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

Comparing jsr166/src/jsr166y/LinkedTransferQueue.java (file contents):
Revision 1.35 by jsr166, Thu Jul 30 22:45:39 2009 UTC vs.
Revision 1.63 by jsr166, Mon Nov 2 05:08:48 2009 UTC

# Line 15 | Line 15 | import java.util.Iterator;
15   import java.util.NoSuchElementException;
16   import java.util.Queue;
17   import java.util.concurrent.locks.LockSupport;
18 import java.util.concurrent.atomic.AtomicReference;
19
18   /**
19 < * An unbounded {@linkplain TransferQueue} based on linked nodes.
19 > * An unbounded {@link TransferQueue} based on linked nodes.
20   * This queue orders elements FIFO (first-in-first-out) with respect
21   * to any given producer.  The <em>head</em> of the queue is that
22   * element that has been on the queue the longest time for some
# Line 54 | Line 52 | public class LinkedTransferQueue<E> exte
52      private static final long serialVersionUID = -3223113410248163686L;
53  
54      /*
55 <     * This class extends the approach used in FIFO-mode
58 <     * SynchronousQueues. See the internal documentation, as well as
59 <     * the PPoPP 2006 paper "Scalable Synchronous Queues" by Scherer,
60 <     * Lea & Scott
61 <     * (http://www.cs.rice.edu/~wns1/papers/2006-PPoPP-SQ.pdf)
55 >     * *** Overview of Dual Queues with Slack ***
56       *
57 <     * The main extension is to provide different Wait modes for the
58 <     * main "xfer" method that puts or takes items.  These don't
59 <     * impact the basic dual-queue logic, but instead control whether
60 <     * or how threads block upon insertion of request or data nodes
61 <     * into the dual queue. It also uses slightly different
62 <     * conventions for tracking whether nodes are off-list or
63 <     * cancelled.
64 <     */
65 <
66 <    // Wait modes for xfer method
67 <    static final int NOWAIT  = 0;
68 <    static final int TIMEOUT = 1;
69 <    static final int WAIT    = 2;
70 <
71 <    /** The number of CPUs, for spin control */
72 <    static final int NCPUS = Runtime.getRuntime().availableProcessors();
57 >     * Dual Queues, introduced by Scherer and Scott
58 >     * (http://www.cs.rice.edu/~wns1/papers/2004-DISC-DDS.pdf) are
59 >     * (linked) queues in which nodes may represent either data or
60 >     * requests.  When a thread tries to enqueue a data node, but
61 >     * encounters a request node, it instead "matches" and removes it;
62 >     * and vice versa for enqueuing requests. Blocking Dual Queues
63 >     * arrange that threads enqueuing unmatched requests block until
64 >     * other threads provide the match. Dual Synchronous Queues (see
65 >     * Scherer, Lea, & Scott
66 >     * http://www.cs.rochester.edu/u/scott/papers/2009_Scherer_CACM_SSQ.pdf)
67 >     * additionally arrange that threads enqueuing unmatched data also
68 >     * block.  Dual Transfer Queues support all of these modes, as
69 >     * dictated by callers.
70 >     *
71 >     * A FIFO dual queue may be implemented using a variation of the
72 >     * Michael & Scott (M&S) lock-free queue algorithm
73 >     * (http://www.cs.rochester.edu/u/scott/papers/1996_PODC_queues.pdf).
74 >     * It maintains two pointer fields, "head", pointing to a
75 >     * (matched) node that in turn points to the first actual
76 >     * (unmatched) queue node (or null if empty); and "tail" that
77 >     * points to the last node on the queue (or again null if
78 >     * empty). For example, here is a possible queue with four data
79 >     * elements:
80 >     *
81 >     *  head                tail
82 >     *    |                   |
83 >     *    v                   v
84 >     *    M -> U -> U -> U -> U
85 >     *
86 >     * The M&S queue algorithm is known to be prone to scalability and
87 >     * overhead limitations when maintaining (via CAS) these head and
88 >     * tail pointers. This has led to the development of
89 >     * contention-reducing variants such as elimination arrays (see
90 >     * Moir et al http://portal.acm.org/citation.cfm?id=1074013) and
91 >     * optimistic back pointers (see Ladan-Mozes & Shavit
92 >     * http://people.csail.mit.edu/edya/publications/OptimisticFIFOQueue-journal.pdf).
93 >     * However, the nature of dual queues enables a simpler tactic for
94 >     * improving M&S-style implementations when dual-ness is needed.
95 >     *
96 >     * In a dual queue, each node must atomically maintain its match
97 >     * status. While there are other possible variants, we implement
98 >     * this here as: for a data-mode node, matching entails CASing an
99 >     * "item" field from a non-null data value to null upon match, and
100 >     * vice-versa for request nodes, CASing from null to a data
101 >     * value. (Note that the linearization properties of this style of
102 >     * queue are easy to verify -- elements are made available by
103 >     * linking, and unavailable by matching.) Compared to plain M&S
104 >     * queues, this property of dual queues requires one additional
105 >     * successful atomic operation per enq/deq pair. But it also
106 >     * enables lower cost variants of queue maintenance mechanics. (A
107 >     * variation of this idea applies even for non-dual queues that
108 >     * support deletion of interior elements, such as
109 >     * j.u.c.ConcurrentLinkedQueue.)
110 >     *
111 >     * Once a node is matched, its match status can never again
112 >     * change.  We may thus arrange that the linked list of them
113 >     * contain a prefix of zero or more matched nodes, followed by a
114 >     * suffix of zero or more unmatched nodes. (Note that we allow
115 >     * both the prefix and suffix to be zero length, which in turn
116 >     * means that we do not use a dummy header.)  If we were not
117 >     * concerned with either time or space efficiency, we could
118 >     * correctly perform enqueue and dequeue operations by traversing
119 >     * from a pointer to the initial node; CASing the item of the
120 >     * first unmatched node on match and CASing the next field of the
121 >     * trailing node on appends. (Plus some special-casing when
122 >     * initially empty).  While this would be a terrible idea in
123 >     * itself, it does have the benefit of not requiring ANY atomic
124 >     * updates on head/tail fields.
125 >     *
126 >     * We introduce here an approach that lies between the extremes of
127 >     * never versus always updating queue (head and tail) pointers.
128 >     * This offers a tradeoff between sometimes requiring extra
129 >     * traversal steps to locate the first and/or last unmatched
130 >     * nodes, versus the reduced overhead and contention of fewer
131 >     * updates to queue pointers. For example, a possible snapshot of
132 >     * a queue is:
133 >     *
134 >     *  head           tail
135 >     *    |              |
136 >     *    v              v
137 >     *    M -> M -> U -> U -> U -> U
138 >     *
139 >     * The best value for this "slack" (the targeted maximum distance
140 >     * between the value of "head" and the first unmatched node, and
141 >     * similarly for "tail") is an empirical matter. We have found
142 >     * that using very small constants in the range of 1-3 work best
143 >     * over a range of platforms. Larger values introduce increasing
144 >     * costs of cache misses and risks of long traversal chains, while
145 >     * smaller values increase CAS contention and overhead.
146 >     *
147 >     * Dual queues with slack differ from plain M&S dual queues by
148 >     * virtue of only sometimes updating head or tail pointers when
149 >     * matching, appending, or even traversing nodes; in order to
150 >     * maintain a targeted slack.  The idea of "sometimes" may be
151 >     * operationalized in several ways. The simplest is to use a
152 >     * per-operation counter incremented on each traversal step, and
153 >     * to try (via CAS) to update the associated queue pointer
154 >     * whenever the count exceeds a threshold. Another, that requires
155 >     * more overhead, is to use random number generators to update
156 >     * with a given probability per traversal step.
157 >     *
158 >     * In any strategy along these lines, because CASes updating
159 >     * fields may fail, the actual slack may exceed targeted
160 >     * slack. However, they may be retried at any time to maintain
161 >     * targets.  Even when using very small slack values, this
162 >     * approach works well for dual queues because it allows all
163 >     * operations up to the point of matching or appending an item
164 >     * (hence potentially allowing progress by another thread) to be
165 >     * read-only, thus not introducing any further contention. As
166 >     * described below, we implement this by performing slack
167 >     * maintenance retries only after these points.
168 >     *
169 >     * As an accompaniment to such techniques, traversal overhead can
170 >     * be further reduced without increasing contention of head
171 >     * pointer updates: Threads may sometimes shortcut the "next" link
172 >     * path from the current "head" node to be closer to the currently
173 >     * known first unmatched node, and similarly for tail. Again, this
174 >     * may be triggered with using thresholds or randomization.
175 >     *
176 >     * These ideas must be further extended to avoid unbounded amounts
177 >     * of costly-to-reclaim garbage caused by the sequential "next"
178 >     * links of nodes starting at old forgotten head nodes: As first
179 >     * described in detail by Boehm
180 >     * (http://portal.acm.org/citation.cfm?doid=503272.503282) if a GC
181 >     * delays noticing that any arbitrarily old node has become
182 >     * garbage, all newer dead nodes will also be unreclaimed.
183 >     * (Similar issues arise in non-GC environments.)  To cope with
184 >     * this in our implementation, upon CASing to advance the head
185 >     * pointer, we set the "next" link of the previous head to point
186 >     * only to itself; thus limiting the length of connected dead lists.
187 >     * (We also take similar care to wipe out possibly garbage
188 >     * retaining values held in other Node fields.)  However, doing so
189 >     * adds some further complexity to traversal: If any "next"
190 >     * pointer links to itself, it indicates that the current thread
191 >     * has lagged behind a head-update, and so the traversal must
192 >     * continue from the "head".  Traversals trying to find the
193 >     * current tail starting from "tail" may also encounter
194 >     * self-links, in which case they also continue at "head".
195 >     *
196 >     * It is tempting in slack-based scheme to not even use CAS for
197 >     * updates (similarly to Ladan-Mozes & Shavit). However, this
198 >     * cannot be done for head updates under the above link-forgetting
199 >     * mechanics because an update may leave head at a detached node.
200 >     * And while direct writes are possible for tail updates, they
201 >     * increase the risk of long retraversals, and hence long garbage
202 >     * chains, which can be much more costly than is worthwhile
203 >     * considering that the cost difference of performing a CAS vs
204 >     * write is smaller when they are not triggered on each operation
205 >     * (especially considering that writes and CASes equally require
206 >     * additional GC bookkeeping ("write barriers") that are sometimes
207 >     * more costly than the writes themselves because of contention).
208 >     *
209 >     * Removal of interior nodes (due to timed out or interrupted
210 >     * waits, or calls to remove(x) or Iterator.remove) can use a
211 >     * scheme roughly similar to that described in Scherer, Lea, and
212 >     * Scott's SynchronousQueue. Given a predecessor, we can unsplice
213 >     * any node except the (actual) tail of the queue. To avoid
214 >     * build-up of cancelled trailing nodes, upon a request to remove
215 >     * a trailing node, it is placed in field "cleanMe" to be
216 >     * unspliced upon the next call to unsplice any other node.
217 >     * Situations needing such mechanics are not common but do occur
218 >     * in practice; for example when an unbounded series of short
219 >     * timed calls to poll repeatedly time out but never otherwise
220 >     * fall off the list because of an untimed call to take at the
221 >     * front of the queue. Note that maintaining field cleanMe does
222 >     * not otherwise much impact garbage retention even if never
223 >     * cleared by some other call because the held node will
224 >     * eventually either directly or indirectly lead to a self-link
225 >     * once off the list.
226 >     *
227 >     * *** Overview of implementation ***
228 >     *
229 >     * We use a threshold-based approach to updates, with a slack
230 >     * threshold of two -- that is, we update head/tail when the
231 >     * current pointer appears to be two or more steps away from the
232 >     * first/last node. The slack value is hard-wired: a path greater
233 >     * than one is naturally implemented by checking equality of
234 >     * traversal pointers except when the list has only one element,
235 >     * in which case we keep slack threshold at one. Avoiding tracking
236 >     * explicit counts across method calls slightly simplifies an
237 >     * already-messy implementation. Using randomization would
238 >     * probably work better if there were a low-quality dirt-cheap
239 >     * per-thread one available, but even ThreadLocalRandom is too
240 >     * heavy for these purposes.
241 >     *
242 >     * With such a small slack threshold value, it is rarely
243 >     * worthwhile to augment this with path short-circuiting; i.e.,
244 >     * unsplicing nodes between head and the first unmatched node, or
245 >     * similarly for tail, rather than advancing head or tail
246 >     * proper. However, it is used (in awaitMatch) immediately before
247 >     * a waiting thread starts to block, as a final bit of helping at
248 >     * a point when contention with others is extremely unlikely
249 >     * (since if other threads that could release it are operating,
250 >     * then the current thread wouldn't be blocking).
251 >     *
252 >     * We allow both the head and tail fields to be null before any
253 >     * nodes are enqueued; initializing upon first append.  This
254 >     * simplifies some other logic, as well as providing more
255 >     * efficient explicit control paths instead of letting JVMs insert
256 >     * implicit NullPointerExceptions when they are null.  While not
257 >     * currently fully implemented, we also leave open the possibility
258 >     * of re-nulling these fields when empty (which is complicated to
259 >     * arrange, for little benefit.)
260 >     *
261 >     * All enqueue/dequeue operations are handled by the single method
262 >     * "xfer" with parameters indicating whether to act as some form
263 >     * of offer, put, poll, take, or transfer (each possibly with
264 >     * timeout). The relative complexity of using one monolithic
265 >     * method outweighs the code bulk and maintenance problems of
266 >     * using separate methods for each case.
267 >     *
268 >     * Operation consists of up to three phases. The first is
269 >     * implemented within method xfer, the second in tryAppend, and
270 >     * the third in method awaitMatch.
271 >     *
272 >     * 1. Try to match an existing node
273 >     *
274 >     *    Starting at head, skip already-matched nodes until finding
275 >     *    an unmatched node of opposite mode, if one exists, in which
276 >     *    case matching it and returning, also if necessary updating
277 >     *    head to one past the matched node (or the node itself if the
278 >     *    list has no other unmatched nodes). If the CAS misses, then
279 >     *    a loop retries advancing head by two steps until either
280 >     *    success or the slack is at most two. By requiring that each
281 >     *    attempt advances head by two (if applicable), we ensure that
282 >     *    the slack does not grow without bound. Traversals also check
283 >     *    if the initial head is now off-list, in which case they
284 >     *    start at the new head.
285 >     *
286 >     *    If no candidates are found and the call was untimed
287 >     *    poll/offer, (argument "how" is NOW) return.
288 >     *
289 >     * 2. Try to append a new node (method tryAppend)
290 >     *
291 >     *    Starting at current tail pointer, find the actual last node
292 >     *    and try to append a new node (or if head was null, establish
293 >     *    the first node). Nodes can be appended only if their
294 >     *    predecessors are either already matched or are of the same
295 >     *    mode. If we detect otherwise, then a new node with opposite
296 >     *    mode must have been appended during traversal, so we must
297 >     *    restart at phase 1. The traversal and update steps are
298 >     *    otherwise similar to phase 1: Retrying upon CAS misses and
299 >     *    checking for staleness.  In particular, if a self-link is
300 >     *    encountered, then we can safely jump to a node on the list
301 >     *    by continuing the traversal at current head.
302 >     *
303 >     *    On successful append, if the call was ASYNC, return.
304 >     *
305 >     * 3. Await match or cancellation (method awaitMatch)
306 >     *
307 >     *    Wait for another thread to match node; instead cancelling if
308 >     *    the current thread was interrupted or the wait timed out. On
309 >     *    multiprocessors, we use front-of-queue spinning: If a node
310 >     *    appears to be the first unmatched node in the queue, it
311 >     *    spins a bit before blocking. In either case, before blocking
312 >     *    it tries to unsplice any nodes between the current "head"
313 >     *    and the first unmatched node.
314 >     *
315 >     *    Front-of-queue spinning vastly improves performance of
316 >     *    heavily contended queues. And so long as it is relatively
317 >     *    brief and "quiet", spinning does not much impact performance
318 >     *    of less-contended queues.  During spins threads check their
319 >     *    interrupt status and generate a thread-local random number
320 >     *    to decide to occasionally perform a Thread.yield. While
321 >     *    yield has underdefined specs, we assume that might it help,
322 >     *    and will not hurt in limiting impact of spinning on busy
323 >     *    systems.  We also use smaller (1/2) spins for nodes that are
324 >     *    not known to be front but whose predecessors have not
325 >     *    blocked -- these "chained" spins avoid artifacts of
326 >     *    front-of-queue rules which otherwise lead to alternating
327 >     *    nodes spinning vs blocking. Further, front threads that
328 >     *    represent phase changes (from data to request node or vice
329 >     *    versa) compared to their predecessors receive additional
330 >     *    chained spins, reflecting longer paths typically required to
331 >     *    unblock threads during phase changes.
332 >     */
333 >
334 >    /** True if on multiprocessor */
335 >    private static final boolean MP =
336 >        Runtime.getRuntime().availableProcessors() > 1;
337 >
338 >    /**
339 >     * The number of times to spin (with randomly interspersed calls
340 >     * to Thread.yield) on multiprocessor before blocking when a node
341 >     * is apparently the first waiter in the queue.  See above for
342 >     * explanation. Must be a power of two. The value is empirically
343 >     * derived -- it works pretty well across a variety of processors,
344 >     * numbers of CPUs, and OSes.
345 >     */
346 >    private static final int FRONT_SPINS   = 1 << 7;
347 >
348 >    /**
349 >     * The number of times to spin before blocking when a node is
350 >     * preceded by another node that is apparently spinning.  Also
351 >     * serves as an increment to FRONT_SPINS on phase changes, and as
352 >     * base average frequency for yielding during spins. Must be a
353 >     * power of two.
354 >     */
355 >    private static final int CHAINED_SPINS = FRONT_SPINS >>> 1;
356 >
357 >    /**
358 >     * Queue nodes. Uses Object, not E, for items to allow forgetting
359 >     * them after use.  Relies heavily on Unsafe mechanics to minimize
360 >     * unnecessary ordering constraints: Writes that intrinsically
361 >     * precede or follow CASes use simple relaxed forms.  Other
362 >     * cleanups use releasing/lazy writes.
363 >     */
364 >    static final class Node {
365 >        final boolean isData;   // false if this is a request node
366 >        volatile Object item;   // initially non-null if isData; CASed to match
367 >        volatile Node next;
368 >        volatile Thread waiter; // null until waiting
369  
370 <    /**
371 <     * The number of times to spin before blocking in timed waits.
372 <     * The value is empirically derived -- it works well across a
373 <     * variety of processors and OSes. Empirically, the best value
84 <     * seems not to vary with number of CPUs (beyond 2) so is just
85 <     * a constant.
86 <     */
87 <    static final int maxTimedSpins = (NCPUS < 2) ? 0 : 32;
88 <
89 <    /**
90 <     * The number of times to spin before blocking in untimed waits.
91 <     * This is greater than timed value because untimed waits spin
92 <     * faster since they don't need to check times on each spin.
93 <     */
94 <    static final int maxUntimedSpins = maxTimedSpins * 16;
95 <
96 <    /**
97 <     * The number of nanoseconds for which it is faster to spin
98 <     * rather than to use timed park. A rough estimate suffices.
99 <     */
100 <    static final long spinForTimeoutThreshold = 1000L;
370 >        // CAS methods for fields
371 >        final boolean casNext(Node cmp, Node val) {
372 >            return UNSAFE.compareAndSwapObject(this, nextOffset, cmp, val);
373 >        }
374  
375 <    /**
376 <     * Node class for LinkedTransferQueue. Opportunistically
377 <     * subclasses from AtomicReference to represent item. Uses Object,
378 <     * not E, to allow setting item to "this" after use, to avoid
106 <     * garbage retention. Similarly, setting the next field to this is
107 <     * used as sentinel that node is off list.
108 <     */
109 <    static final class Node<E> extends AtomicReference<Object> {
110 <        volatile Node<E> next;
111 <        volatile Thread waiter;       // to control park/unpark
112 <        final boolean isData;
375 >        final boolean casItem(Object cmp, Object val) {
376 >            assert cmp == null || cmp.getClass() != Node.class;
377 >            return UNSAFE.compareAndSwapObject(this, itemOffset, cmp, val);
378 >        }
379  
380 <        Node(E item, boolean isData) {
381 <            super(item);
380 >        /**
381 >         * Creates a new node. Uses relaxed write because item can only
382 >         * be seen if followed by CAS.
383 >         */
384 >        Node(Object item, boolean isData) {
385 >            UNSAFE.putObject(this, itemOffset, item); // relaxed write
386              this.isData = isData;
387          }
388  
389 <        // Unsafe mechanics
389 >        /**
390 >         * Links node to itself to avoid garbage retention.  Called
391 >         * only after CASing head field, so uses relaxed write.
392 >         */
393 >        final void forgetNext() {
394 >            UNSAFE.putObject(this, nextOffset, this);
395 >        }
396  
397 <        private static final sun.misc.Unsafe UNSAFE = getUnsafe();
398 <        private static final long nextOffset =
399 <            objectFieldOffset(UNSAFE, "next", Node.class);
397 >        /**
398 >         * Sets item to self (using a releasing/lazy write) and waiter
399 >         * to null, to avoid garbage retention after extracting or
400 >         * cancelling.
401 >         */
402 >        final void forgetContents() {
403 >            UNSAFE.putOrderedObject(this, itemOffset, this);
404 >            UNSAFE.putOrderedObject(this, waiterOffset, null);
405 >        }
406  
407 <        final boolean casNext(Node<E> cmp, Node<E> val) {
408 <            return UNSAFE.compareAndSwapObject(this, nextOffset, cmp, val);
407 >        /**
408 >         * Returns true if this node has been matched, including the
409 >         * case of artificial matches due to cancellation.
410 >         */
411 >        final boolean isMatched() {
412 >            Object x = item;
413 >            return (x == this) || ((x == null) == isData);
414          }
415  
416 <        final void clearNext() {
417 <            UNSAFE.putOrderedObject(this, nextOffset, this);
416 >        /**
417 >         * Returns true if this is an unmatched request node.
418 >         */
419 >        final boolean isUnmatchedRequest() {
420 >            return !isData && item == null;
421          }
422  
423          /**
424 <         * Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
425 <         * Replace with a simple call to Unsafe.getUnsafe when integrating
426 <         * into a jdk.
137 <         *
138 <         * @return a sun.misc.Unsafe
424 >         * Returns true if a node with the given mode cannot be
425 >         * appended to this node because this node is unmatched and
426 >         * has opposite data mode.
427           */
428 <        private static sun.misc.Unsafe getUnsafe() {
429 <            try {
430 <                return sun.misc.Unsafe.getUnsafe();
431 <            } catch (SecurityException se) {
432 <                try {
433 <                    return java.security.AccessController.doPrivileged
434 <                        (new java.security
435 <                         .PrivilegedExceptionAction<sun.misc.Unsafe>() {
436 <                            public sun.misc.Unsafe run() throws Exception {
437 <                                java.lang.reflect.Field f = sun.misc
438 <                                    .Unsafe.class.getDeclaredField("theUnsafe");
439 <                                f.setAccessible(true);
440 <                                return (sun.misc.Unsafe) f.get(null);
441 <                            }});
442 <                } catch (java.security.PrivilegedActionException e) {
155 <                    throw new RuntimeException("Could not initialize intrinsics",
156 <                                               e.getCause());
157 <                }
428 >        final boolean cannotPrecede(boolean haveData) {
429 >            boolean d = isData;
430 >            Object x;
431 >            return d != haveData && (x = item) != this && (x != null) == d;
432 >        }
433 >
434 >        /**
435 >         * Tries to artificially match a data node -- used by remove.
436 >         */
437 >        final boolean tryMatchData() {
438 >            assert isData;
439 >            Object x = item;
440 >            if (x != null && x != this && casItem(x, null)) {
441 >                LockSupport.unpark(waiter);
442 >                return true;
443              }
444 +            return false;
445          }
446  
447 +        // Unsafe mechanics
448 +        private static final sun.misc.Unsafe UNSAFE = getUnsafe();
449 +        private static final long nextOffset =
450 +            objectFieldOffset(UNSAFE, "next", Node.class);
451 +        private static final long itemOffset =
452 +            objectFieldOffset(UNSAFE, "item", Node.class);
453 +        private static final long waiterOffset =
454 +            objectFieldOffset(UNSAFE, "waiter", Node.class);
455 +
456          private static final long serialVersionUID = -3375979862319811754L;
457      }
458  
459 <    /**
460 <     * Padded version of AtomicReference used for head, tail and
166 <     * cleanMe, to alleviate contention across threads CASing one vs
167 <     * the other.
168 <     */
169 <    static final class PaddedAtomicReference<T> extends AtomicReference<T> {
170 <        // enough padding for 64bytes with 4byte refs
171 <        Object p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, pa, pb, pc, pd, pe;
172 <        PaddedAtomicReference(T r) { super(r); }
173 <        private static final long serialVersionUID = 8170090609809740854L;
174 <    }
459 >    /** head of the queue; null until first enqueue */
460 >    transient volatile Node head;
461  
462 +    /** predecessor of dangling unspliceable node */
463 +    private transient volatile Node cleanMe; // decl here reduces contention
464  
465 <    /** head of the queue */
466 <    private transient final PaddedAtomicReference<Node<E>> head;
465 >    /** tail of the queue; null until first append */
466 >    private transient volatile Node tail;
467  
468 <    /** tail of the queue */
469 <    private transient final PaddedAtomicReference<Node<E>> tail;
468 >    // CAS methods for fields
469 >    private boolean casTail(Node cmp, Node val) {
470 >        return UNSAFE.compareAndSwapObject(this, tailOffset, cmp, val);
471 >    }
472  
473 <    /**
474 <     * Reference to a cancelled node that might not yet have been
475 <     * unlinked from queue because it was the last inserted node
186 <     * when it cancelled.
187 <     */
188 <    private transient final PaddedAtomicReference<Node<E>> cleanMe;
473 >    private boolean casHead(Node cmp, Node val) {
474 >        return UNSAFE.compareAndSwapObject(this, headOffset, cmp, val);
475 >    }
476  
477 <    /**
478 <     * Tries to cas nh as new head; if successful, unlink
479 <     * old head's next node to avoid garbage retention.
477 >    private boolean casCleanMe(Node cmp, Node val) {
478 >        return UNSAFE.compareAndSwapObject(this, cleanMeOffset, cmp, val);
479 >    }
480 >
481 >    /*
482 >     * Possible values for "how" argument in xfer method. Beware that
483 >     * the order of assigned numerical values matters.
484       */
485 <    private boolean advanceHead(Node<E> h, Node<E> nh) {
486 <        if (h == head.get() && head.compareAndSet(h, nh)) {
487 <            h.clearNext(); // forget old next
488 <            return true;
489 <        }
490 <        return false;
485 >    private static final int NOW     = 0; // for untimed poll, tryTransfer
486 >    private static final int ASYNC   = 1; // for offer, put, add
487 >    private static final int SYNC    = 2; // for transfer, take
488 >    private static final int TIMEOUT = 3; // for timed poll, tryTransfer
489 >
490 >    @SuppressWarnings("unchecked")
491 >    static <E> E cast(Object item) {
492 >        assert item == null || item.getClass() != Node.class;
493 >        return (E) item;
494      }
495  
496      /**
497 <     * Puts or takes an item. Used for most queue operations (except
204 <     * poll() and tryTransfer()). See the similar code in
205 <     * SynchronousQueue for detailed explanation.
497 >     * Implements all queuing methods. See above for explanation.
498       *
499 <     * @param e the item or if null, signifies that this is a take
500 <     * @param mode the wait mode: NOWAIT, TIMEOUT, WAIT
499 >     * @param e the item or null for take
500 >     * @param haveData true if this is a put, else a take
501 >     * @param how NOW, ASYNC, SYNC, or TIMEOUT
502       * @param nanos timeout in nanosecs, used only if mode is TIMEOUT
503 <     * @return an item, or null on failure
503 >     * @return an item if matched, else e
504 >     * @throws NullPointerException if haveData mode but e is null
505       */
506 <    private E xfer(E e, int mode, long nanos) {
507 <        boolean isData = (e != null);
508 <        Node<E> s = null;
509 <        final PaddedAtomicReference<Node<E>> head = this.head;
216 <        final PaddedAtomicReference<Node<E>> tail = this.tail;
506 >    private E xfer(E e, boolean haveData, int how, long nanos) {
507 >        if (haveData && (e == null))
508 >            throw new NullPointerException();
509 >        Node s = null;                        // the node to append, if needed
510  
511 <        for (;;) {
219 <            Node<E> t = tail.get();
220 <            Node<E> h = head.get();
511 >        retry: for (;;) {                     // restart on append race
512  
513 <            if (t != null && (t == h || t.isData == isData)) {
514 <                if (s == null)
515 <                    s = new Node<E>(e, isData);
516 <                Node<E> last = t.next;
517 <                if (last != null) {
518 <                    if (t == tail.get())
519 <                        tail.compareAndSet(t, last);
520 <                }
521 <                else if (t.casNext(null, s)) {
522 <                    tail.compareAndSet(t, s);
523 <                    return awaitFulfill(t, s, e, mode, nanos);
524 <                }
525 <            }
526 <
527 <            else if (h != null) {
528 <                Node<E> first = h.next;
529 <                if (t == tail.get() && first != null &&
530 <                    advanceHead(h, first)) {
531 <                    Object x = first.get();
532 <                    if (x != first && first.compareAndSet(x, e)) {
533 <                        LockSupport.unpark(first.waiter);
243 <                        return isData ? e : (E) x;
513 >            for (Node h = head, p = h; p != null;) { // find & match first node
514 >                boolean isData = p.isData;
515 >                Object item = p.item;
516 >                if (item != p && (item != null) == isData) { // unmatched
517 >                    if (isData == haveData)   // can't match
518 >                        break;
519 >                    if (p.casItem(item, e)) { // match
520 >                        for (Node q = p; q != h;) {
521 >                            Node n = q.next;  // update head by 2
522 >                            if (n != null)    // unless singleton
523 >                                q = n;
524 >                            if (head == h && casHead(h, q)) {
525 >                                h.forgetNext();
526 >                                break;
527 >                            }                 // advance and retry
528 >                            if ((h = head)   == null ||
529 >                                (q = h.next) == null || !q.isMatched())
530 >                                break;        // unless slack < 2
531 >                        }
532 >                        LockSupport.unpark(p.waiter);
533 >                        return this.<E>cast(item);
534                      }
535                  }
536 +                Node n = p.next;
537 +                p = (p != n) ? n : (h = head); // Use head if p offlist
538              }
539 +
540 +            if (how != NOW) {                  // No matches available
541 +                if (s == null)
542 +                    s = new Node(e, haveData);
543 +                Node pred = tryAppend(s, haveData);
544 +                if (pred == null)
545 +                    continue retry;           // lost race vs opposite mode
546 +                if (how >= SYNC)
547 +                    return awaitMatch(s, pred, e, how, nanos);
548 +            }
549 +            return e; // not waiting
550          }
551      }
552  
250
553      /**
554 <     * Version of xfer for poll() and tryTransfer, which
555 <     * simplifies control paths both here and in xfer.
556 <     */
557 <    private E fulfill(E e) {
558 <        boolean isData = (e != null);
559 <        final PaddedAtomicReference<Node<E>> head = this.head;
560 <        final PaddedAtomicReference<Node<E>> tail = this.tail;
561 <
562 <        for (;;) {
563 <            Node<E> t = tail.get();
564 <            Node<E> h = head.get();
565 <
566 <            if (t != null && (t == h || t.isData == isData)) {
567 <                Node<E> last = t.next;
568 <                if (t == tail.get()) {
569 <                    if (last != null)
570 <                        tail.compareAndSet(t, last);
571 <                    else
572 <                        return null;
573 <                }
574 <            }
575 <            else if (h != null) {
576 <                Node<E> first = h.next;
577 <                if (t == tail.get() &&
578 <                    first != null &&
579 <                    advanceHead(h, first)) {
580 <                    Object x = first.get();
581 <                    if (x != first && first.compareAndSet(x, e)) {
280 <                        LockSupport.unpark(first.waiter);
281 <                        return isData ? e : (E) x;
282 <                    }
554 >     * Tries to append node s as tail.
555 >     *
556 >     * @param s the node to append
557 >     * @param haveData true if appending in data mode
558 >     * @return null on failure due to losing race with append in
559 >     * different mode, else s's predecessor, or s itself if no
560 >     * predecessor
561 >     */
562 >    private Node tryAppend(Node s, boolean haveData) {
563 >        for (Node t = tail, p = t;;) {        // move p to last node and append
564 >            Node n, u;                        // temps for reads of next & tail
565 >            if (p == null && (p = head) == null) {
566 >                if (casHead(null, s))
567 >                    return s;                 // initialize
568 >            }
569 >            else if (p.cannotPrecede(haveData))
570 >                return null;                  // lost race vs opposite mode
571 >            else if ((n = p.next) != null)    // not last; keep traversing
572 >                p = p != t && t != (u = tail) ? (t = u) : // stale tail
573 >                    (p != n) ? n : null;      // restart if off list
574 >            else if (!p.casNext(null, s))
575 >                p = p.next;                   // re-read on CAS failure
576 >            else {
577 >                if (p != t) {                 // update if slack now >= 2
578 >                    while ((tail != t || !casTail(t, s)) &&
579 >                           (t = tail)   != null &&
580 >                           (s = t.next) != null && // advance and retry
581 >                           (s = s.next) != null && s != t);
582                  }
583 +                return p;
584              }
585          }
586      }
587  
588      /**
589 <     * Spins/blocks until node s is fulfilled or caller gives up,
290 <     * depending on wait mode.
589 >     * Spins/yields/blocks until node s is matched or caller gives up.
590       *
292     * @param pred the predecessor of waiting node
591       * @param s the waiting node
592 +     * @param pred the predecessor of s, or s itself if it has no
593 +     * predecessor, or null if unknown (the null case does not occur
594 +     * in any current calls but may in possible future extensions)
595       * @param e the comparison value for checking match
596 <     * @param mode mode
596 >     * @param how either SYNC or TIMEOUT
597       * @param nanos timeout value
598 <     * @return matched item, or s if cancelled
598 >     * @return matched item, or e if unmatched on interrupt or timeout
599       */
600 <    private E awaitFulfill(Node<E> pred, Node<E> s, E e,
601 <                           int mode, long nanos) {
301 <        if (mode == NOWAIT)
302 <            return null;
303 <
304 <        long lastTime = (mode == TIMEOUT) ? System.nanoTime() : 0;
600 >    private E awaitMatch(Node s, Node pred, E e, int how, long nanos) {
601 >        long lastTime = (how == TIMEOUT) ? System.nanoTime() : 0L;
602          Thread w = Thread.currentThread();
603 <        int spins = -1; // set to desired spin count below
603 >        int spins = -1; // initialized after first item and cancel checks
604 >        ThreadLocalRandom randomYields = null; // bound if needed
605 >
606          for (;;) {
607 <            if (w.isInterrupted())
608 <                s.compareAndSet(e, s);
609 <            Object x = s.get();
610 <            if (x != e) {                 // Node was matched or cancelled
611 <                advanceHead(pred, s);     // unlink if head
612 <                if (x == s) {             // was cancelled
613 <                    clean(pred, s);
614 <                    return null;
615 <                }
616 <                else if (x != null) {
617 <                    s.set(s);             // avoid garbage retention
618 <                    return (E) x;
619 <                }
620 <                else
621 <                    return e;
607 >            Object item = s.item;
608 >            if (item != e) {                  // matched
609 >                assert item != s;
610 >                s.forgetContents();           // avoid garbage
611 >                return this.<E>cast(item);
612 >            }
613 >            if ((w.isInterrupted() || (how == TIMEOUT && nanos <= 0)) &&
614 >                    s.casItem(e, s)) {       // cancel
615 >                unsplice(pred, s);
616 >                return e;
617 >            }
618 >
619 >            if (spins < 0) {                  // establish spins at/near front
620 >                if ((spins = spinsFor(pred, s.isData)) > 0)
621 >                    randomYields = ThreadLocalRandom.current();
622 >            }
623 >            else if (spins > 0) {             // spin
624 >                if (--spins == 0)
625 >                    shortenHeadPath();        // reduce slack before blocking
626 >                else if (randomYields.nextInt(CHAINED_SPINS) == 0)
627 >                    Thread.yield();           // occasionally yield
628              }
629 <            if (mode == TIMEOUT) {
629 >            else if (s.waiter == null) {
630 >                s.waiter = w;                 // request unpark then recheck
631 >            }
632 >            else if (how == TIMEOUT) {
633                  long now = System.nanoTime();
634 <                nanos -= now - lastTime;
634 >                if ((nanos -= now - lastTime) > 0)
635 >                    LockSupport.parkNanos(this, nanos);
636                  lastTime = now;
328                if (nanos <= 0) {
329                    s.compareAndSet(e, s); // try to cancel
330                    continue;
331                }
637              }
638 <            if (spins < 0) {
334 <                Node<E> h = head.get(); // only spin if at head
335 <                spins = ((h != null && h.next == s) ?
336 <                         ((mode == TIMEOUT) ?
337 <                          maxTimedSpins : maxUntimedSpins) : 0);
338 <            }
339 <            if (spins > 0)
340 <                --spins;
341 <            else if (s.waiter == null)
342 <                s.waiter = w;
343 <            else if (mode != TIMEOUT) {
638 >            else {
639                  LockSupport.park(this);
640                  s.waiter = null;
641 <                spins = -1;
347 <            }
348 <            else if (nanos > spinForTimeoutThreshold) {
349 <                LockSupport.parkNanos(this, nanos);
350 <                s.waiter = null;
351 <                spins = -1;
641 >                spins = -1;                   // spin if front upon wakeup
642              }
643          }
644      }
645  
646      /**
647 <     * Returns validated tail for use in cleaning methods.
647 >     * Returns spin/yield value for a node with given predecessor and
648 >     * data mode. See above for explanation.
649       */
650 <    private Node<E> getValidatedTail() {
651 <        for (;;) {
652 <            Node<E> h = head.get();
653 <            Node<E> first = h.next;
654 <            if (first != null && first.next == first) { // help advance
655 <                advanceHead(h, first);
656 <                continue;
650 >    private static int spinsFor(Node pred, boolean haveData) {
651 >        if (MP && pred != null) {
652 >            if (pred.isData != haveData)      // phase change
653 >                return FRONT_SPINS + CHAINED_SPINS;
654 >            if (pred.isMatched())             // probably at front
655 >                return FRONT_SPINS;
656 >            if (pred.waiter == null)          // pred apparently spinning
657 >                return CHAINED_SPINS;
658 >        }
659 >        return 0;
660 >    }
661 >
662 >    /**
663 >     * Tries (once) to unsplice nodes between head and first unmatched
664 >     * or trailing node; failing on contention.
665 >     */
666 >    private void shortenHeadPath() {
667 >        Node h, hn, p, q;
668 >        if ((p = h = head) != null && h.isMatched() &&
669 >            (q = hn = h.next) != null) {
670 >            Node n;
671 >            while ((n = q.next) != q) {
672 >                if (n == null || !q.isMatched()) {
673 >                    if (hn != q && h.next == hn)
674 >                        h.casNext(hn, q);
675 >                    break;
676 >                }
677 >                p = q;
678 >                q = n;
679              }
680 <            Node<E> t = tail.get();
681 <            Node<E> last = t.next;
682 <            if (t == tail.get()) {
683 <                if (last != null)
684 <                    tail.compareAndSet(t, last); // help advance
685 <                else
686 <                    return t;
680 >        }
681 >    }
682 >
683 >    /* -------------- Traversal methods -------------- */
684 >
685 >    /**
686 >     * Returns the successor of p, or the head node if p.next has been
687 >     * linked to self, which will only be true if traversing with a
688 >     * stale pointer that is now off the list.
689 >     */
690 >    final Node succ(Node p) {
691 >        Node next = p.next;
692 >        return (p == next) ? head : next;
693 >    }
694 >
695 >    /**
696 >     * Returns the first unmatched node of the given mode, or null if
697 >     * none.  Used by methods isEmpty, hasWaitingConsumer.
698 >     */
699 >    private Node firstOfMode(boolean isData) {
700 >        for (Node p = head; p != null; p = succ(p)) {
701 >            if (!p.isMatched())
702 >                return (p.isData == isData) ? p : null;
703 >        }
704 >        return null;
705 >    }
706 >
707 >    /**
708 >     * Returns the item in the first unmatched node with isData; or
709 >     * null if none.  Used by peek.
710 >     */
711 >    private E firstDataItem() {
712 >        for (Node p = head; p != null; p = succ(p)) {
713 >            Object item = p.item;
714 >            if (p.isData) {
715 >                if (item != null && item != p)
716 >                    return this.<E>cast(item);
717              }
718 +            else if (item == null)
719 +                return null;
720          }
721 +        return null;
722      }
723  
724      /**
725 <     * Gets rid of cancelled node s with original predecessor pred.
726 <     *
381 <     * @param pred predecessor of cancelled node
382 <     * @param s the cancelled node
725 >     * Traverses and counts unmatched nodes of the given mode.
726 >     * Used by methods size and getWaitingConsumerCount.
727       */
728 <    private void clean(Node<E> pred, Node<E> s) {
729 <        Thread w = s.waiter;
730 <        if (w != null) {             // Wake up thread
731 <            s.waiter = null;
732 <            if (w != Thread.currentThread())
733 <                LockSupport.unpark(w);
728 >    private int countOfMode(boolean data) {
729 >        int count = 0;
730 >        for (Node p = head; p != null; ) {
731 >            if (!p.isMatched()) {
732 >                if (p.isData != data)
733 >                    return 0;
734 >                if (++count == Integer.MAX_VALUE) // saturated
735 >                    break;
736 >            }
737 >            Node n = p.next;
738 >            if (n != p)
739 >                p = n;
740 >            else {
741 >                count = 0;
742 >                p = head;
743 >            }
744 >        }
745 >        return count;
746 >    }
747 >
748 >    final class Itr implements Iterator<E> {
749 >        private Node nextNode;   // next node to return item for
750 >        private E nextItem;      // the corresponding item
751 >        private Node lastRet;    // last returned node, to support remove
752 >        private Node lastPred;   // predecessor to unlink lastRet
753 >
754 >        /**
755 >         * Moves to next node after prev, or first node if prev null.
756 >         */
757 >        private void advance(Node prev) {
758 >            lastPred = lastRet;
759 >            lastRet = prev;
760 >            for (Node p = (prev == null) ? head : succ(prev);
761 >                 p != null; p = succ(p)) {
762 >                Object item = p.item;
763 >                if (p.isData) {
764 >                    if (item != null && item != p) {
765 >                        nextItem = LinkedTransferQueue.this.<E>cast(item);
766 >                        nextNode = p;
767 >                        return;
768 >                    }
769 >                }
770 >                else if (item == null)
771 >                    break;
772 >            }
773 >            nextNode = null;
774          }
775  
776 <        if (pred == null)
777 <            return;
776 >        Itr() {
777 >            advance(null);
778 >        }
779 >
780 >        public final boolean hasNext() {
781 >            return nextNode != null;
782 >        }
783 >
784 >        public final E next() {
785 >            Node p = nextNode;
786 >            if (p == null) throw new NoSuchElementException();
787 >            E e = nextItem;
788 >            advance(p);
789 >            return e;
790 >        }
791 >
792 >        public final void remove() {
793 >            Node p = lastRet;
794 >            if (p == null) throw new IllegalStateException();
795 >            findAndRemoveDataNode(lastPred, p);
796 >        }
797 >    }
798 >
799 >    /* -------------- Removal methods -------------- */
800  
801 +    /**
802 +     * Unsplices (now or later) the given deleted/cancelled node with
803 +     * the given predecessor.
804 +     *
805 +     * @param pred predecessor of node to be unspliced
806 +     * @param s the node to be unspliced
807 +     */
808 +    private void unsplice(Node pred, Node s) {
809 +        s.forgetContents(); // clear unneeded fields
810          /*
811           * At any given time, exactly one node on list cannot be
812 <         * deleted -- the last inserted node. To accommodate this, if
813 <         * we cannot delete s, we save its predecessor as "cleanMe",
814 <         * processing the previously saved version first. At least one
815 <         * of node s or the node previously saved can always be
812 >         * unlinked -- the last inserted node. To accommodate this, if
813 >         * we cannot unlink s, we save its predecessor as "cleanMe",
814 >         * processing the previously saved version first. Because only
815 >         * one node in the list can have a null next, at least one of
816 >         * node s or the node previously saved can always be
817           * processed, so this always terminates.
818           */
819 <        while (pred.next == s) {
820 <            Node<E> oldpred = reclean();  // First, help get rid of cleanMe
821 <            Node<E> t = getValidatedTail();
822 <            if (s != t) {               // If not tail, try to unsplice
823 <                Node<E> sn = s.next;      // s.next == s means s already off list
824 <                if (sn == s || pred.casNext(s, sn))
819 >        if (pred != null && pred != s) {
820 >            while (pred.next == s) {
821 >                Node oldpred = (cleanMe == null) ? null : reclean();
822 >                Node n = s.next;
823 >                if (n != null) {
824 >                    if (n != s)
825 >                        pred.casNext(s, n);
826                      break;
827 +                }
828 +                if (oldpred == pred ||      // Already saved
829 +                    ((oldpred == null || oldpred.next == s) &&
830 +                     casCleanMe(oldpred, pred))) {
831 +                    break;
832 +                }
833              }
411            else if (oldpred == pred || // Already saved
412                     (oldpred == null && cleanMe.compareAndSet(null, pred)))
413                break;                  // Postpone cleaning
834          }
835      }
836  
837      /**
838 <     * Tries to unsplice the cancelled node held in cleanMe that was
839 <     * previously uncleanable because it was at tail.
838 >     * Tries to unsplice the deleted/cancelled node held in cleanMe
839 >     * that was previously uncleanable because it was at tail.
840       *
841       * @return current cleanMe node (or null)
842       */
843 <    private Node<E> reclean() {
843 >    private Node reclean() {
844          /*
845 <         * cleanMe is, or at one time was, predecessor of cancelled
846 <         * node s that was the tail so could not be unspliced.  If s
845 >         * cleanMe is, or at one time was, predecessor of a cancelled
846 >         * node s that was the tail so could not be unspliced.  If it
847           * is no longer the tail, try to unsplice if necessary and
848           * make cleanMe slot available.  This differs from similar
849 <         * code in clean() because we must check that pred still
850 <         * points to a cancelled node that must be unspliced -- if
851 <         * not, we can (must) clear cleanMe without unsplicing.
852 <         * This can loop only due to contention on casNext or
433 <         * clearing cleanMe.
849 >         * code in unsplice() because we must check that pred still
850 >         * points to a matched node that can be unspliced -- if not,
851 >         * we can (must) clear cleanMe without unsplicing.  This can
852 >         * loop only due to contention.
853           */
854 <        Node<E> pred;
855 <        while ((pred = cleanMe.get()) != null) {
856 <            Node<E> t = getValidatedTail();
857 <            Node<E> s = pred.next;
858 <            if (s != t) {
859 <                Node<E> sn;
860 <                if (s == null || s == pred || s.get() != s ||
861 <                    (sn = s.next) == s || pred.casNext(s, sn))
862 <                    cleanMe.compareAndSet(pred, null);
854 >        Node pred;
855 >        while ((pred = cleanMe) != null) {
856 >            Node s = pred.next;
857 >            Node n;
858 >            if (s == null || s == pred || !s.isMatched())
859 >                casCleanMe(pred, null); // already gone
860 >            else if ((n = s.next) != null) {
861 >                if (n != s)
862 >                    pred.casNext(s, n);
863 >                casCleanMe(pred, null);
864              }
865 <            else // s is still tail; cannot clean
865 >            else
866                  break;
867          }
868          return pred;
869      }
870  
871      /**
872 +     * Main implementation of Iterator.remove(). Find
873 +     * and unsplice the given data node.
874 +     * @param possiblePred possible predecessor of s
875 +     * @param s the node to remove
876 +     */
877 +    final void findAndRemoveDataNode(Node possiblePred, Node s) {
878 +        assert s.isData;
879 +        if (s.tryMatchData()) {
880 +            if (possiblePred != null && possiblePred.next == s)
881 +                unsplice(possiblePred, s); // was actual predecessor
882 +            else {
883 +                for (Node pred = null, p = head; p != null; ) {
884 +                    if (p == s) {
885 +                        unsplice(pred, p);
886 +                        break;
887 +                    }
888 +                    if (p.isUnmatchedRequest())
889 +                        break;
890 +                    pred = p;
891 +                    if ((p = p.next) == pred) { // stale
892 +                        pred = null;
893 +                        p = head;
894 +                    }
895 +                }
896 +            }
897 +        }
898 +    }
899 +
900 +    /**
901 +     * Main implementation of remove(Object)
902 +     */
903 +    private boolean findAndRemove(Object e) {
904 +        if (e != null) {
905 +            for (Node pred = null, p = head; p != null; ) {
906 +                Object item = p.item;
907 +                if (p.isData) {
908 +                    if (item != null && item != p && e.equals(item) &&
909 +                        p.tryMatchData()) {
910 +                        unsplice(pred, p);
911 +                        return true;
912 +                    }
913 +                }
914 +                else if (item == null)
915 +                    break;
916 +                pred = p;
917 +                if ((p = p.next) == pred) { // stale
918 +                    pred = null;
919 +                    p = head;
920 +                }
921 +            }
922 +        }
923 +        return false;
924 +    }
925 +
926 +
927 +    /**
928       * Creates an initially empty {@code LinkedTransferQueue}.
929       */
930      public LinkedTransferQueue() {
455        Node<E> dummy = new Node<E>(null, false);
456        head = new PaddedAtomicReference<Node<E>>(dummy);
457        tail = new PaddedAtomicReference<Node<E>>(dummy);
458        cleanMe = new PaddedAtomicReference<Node<E>>(null);
931      }
932  
933      /**
# Line 479 | Line 951 | public class LinkedTransferQueue<E> exte
951       * @throws NullPointerException if the specified element is null
952       */
953      public void put(E e) {
954 <        offer(e);
954 >        xfer(e, true, ASYNC, 0);
955      }
956  
957      /**
# Line 492 | Line 964 | public class LinkedTransferQueue<E> exte
964       * @throws NullPointerException if the specified element is null
965       */
966      public boolean offer(E e, long timeout, TimeUnit unit) {
967 <        return offer(e);
967 >        xfer(e, true, ASYNC, 0);
968 >        return true;
969      }
970  
971      /**
# Line 504 | Line 977 | public class LinkedTransferQueue<E> exte
977       * @throws NullPointerException if the specified element is null
978       */
979      public boolean offer(E e) {
980 <        if (e == null) throw new NullPointerException();
508 <        xfer(e, NOWAIT, 0);
980 >        xfer(e, true, ASYNC, 0);
981          return true;
982      }
983  
984      /**
985       * Inserts the specified element at the tail of this queue.
986 <     * As the queue is unbounded this method will never throw
986 >     * As the queue is unbounded, this method will never throw
987       * {@link IllegalStateException} or return {@code false}.
988       *
989       * @return {@code true} (as specified by {@link Collection#add})
990       * @throws NullPointerException if the specified element is null
991       */
992      public boolean add(E e) {
993 <        return offer(e);
993 >        xfer(e, true, ASYNC, 0);
994 >        return true;
995      }
996  
997      /**
998 <     * Transfers the specified element immediately if there exists a
999 <     * consumer already waiting to receive it (in {@link #take} or
1000 <     * timed {@link #poll(Object,long,TimeUnit) poll}), otherwise
1001 <     * returning {@code false} without enqueuing the element.
998 >     * Transfers the element to a waiting consumer immediately, if possible.
999 >     *
1000 >     * <p>More precisely, transfers the specified element immediately
1001 >     * if there exists a consumer already waiting to receive it (in
1002 >     * {@link #take} or timed {@link #poll(long,TimeUnit) poll}),
1003 >     * otherwise returning {@code false} without enqueuing the element.
1004       *
1005       * @throws NullPointerException if the specified element is null
1006       */
1007      public boolean tryTransfer(E e) {
1008 <        if (e == null) throw new NullPointerException();
534 <        return fulfill(e) != null;
1008 >        return xfer(e, true, NOW, 0) == null;
1009      }
1010  
1011      /**
1012 <     * Inserts the specified element at the tail of this queue,
1013 <     * waiting if necessary for the element to be received by a
1014 <     * consumer invoking {@code take} or {@code poll}.
1012 >     * Transfers the element to a consumer, waiting if necessary to do so.
1013 >     *
1014 >     * <p>More precisely, transfers the specified element immediately
1015 >     * if there exists a consumer already waiting to receive it (in
1016 >     * {@link #take} or timed {@link #poll(long,TimeUnit) poll}),
1017 >     * else inserts the specified element at the tail of this queue
1018 >     * and waits until the element is received by a consumer.
1019       *
542     * @throws InterruptedException {@inheritDoc}
1020       * @throws NullPointerException if the specified element is null
1021       */
1022      public void transfer(E e) throws InterruptedException {
1023 <        if (e == null) throw new NullPointerException();
1024 <        if (xfer(e, WAIT, 0) == null) {
548 <            Thread.interrupted();
1023 >        if (xfer(e, true, SYNC, 0) != null) {
1024 >            Thread.interrupted(); // failure possible only due to interrupt
1025              throw new InterruptedException();
1026          }
1027      }
1028  
1029      /**
1030 <     * Inserts the specified element at the tail of this queue,
1031 <     * waiting up to the specified wait time for the element to be
1032 <     * received by a consumer invoking {@code take} or {@code poll}.
1030 >     * Transfers the element to a consumer if it is possible to do so
1031 >     * before the timeout elapses.
1032 >     *
1033 >     * <p>More precisely, transfers the specified element immediately
1034 >     * if there exists a consumer already waiting to receive it (in
1035 >     * {@link #take} or timed {@link #poll(long,TimeUnit) poll}),
1036 >     * else inserts the specified element at the tail of this queue
1037 >     * and waits until the element is received by a consumer,
1038 >     * returning {@code false} if the specified wait time elapses
1039 >     * before the element can be transferred.
1040       *
558     * @throws InterruptedException {@inheritDoc}
1041       * @throws NullPointerException if the specified element is null
1042       */
1043      public boolean tryTransfer(E e, long timeout, TimeUnit unit)
1044          throws InterruptedException {
1045 <        if (e == null) throw new NullPointerException();
564 <        if (xfer(e, TIMEOUT, unit.toNanos(timeout)) != null)
1045 >        if (xfer(e, true, TIMEOUT, unit.toNanos(timeout)) == null)
1046              return true;
1047          if (!Thread.interrupted())
1048              return false;
# Line 569 | Line 1050 | public class LinkedTransferQueue<E> exte
1050      }
1051  
1052      public E take() throws InterruptedException {
1053 <        E e = xfer(null, WAIT, 0);
1053 >        E e = xfer(null, false, SYNC, 0);
1054          if (e != null)
1055              return e;
1056          Thread.interrupted();
1057          throw new InterruptedException();
1058      }
1059  
579    /**
580     * @throws InterruptedException {@inheritDoc}
581     */
1060      public E poll(long timeout, TimeUnit unit) throws InterruptedException {
1061 <        E e = xfer(null, TIMEOUT, unit.toNanos(timeout));
1061 >        E e = xfer(null, false, TIMEOUT, unit.toNanos(timeout));
1062          if (e != null || !Thread.interrupted())
1063              return e;
1064          throw new InterruptedException();
1065      }
1066  
1067      public E poll() {
1068 <        return fulfill(null);
1068 >        return xfer(null, false, NOW, 0);
1069      }
1070  
1071      /**
# Line 626 | Line 1104 | public class LinkedTransferQueue<E> exte
1104          return n;
1105      }
1106  
629    // Traversal-based methods
630
631    /**
632     * Returns head after performing any outstanding helping steps.
633     */
634    private Node<E> traversalHead() {
635        for (;;) {
636            Node<E> t = tail.get();
637            Node<E> h = head.get();
638            if (h != null && t != null) {
639                Node<E> last = t.next;
640                Node<E> first = h.next;
641                if (t == tail.get()) {
642                    if (last != null)
643                        tail.compareAndSet(t, last);
644                    else if (first != null) {
645                        Object x = first.get();
646                        if (x == first)
647                            advanceHead(h, first);
648                        else
649                            return h;
650                    }
651                    else
652                        return h;
653                }
654            }
655            reclean();
656        }
657    }
658
1107      /**
1108       * Returns an iterator over the elements in this queue in proper
1109       * sequence, from head to tail.
# Line 673 | Line 1121 | public class LinkedTransferQueue<E> exte
1121          return new Itr();
1122      }
1123  
676    /**
677     * Iterators. Basic strategy is to traverse list, treating
678     * non-data (i.e., request) nodes as terminating list.
679     * Once a valid data node is found, the item is cached
680     * so that the next call to next() will return it even
681     * if subsequently removed.
682     */
683    class Itr implements Iterator<E> {
684        Node<E> next;        // node to return next
685        Node<E> pnext;       // predecessor of next
686        Node<E> curr;        // last returned node, for remove()
687        Node<E> pcurr;       // predecessor of curr, for remove()
688        E nextItem;          // Cache of next item, once committed to in next
689
690        Itr() {
691            advance();
692        }
693
694        /**
695         * Moves to next valid node and returns item to return for
696         * next(), or null if no such.
697         */
698        private E advance() {
699            pcurr = pnext;
700            curr = next;
701            E item = nextItem;
702
703            for (;;) {
704                pnext = (next == null) ? traversalHead() : next;
705                next = pnext.next;
706                if (next == pnext) {
707                    next = null;
708                    continue;  // restart
709                }
710                if (next == null)
711                    break;
712                Object x = next.get();
713                if (x != null && x != next) {
714                    nextItem = (E) x;
715                    break;
716                }
717            }
718            return item;
719        }
720
721        public boolean hasNext() {
722            return next != null;
723        }
724
725        public E next() {
726            if (next == null)
727                throw new NoSuchElementException();
728            return advance();
729        }
730
731        public void remove() {
732            Node<E> p = curr;
733            if (p == null)
734                throw new IllegalStateException();
735            Object x = p.get();
736            if (x != null && x != p && p.compareAndSet(x, p))
737                clean(pcurr, p);
738        }
739    }
740
1124      public E peek() {
1125 <        for (;;) {
743 <            Node<E> h = traversalHead();
744 <            Node<E> p = h.next;
745 <            if (p == null)
746 <                return null;
747 <            Object x = p.get();
748 <            if (p != x) {
749 <                if (!p.isData)
750 <                    return null;
751 <                if (x != null)
752 <                    return (E) x;
753 <            }
754 <        }
1125 >        return firstDataItem();
1126      }
1127  
1128 +    /**
1129 +     * Returns {@code true} if this queue contains no elements.
1130 +     *
1131 +     * @return {@code true} if this queue contains no elements
1132 +     */
1133      public boolean isEmpty() {
1134 <        for (;;) {
759 <            Node<E> h = traversalHead();
760 <            Node<E> p = h.next;
761 <            if (p == null)
762 <                return true;
763 <            Object x = p.get();
764 <            if (p != x) {
765 <                if (!p.isData)
766 <                    return true;
767 <                if (x != null)
768 <                    return false;
769 <            }
770 <        }
1134 >        return firstOfMode(true) == null;
1135      }
1136  
1137      public boolean hasWaitingConsumer() {
1138 <        for (;;) {
775 <            Node<E> h = traversalHead();
776 <            Node<E> p = h.next;
777 <            if (p == null)
778 <                return false;
779 <            Object x = p.get();
780 <            if (p != x)
781 <                return !p.isData;
782 <        }
1138 >        return firstOfMode(false) != null;
1139      }
1140  
1141      /**
# Line 795 | Line 1151 | public class LinkedTransferQueue<E> exte
1151       * @return the number of elements in this queue
1152       */
1153      public int size() {
1154 <        for (;;) {
799 <            int count = 0;
800 <            Node<E> pred = traversalHead();
801 <            for (;;) {
802 <                Node<E> q = pred.next;
803 <                if (q == pred) // restart
804 <                    break;
805 <                if (q == null || !q.isData)
806 <                    return count;
807 <                Object x = q.get();
808 <                if (x != null && x != q) {
809 <                    if (++count == Integer.MAX_VALUE) // saturated
810 <                        return count;
811 <                }
812 <                pred = q;
813 <            }
814 <        }
1154 >        return countOfMode(true);
1155      }
1156  
1157      public int getWaitingConsumerCount() {
1158 <        // converse of size -- count valid non-data nodes
819 <        for (;;) {
820 <            int count = 0;
821 <            Node<E> pred = traversalHead();
822 <            for (;;) {
823 <                Node<E> q = pred.next;
824 <                if (q == pred) // restart
825 <                    break;
826 <                if (q == null || q.isData)
827 <                    return count;
828 <                Object x = q.get();
829 <                if (x == null) {
830 <                    if (++count == Integer.MAX_VALUE) // saturated
831 <                        return count;
832 <                }
833 <                pred = q;
834 <            }
835 <        }
1158 >        return countOfMode(false);
1159      }
1160  
1161 +    /**
1162 +     * Removes a single instance of the specified element from this queue,
1163 +     * if it is present.  More formally, removes an element {@code e} such
1164 +     * that {@code o.equals(e)}, if this queue contains one or more such
1165 +     * elements.
1166 +     * Returns {@code true} if this queue contained the specified element
1167 +     * (or equivalently, if this queue changed as a result of the call).
1168 +     *
1169 +     * @param o element to be removed from this queue, if present
1170 +     * @return {@code true} if this queue changed as a result of the call
1171 +     */
1172      public boolean remove(Object o) {
1173 <        if (o == null)
840 <            return false;
841 <        for (;;) {
842 <            Node<E> pred = traversalHead();
843 <            for (;;) {
844 <                Node<E> q = pred.next;
845 <                if (q == pred) // restart
846 <                    break;
847 <                if (q == null || !q.isData)
848 <                    return false;
849 <                Object x = q.get();
850 <                if (x != null && x != q && o.equals(x) &&
851 <                    q.compareAndSet(x, q)) {
852 <                    clean(pred, q);
853 <                    return true;
854 <                }
855 <                pred = q;
856 <            }
857 <        }
1173 >        return findAndRemove(o);
1174      }
1175  
1176      /**
# Line 869 | Line 1185 | public class LinkedTransferQueue<E> exte
1185      }
1186  
1187      /**
1188 <     * Save the state to a stream (that is, serialize it).
1188 >     * Saves the state to a stream (that is, serializes it).
1189       *
1190       * @serialData All of the elements (each an {@code E}) in
1191       * the proper order, followed by a null
# Line 885 | Line 1201 | public class LinkedTransferQueue<E> exte
1201      }
1202  
1203      /**
1204 <     * Reconstitute the Queue instance from a stream (that is,
1205 <     * deserialize it).
1204 >     * Reconstitutes the Queue instance from a stream (that is,
1205 >     * deserializes it).
1206       *
1207       * @param s the stream
1208       */
1209      private void readObject(java.io.ObjectInputStream s)
1210          throws java.io.IOException, ClassNotFoundException {
1211          s.defaultReadObject();
896        resetHeadAndTail();
1212          for (;;) {
1213              @SuppressWarnings("unchecked") E item = (E) s.readObject();
1214              if (item == null)
# Line 903 | Line 1218 | public class LinkedTransferQueue<E> exte
1218          }
1219      }
1220  
906    // Support for resetting head/tail while deserializing
907    private void resetHeadAndTail() {
908        Node<E> dummy = new Node<E>(null, false);
909        UNSAFE.putObjectVolatile(this, headOffset,
910                                 new PaddedAtomicReference<Node<E>>(dummy));
911        UNSAFE.putObjectVolatile(this, tailOffset,
912                                 new PaddedAtomicReference<Node<E>>(dummy));
913        UNSAFE.putObjectVolatile(this, cleanMeOffset,
914                                 new PaddedAtomicReference<Node<E>>(null));
915    }
916
1221      // Unsafe mechanics
1222  
1223      private static final sun.misc.Unsafe UNSAFE = getUnsafe();
# Line 924 | Line 1228 | public class LinkedTransferQueue<E> exte
1228      private static final long cleanMeOffset =
1229          objectFieldOffset(UNSAFE, "cleanMe", LinkedTransferQueue.class);
1230  
927
1231      static long objectFieldOffset(sun.misc.Unsafe UNSAFE,
1232                                    String field, Class<?> klazz) {
1233          try {
# Line 944 | Line 1247 | public class LinkedTransferQueue<E> exte
1247       *
1248       * @return a sun.misc.Unsafe
1249       */
1250 <    private static sun.misc.Unsafe getUnsafe() {
1250 >    static sun.misc.Unsafe getUnsafe() {
1251          try {
1252              return sun.misc.Unsafe.getUnsafe();
1253          } catch (SecurityException se) {
# Line 964 | Line 1267 | public class LinkedTransferQueue<E> exte
1267              }
1268          }
1269      }
1270 +
1271   }

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines