Returns an iterator over the elements in this queue in proper sequence. 1160 * The elements will be returned in order from first (head) to last (tail). 1161 * 1162 * <p>The returned iterator is a "weakly consistent" iterator that 1163 * will never throw {@link java.util.ConcurrentModificationException 1164 * ConcurrentModificationException}, and guarantees to traverse 1165 * elements as they existed upon construction of the iterator, and 1166 * may (but is not guaranteed to) reflect any modifications 1167 * subsequent to construction. 1168 * 1169 * @return an iterator over the elements in this queue in proper sequence 1170 */ 1171 public Iterator<E> iterator() { 1172 return new Itr(); 1173 } 1174 1175 public E peek() { 1176 return firstDataItem(); 1177 } 1178 1179 /** 1180 * Returns {@code true} if this queue contains no elements. 1181 * 1182 * @return {@code true} if this queue contains no elements 1183 */ 1184 public boolean isEmpty() { 1185 for (Node p = head; p != null; p = succ(p)) { 1186 if (!p.isMatched()) 1187 return !p.isData; 1188 } 1189 return true; 1190 } 1191 1192 public boolean hasWaitingConsumer() { 1193 return firstOfMode(false) != null; 1194 } 1195 1196 /** 1197 * Returns the number of elements in this queue. If this queue 1198 * contains more than {@code Integer.MAX_VALUE} elements, returns 1199 * {@code Integer.MAX_VALUE}. 1200 * 1201 * <p>Beware that, unlike in most collections, this method is 1202 * <em>NOT</em> a constant-time operation. Because of the 1203 * asynchronous nature of these queues, determining the current 1204 * number of elements requires an O(n) traversal. 1205 * 1206 * @return the number of elements in this queue 1207 */ 1208 public int size() { 1209 return countOfMode(true); 1210 } 1211 1212 public int getWaitingConsumerCount() { 1213 return countOfMode(false); 1214 } 1215 1216 /** 1217 * Removes a single instance of the specified element from this queue, 1218 * if it is present. More formally, removes an element {@code e} such 1219 * that {@code o.equals(e)}, if this queue contains one or more such 1220 * elements. 1221 * Returns {@code true} if this queue contained the specified element 1222 * (or equivalently, if this queue changed as a result of the call). 1223 * 1224 * @param o element to be removed from this queue, if present 1225 * @return {@code true} if this queue changed as a result of the call 1226 */ 1227 public boolean remove(Object o) { 1228 return findAndRemove(o); 1229 } 1230 1231 /** 1232 * Returns {@code true} if this queue contains the specified element. 1233 * More formally, returns {@code true} if and only if this queue contains 1234 * at least one element {@code e} such that {@code o.equals(e)}. 1235 * 1236 * @param o object to be checked for containment in this queue 1237 * @return {@code true} if this queue contains the specified element 1238 */ 1239 public boolean contains(Object o) { 1240 if (o == null) return false; 1241 for (Node p = head; p != null; p = succ(p)) { 1242 Object item = p.item; 1243 if (p.isData) { 1244 if (item != null && item != p && o.equals(item)) 1245 return true; 1246 } 1247 else if (item == null) 1248 break; 1249 } 1250 return false; 1251 } 1252 1253 /** 1254 * Always returns {@code Integer.MAX_VALUE} because a 1255 * {@code LinkedTransferQueue} is not capacity constrained. 1256 * 1257 * @return {@code Integer.MAX_VALUE} (as specified by 1258 * {@link java.util.concurrent.BlockingQueue#remainingCapacity() 1259 * BlockingQueue.remainingCapacity}) 1260 */ 1261 public int remainingCapacity() { 1262 return Integer.MAX_VALUE; 1263 } 1264 1265 /** 1266 * Saves this queue to a stream (that is, serializes it). 1267 * 1268 * @serialData All of the elements (each an {@code E}) in 1269 * the proper order, followed by a null 1270 */ 1271 private void writeObject(java.io.ObjectOutputStream s) 1272 throws java.io.IOException { 1273 s.defaultWriteObject(); 1274 for (E e : this) 1275 s.writeObject(e); 1276 // Use trailing null as sentinel 1277 s.writeObject(null); 1278 } 1279 1280 /** 1281 * Reconstitutes this queue from a stream (that is, deserializes it). 1282 */ 1283 private void readObject(java.io.ObjectInputStream s) 1284 throws java.io.IOException, ClassNotFoundException { 1285 s.defaultReadObject(); 1286 for (;;) { 1287 @SuppressWarnings("unchecked") 1288 E item = (E) s.readObject(); 1289 if (item == null) 1290 break; 1291 else 1292 offer(item); 1293 } 1294 } 1295 1296 // Unsafe mechanics 1297 1298 private static final sun.misc.Unsafe UNSAFE; 1299 private static final long headOffset; 1300 private static final long tailOffset; 1301 private static final long sweepVotesOffset; 1302 static { 1303 try { 1304 UNSAFE = sun.misc.Unsafe.getUnsafe(); 1305 Class<?> k = LinkedTransferQueue.class; 1306 headOffset = UNSAFE.objectFieldOffset 1307 (k.getDeclaredField("head")); 1308 tailOffset = UNSAFE.objectFieldOffset 1309 (k.getDeclaredField("tail")); 1310 sweepVotesOffset = UNSAFE.objectFieldOffset 1311 (k.getDeclaredField("sweepVotes")); 1312 } catch (Exception e) { 1313 throw new Error(e); 1314 } 1315 1316 // Reduce the risk of rare disastrous classloading in first call to 1317 // LockSupport.park: https://bugs.openjdk.java.net/browse/JDK-8074773 1318 Class<?> ensureLoaded = LockSupport.class; 1319 } 1320} 1321