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 with assistance from members of JCP JSR-166 32 * Expert Group and released to the public domain, as explained at 33 * http://creativecommons.org/publicdomain/zero/1.0/ 34 */ 35 36 package java.util.concurrent; 37 38 import java.util.AbstractQueue; 39 import java.util.Collection; 40 import java.util.Iterator; 41 import java.util.NoSuchElementException; 42 import java.util.Queue; 43 import java.util.concurrent.TimeUnit; 44 import java.util.concurrent.locks.LockSupport; 45 46 /** 47 * An unbounded {@link TransferQueue} based on linked nodes. 48 * This queue orders elements FIFO (first-in-first-out) with respect 49 * to any given producer. The <em>head</em> of the queue is that 50 * element that has been on the queue the longest time for some 51 * producer. The <em>tail</em> of the queue is that element that has 52 * been on the queue the shortest time for some producer. 53 * 54 * <p>Beware that, unlike in most collections, the {@code size} method 55 * is <em>NOT</em> a constant-time operation. Because of the 56 * asynchronous nature of these queues, determining the current number 57 * of elements requires a traversal of the elements, and so may report 58 * inaccurate results if this collection is modified during traversal. 59 * Additionally, the bulk operations {@code addAll}, 60 * {@code removeAll}, {@code retainAll}, {@code containsAll}, 61 * {@code equals}, and {@code toArray} are <em>not</em> guaranteed 62 * to be performed atomically. For example, an iterator operating 63 * concurrently with an {@code addAll} operation might view only some 64 * of the added elements. 65 * 66 * <p>This class and its iterator implement all of the 67 * <em>optional</em> methods of the {@link Collection} and {@link 68 * Iterator} interfaces. 69 * 70 * <p>Memory consistency effects: As with other concurrent 71 * collections, actions in a thread prior to placing an object into a 72 * {@code LinkedTransferQueue} 73 * <a href="package-summary.html#MemoryVisibility"><i>happen-before</i></a> 74 * actions subsequent to the access or removal of that element from 75 * the {@code LinkedTransferQueue} in another thread. 76 * 77 * <p>This class is a member of the 78 * <a href="{@docRoot}/../technotes/guides/collections/index.html"> 79 * Java Collections Framework</a>. 80 * 81 * @since 1.7 82 * @author Doug Lea 83 * @param <E> the type of elements held in this collection 84 */ 85 public class LinkedTransferQueue<E> extends AbstractQueue<E> 86 implements TransferQueue<E>, java.io.Serializable { 87 private static final long serialVersionUID = -3223113410248163686L; 88 89 /* 90 * *** Overview of Dual Queues with Slack *** 91 * 92 * Dual Queues, introduced by Scherer and Scott 93 * (http://www.cs.rice.edu/~wns1/papers/2004-DISC-DDS.pdf) are 94 * (linked) queues in which nodes may represent either data or 95 * requests. When a thread tries to enqueue a data node, but 96 * encounters a request node, it instead "matches" and removes it; 97 * and vice versa for enqueuing requests. Blocking Dual Queues 98 * arrange that threads enqueuing unmatched requests block until 99 * other threads provide the match. Dual Synchronous Queues (see 100 * Scherer, Lea, & Scott 101 * http://www.cs.rochester.edu/u/scott/papers/2009_Scherer_CACM_SSQ.pdf) 102 * additionally arrange that threads enqueuing unmatched data also 103 * block. Dual Transfer Queues support all of these modes, as 104 * dictated by callers. 105 * 106 * A FIFO dual queue may be implemented using a variation of the 107 * Michael & Scott (M&S) lock-free queue algorithm 108 * (http://www.cs.rochester.edu/u/scott/papers/1996_PODC_queues.pdf). 109 * It maintains two pointer fields, "head", pointing to a 110 * (matched) node that in turn points to the first actual 111 * (unmatched) queue node (or null if empty); and "tail" that 112 * points to the last node on the queue (or again null if 113 * empty). For example, here is a possible queue with four data 114 * elements: 115 * 116 * head tail 117 * | | 118 * v v 119 * M -> U -> U -> U -> U 120 * 121 * The M&S queue algorithm is known to be prone to scalability and 122 * overhead limitations when maintaining (via CAS) these head and 123 * tail pointers. This has led to the development of 124 * contention-reducing variants such as elimination arrays (see 125 * Moir et al http://portal.acm.org/citation.cfm?id=1074013) and 126 * optimistic back pointers (see Ladan-Mozes & Shavit 127 * http://people.csail.mit.edu/edya/publications/OptimisticFIFOQueue-journal.pdf). 128 * However, the nature of dual queues enables a simpler tactic for 129 * improving M&S-style implementations when dual-ness is needed. 130 * 131 * In a dual queue, each node must atomically maintain its match 132 * status. While there are other possible variants, we implement 133 * this here as: for a data-mode node, matching entails CASing an 134 * "item" field from a non-null data value to null upon match, and 135 * vice-versa for request nodes, CASing from null to a data 136 * value. (Note that the linearization properties of this style of 137 * queue are easy to verify -- elements are made available by 138 * linking, and unavailable by matching.) Compared to plain M&S 139 * queues, this property of dual queues requires one additional 140 * successful atomic operation per enq/deq pair. But it also 141 * enables lower cost variants of queue maintenance mechanics. (A 142 * variation of this idea applies even for non-dual queues that 143 * support deletion of interior elements, such as 144 * j.u.c.ConcurrentLinkedQueue.) 145 * 146 * Once a node is matched, its match status can never again 147 * change. We may thus arrange that the linked list of them 148 * contain a prefix of zero or more matched nodes, followed by a 149 * suffix of zero or more unmatched nodes. (Note that we allow 150 * both the prefix and suffix to be zero length, which in turn 151 * means that we do not use a dummy header.) If we were not 152 * concerned with either time or space efficiency, we could 153 * correctly perform enqueue and dequeue operations by traversing 154 * from a pointer to the initial node; CASing the item of the 155 * first unmatched node on match and CASing the next field of the 156 * trailing node on appends. (Plus some special-casing when 157 * initially empty). While this would be a terrible idea in 158 * itself, it does have the benefit of not requiring ANY atomic 159 * updates on head/tail fields. 160 * 161 * We introduce here an approach that lies between the extremes of 162 * never versus always updating queue (head and tail) pointers. 163 * This offers a tradeoff between sometimes requiring extra 164 * traversal steps to locate the first and/or last unmatched 165 * nodes, versus the reduced overhead and contention of fewer 166 * updates to queue pointers. For example, a possible snapshot of 167 * a queue is: 168 * 169 * head tail 170 * | | 171 * v v 172 * M -> M -> U -> U -> U -> U 173 * 174 * The best value for this "slack" (the targeted maximum distance 175 * between the value of "head" and the first unmatched node, and 176 * similarly for "tail") is an empirical matter. We have found 177 * that using very small constants in the range of 1-3 work best 178 * over a range of platforms. Larger values introduce increasing 179 * costs of cache misses and risks of long traversal chains, while 180 * smaller values increase CAS contention and overhead. 181 * 182 * Dual queues with slack differ from plain M&S dual queues by 183 * virtue of only sometimes updating head or tail pointers when 184 * matching, appending, or even traversing nodes; in order to 185 * maintain a targeted slack. The idea of "sometimes" may be 186 * operationalized in several ways. The simplest is to use a 187 * per-operation counter incremented on each traversal step, and 188 * to try (via CAS) to update the associated queue pointer 189 * whenever the count exceeds a threshold. Another, that requires 190 * more overhead, is to use random number generators to update 191 * with a given probability per traversal step. 192 * 193 * In any strategy along these lines, because CASes updating 194 * fields may fail, the actual slack may exceed targeted 195 * slack. However, they may be retried at any time to maintain 196 * targets. Even when using very small slack values, this 197 * approach works well for dual queues because it allows all 198 * operations up to the point of matching or appending an item 199 * (hence potentially allowing progress by another thread) to be 200 * read-only, thus not introducing any further contention. As 201 * described below, we implement this by performing slack 202 * maintenance retries only after these points. 203 * 204 * As an accompaniment to such techniques, traversal overhead can 205 * be further reduced without increasing contention of head 206 * pointer updates: Threads may sometimes shortcut the "next" link 207 * path from the current "head" node to be closer to the currently 208 * known first unmatched node, and similarly for tail. Again, this 209 * may be triggered with using thresholds or randomization. 210 * 211 * These ideas must be further extended to avoid unbounded amounts 212 * of costly-to-reclaim garbage caused by the sequential "next" 213 * links of nodes starting at old forgotten head nodes: As first 214 * described in detail by Boehm 215 * (http://portal.acm.org/citation.cfm?doid=503272.503282) if a GC 216 * delays noticing that any arbitrarily old node has become 217 * garbage, all newer dead nodes will also be unreclaimed. 218 * (Similar issues arise in non-GC environments.) To cope with 219 * this in our implementation, upon CASing to advance the head 220 * pointer, we set the "next" link of the previous head to point 221 * only to itself; thus limiting the length of connected dead lists. 222 * (We also take similar care to wipe out possibly garbage 223 * retaining values held in other Node fields.) However, doing so 224 * adds some further complexity to traversal: If any "next" 225 * pointer links to itself, it indicates that the current thread 226 * has lagged behind a head-update, and so the traversal must 227 * continue from the "head". Traversals trying to find the 228 * current tail starting from "tail" may also encounter 229 * self-links, in which case they also continue at "head". 230 * 231 * It is tempting in slack-based scheme to not even use CAS for 232 * updates (similarly to Ladan-Mozes & Shavit). However, this 233 * cannot be done for head updates under the above link-forgetting 234 * mechanics because an update may leave head at a detached node. 235 * And while direct writes are possible for tail updates, they 236 * increase the risk of long retraversals, and hence long garbage 237 * chains, which can be much more costly than is worthwhile 238 * considering that the cost difference of performing a CAS vs 239 * write is smaller when they are not triggered on each operation 240 * (especially considering that writes and CASes equally require 241 * additional GC bookkeeping ("write barriers") that are sometimes 242 * more costly than the writes themselves because of contention). 243 * 244 * *** Overview of implementation *** 245 * 246 * We use a threshold-based approach to updates, with a slack 247 * threshold of two -- that is, we update head/tail when the 248 * current pointer appears to be two or more steps away from the 249 * first/last node. The slack value is hard-wired: a path greater 250 * than one is naturally implemented by checking equality of 251 * traversal pointers except when the list has only one element, 252 * in which case we keep slack threshold at one. Avoiding tracking 253 * explicit counts across method calls slightly simplifies an 254 * already-messy implementation. Using randomization would 255 * probably work better if there were a low-quality dirt-cheap 256 * per-thread one available, but even ThreadLocalRandom is too 257 * heavy for these purposes. 258 * 259 * With such a small slack threshold value, it is not worthwhile 260 * to augment this with path short-circuiting (i.e., unsplicing 261 * interior nodes) except in the case of cancellation/removal (see 262 * below). 263 * 264 * We allow both the head and tail fields to be null before any 265 * nodes are enqueued; initializing upon first append. This 266 * simplifies some other logic, as well as providing more 267 * efficient explicit control paths instead of letting JVMs insert 268 * implicit NullPointerExceptions when they are null. While not 269 * currently fully implemented, we also leave open the possibility 270 * of re-nulling these fields when empty (which is complicated to 271 * arrange, for little benefit.) 272 * 273 * All enqueue/dequeue operations are handled by the single method 274 * "xfer" with parameters indicating whether to act as some form 275 * of offer, put, poll, take, or transfer (each possibly with 276 * timeout). The relative complexity of using one monolithic 277 * method outweighs the code bulk and maintenance problems of 278 * using separate methods for each case. 279 * 280 * Operation consists of up to three phases. The first is 281 * implemented within method xfer, the second in tryAppend, and 282 * the third in method awaitMatch. 283 * 284 * 1. Try to match an existing node 285 * 286 * Starting at head, skip already-matched nodes until finding 287 * an unmatched node of opposite mode, if one exists, in which 288 * case matching it and returning, also if necessary updating 289 * head to one past the matched node (or the node itself if the 290 * list has no other unmatched nodes). If the CAS misses, then 291 * a loop retries advancing head by two steps until either 292 * success or the slack is at most two. By requiring that each 293 * attempt advances head by two (if applicable), we ensure that 294 * the slack does not grow without bound. Traversals also check 295 * if the initial head is now off-list, in which case they 296 * start at the new head. 297 * 298 * If no candidates are found and the call was untimed 299 * poll/offer, (argument "how" is NOW) return. 300 * 301 * 2. Try to append a new node (method tryAppend) 302 * 303 * Starting at current tail pointer, find the actual last node 304 * and try to append a new node (or if head was null, establish 305 * the first node). Nodes can be appended only if their 306 * predecessors are either already matched or are of the same 307 * mode. If we detect otherwise, then a new node with opposite 308 * mode must have been appended during traversal, so we must 309 * restart at phase 1. The traversal and update steps are 310 * otherwise similar to phase 1: Retrying upon CAS misses and 311 * checking for staleness. In particular, if a self-link is 312 * encountered, then we can safely jump to a node on the list 313 * by continuing the traversal at current head. 314 * 315 * On successful append, if the call was ASYNC, return. 316 * 317 * 3. Await match or cancellation (method awaitMatch) 318 * 319 * Wait for another thread to match node; instead cancelling if 320 * the current thread was interrupted or the wait timed out. On 321 * multiprocessors, we use front-of-queue spinning: If a node 322 * appears to be the first unmatched node in the queue, it 323 * spins a bit before blocking. In either case, before blocking 324 * it tries to unsplice any nodes between the current "head" 325 * and the first unmatched node. 326 * 327 * Front-of-queue spinning vastly improves performance of 328 * heavily contended queues. And so long as it is relatively 329 * brief and "quiet", spinning does not much impact performance 330 * of less-contended queues. During spins threads check their 331 * interrupt status and generate a thread-local random number 332 * to decide to occasionally perform a Thread.yield. While 333 * yield has underdefined specs, we assume that it might help, 334 * and will not hurt, in limiting impact of spinning on busy 335 * systems. We also use smaller (1/2) spins for nodes that are 336 * not known to be front but whose predecessors have not 337 * blocked -- these "chained" spins avoid artifacts of 338 * front-of-queue rules which otherwise lead to alternating 339 * nodes spinning vs blocking. Further, front threads that 340 * represent phase changes (from data to request node or vice 341 * versa) compared to their predecessors receive additional 342 * chained spins, reflecting longer paths typically required to 343 * unblock threads during phase changes. 344 * 345 * 346 * ** Unlinking removed interior nodes ** 347 * 348 * In addition to minimizing garbage retention via self-linking 349 * described above, we also unlink removed interior nodes. These 350 * may arise due to timed out or interrupted waits, or calls to 351 * remove(x) or Iterator.remove. Normally, given a node that was 352 * at one time known to be the predecessor of some node s that is 353 * to be removed, we can unsplice s by CASing the next field of 354 * its predecessor if it still points to s (otherwise s must 355 * already have been removed or is now offlist). But there are two 356 * situations in which we cannot guarantee to make node s 357 * unreachable in this way: (1) If s is the trailing node of list 358 * (i.e., with null next), then it is pinned as the target node 359 * for appends, so can only be removed later after other nodes are 360 * appended. (2) We cannot necessarily unlink s given a 361 * predecessor node that is matched (including the case of being 362 * cancelled): the predecessor may already be unspliced, in which 363 * case some previous reachable node may still point to s. 364 * (For further explanation see Herlihy & Shavit "The Art of 365 * Multiprocessor Programming" chapter 9). Although, in both 366 * cases, we can rule out the need for further action if either s 367 * or its predecessor are (or can be made to be) at, or fall off 368 * from, the head of list. 369 * 370 * Without taking these into account, it would be possible for an 371 * unbounded number of supposedly removed nodes to remain 372 * reachable. Situations leading to such buildup are uncommon but 373 * can occur in practice; for example when a series of short timed 374 * calls to poll repeatedly time out but never otherwise fall off 375 * the list because of an untimed call to take at the front of the 376 * queue. 377 * 378 * When these cases arise, rather than always retraversing the 379 * entire list to find an actual predecessor to unlink (which 380 * won't help for case (1) anyway), we record a conservative 381 * estimate of possible unsplice failures (in "sweepVotes"). 382 * We trigger a full sweep when the estimate exceeds a threshold 383 * ("SWEEP_THRESHOLD") indicating the maximum number of estimated 384 * removal failures to tolerate before sweeping through, unlinking 385 * cancelled nodes that were not unlinked upon initial removal. 386 * We perform sweeps by the thread hitting threshold (rather than 387 * background threads or by spreading work to other threads) 388 * because in the main contexts in which removal occurs, the 389 * caller is already timed-out, cancelled, or performing a 390 * potentially O(n) operation (e.g. remove(x)), none of which are 391 * time-critical enough to warrant the overhead that alternatives 392 * would impose on other threads. 393 * 394 * Because the sweepVotes estimate is conservative, and because 395 * nodes become unlinked "naturally" as they fall off the head of 396 * the queue, and because we allow votes to accumulate even while 397 * sweeps are in progress, there are typically significantly fewer 398 * such nodes than estimated. Choice of a threshold value 399 * balances the likelihood of wasted effort and contention, versus 400 * providing a worst-case bound on retention of interior nodes in 401 * quiescent queues. The value defined below was chosen 402 * empirically to balance these under various timeout scenarios. 403 * 404 * Note that we cannot self-link unlinked interior nodes during 405 * sweeps. However, the associated garbage chains terminate when 406 * some successor ultimately falls off the head of the list and is 407 * self-linked. 408 */ 409 410 /** True if on multiprocessor */ 411 private static final boolean MP = 412 Runtime.getRuntime().availableProcessors() > 1; 413 414 /** 415 * The number of times to spin (with randomly interspersed calls 416 * to Thread.yield) on multiprocessor before blocking when a node 417 * is apparently the first waiter in the queue. See above for 418 * explanation. Must be a power of two. The value is empirically 419 * derived -- it works pretty well across a variety of processors, 420 * numbers of CPUs, and OSes. 421 */ 422 private static final int FRONT_SPINS = 1 << 7; 423 424 /** 425 * The number of times to spin before blocking when a node is 426 * preceded by another node that is apparently spinning. Also 427 * serves as an increment to FRONT_SPINS on phase changes, and as 428 * base average frequency for yielding during spins. Must be a 429 * power of two. 430 */ 431 private static final int CHAINED_SPINS = FRONT_SPINS >>> 1; 432 433 /** 434 * The maximum number of estimated removal failures (sweepVotes) 435 * to tolerate before sweeping through the queue unlinking 436 * cancelled nodes that were not unlinked upon initial 437 * removal. See above for explanation. The value must be at least 438 * two to avoid useless sweeps when removing trailing nodes. 439 */ 440 static final int SWEEP_THRESHOLD = 32; 441 442 /** 443 * Queue nodes. Uses Object, not E, for items to allow forgetting 444 * them after use. Relies heavily on Unsafe mechanics to minimize 445 * unnecessary ordering constraints: Writes that are intrinsically 446 * ordered wrt other accesses or CASes use simple relaxed forms. 447 */ 448 static final class Node { 449 final boolean isData; // false if this is a request node 450 volatile Object item; // initially non-null if isData; CASed to match 451 volatile Node next; 452 volatile Thread waiter; // null until waiting 453 454 // CAS methods for fields 455 final boolean casNext(Node cmp, Node val) { 456 return UNSAFE.compareAndSwapObject(this, nextOffset, cmp, val); 457 } 458 459 final boolean casItem(Object cmp, Object val) { 460 // assert cmp == null || cmp.getClass() != Node.class; 461 return UNSAFE.compareAndSwapObject(this, itemOffset, cmp, val); 462 } 463 464 /** 465 * Constructs a new node. Uses relaxed write because item can 466 * only be seen after publication via casNext. 467 */ 468 Node(Object item, boolean isData) { 469 UNSAFE.putObject(this, itemOffset, item); // relaxed write 470 this.isData = isData; 471 } 472 473 /** 474 * Links node to itself to avoid garbage retention. Called 475 * only after CASing head field, so uses relaxed write. 476 */ 477 final void forgetNext() { 478 UNSAFE.putObject(this, nextOffset, this); 479 } 480 481 /** 482 * Sets item to self and waiter to null, to avoid garbage 483 * retention after matching or cancelling. Uses relaxed writes 484 * because order is already constrained in the only calling 485 * contexts: item is forgotten only after volatile/atomic 486 * mechanics that extract items. Similarly, clearing waiter 487 * follows either CAS or return from park (if ever parked; 488 * else we don't care). 489 */ 490 final void forgetContents() { 491 UNSAFE.putObject(this, itemOffset, this); 492 UNSAFE.putObject(this, waiterOffset, null); 493 } 494 495 /** 496 * Returns true if this node has been matched, including the 497 * case of artificial matches due to cancellation. 498 */ 499 final boolean isMatched() { 500 Object x = item; 501 return (x == this) || ((x == null) == isData); 502 } 503 504 /** 505 * Returns true if this is an unmatched request node. 506 */ 507 final boolean isUnmatchedRequest() { 508 return !isData && item == null; 509 } 510 511 /** 512 * Returns true if a node with the given mode cannot be 513 * appended to this node because this node is unmatched and 514 * has opposite data mode. 515 */ 516 final boolean cannotPrecede(boolean haveData) { 517 boolean d = isData; 518 Object x; 519 return d != haveData && (x = item) != this && (x != null) == d; 520 } 521 522 /** 523 * Tries to artificially match a data node -- used by remove. 524 */ 525 final boolean tryMatchData() { 526 // assert isData; 527 Object x = item; 528 if (x != null && x != this && casItem(x, null)) { 529 LockSupport.unpark(waiter); 530 return true; 531 } 532 return false; 533 } 534 535 private static final long serialVersionUID = -3375979862319811754L; 536 537 // Unsafe mechanics 538 private static final sun.misc.Unsafe UNSAFE; 539 private static final long itemOffset; 540 private static final long nextOffset; 541 private static final long waiterOffset; 542 static { 543 try { 544 UNSAFE = sun.misc.Unsafe.getUnsafe(); 545 Class<?> k = Node.class; 546 itemOffset = UNSAFE.objectFieldOffset 547 (k.getDeclaredField("item")); 548 nextOffset = UNSAFE.objectFieldOffset 549 (k.getDeclaredField("next")); 550 waiterOffset = UNSAFE.objectFieldOffset 551 (k.getDeclaredField("waiter")); 552 } catch (Exception e) { 553 throw new Error(e); 554 } 555 } 556 } 557 558 /** head of the queue; null until first enqueue */ 559 transient volatile Node head; 560 561 /** tail of the queue; null until first append */ 562 private transient volatile Node tail; 563 564 /** The number of apparent failures to unsplice removed nodes */ 565 private transient volatile int sweepVotes; 566 567 // CAS methods for fields 568 private boolean casTail(Node cmp, Node val) { 569 return UNSAFE.compareAndSwapObject(this, tailOffset, cmp, val); 570 } 571 572 private boolean casHead(Node cmp, Node val) { 573 return UNSAFE.compareAndSwapObject(this, headOffset, cmp, val); 574 } 575 576 private boolean casSweepVotes(int cmp, int val) { 577 return UNSAFE.compareAndSwapInt(this, sweepVotesOffset, cmp, val); 578 } 579 580 /* 581 * Possible values for "how" argument in xfer method. 582 */ 583 private static final int NOW = 0; // for untimed poll, tryTransfer 584 private static final int ASYNC = 1; // for offer, put, add 585 private static final int SYNC = 2; // for transfer, take 586 private static final int TIMED = 3; // for timed poll, tryTransfer 587 588 @SuppressWarnings("unchecked") 589 static <E> E cast(Object item) { 590 // assert item == null || item.getClass() != Node.class; 591 return (E) item; 592 } 593 594 /** 595 * Implements all queuing methods. See above for explanation. 596 * 597 * @param e the item or null for take 598 * @param haveData true if this is a put, else a take 599 * @param how NOW, ASYNC, SYNC, or TIMED 600 * @param nanos timeout in nanosecs, used only if mode is TIMED 601 * @return an item if matched, else e 602 * @throws NullPointerException if haveData mode but e is null 603 */ 604 private E xfer(E e, boolean haveData, int how, long nanos) { 605 if (haveData && (e == null)) 606 throw new NullPointerException(); 607 Node s = null; // the node to append, if needed 608 609 retry: 610 for (;;) { // restart on append race 611 612 for (Node h = head, p = h; p != null;) { // find & match first node 613 boolean isData = p.isData; 614 Object item = p.item; 615 if (item != p && (item != null) == isData) { // unmatched 616 if (isData == haveData) // can't match 617 break; 618 if (p.casItem(item, e)) { // match 619 for (Node q = p; q != h;) { 620 Node n = q.next; // update by 2 unless singleton 621 if (head == h && casHead(h, n == null ? q : n)) { 622 h.forgetNext(); 623 break; 624 } // advance and retry 625 if ((h = head) == null || 626 (q = h.next) == null || !q.isMatched()) 627 break; // unless slack < 2 628 } 629 LockSupport.unpark(p.waiter); 630 return LinkedTransferQueue.<E>cast(item); 631 } 632 } 633 Node n = p.next; 634 p = (p != n) ? n : (h = head); // Use head if p offlist 635 } 636 637 if (how != NOW) { // No matches available 638 if (s == null) 639 s = new Node(e, haveData); 640 Node pred = tryAppend(s, haveData); 641 if (pred == null) 642 continue retry; // lost race vs opposite mode 643 if (how != ASYNC) 644 return awaitMatch(s, pred, e, (how == TIMED), nanos); 645 } 646 return e; // not waiting 647 } 648 } 649 650 /** 651 * Tries to append node s as tail. 652 * 653 * @param s the node to append 654 * @param haveData true if appending in data mode 655 * @return null on failure due to losing race with append in 656 * different mode, else s's predecessor, or s itself if no 657 * predecessor 658 */ 659 private Node tryAppend(Node s, boolean haveData) { 660 for (Node t = tail, p = t;;) { // move p to last node and append 661 Node n, u; // temps for reads of next & tail 662 if (p == null && (p = head) == null) { 663 if (casHead(null, s)) 664 return s; // initialize 665 } 666 else if (p.cannotPrecede(haveData)) 667 return null; // lost race vs opposite mode 668 else if ((n = p.next) != null) // not last; keep traversing 669 p = p != t && t != (u = tail) ? (t = u) : // stale tail 670 (p != n) ? n : null; // restart if off list 671 else if (!p.casNext(null, s)) 672 p = p.next; // re-read on CAS failure 673 else { 674 if (p != t) { // update if slack now >= 2 675 while ((tail != t || !casTail(t, s)) && 676 (t = tail) != null && 677 (s = t.next) != null && // advance and retry 678 (s = s.next) != null && s != t); 679 } 680 return p; 681 } 682 } 683 } 684 685 /** 686 * Spins/yields/blocks until node s is matched or caller gives up. 687 * 688 * @param s the waiting node 689 * @param pred the predecessor of s, or s itself if it has no 690 * predecessor, or null if unknown (the null case does not occur 691 * in any current calls but may in possible future extensions) 692 * @param e the comparison value for checking match 693 * @param timed if true, wait only until timeout elapses 694 * @param nanos timeout in nanosecs, used only if timed is true 695 * @return matched item, or e if unmatched on interrupt or timeout 696 */ 697 private E awaitMatch(Node s, Node pred, E e, boolean timed, long nanos) { 698 long lastTime = timed ? System.nanoTime() : 0L; 699 Thread w = Thread.currentThread(); 700 int spins = -1; // initialized after first item and cancel checks 701 ThreadLocalRandom randomYields = null; // bound if needed 702 703 for (;;) { 704 Object item = s.item; 705 if (item != e) { // matched 706 // assert item != s; 707 s.forgetContents(); // avoid garbage 708 return LinkedTransferQueue.<E>cast(item); 709 } 710 if ((w.isInterrupted() || (timed && nanos <= 0)) && 711 s.casItem(e, s)) { // cancel 712 unsplice(pred, s); 713 return e; 714 } 715 716 if (spins < 0) { // establish spins at/near front 717 if ((spins = spinsFor(pred, s.isData)) > 0) 718 randomYields = ThreadLocalRandom.current(); 719 } 720 else if (spins > 0) { // spin 721 --spins; 722 if (randomYields.nextInt(CHAINED_SPINS) == 0) 723 Thread.yield(); // occasionally yield 724 } 725 else if (s.waiter == null) { 726 s.waiter = w; // request unpark then recheck 727 } 728 else if (timed) { 729 long now = System.nanoTime(); 730 if ((nanos -= now - lastTime) > 0) 731 LockSupport.parkNanos(this, nanos); 732 lastTime = now; 733 } 734 else { 735 LockSupport.park(this); 736 } 737 } 738 } 739 740 /** 741 * Returns spin/yield value for a node with given predecessor and 742 * data mode. See above for explanation. 743 */ 744 private static int spinsFor(Node pred, boolean haveData) { 745 if (MP && pred != null) { 746 if (pred.isData != haveData) // phase change 747 return FRONT_SPINS + CHAINED_SPINS; 748 if (pred.isMatched()) // probably at front 749 return FRONT_SPINS; 750 if (pred.waiter == null) // pred apparently spinning 751 return CHAINED_SPINS; 752 } 753 return 0; 754 } 755 756 /* -------------- Traversal methods -------------- */ 757 758 /** 759 * Returns the successor of p, or the head node if p.next has been 760 * linked to self, which will only be true if traversing with a 761 * stale pointer that is now off the list. 762 */ 763 final Node succ(Node p) { 764 Node next = p.next; 765 return (p == next) ? head : next; 766 } 767 768 /** 769 * Returns the first unmatched node of the given mode, or null if 770 * none. Used by methods isEmpty, hasWaitingConsumer. 771 */ 772 private Node firstOfMode(boolean isData) { 773 for (Node p = head; p != null; p = succ(p)) { 774 if (!p.isMatched()) 775 return (p.isData == isData) ? p : null; 776 } 777 return null; 778 } 779 780 /** 781 * Returns the item in the first unmatched node with isData; or 782 * null if none. Used by peek. 783 */ 784 private E firstDataItem() { 785 for (Node p = head; p != null; p = succ(p)) { 786 Object item = p.item; 787 if (p.isData) { 788 if (item != null && item != p) 789 return LinkedTransferQueue.<E>cast(item); 790 } 791 else if (item == null) 792 return null; 793 } 794 return null; 795 } 796 797 /** 798 * Traverses and counts unmatched nodes of the given mode. 799 * Used by methods size and getWaitingConsumerCount. 800 */ 801 private int countOfMode(boolean data) { 802 int count = 0; 803 for (Node p = head; p != null; ) { 804 if (!p.isMatched()) { 805 if (p.isData != data) 806 return 0; 807 if (++count == Integer.MAX_VALUE) // saturated 808 break; 809 } 810 Node n = p.next; 811 if (n != p) 812 p = n; 813 else { 814 count = 0; 815 p = head; 816 } 817 } 818 return count; 819 } 820 821 final class Itr implements Iterator<E> { 822 private Node nextNode; // next node to return item for 823 private E nextItem; // the corresponding item 824 private Node lastRet; // last returned node, to support remove 825 private Node lastPred; // predecessor to unlink lastRet 826 827 /** 828 * Moves to next node after prev, or first node if prev null. 829 */ 830 private void advance(Node prev) { 831 /* 832 * To track and avoid buildup of deleted nodes in the face 833 * of calls to both Queue.remove and Itr.remove, we must 834 * include variants of unsplice and sweep upon each 835 * advance: Upon Itr.remove, we may need to catch up links 836 * from lastPred, and upon other removes, we might need to 837 * skip ahead from stale nodes and unsplice deleted ones 838 * found while advancing. 839 */ 840 841 Node r, b; // reset lastPred upon possible deletion of lastRet 842 if ((r = lastRet) != null && !r.isMatched()) 843 lastPred = r; // next lastPred is old lastRet 844 else if ((b = lastPred) == null || b.isMatched()) 845 lastPred = null; // at start of list 846 else { 847 Node s, n; // help with removal of lastPred.next 848 while ((s = b.next) != null && 849 s != b && s.isMatched() && 850 (n = s.next) != null && n != s) 851 b.casNext(s, n); 852 } 853 854 this.lastRet = prev; 855 856 for (Node p = prev, s, n;;) { 857 s = (p == null) ? head : p.next; 858 if (s == null) 859 break; 860 else if (s == p) { 861 p = null; 862 continue; 863 } 864 Object item = s.item; 865 if (s.isData) { 866 if (item != null && item != s) { 867 nextItem = LinkedTransferQueue.<E>cast(item); 868 nextNode = s; 869 return; 870 } 871 } 872 else if (item == null) 873 break; 874 // assert s.isMatched(); 875 if (p == null) 876 p = s; 877 else if ((n = s.next) == null) 878 break; 879 else if (s == n) 880 p = null; 881 else 882 p.casNext(s, n); 883 } 884 nextNode = null; 885 nextItem = null; 886 } 887 888 Itr() { 889 advance(null); 890 } 891 892 public final boolean hasNext() { 893 return nextNode != null; 894 } 895 896 public final E next() { 897 Node p = nextNode; 898 if (p == null) throw new NoSuchElementException(); 899 E e = nextItem; 900 advance(p); 901 return e; 902 } 903 904 public final void remove() { 905 final Node lastRet = this.lastRet; 906 if (lastRet == null) 907 throw new IllegalStateException(); 908 this.lastRet = null; 909 if (lastRet.tryMatchData()) 910 unsplice(lastPred, lastRet); 911 } 912 } 913 914 /* -------------- Removal methods -------------- */ 915 916 /** 917 * Unsplices (now or later) the given deleted/cancelled node with 918 * the given predecessor. 919 * 920 * @param pred a node that was at one time known to be the 921 * predecessor of s, or null or s itself if s is/was at head 922 * @param s the node to be unspliced 923 */ 924 final void unsplice(Node pred, Node s) { 925 s.forgetContents(); // forget unneeded fields 926 /* 927 * See above for rationale. Briefly: if pred still points to 928 * s, try to unlink s. If s cannot be unlinked, because it is 929 * trailing node or pred might be unlinked, and neither pred 930 * nor s are head or offlist, add to sweepVotes, and if enough 931 * votes have accumulated, sweep. 932 */ 933 if (pred != null && pred != s && pred.next == s) { 934 Node n = s.next; 935 if (n == null || 936 (n != s && pred.casNext(s, n) && pred.isMatched())) { 937 for (;;) { // check if at, or could be, head 938 Node h = head; 939 if (h == pred || h == s || h == null) 940 return; // at head or list empty 941 if (!h.isMatched()) 942 break; 943 Node hn = h.next; 944 if (hn == null) 945 return; // now empty 946 if (hn != h && casHead(h, hn)) 947 h.forgetNext(); // advance head 948 } 949 if (pred.next != pred && s.next != s) { // recheck if offlist 950 for (;;) { // sweep now if enough votes 951 int v = sweepVotes; 952 if (v < SWEEP_THRESHOLD) { 953 if (casSweepVotes(v, v + 1)) 954 break; 955 } 956 else if (casSweepVotes(v, 0)) { 957 sweep(); 958 break; 959 } 960 } 961 } 962 } 963 } 964 } 965 966 /** 967 * Unlinks matched (typically cancelled) nodes encountered in a 968 * traversal from head. 969 */ 970 private void sweep() { 971 for (Node p = head, s, n; p != null && (s = p.next) != null; ) { 972 if (!s.isMatched()) 973 // Unmatched nodes are never self-linked 974 p = s; 975 else if ((n = s.next) == null) // trailing node is pinned 976 break; 977 else if (s == n) // stale 978 // No need to also check for p == s, since that implies s == n 979 p = head; 980 else 981 p.casNext(s, n); 982 } 983 } 984 985 /** 986 * Main implementation of remove(Object) 987 */ 988 private boolean findAndRemove(Object e) { 989 if (e != null) { 990 for (Node pred = null, p = head; p != null; ) { 991 Object item = p.item; 992 if (p.isData) { 993 if (item != null && item != p && e.equals(item) && 994 p.tryMatchData()) { 995 unsplice(pred, p); 996 return true; 997 } 998 } 999 else if (item == null) 1000 break; 1001 pred = p; 1002 if ((p = p.next) == pred) { // stale 1003 pred = null; 1004 p = head; 1005 } 1006 } 1007 } 1008 return false; 1009 } 1010 1011 /** 1012 * Creates an initially empty {@code LinkedTransferQueue}. 1013 */ 1014 public LinkedTransferQueue() { 1015 } 1016 1017 /** 1018 * Creates a {@code LinkedTransferQueue} 1019 * initially containing the elements of the given collection, 1020 * added in traversal order of the collection's iterator. 1021 * 1022 * @param c the collection of elements to initially contain 1023 * @throws NullPointerException if the specified collection or any 1024 * of its elements are null 1025 */ 1026 public LinkedTransferQueue(Collection<? extends E> c) { 1027 this(); 1028 addAll(c); 1029 } 1030 1031 /** 1032 * Inserts the specified element at the tail of this queue. 1033 * As the queue is unbounded, this method will never block. 1034 * 1035 * @throws NullPointerException if the specified element is null 1036 */ 1037 public void put(E e) { 1038 xfer(e, true, ASYNC, 0); 1039 } 1040 1041 /** 1042 * Inserts the specified element at the tail of this queue. 1043 * As the queue is unbounded, this method will never block or 1044 * return {@code false}. 1045 * 1046 * @return {@code true} (as specified by 1047 * {@link java.util.concurrent.BlockingQueue#offer(Object,long,TimeUnit) 1048 * BlockingQueue.offer}) 1049 * @throws NullPointerException if the specified element is null 1050 */ 1051 public boolean offer(E e, long timeout, TimeUnit unit) { 1052 xfer(e, true, ASYNC, 0); 1053 return true; 1054 } 1055 1056 /** 1057 * Inserts the specified element at the tail of this queue. 1058 * As the queue is unbounded, this method will never return {@code false}. 1059 * 1060 * @return {@code true} (as specified by {@link Queue#offer}) 1061 * @throws NullPointerException if the specified element is null 1062 */ 1063 public boolean offer(E e) { 1064 xfer(e, true, ASYNC, 0); 1065 return true; 1066 } 1067 1068 /** 1069 * Inserts the specified element at the tail of this queue. 1070 * As the queue is unbounded, this method will never throw 1071 * {@link IllegalStateException} or return {@code false}. 1072 * 1073 * @return {@code true} (as specified by {@link Collection#add}) 1074 * @throws NullPointerException if the specified element is null 1075 */ 1076 public boolean add(E e) { 1077 xfer(e, true, ASYNC, 0); 1078 return true; 1079 } 1080 1081 /** 1082 * Transfers the element to a waiting consumer immediately, if possible. 1083 * 1084 * <p>More precisely, transfers the specified element immediately 1085 * if there exists a consumer already waiting to receive it (in 1086 * {@link #take} or timed {@link #poll(long,TimeUnit) poll}), 1087 * otherwise returning {@code false} without enqueuing the element. 1088 * 1089 * @throws NullPointerException if the specified element is null 1090 */ 1091 public boolean tryTransfer(E e) { 1092 return xfer(e, true, NOW, 0) == null; 1093 } 1094 1095 /** 1096 * Transfers the element to a consumer, waiting if necessary to do so. 1097 * 1098 * <p>More precisely, transfers the specified element immediately 1099 * if there exists a consumer already waiting to receive it (in 1100 * {@link #take} or timed {@link #poll(long,TimeUnit) poll}), 1101 * else inserts the specified element at the tail of this queue 1102 * and waits until the element is received by a consumer. 1103 * 1104 * @throws NullPointerException if the specified element is null 1105 */ 1106 public void transfer(E e) throws InterruptedException { 1107 if (xfer(e, true, SYNC, 0) != null) { 1108 Thread.interrupted(); // failure possible only due to interrupt 1109 throw new InterruptedException(); 1110 } 1111 } 1112 1113 /** 1114 * Transfers the element to a consumer if it is possible to do so 1115 * before the timeout elapses. 1116 * 1117 * <p>More precisely, transfers the specified element immediately 1118 * if there exists a consumer already waiting to receive it (in 1119 * {@link #take} or timed {@link #poll(long,TimeUnit) poll}), 1120 * else inserts the specified element at the tail of this queue 1121 * and waits until the element is received by a consumer, 1122 * returning {@code false} if the specified wait time elapses 1123 * before the element can be transferred. 1124 * 1125 * @throws NullPointerException if the specified element is null 1126 */ 1127 public boolean tryTransfer(E e, long timeout, TimeUnit unit) 1128 throws InterruptedException { 1129 if (xfer(e, true, TIMED, unit.toNanos(timeout)) == null) 1130 return true; 1131 if (!Thread.interrupted()) 1132 return false; 1133 throw new InterruptedException(); 1134 } 1135 1136 public E take() throws InterruptedException { 1137 E e = xfer(null, false, SYNC, 0); 1138 if (e != null) 1139 return e; 1140 Thread.interrupted(); 1141 throw new InterruptedException(); 1142 } 1143 1144 public E poll(long timeout, TimeUnit unit) throws InterruptedException { 1145 E e = xfer(null, false, TIMED, unit.toNanos(timeout)); 1146 if (e != null || !Thread.interrupted()) 1147 return e; 1148 throw new InterruptedException(); 1149 } 1150 1151 public E poll() { 1152 return xfer(null, false, NOW, 0); 1153 } 1154 1155 /** 1156 * @throws NullPointerException {@inheritDoc} 1157 * @throws IllegalArgumentException {@inheritDoc} 1158 */ 1159 public int drainTo(Collection<? super E> c) { 1160 if (c == null) 1161 throw new NullPointerException(); 1162 if (c == this) 1163 throw new IllegalArgumentException(); 1164 int n = 0; 1165 for (E e; (e = poll()) != null;) { 1166 c.add(e); 1167 ++n; 1168 } 1169 return n; 1170 } 1171 1172 /** 1173 * @throws NullPointerException {@inheritDoc} 1174 * @throws IllegalArgumentException {@inheritDoc} 1175 */ 1176 public int drainTo(Collection<? super E> c, int maxElements) { 1177 if (c == null) 1178 throw new NullPointerException(); 1179 if (c == this) 1180 throw new IllegalArgumentException(); 1181 int n = 0; 1182 for (E e; n < maxElements && (e = poll()) != null;) { 1183 c.add(e); 1184 ++n; 1185 } 1186 return n; 1187 } 1188 1189 /** 1190 * Returns an iterator over the elements in this queue in proper sequence. 1191 * The elements will be returned in order from first (head) to last (tail). 1192 * 1193 * <p>The returned iterator is a "weakly consistent" iterator that 1194 * will never throw {@link java.util.ConcurrentModificationException 1195 * ConcurrentModificationException}, and guarantees to traverse 1196 * elements as they existed upon construction of the iterator, and 1197 * may (but is not guaranteed to) reflect any modifications 1198 * subsequent to construction. 1199 * 1200 * @return an iterator over the elements in this queue in proper sequence 1201 */ 1202 public Iterator<E> iterator() { 1203 return new Itr(); 1204 } 1205 1206 public E peek() { 1207 return firstDataItem(); 1208 } 1209 1210 /** 1211 * Returns {@code true} if this queue contains no elements. 1212 * 1213 * @return {@code true} if this queue contains no elements 1214 */ 1215 public boolean isEmpty() { 1216 for (Node p = head; p != null; p = succ(p)) { 1217 if (!p.isMatched()) 1218 return !p.isData; 1219 } 1220 return true; 1221 } 1222 1223 public boolean hasWaitingConsumer() { 1224 return firstOfMode(false) != null; 1225 } 1226 1227 /** 1228 * Returns the number of elements in this queue. If this queue 1229 * contains more than {@code Integer.MAX_VALUE} elements, returns 1230 * {@code Integer.MAX_VALUE}. 1231 * 1232 * <p>Beware that, unlike in most collections, this method is 1233 * <em>NOT</em> a constant-time operation. Because of the 1234 * asynchronous nature of these queues, determining the current 1235 * number of elements requires an O(n) traversal. 1236 * 1237 * @return the number of elements in this queue 1238 */ 1239 public int size() { 1240 return countOfMode(true); 1241 } 1242 1243 public int getWaitingConsumerCount() { 1244 return countOfMode(false); 1245 } 1246 1247 /** 1248 * Removes a single instance of the specified element from this queue, 1249 * if it is present. More formally, removes an element {@code e} such 1250 * that {@code o.equals(e)}, if this queue contains one or more such 1251 * elements. 1252 * Returns {@code true} if this queue contained the specified element 1253 * (or equivalently, if this queue changed as a result of the call). 1254 * 1255 * @param o element to be removed from this queue, if present 1256 * @return {@code true} if this queue changed as a result of the call 1257 */ 1258 public boolean remove(Object o) { 1259 return findAndRemove(o); 1260 } 1261 1262 /** 1263 * Returns {@code true} if this queue contains the specified element. 1264 * More formally, returns {@code true} if and only if this queue contains 1265 * at least one element {@code e} such that {@code o.equals(e)}. 1266 * 1267 * @param o object to be checked for containment in this queue 1268 * @return {@code true} if this queue contains the specified element 1269 */ 1270 public boolean contains(Object o) { 1271 if (o == null) return false; 1272 for (Node p = head; p != null; p = succ(p)) { 1273 Object item = p.item; 1274 if (p.isData) { 1275 if (item != null && item != p && o.equals(item)) 1276 return true; 1277 } 1278 else if (item == null) 1279 break; 1280 } 1281 return false; 1282 } 1283 1284 /** 1285 * Always returns {@code Integer.MAX_VALUE} because a 1286 * {@code LinkedTransferQueue} is not capacity constrained. 1287 * 1288 * @return {@code Integer.MAX_VALUE} (as specified by 1289 * {@link java.util.concurrent.BlockingQueue#remainingCapacity() 1290 * BlockingQueue.remainingCapacity}) 1291 */ 1292 public int remainingCapacity() { 1293 return Integer.MAX_VALUE; 1294 } 1295 1296 /** 1297 * Saves the state to a stream (that is, serializes it). 1298 * 1299 * @serialData All of the elements (each an {@code E}) in 1300 * the proper order, followed by a null 1301 * @param s the stream 1302 */ 1303 private void writeObject(java.io.ObjectOutputStream s) 1304 throws java.io.IOException { 1305 s.defaultWriteObject(); 1306 for (E e : this) 1307 s.writeObject(e); 1308 // Use trailing null as sentinel 1309 s.writeObject(null); 1310 } 1311 1312 /** 1313 * Reconstitutes the Queue instance from a stream (that is, 1314 * deserializes it). 1315 * 1316 * @param s the stream 1317 */ 1318 private void readObject(java.io.ObjectInputStream s) 1319 throws java.io.IOException, ClassNotFoundException { 1320 s.defaultReadObject(); 1321 for (;;) { 1322 @SuppressWarnings("unchecked") E item = (E) s.readObject(); 1323 if (item == null) 1324 break; 1325 else 1326 offer(item); 1327 } 1328 } 1329 1330 // Unsafe mechanics 1331 1332 private static final sun.misc.Unsafe UNSAFE; 1333 private static final long headOffset; 1334 private static final long tailOffset; 1335 private static final long sweepVotesOffset; 1336 static { 1337 try { 1338 UNSAFE = sun.misc.Unsafe.getUnsafe(); 1339 Class<?> k = LinkedTransferQueue.class; 1340 headOffset = UNSAFE.objectFieldOffset 1341 (k.getDeclaredField("head")); 1342 tailOffset = UNSAFE.objectFieldOffset 1343 (k.getDeclaredField("tail")); 1344 sweepVotesOffset = UNSAFE.objectFieldOffset 1345 (k.getDeclaredField("sweepVotes")); 1346 } catch (Exception e) { 1347 throw new Error(e); 1348 } 1349 } 1350 }