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 1.43 by jsr166, Sat Aug 1 21:24:01 2009 UTC vs.
Revision 1.78 by dl, Thu Sep 9 16:52:49 2010 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;
18  
19   /**
20   * An unbounded {@link TransferQueue} based on linked nodes.
# Line 54 | Line 53 | public class LinkedTransferQueue<E> exte
53      private static final long serialVersionUID = -3223113410248163686L;
54  
55      /*
56 <     * 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)
56 >     * *** Overview of Dual Queues with Slack ***
57       *
58 <     * The main extension is to provide different Wait modes for the
59 <     * main "xfer" method that puts or takes items.  These don't
60 <     * impact the basic dual-queue logic, but instead control whether
61 <     * or how threads block upon insertion of request or data nodes
62 <     * into the dual queue. It also uses slightly different
63 <     * conventions for tracking whether nodes are off-list or
64 <     * cancelled.
65 <     */
66 <
67 <    // Wait modes for xfer method
68 <    static final int NOWAIT  = 0;
69 <    static final int TIMEOUT = 1;
70 <    static final int WAIT    = 2;
71 <
72 <    /** The number of CPUs, for spin control */
73 <    static final int NCPUS = Runtime.getRuntime().availableProcessors();
74 <
75 <    /**
76 <     * The number of times to spin before blocking in timed waits.
77 <     * The value is empirically derived -- it works well across a
78 <     * variety of processors and OSes. Empirically, the best value
79 <     * seems not to vary with number of CPUs (beyond 2) so is just
80 <     * a constant.
81 <     */
82 <    static final int maxTimedSpins = (NCPUS < 2) ? 0 : 32;
83 <
84 <    /**
85 <     * The number of times to spin before blocking in untimed waits.
86 <     * This is greater than timed value because untimed waits spin
87 <     * faster since they don't need to check times on each spin.
88 <     */
89 <    static final int maxUntimedSpins = maxTimedSpins * 16;
58 >     * Dual Queues, introduced by Scherer and Scott
59 >     * (http://www.cs.rice.edu/~wns1/papers/2004-DISC-DDS.pdf) are
60 >     * (linked) queues in which nodes may represent either data or
61 >     * requests.  When a thread tries to enqueue a data node, but
62 >     * encounters a request node, it instead "matches" and removes it;
63 >     * and vice versa for enqueuing requests. Blocking Dual Queues
64 >     * arrange that threads enqueuing unmatched requests block until
65 >     * other threads provide the match. Dual Synchronous Queues (see
66 >     * Scherer, Lea, & Scott
67 >     * http://www.cs.rochester.edu/u/scott/papers/2009_Scherer_CACM_SSQ.pdf)
68 >     * additionally arrange that threads enqueuing unmatched data also
69 >     * block.  Dual Transfer Queues support all of these modes, as
70 >     * dictated by callers.
71 >     *
72 >     * A FIFO dual queue may be implemented using a variation of the
73 >     * Michael & Scott (M&S) lock-free queue algorithm
74 >     * (http://www.cs.rochester.edu/u/scott/papers/1996_PODC_queues.pdf).
75 >     * It maintains two pointer fields, "head", pointing to a
76 >     * (matched) node that in turn points to the first actual
77 >     * (unmatched) queue node (or null if empty); and "tail" that
78 >     * points to the last node on the queue (or again null if
79 >     * empty). For example, here is a possible queue with four data
80 >     * elements:
81 >     *
82 >     *  head                tail
83 >     *    |                   |
84 >     *    v                   v
85 >     *    M -> U -> U -> U -> U
86 >     *
87 >     * The M&S queue algorithm is known to be prone to scalability and
88 >     * overhead limitations when maintaining (via CAS) these head and
89 >     * tail pointers. This has led to the development of
90 >     * contention-reducing variants such as elimination arrays (see
91 >     * Moir et al http://portal.acm.org/citation.cfm?id=1074013) and
92 >     * optimistic back pointers (see Ladan-Mozes & Shavit
93 >     * http://people.csail.mit.edu/edya/publications/OptimisticFIFOQueue-journal.pdf).
94 >     * However, the nature of dual queues enables a simpler tactic for
95 >     * improving M&S-style implementations when dual-ness is needed.
96 >     *
97 >     * In a dual queue, each node must atomically maintain its match
98 >     * status. While there are other possible variants, we implement
99 >     * this here as: for a data-mode node, matching entails CASing an
100 >     * "item" field from a non-null data value to null upon match, and
101 >     * vice-versa for request nodes, CASing from null to a data
102 >     * value. (Note that the linearization properties of this style of
103 >     * queue are easy to verify -- elements are made available by
104 >     * linking, and unavailable by matching.) Compared to plain M&S
105 >     * queues, this property of dual queues requires one additional
106 >     * successful atomic operation per enq/deq pair. But it also
107 >     * enables lower cost variants of queue maintenance mechanics. (A
108 >     * variation of this idea applies even for non-dual queues that
109 >     * support deletion of interior elements, such as
110 >     * j.u.c.ConcurrentLinkedQueue.)
111 >     *
112 >     * Once a node is matched, its match status can never again
113 >     * change.  We may thus arrange that the linked list of them
114 >     * contain a prefix of zero or more matched nodes, followed by a
115 >     * suffix of zero or more unmatched nodes. (Note that we allow
116 >     * both the prefix and suffix to be zero length, which in turn
117 >     * means that we do not use a dummy header.)  If we were not
118 >     * concerned with either time or space efficiency, we could
119 >     * correctly perform enqueue and dequeue operations by traversing
120 >     * from a pointer to the initial node; CASing the item of the
121 >     * first unmatched node on match and CASing the next field of the
122 >     * trailing node on appends. (Plus some special-casing when
123 >     * initially empty).  While this would be a terrible idea in
124 >     * itself, it does have the benefit of not requiring ANY atomic
125 >     * updates on head/tail fields.
126 >     *
127 >     * We introduce here an approach that lies between the extremes of
128 >     * never versus always updating queue (head and tail) pointers.
129 >     * This offers a tradeoff between sometimes requiring extra
130 >     * traversal steps to locate the first and/or last unmatched
131 >     * nodes, versus the reduced overhead and contention of fewer
132 >     * updates to queue pointers. For example, a possible snapshot of
133 >     * a queue is:
134 >     *
135 >     *  head           tail
136 >     *    |              |
137 >     *    v              v
138 >     *    M -> M -> U -> U -> U -> U
139 >     *
140 >     * The best value for this "slack" (the targeted maximum distance
141 >     * between the value of "head" and the first unmatched node, and
142 >     * similarly for "tail") is an empirical matter. We have found
143 >     * that using very small constants in the range of 1-3 work best
144 >     * over a range of platforms. Larger values introduce increasing
145 >     * costs of cache misses and risks of long traversal chains, while
146 >     * smaller values increase CAS contention and overhead.
147 >     *
148 >     * Dual queues with slack differ from plain M&S dual queues by
149 >     * virtue of only sometimes updating head or tail pointers when
150 >     * matching, appending, or even traversing nodes; in order to
151 >     * maintain a targeted slack.  The idea of "sometimes" may be
152 >     * operationalized in several ways. The simplest is to use a
153 >     * per-operation counter incremented on each traversal step, and
154 >     * to try (via CAS) to update the associated queue pointer
155 >     * whenever the count exceeds a threshold. Another, that requires
156 >     * more overhead, is to use random number generators to update
157 >     * with a given probability per traversal step.
158 >     *
159 >     * In any strategy along these lines, because CASes updating
160 >     * fields may fail, the actual slack may exceed targeted
161 >     * slack. However, they may be retried at any time to maintain
162 >     * targets.  Even when using very small slack values, this
163 >     * approach works well for dual queues because it allows all
164 >     * operations up to the point of matching or appending an item
165 >     * (hence potentially allowing progress by another thread) to be
166 >     * read-only, thus not introducing any further contention. As
167 >     * described below, we implement this by performing slack
168 >     * maintenance retries only after these points.
169 >     *
170 >     * As an accompaniment to such techniques, traversal overhead can
171 >     * be further reduced without increasing contention of head
172 >     * pointer updates: Threads may sometimes shortcut the "next" link
173 >     * path from the current "head" node to be closer to the currently
174 >     * known first unmatched node, and similarly for tail. Again, this
175 >     * may be triggered with using thresholds or randomization.
176 >     *
177 >     * These ideas must be further extended to avoid unbounded amounts
178 >     * of costly-to-reclaim garbage caused by the sequential "next"
179 >     * links of nodes starting at old forgotten head nodes: As first
180 >     * described in detail by Boehm
181 >     * (http://portal.acm.org/citation.cfm?doid=503272.503282) if a GC
182 >     * delays noticing that any arbitrarily old node has become
183 >     * garbage, all newer dead nodes will also be unreclaimed.
184 >     * (Similar issues arise in non-GC environments.)  To cope with
185 >     * this in our implementation, upon CASing to advance the head
186 >     * pointer, we set the "next" link of the previous head to point
187 >     * only to itself; thus limiting the length of connected dead lists.
188 >     * (We also take similar care to wipe out possibly garbage
189 >     * retaining values held in other Node fields.)  However, doing so
190 >     * adds some further complexity to traversal: If any "next"
191 >     * pointer links to itself, it indicates that the current thread
192 >     * has lagged behind a head-update, and so the traversal must
193 >     * continue from the "head".  Traversals trying to find the
194 >     * current tail starting from "tail" may also encounter
195 >     * self-links, in which case they also continue at "head".
196 >     *
197 >     * It is tempting in slack-based scheme to not even use CAS for
198 >     * updates (similarly to Ladan-Mozes & Shavit). However, this
199 >     * cannot be done for head updates under the above link-forgetting
200 >     * mechanics because an update may leave head at a detached node.
201 >     * And while direct writes are possible for tail updates, they
202 >     * increase the risk of long retraversals, and hence long garbage
203 >     * chains, which can be much more costly than is worthwhile
204 >     * considering that the cost difference of performing a CAS vs
205 >     * write is smaller when they are not triggered on each operation
206 >     * (especially considering that writes and CASes equally require
207 >     * additional GC bookkeeping ("write barriers") that are sometimes
208 >     * more costly than the writes themselves because of contention).
209 >     *
210 >     * *** Overview of implementation ***
211 >     *
212 >     * We use a threshold-based approach to updates, with a slack
213 >     * threshold of two -- that is, we update head/tail when the
214 >     * current pointer appears to be two or more steps away from the
215 >     * first/last node. The slack value is hard-wired: a path greater
216 >     * than one is naturally implemented by checking equality of
217 >     * traversal pointers except when the list has only one element,
218 >     * in which case we keep slack threshold at one. Avoiding tracking
219 >     * explicit counts across method calls slightly simplifies an
220 >     * already-messy implementation. Using randomization would
221 >     * probably work better if there were a low-quality dirt-cheap
222 >     * per-thread one available, but even ThreadLocalRandom is too
223 >     * heavy for these purposes.
224 >     *
225 >     * With such a small slack threshold value, it is not worthwhile
226 >     * to augment this with path short-circuiting (i.e., unsplicing
227 >     * interior nodes) except in the case of cancellation/removal (see
228 >     * below).
229 >     *
230 >     * We allow both the head and tail fields to be null before any
231 >     * nodes are enqueued; initializing upon first append.  This
232 >     * simplifies some other logic, as well as providing more
233 >     * efficient explicit control paths instead of letting JVMs insert
234 >     * implicit NullPointerExceptions when they are null.  While not
235 >     * currently fully implemented, we also leave open the possibility
236 >     * of re-nulling these fields when empty (which is complicated to
237 >     * arrange, for little benefit.)
238 >     *
239 >     * All enqueue/dequeue operations are handled by the single method
240 >     * "xfer" with parameters indicating whether to act as some form
241 >     * of offer, put, poll, take, or transfer (each possibly with
242 >     * timeout). The relative complexity of using one monolithic
243 >     * method outweighs the code bulk and maintenance problems of
244 >     * using separate methods for each case.
245 >     *
246 >     * Operation consists of up to three phases. The first is
247 >     * implemented within method xfer, the second in tryAppend, and
248 >     * the third in method awaitMatch.
249 >     *
250 >     * 1. Try to match an existing node
251 >     *
252 >     *    Starting at head, skip already-matched nodes until finding
253 >     *    an unmatched node of opposite mode, if one exists, in which
254 >     *    case matching it and returning, also if necessary updating
255 >     *    head to one past the matched node (or the node itself if the
256 >     *    list has no other unmatched nodes). If the CAS misses, then
257 >     *    a loop retries advancing head by two steps until either
258 >     *    success or the slack is at most two. By requiring that each
259 >     *    attempt advances head by two (if applicable), we ensure that
260 >     *    the slack does not grow without bound. Traversals also check
261 >     *    if the initial head is now off-list, in which case they
262 >     *    start at the new head.
263 >     *
264 >     *    If no candidates are found and the call was untimed
265 >     *    poll/offer, (argument "how" is NOW) return.
266 >     *
267 >     * 2. Try to append a new node (method tryAppend)
268 >     *
269 >     *    Starting at current tail pointer, find the actual last node
270 >     *    and try to append a new node (or if head was null, establish
271 >     *    the first node). Nodes can be appended only if their
272 >     *    predecessors are either already matched or are of the same
273 >     *    mode. If we detect otherwise, then a new node with opposite
274 >     *    mode must have been appended during traversal, so we must
275 >     *    restart at phase 1. The traversal and update steps are
276 >     *    otherwise similar to phase 1: Retrying upon CAS misses and
277 >     *    checking for staleness.  In particular, if a self-link is
278 >     *    encountered, then we can safely jump to a node on the list
279 >     *    by continuing the traversal at current head.
280 >     *
281 >     *    On successful append, if the call was ASYNC, return.
282 >     *
283 >     * 3. Await match or cancellation (method awaitMatch)
284 >     *
285 >     *    Wait for another thread to match node; instead cancelling if
286 >     *    the current thread was interrupted or the wait timed out. On
287 >     *    multiprocessors, we use front-of-queue spinning: If a node
288 >     *    appears to be the first unmatched node in the queue, it
289 >     *    spins a bit before blocking. In either case, before blocking
290 >     *    it tries to unsplice any nodes between the current "head"
291 >     *    and the first unmatched node.
292 >     *
293 >     *    Front-of-queue spinning vastly improves performance of
294 >     *    heavily contended queues. And so long as it is relatively
295 >     *    brief and "quiet", spinning does not much impact performance
296 >     *    of less-contended queues.  During spins threads check their
297 >     *    interrupt status and generate a thread-local random number
298 >     *    to decide to occasionally perform a Thread.yield. While
299 >     *    yield has underdefined specs, we assume that might it help,
300 >     *    and will not hurt in limiting impact of spinning on busy
301 >     *    systems.  We also use smaller (1/2) spins for nodes that are
302 >     *    not known to be front but whose predecessors have not
303 >     *    blocked -- these "chained" spins avoid artifacts of
304 >     *    front-of-queue rules which otherwise lead to alternating
305 >     *    nodes spinning vs blocking. Further, front threads that
306 >     *    represent phase changes (from data to request node or vice
307 >     *    versa) compared to their predecessors receive additional
308 >     *    chained spins, reflecting longer paths typically required to
309 >     *    unblock threads during phase changes.
310 >     *
311 >     *
312 >     * ** Unlinking removed interior nodes **
313 >     *
314 >     * In addition to minimizing garbage retention via self-linking
315 >     * described above, we also unlink removed interior nodes. These
316 >     * may arise due to timed out or interrupted waits, or calls to
317 >     * remove(x) or Iterator.remove.  Normally, given a node that was
318 >     * at one time known to be the predecessor of some node s that is
319 >     * to be removed, we can unsplice s by CASing the next field of
320 >     * its predecessor if it still points to s (otherwise s must
321 >     * already have been removed or is now offlist). But there are two
322 >     * situations in which we cannot guarantee to make node s
323 >     * unreachable in this way: (1) If s is the trailing node of list
324 >     * (i.e., with null next), then it is pinned as the target node
325 >     * for appends, so can only be removed later after other nodes are
326 >     * appended. (2) We cannot necessarily unlink s given a
327 >     * predecessor node that is matched (including the case of being
328 >     * cancelled): the predecessor may already be unspliced, in which
329 >     * case some previous reachable node may still point to s.
330 >     * (For further explanation see Herlihy & Shavit "The Art of
331 >     * Multiprocessor Programming" chapter 9).  Although, in both
332 >     * cases, we can rule out the need for further action if either s
333 >     * or its predecessor are (or can be made to be) at, or fall off
334 >     * from, the head of list.
335 >     *
336 >     * Without taking these into account, it would be possible for an
337 >     * unbounded number of supposedly removed nodes to remain
338 >     * reachable.  Situations leading to such buildup are uncommon but
339 >     * can occur in practice; for example when a series of short timed
340 >     * calls to poll repeatedly time out but never otherwise fall off
341 >     * the list because of an untimed call to take at the front of the
342 >     * queue.
343 >     *
344 >     * When these cases arise, rather than always retraversing the
345 >     * entire list to find an actual predecessor to unlink (which
346 >     * won't help for case (1) anyway), we record a conservative
347 >     * estimate of possible unsplice failures (in "sweepVotes").
348 >     * We trigger a full sweep when the estimate exceeds a threshold
349 >     * ("SWEEP_THRESHOLD") indicating the maximum number of estimated
350 >     * removal failures to tolerate before sweeping through, unlinking
351 >     * cancelled nodes that were not unlinked upon initial removal.
352 >     * We perform sweeps by the thread hitting threshold (rather than
353 >     * background threads or by spreading work to other threads)
354 >     * because in the main contexts in which removal occurs, the
355 >     * caller is already timed-out, cancelled, or performing a
356 >     * potentially O(n) operation (e.g. remove(x)), none of which are
357 >     * time-critical enough to warrant the overhead that alternatives
358 >     * would impose on other threads.
359 >     *
360 >     * Because the sweepVotes estimate is conservative, and because
361 >     * nodes become unlinked "naturally" as they fall off the head of
362 >     * the queue, and because we allow votes to accumulate even while
363 >     * sweeps are in progress, there are typically significantly fewer
364 >     * such nodes than estimated.  Choice of a threshold value
365 >     * balances the likelihood of wasted effort and contention, versus
366 >     * providing a worst-case bound on retention of interior nodes in
367 >     * quiescent queues. The value defined below was chosen
368 >     * empirically to balance these under various timeout scenarios.
369 >     *
370 >     * Note that we cannot self-link unlinked interior nodes during
371 >     * sweeps. However, the associated garbage chains terminate when
372 >     * some successor ultimately falls off the head of the list and is
373 >     * self-linked.
374 >     */
375 >
376 >    /** True if on multiprocessor */
377 >    private static final boolean MP =
378 >        Runtime.getRuntime().availableProcessors() > 1;
379 >
380 >    /**
381 >     * The number of times to spin (with randomly interspersed calls
382 >     * to Thread.yield) on multiprocessor before blocking when a node
383 >     * is apparently the first waiter in the queue.  See above for
384 >     * explanation. Must be a power of two. The value is empirically
385 >     * derived -- it works pretty well across a variety of processors,
386 >     * numbers of CPUs, and OSes.
387 >     */
388 >    private static final int FRONT_SPINS   = 1 << 7;
389 >
390 >    /**
391 >     * The number of times to spin before blocking when a node is
392 >     * preceded by another node that is apparently spinning.  Also
393 >     * serves as an increment to FRONT_SPINS on phase changes, and as
394 >     * base average frequency for yielding during spins. Must be a
395 >     * power of two.
396 >     */
397 >    private static final int CHAINED_SPINS = FRONT_SPINS >>> 1;
398 >
399 >    /**
400 >     * The maximum number of estimated removal failures (sweepVotes)
401 >     * to tolerate before sweeping through the queue unlinking
402 >     * cancelled nodes that were not unlinked upon initial
403 >     * removal. See above for explanation. The value must be at least
404 >     * two to avoid useless sweeps when removing trailing nodes.
405 >     */
406 >    static final int SWEEP_THRESHOLD = 32;
407 >
408 >    /**
409 >     * Queue nodes. Uses Object, not E, for items to allow forgetting
410 >     * them after use.  Relies heavily on Unsafe mechanics to minimize
411 >     * unnecessary ordering constraints: Writes that are intrinsically
412 >     * ordered wrt other accesses or CASes use simple relaxed forms.
413 >     */
414 >    static final class Node {
415 >        final boolean isData;   // false if this is a request node
416 >        volatile Object item;   // initially non-null if isData; CASed to match
417 >        volatile Node next;
418 >        volatile Thread waiter; // null until waiting
419  
420 <    /**
421 <     * The number of nanoseconds for which it is faster to spin
422 <     * rather than to use timed park. A rough estimate suffices.
423 <     */
100 <    static final long spinForTimeoutThreshold = 1000L;
420 >        // CAS methods for fields
421 >        final boolean casNext(Node cmp, Node val) {
422 >            return UNSAFE.compareAndSwapObject(this, nextOffset, cmp, val);
423 >        }
424  
425 <    /**
426 <     * Node class for LinkedTransferQueue. Opportunistically
427 <     * subclasses from AtomicReference to represent item. Uses Object,
428 <     * 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;
425 >        final boolean casItem(Object cmp, Object val) {
426 >            //            assert cmp == null || cmp.getClass() != Node.class;
427 >            return UNSAFE.compareAndSwapObject(this, itemOffset, cmp, val);
428 >        }
429  
430 <        Node(E item, boolean isData) {
431 <            super(item);
430 >        /**
431 >         * Constructs a new node.  Uses relaxed write because item can
432 >         * only be seen after publication via casNext.
433 >         */
434 >        Node(Object item, boolean isData) {
435 >            UNSAFE.putObject(this, itemOffset, item); // relaxed write
436              this.isData = isData;
437          }
438  
439 <        // Unsafe mechanics
439 >        /**
440 >         * Links node to itself to avoid garbage retention.  Called
441 >         * only after CASing head field, so uses relaxed write.
442 >         */
443 >        final void forgetNext() {
444 >            UNSAFE.putObject(this, nextOffset, this);
445 >        }
446  
447 <        private static final sun.misc.Unsafe UNSAFE = getUnsafe();
448 <        private static final long nextOffset =
449 <            objectFieldOffset(UNSAFE, "next", Node.class);
447 >        /**
448 >         * Sets item to self and waiter to null, to avoid garbage
449 >         * retention after matching or cancelling. Uses relaxed writes
450 >         * because order is already constrained in the only calling
451 >         * contexts: item is forgotten only after volatile/atomic
452 >         * mechanics that extract items.  Similarly, clearing waiter
453 >         * follows either CAS or return from park (if ever parked;
454 >         * else we don't care).
455 >         */
456 >        final void forgetContents() {
457 >            UNSAFE.putObject(this, itemOffset, this);
458 >            UNSAFE.putObject(this, waiterOffset, null);
459 >        }
460  
461 <        final boolean casNext(Node<E> cmp, Node<E> val) {
462 <            return UNSAFE.compareAndSwapObject(this, nextOffset, cmp, val);
461 >        /**
462 >         * Returns true if this node has been matched, including the
463 >         * case of artificial matches due to cancellation.
464 >         */
465 >        final boolean isMatched() {
466 >            Object x = item;
467 >            return (x == this) || ((x == null) == isData);
468          }
469  
470 <        final void clearNext() {
471 <            UNSAFE.putOrderedObject(this, nextOffset, this);
470 >        /**
471 >         * Returns true if this is an unmatched request node.
472 >         */
473 >        final boolean isUnmatchedRequest() {
474 >            return !isData && item == null;
475          }
476  
477          /**
478 <         * Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
479 <         * Replace with a simple call to Unsafe.getUnsafe when integrating
480 <         * into a jdk.
137 <         *
138 <         * @return a sun.misc.Unsafe
478 >         * Returns true if a node with the given mode cannot be
479 >         * appended to this node because this node is unmatched and
480 >         * has opposite data mode.
481           */
482 <        private static sun.misc.Unsafe getUnsafe() {
483 <            try {
484 <                return sun.misc.Unsafe.getUnsafe();
485 <            } catch (SecurityException se) {
486 <                try {
487 <                    return java.security.AccessController.doPrivileged
488 <                        (new java.security
489 <                         .PrivilegedExceptionAction<sun.misc.Unsafe>() {
490 <                            public sun.misc.Unsafe run() throws Exception {
491 <                                java.lang.reflect.Field f = sun.misc
492 <                                    .Unsafe.class.getDeclaredField("theUnsafe");
493 <                                f.setAccessible(true);
494 <                                return (sun.misc.Unsafe) f.get(null);
495 <                            }});
496 <                } catch (java.security.PrivilegedActionException e) {
155 <                    throw new RuntimeException("Could not initialize intrinsics",
156 <                                               e.getCause());
157 <                }
482 >        final boolean cannotPrecede(boolean haveData) {
483 >            boolean d = isData;
484 >            Object x;
485 >            return d != haveData && (x = item) != this && (x != null) == d;
486 >        }
487 >
488 >        /**
489 >         * Tries to artificially match a data node -- used by remove.
490 >         */
491 >        final boolean tryMatchData() {
492 >            //            assert isData;
493 >            Object x = item;
494 >            if (x != null && x != this && casItem(x, null)) {
495 >                LockSupport.unpark(waiter);
496 >                return true;
497              }
498 +            return false;
499          }
500  
501 +        // Unsafe mechanics
502 +        private static final sun.misc.Unsafe UNSAFE = getUnsafe();
503 +        private static final long nextOffset =
504 +            objectFieldOffset(UNSAFE, "next", Node.class);
505 +        private static final long itemOffset =
506 +            objectFieldOffset(UNSAFE, "item", Node.class);
507 +        private static final long waiterOffset =
508 +            objectFieldOffset(UNSAFE, "waiter", Node.class);
509 +
510          private static final long serialVersionUID = -3375979862319811754L;
511      }
512  
513 <    /**
514 <     * 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 <    }
513 >    /** head of the queue; null until first enqueue */
514 >    transient volatile Node head;
515  
516 +    /** tail of the queue; null until first append */
517 +    private transient volatile Node tail;
518  
519 <    /** head of the queue */
520 <    private transient final PaddedAtomicReference<Node<E>> head;
519 >    /** The number of apparent failures to unsplice removed nodes */
520 >    private transient volatile int sweepVotes;
521  
522 <    /** tail of the queue */
523 <    private transient final PaddedAtomicReference<Node<E>> tail;
522 >    // CAS methods for fields
523 >    private boolean casTail(Node cmp, Node val) {
524 >        return UNSAFE.compareAndSwapObject(this, tailOffset, cmp, val);
525 >    }
526  
527 <    /**
528 <     * Reference to a cancelled node that might not yet have been
529 <     * unlinked from queue because it was the last inserted node
186 <     * when it cancelled.
187 <     */
188 <    private transient final PaddedAtomicReference<Node<E>> cleanMe;
527 >    private boolean casHead(Node cmp, Node val) {
528 >        return UNSAFE.compareAndSwapObject(this, headOffset, cmp, val);
529 >    }
530  
531 <    /**
532 <     * Tries to cas nh as new head; if successful, unlink
533 <     * old head's next node to avoid garbage retention.
531 >    private boolean casSweepVotes(int cmp, int val) {
532 >        return UNSAFE.compareAndSwapInt(this, sweepVotesOffset, cmp, val);
533 >    }
534 >
535 >    /*
536 >     * Possible values for "how" argument in xfer method.
537       */
538 <    private boolean advanceHead(Node<E> h, Node<E> nh) {
539 <        if (h == head.get() && head.compareAndSet(h, nh)) {
540 <            h.clearNext(); // forget old next
541 <            return true;
542 <        }
543 <        return false;
538 >    private static final int NOW   = 0; // for untimed poll, tryTransfer
539 >    private static final int ASYNC = 1; // for offer, put, add
540 >    private static final int SYNC  = 2; // for transfer, take
541 >    private static final int TIMED = 3; // for timed poll, tryTransfer
542 >
543 >    @SuppressWarnings("unchecked")
544 >    static <E> E cast(Object item) {
545 >        //        assert item == null || item.getClass() != Node.class;
546 >        return (E) item;
547      }
548  
549      /**
550 <     * Puts or takes an item. Used for most queue operations (except
551 <     * poll() and tryTransfer()). See the similar code in
552 <     * SynchronousQueue for detailed explanation.
553 <     *
554 <     * @param e the item or if null, signifies that this is a take
555 <     * @param mode the wait mode: NOWAIT, TIMEOUT, WAIT
556 <     * @param nanos timeout in nanosecs, used only if mode is TIMEOUT
557 <     * @return an item, or null on failure
558 <     */
559 <    private E xfer(E e, int mode, long nanos) {
560 <        boolean isData = (e != null);
561 <        Node<E> s = null;
562 <        final PaddedAtomicReference<Node<E>> head = this.head;
216 <        final PaddedAtomicReference<Node<E>> tail = this.tail;
550 >     * Implements all queuing methods. See above for explanation.
551 >     *
552 >     * @param e the item or null for take
553 >     * @param haveData true if this is a put, else a take
554 >     * @param how NOW, ASYNC, SYNC, or TIMED
555 >     * @param nanos timeout in nanosecs, used only if mode is TIMED
556 >     * @return an item if matched, else e
557 >     * @throws NullPointerException if haveData mode but e is null
558 >     */
559 >    private E xfer(E e, boolean haveData, int how, long nanos) {
560 >        if (haveData && (e == null))
561 >            throw new NullPointerException();
562 >        Node s = null;                        // the node to append, if needed
563  
564 <        for (;;) {
219 <            Node<E> t = tail.get();
220 <            Node<E> h = head.get();
564 >        retry: for (;;) {                     // restart on append race
565  
566 <            if (t != null && (t == h || t.isData == isData)) {
567 <                if (s == null)
568 <                    s = new Node<E>(e, isData);
569 <                Node<E> last = t.next;
570 <                if (last != null) {
571 <                    if (t == tail.get())
572 <                        tail.compareAndSet(t, last);
573 <                }
574 <                else if (t.casNext(null, s)) {
575 <                    tail.compareAndSet(t, s);
576 <                    return awaitFulfill(t, s, e, mode, nanos);
577 <                }
578 <            }
579 <
580 <            else if (h != null) {
581 <                Node<E> first = h.next;
582 <                if (t == tail.get() && first != null &&
583 <                    advanceHead(h, first)) {
584 <                    Object x = first.get();
241 <                    if (x != first && first.compareAndSet(x, e)) {
242 <                        LockSupport.unpark(first.waiter);
243 <                        return isData ? e : (E) x;
566 >            for (Node h = head, p = h; p != null;) { // find & match first node
567 >                boolean isData = p.isData;
568 >                Object item = p.item;
569 >                if (item != p && (item != null) == isData) { // unmatched
570 >                    if (isData == haveData)   // can't match
571 >                        break;
572 >                    if (p.casItem(item, e)) { // match
573 >                        for (Node q = p; q != h;) {
574 >                            Node n = q.next;  // update by 2 unless singleton
575 >                            if (head == h && casHead(h, n == null? q : n)) {
576 >                                h.forgetNext();
577 >                                break;
578 >                            }                 // advance and retry
579 >                            if ((h = head)   == null ||
580 >                                (q = h.next) == null || !q.isMatched())
581 >                                break;        // unless slack < 2
582 >                        }
583 >                        LockSupport.unpark(p.waiter);
584 >                        return this.<E>cast(item);
585                      }
586                  }
587 +                Node n = p.next;
588 +                p = (p != n) ? n : (h = head); // Use head if p offlist
589              }
590 +
591 +            if (how != NOW) {                 // No matches available
592 +                if (s == null)
593 +                    s = new Node(e, haveData);
594 +                Node pred = tryAppend(s, haveData);
595 +                if (pred == null)
596 +                    continue retry;           // lost race vs opposite mode
597 +                if (how != ASYNC)
598 +                    return awaitMatch(s, pred, e, (how == TIMED), nanos);
599 +            }
600 +            return e; // not waiting
601          }
602      }
603  
250
604      /**
605 <     * Version of xfer for poll() and tryTransfer, which
606 <     * simplifies control paths both here and in xfer.
607 <     */
608 <    private E fulfill(E e) {
609 <        boolean isData = (e != null);
610 <        final PaddedAtomicReference<Node<E>> head = this.head;
611 <        final PaddedAtomicReference<Node<E>> tail = this.tail;
612 <
613 <        for (;;) {
614 <            Node<E> t = tail.get();
615 <            Node<E> h = head.get();
616 <
617 <            if (t != null && (t == h || t.isData == isData)) {
618 <                Node<E> last = t.next;
619 <                if (t == tail.get()) {
620 <                    if (last != null)
621 <                        tail.compareAndSet(t, last);
622 <                    else
623 <                        return null;
624 <                }
625 <            }
626 <            else if (h != null) {
627 <                Node<E> first = h.next;
628 <                if (t == tail.get() &&
629 <                    first != null &&
630 <                    advanceHead(h, first)) {
631 <                    Object x = first.get();
632 <                    if (x != first && first.compareAndSet(x, e)) {
280 <                        LockSupport.unpark(first.waiter);
281 <                        return isData ? e : (E) x;
282 <                    }
605 >     * Tries to append node s as tail.
606 >     *
607 >     * @param s the node to append
608 >     * @param haveData true if appending in data mode
609 >     * @return null on failure due to losing race with append in
610 >     * different mode, else s's predecessor, or s itself if no
611 >     * predecessor
612 >     */
613 >    private Node tryAppend(Node s, boolean haveData) {
614 >        for (Node t = tail, p = t;;) {        // move p to last node and append
615 >            Node n, u;                        // temps for reads of next & tail
616 >            if (p == null && (p = head) == null) {
617 >                if (casHead(null, s))
618 >                    return s;                 // initialize
619 >            }
620 >            else if (p.cannotPrecede(haveData))
621 >                return null;                  // lost race vs opposite mode
622 >            else if ((n = p.next) != null)    // not last; keep traversing
623 >                p = p != t && t != (u = tail) ? (t = u) : // stale tail
624 >                    (p != n) ? n : null;      // restart if off list
625 >            else if (!p.casNext(null, s))
626 >                p = p.next;                   // re-read on CAS failure
627 >            else {
628 >                if (p != t) {                 // update if slack now >= 2
629 >                    while ((tail != t || !casTail(t, s)) &&
630 >                           (t = tail)   != null &&
631 >                           (s = t.next) != null && // advance and retry
632 >                           (s = s.next) != null && s != t);
633                  }
634 +                return p;
635              }
636          }
637      }
638  
639      /**
640 <     * Spins/blocks until node s is fulfilled or caller gives up,
290 <     * depending on wait mode.
640 >     * Spins/yields/blocks until node s is matched or caller gives up.
641       *
292     * @param pred the predecessor of waiting node
642       * @param s the waiting node
643 +     * @param pred the predecessor of s, or s itself if it has no
644 +     * predecessor, or null if unknown (the null case does not occur
645 +     * in any current calls but may in possible future extensions)
646       * @param e the comparison value for checking match
647 <     * @param mode mode
648 <     * @param nanos timeout value
649 <     * @return matched item, or null if cancelled
650 <     */
651 <    private E awaitFulfill(Node<E> pred, Node<E> s, E e,
652 <                           int mode, long nanos) {
301 <        if (mode == NOWAIT)
302 <            return null;
303 <
304 <        long lastTime = (mode == TIMEOUT) ? System.nanoTime() : 0;
647 >     * @param timed if true, wait only until timeout elapses
648 >     * @param nanos timeout in nanosecs, used only if timed is true
649 >     * @return matched item, or e if unmatched on interrupt or timeout
650 >     */
651 >    private E awaitMatch(Node s, Node pred, E e, boolean timed, long nanos) {
652 >        long lastTime = timed ? System.nanoTime() : 0L;
653          Thread w = Thread.currentThread();
654 <        int spins = -1; // set to desired spin count below
654 >        int spins = -1; // initialized after first item and cancel checks
655 >        ThreadLocalRandom randomYields = null; // bound if needed
656 >
657          for (;;) {
658 <            if (w.isInterrupted())
659 <                s.compareAndSet(e, s);
660 <            Object x = s.get();
661 <            if (x != e) {                 // Node was matched or cancelled
662 <                advanceHead(pred, s);     // unlink if head
663 <                if (x == s) {             // was cancelled
664 <                    clean(pred, s);
665 <                    return null;
666 <                }
667 <                else if (x != null) {
668 <                    s.set(s);             // avoid garbage retention
669 <                    return (E) x;
670 <                }
671 <                else
672 <                    return e;
658 >            Object item = s.item;
659 >            if (item != e) {                  // matched
660 >                //                assert item != s;
661 >                s.forgetContents();           // avoid garbage
662 >                return this.<E>cast(item);
663 >            }
664 >            if ((w.isInterrupted() || (timed && nanos <= 0)) &&
665 >                    s.casItem(e, s)) {        // cancel
666 >                unsplice(pred, s);
667 >                return e;
668 >            }
669 >
670 >            if (spins < 0) {                  // establish spins at/near front
671 >                if ((spins = spinsFor(pred, s.isData)) > 0)
672 >                    randomYields = ThreadLocalRandom.current();
673 >            }
674 >            else if (spins > 0) {             // spin
675 >                --spins;
676 >                if (randomYields.nextInt(CHAINED_SPINS) == 0)
677 >                    Thread.yield();           // occasionally yield
678 >            }
679 >            else if (s.waiter == null) {
680 >                s.waiter = w;                 // request unpark then recheck
681              }
682 <            if (mode == TIMEOUT) {
682 >            else if (timed) {
683                  long now = System.nanoTime();
684 <                nanos -= now - lastTime;
684 >                if ((nanos -= now - lastTime) > 0)
685 >                    LockSupport.parkNanos(this, nanos);
686                  lastTime = now;
328                if (nanos <= 0) {
329                    s.compareAndSet(e, s); // try to cancel
330                    continue;
331                }
332            }
333            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);
687              }
688 <            if (spins > 0)
340 <                --spins;
341 <            else if (s.waiter == null)
342 <                s.waiter = w;
343 <            else if (mode != TIMEOUT) {
688 >            else {
689                  LockSupport.park(this);
345                s.waiter = null;
346                spins = -1;
347            }
348            else if (nanos > spinForTimeoutThreshold) {
349                LockSupport.parkNanos(this, nanos);
350                s.waiter = null;
351                spins = -1;
690              }
691          }
692      }
693  
694      /**
695 <     * Returns validated tail for use in cleaning methods.
695 >     * Returns spin/yield value for a node with given predecessor and
696 >     * data mode. See above for explanation.
697       */
698 <    private Node<E> getValidatedTail() {
699 <        for (;;) {
700 <            Node<E> h = head.get();
701 <            Node<E> first = h.next;
702 <            if (first != null && first.get() == first) { // help advance
703 <                advanceHead(h, first);
704 <                continue;
705 <            }
367 <            Node<E> t = tail.get();
368 <            Node<E> last = t.next;
369 <            if (t == tail.get()) {
370 <                if (last != null)
371 <                    tail.compareAndSet(t, last); // help advance
372 <                else
373 <                    return t;
374 <            }
698 >    private static int spinsFor(Node pred, boolean haveData) {
699 >        if (MP && pred != null) {
700 >            if (pred.isData != haveData)      // phase change
701 >                return FRONT_SPINS + CHAINED_SPINS;
702 >            if (pred.isMatched())             // probably at front
703 >                return FRONT_SPINS;
704 >            if (pred.waiter == null)          // pred apparently spinning
705 >                return CHAINED_SPINS;
706          }
707 +        return 0;
708      }
709  
710 +    /* -------------- Traversal methods -------------- */
711 +
712      /**
713 <     * Gets rid of cancelled node s with original predecessor pred.
714 <     *
715 <     * @param pred predecessor of cancelled node
716 <     * @param s the cancelled node
713 >     * Returns the successor of p, or the head node if p.next has been
714 >     * linked to self, which will only be true if traversing with a
715 >     * stale pointer that is now off the list.
716 >     */
717 >    final Node succ(Node p) {
718 >        Node next = p.next;
719 >        return (p == next) ? head : next;
720 >    }
721 >
722 >    /**
723 >     * Returns the first unmatched node of the given mode, or null if
724 >     * none.  Used by methods isEmpty, hasWaitingConsumer.
725       */
726 <    private void clean(Node<E> pred, Node<E> s) {
727 <        Thread w = s.waiter;
728 <        if (w != null) {             // Wake up thread
729 <            s.waiter = null;
730 <            if (w != Thread.currentThread())
731 <                LockSupport.unpark(w);
726 >    private Node firstOfMode(boolean isData) {
727 >        for (Node p = head; p != null; p = succ(p)) {
728 >            if (!p.isMatched())
729 >                return (p.isData == isData) ? p : null;
730 >        }
731 >        return null;
732 >    }
733 >
734 >    /**
735 >     * Returns the item in the first unmatched node with isData; or
736 >     * null if none.  Used by peek.
737 >     */
738 >    private E firstDataItem() {
739 >        for (Node p = head; p != null; p = succ(p)) {
740 >            Object item = p.item;
741 >            if (p.isData) {
742 >                if (item != null && item != p)
743 >                    return this.<E>cast(item);
744 >            }
745 >            else if (item == null)
746 >                return null;
747          }
748 +        return null;
749 +    }
750  
751 <        if (pred == null)
752 <            return;
751 >    /**
752 >     * Traverses and counts unmatched nodes of the given mode.
753 >     * Used by methods size and getWaitingConsumerCount.
754 >     */
755 >    private int countOfMode(boolean data) {
756 >        int count = 0;
757 >        for (Node p = head; p != null; ) {
758 >            if (!p.isMatched()) {
759 >                if (p.isData != data)
760 >                    return 0;
761 >                if (++count == Integer.MAX_VALUE) // saturated
762 >                    break;
763 >            }
764 >            Node n = p.next;
765 >            if (n != p)
766 >                p = n;
767 >            else {
768 >                count = 0;
769 >                p = head;
770 >            }
771 >        }
772 >        return count;
773 >    }
774 >
775 >    final class Itr implements Iterator<E> {
776 >        private Node nextNode;   // next node to return item for
777 >        private E nextItem;      // the corresponding item
778 >        private Node lastRet;    // last returned node, to support remove
779 >        private Node lastPred;   // predecessor to unlink lastRet
780  
781 <        /*
782 <         * At any given time, exactly one node on list cannot be
397 <         * deleted -- the last inserted node. To accommodate this, if
398 <         * we cannot delete s, we save its predecessor as "cleanMe",
399 <         * processing the previously saved version first. At least one
400 <         * of node s or the node previously saved can always be
401 <         * processed, so this always terminates.
781 >        /**
782 >         * Moves to next node after prev, or first node if prev null.
783           */
784 <        while (pred.next == s) {
785 <            Node<E> oldpred = reclean();  // First, help get rid of cleanMe
786 <            Node<E> t = getValidatedTail();
787 <            if (s != t) {               // If not tail, try to unsplice
788 <                Node<E> sn = s.next;      // s.next == s means s already off list
789 <                if (sn == s || pred.casNext(s, sn))
784 >        private void advance(Node prev) {
785 >            lastPred = lastRet;
786 >            lastRet = prev;
787 >            for (Node p = (prev == null) ? head : succ(prev);
788 >                 p != null; p = succ(p)) {
789 >                Object item = p.item;
790 >                if (p.isData) {
791 >                    if (item != null && item != p) {
792 >                        nextItem = LinkedTransferQueue.this.<E>cast(item);
793 >                        nextNode = p;
794 >                        return;
795 >                    }
796 >                }
797 >                else if (item == null)
798                      break;
799              }
800 <            else if (oldpred == pred || // Already saved
801 <                     (oldpred == null && cleanMe.compareAndSet(null, pred)))
802 <                break;                  // Postpone cleaning
800 >            nextNode = null;
801 >        }
802 >
803 >        Itr() {
804 >            advance(null);
805 >        }
806 >
807 >        public final boolean hasNext() {
808 >            return nextNode != null;
809 >        }
810 >
811 >        public final E next() {
812 >            Node p = nextNode;
813 >            if (p == null) throw new NoSuchElementException();
814 >            E e = nextItem;
815 >            advance(p);
816 >            return e;
817 >        }
818 >
819 >        public final void remove() {
820 >            Node p = lastRet;
821 >            if (p == null) throw new IllegalStateException();
822 >            if (p.tryMatchData())
823 >                unsplice(lastPred, p);
824          }
825      }
826  
827 +    /* -------------- Removal methods -------------- */
828 +
829      /**
830 <     * Tries to unsplice the cancelled node held in cleanMe that was
831 <     * previously uncleanable because it was at tail.
830 >     * Unsplices (now or later) the given deleted/cancelled node with
831 >     * the given predecessor.
832       *
833 <     * @return current cleanMe node (or null)
833 >     * @param pred a node that was at one time known to be the
834 >     * predecessor of s, or null or s itself if s is/was at head
835 >     * @param s the node to be unspliced
836       */
837 <    private Node<E> reclean() {
837 >    final void unsplice(Node pred, Node s) {
838 >        s.forgetContents(); // forget unneeded fields
839          /*
840 <         * cleanMe is, or at one time was, predecessor of cancelled
841 <         * node s that was the tail so could not be unspliced.  If s
842 <         * is no longer the tail, try to unsplice if necessary and
843 <         * make cleanMe slot available.  This differs from similar
844 <         * code in clean() because we must check that pred still
430 <         * points to a cancelled node that must be unspliced -- if
431 <         * not, we can (must) clear cleanMe without unsplicing.
432 <         * This can loop only due to contention on casNext or
433 <         * clearing cleanMe.
840 >         * See above for rationale. Briefly: if pred still points to
841 >         * s, try to unlink s.  If s cannot be unlinked, because it is
842 >         * trailing node or pred might be unlinked, and neither pred
843 >         * nor s are head or offlist, add to sweepVotes, and if enough
844 >         * votes have accumulated, sweep.
845           */
846 <        Node<E> pred;
847 <        while ((pred = cleanMe.get()) != null) {
848 <            Node<E> t = getValidatedTail();
849 <            Node<E> s = pred.next;
850 <            if (s != t) {
851 <                Node<E> sn;
852 <                if (s == null || s == pred || s.get() != s ||
853 <                    (sn = s.next) == s || pred.casNext(s, sn))
854 <                    cleanMe.compareAndSet(pred, null);
846 >        if (pred != null && pred != s && pred.next == s) {
847 >            Node n = s.next;
848 >            if (n == null ||
849 >                (n != s && pred.casNext(s, n) && pred.isMatched())) {
850 >                for (;;) {               // check if at, or could be, head
851 >                    Node h = head;
852 >                    if (h == pred || h == s || h == null)
853 >                        return;          // at head or list empty
854 >                    if (!h.isMatched())
855 >                        break;
856 >                    Node hn = h.next;
857 >                    if (hn == null)
858 >                        return;          // now empty
859 >                    if (hn != h && casHead(h, hn))
860 >                        h.forgetNext();  // advance head
861 >                }
862 >                if (pred.next != pred && s.next != s) { // recheck if offlist
863 >                    for (;;) {           // sweep now if enough votes
864 >                        int v = sweepVotes;
865 >                        if (v < SWEEP_THRESHOLD) {
866 >                            if (casSweepVotes(v, v + 1))
867 >                                break;
868 >                        }
869 >                        else if (casSweepVotes(v, 0)) {
870 >                            sweep();
871 >                            break;
872 >                        }
873 >                    }
874 >                }
875              }
876 <            else // s is still tail; cannot clean
876 >        }
877 >    }
878 >
879 >    /**
880 >     * Unlinks matched (typically cancelled) nodes encountered in a
881 >     * traversal from head.
882 >     */
883 >    private void sweep() {
884 >        for (Node p = head, s, n; p != null && (s = p.next) != null; ) {
885 >            if (p == s)                    // stale
886 >                p = head;
887 >            else if (!s.isMatched())
888 >                p = s;
889 >            else if ((n = s.next) == null || s == n) // trailing node is pinned
890                  break;
891 +            else
892 +                p.casNext(s, n);
893          }
448        return pred;
894      }
895  
896      /**
897 +     * Main implementation of remove(Object)
898 +     */
899 +    private boolean findAndRemove(Object e) {
900 +        if (e != null) {
901 +            for (Node pred = null, p = head; p != null; ) {
902 +                Object item = p.item;
903 +                if (p.isData) {
904 +                    if (item != null && item != p && e.equals(item) &&
905 +                        p.tryMatchData()) {
906 +                        unsplice(pred, p);
907 +                        return true;
908 +                    }
909 +                }
910 +                else if (item == null)
911 +                    break;
912 +                pred = p;
913 +                if ((p = p.next) == pred) { // stale
914 +                    pred = null;
915 +                    p = head;
916 +                }
917 +            }
918 +        }
919 +        return false;
920 +    }
921 +
922 +
923 +    /**
924       * Creates an initially empty {@code LinkedTransferQueue}.
925       */
926      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);
927      }
928  
929      /**
# Line 479 | Line 947 | public class LinkedTransferQueue<E> exte
947       * @throws NullPointerException if the specified element is null
948       */
949      public void put(E e) {
950 <        offer(e);
950 >        xfer(e, true, ASYNC, 0);
951      }
952  
953      /**
# Line 492 | Line 960 | public class LinkedTransferQueue<E> exte
960       * @throws NullPointerException if the specified element is null
961       */
962      public boolean offer(E e, long timeout, TimeUnit unit) {
963 <        return offer(e);
963 >        xfer(e, true, ASYNC, 0);
964 >        return true;
965      }
966  
967      /**
# Line 504 | Line 973 | public class LinkedTransferQueue<E> exte
973       * @throws NullPointerException if the specified element is null
974       */
975      public boolean offer(E e) {
976 <        if (e == null) throw new NullPointerException();
508 <        xfer(e, NOWAIT, 0);
976 >        xfer(e, true, ASYNC, 0);
977          return true;
978      }
979  
# Line 518 | Line 986 | public class LinkedTransferQueue<E> exte
986       * @throws NullPointerException if the specified element is null
987       */
988      public boolean add(E e) {
989 <        return offer(e);
989 >        xfer(e, true, ASYNC, 0);
990 >        return true;
991      }
992  
993      /**
# Line 532 | Line 1001 | public class LinkedTransferQueue<E> exte
1001       * @throws NullPointerException if the specified element is null
1002       */
1003      public boolean tryTransfer(E e) {
1004 <        if (e == null) throw new NullPointerException();
536 <        return fulfill(e) != null;
1004 >        return xfer(e, true, NOW, 0) == null;
1005      }
1006  
1007      /**
# Line 548 | Line 1016 | public class LinkedTransferQueue<E> exte
1016       * @throws NullPointerException if the specified element is null
1017       */
1018      public void transfer(E e) throws InterruptedException {
1019 <        if (e == null) throw new NullPointerException();
1020 <        if (xfer(e, WAIT, 0) == null) {
553 <            Thread.interrupted();
1019 >        if (xfer(e, true, SYNC, 0) != null) {
1020 >            Thread.interrupted(); // failure possible only due to interrupt
1021              throw new InterruptedException();
1022          }
1023      }
# Line 571 | Line 1038 | public class LinkedTransferQueue<E> exte
1038       */
1039      public boolean tryTransfer(E e, long timeout, TimeUnit unit)
1040          throws InterruptedException {
1041 <        if (e == null) throw new NullPointerException();
575 <        if (xfer(e, TIMEOUT, unit.toNanos(timeout)) != null)
1041 >        if (xfer(e, true, TIMED, unit.toNanos(timeout)) == null)
1042              return true;
1043          if (!Thread.interrupted())
1044              return false;
# Line 580 | Line 1046 | public class LinkedTransferQueue<E> exte
1046      }
1047  
1048      public E take() throws InterruptedException {
1049 <        E e = xfer(null, WAIT, 0);
1049 >        E e = xfer(null, false, SYNC, 0);
1050          if (e != null)
1051              return e;
1052          Thread.interrupted();
# Line 588 | Line 1054 | public class LinkedTransferQueue<E> exte
1054      }
1055  
1056      public E poll(long timeout, TimeUnit unit) throws InterruptedException {
1057 <        E e = xfer(null, TIMEOUT, unit.toNanos(timeout));
1057 >        E e = xfer(null, false, TIMED, unit.toNanos(timeout));
1058          if (e != null || !Thread.interrupted())
1059              return e;
1060          throw new InterruptedException();
1061      }
1062  
1063      public E poll() {
1064 <        return fulfill(null);
1064 >        return xfer(null, false, NOW, 0);
1065      }
1066  
1067      /**
# Line 634 | Line 1100 | public class LinkedTransferQueue<E> exte
1100          return n;
1101      }
1102  
637    // Traversal-based methods
638
639    /**
640     * Returns head after performing any outstanding helping steps.
641     */
642    private Node<E> traversalHead() {
643        for (;;) {
644            Node<E> t = tail.get();
645            Node<E> h = head.get();
646            if (h != null && t != null) {
647                Node<E> last = t.next;
648                Node<E> first = h.next;
649                if (t == tail.get()) {
650                    if (last != null)
651                        tail.compareAndSet(t, last);
652                    else if (first != null) {
653                        Object x = first.get();
654                        if (x == first)
655                            advanceHead(h, first);
656                        else
657                            return h;
658                    }
659                    else
660                        return h;
661                }
662            }
663            reclean();
664        }
665    }
666
1103      /**
1104       * Returns an iterator over the elements in this queue in proper
1105       * sequence, from head to tail.
# Line 681 | Line 1117 | public class LinkedTransferQueue<E> exte
1117          return new Itr();
1118      }
1119  
684    /**
685     * Iterators. Basic strategy is to traverse list, treating
686     * non-data (i.e., request) nodes as terminating list.
687     * Once a valid data node is found, the item is cached
688     * so that the next call to next() will return it even
689     * if subsequently removed.
690     */
691    class Itr implements Iterator<E> {
692        Node<E> next;        // node to return next
693        Node<E> pnext;       // predecessor of next
694        Node<E> curr;        // last returned node, for remove()
695        Node<E> pcurr;       // predecessor of curr, for remove()
696        E nextItem;          // Cache of next item, once committed to in next
697
698        Itr() {
699            advance();
700        }
701
702        /**
703         * Moves to next valid node and returns item to return for
704         * next(), or null if no such.
705         */
706        private E advance() {
707            pcurr = pnext;
708            curr = next;
709            E item = nextItem;
710
711            for (;;) {
712                pnext = (next == null) ? traversalHead() : next;
713                next = pnext.next;
714                if (next == pnext) {
715                    next = null;
716                    continue;  // restart
717                }
718                if (next == null)
719                    break;
720                Object x = next.get();
721                if (x != null && x != next) {
722                    nextItem = (E) x;
723                    break;
724                }
725            }
726            return item;
727        }
728
729        public boolean hasNext() {
730            return next != null;
731        }
732
733        public E next() {
734            if (next == null)
735                throw new NoSuchElementException();
736            return advance();
737        }
738
739        public void remove() {
740            Node<E> p = curr;
741            if (p == null)
742                throw new IllegalStateException();
743            Object x = p.get();
744            if (x != null && x != p && p.compareAndSet(x, p))
745                clean(pcurr, p);
746        }
747    }
748
1120      public E peek() {
1121 <        for (;;) {
751 <            Node<E> h = traversalHead();
752 <            Node<E> p = h.next;
753 <            if (p == null)
754 <                return null;
755 <            Object x = p.get();
756 <            if (p != x) {
757 <                if (!p.isData)
758 <                    return null;
759 <                if (x != null)
760 <                    return (E) x;
761 <            }
762 <        }
1121 >        return firstDataItem();
1122      }
1123  
1124      /**
# Line 768 | Line 1127 | public class LinkedTransferQueue<E> exte
1127       * @return {@code true} if this queue contains no elements
1128       */
1129      public boolean isEmpty() {
1130 <        for (;;) {
1131 <            Node<E> h = traversalHead();
1132 <            Node<E> p = h.next;
774 <            if (p == null)
775 <                return true;
776 <            Object x = p.get();
777 <            if (p != x) {
778 <                if (!p.isData)
779 <                    return true;
780 <                if (x != null)
781 <                    return false;
782 <            }
1130 >        for (Node p = head; p != null; p = succ(p)) {
1131 >            if (!p.isMatched())
1132 >                return !p.isData;
1133          }
1134 +        return true;
1135      }
1136  
1137      public boolean hasWaitingConsumer() {
1138 <        for (;;) {
788 <            Node<E> h = traversalHead();
789 <            Node<E> p = h.next;
790 <            if (p == null)
791 <                return false;
792 <            Object x = p.get();
793 <            if (p != x)
794 <                return !p.isData;
795 <        }
1138 >        return firstOfMode(false) != null;
1139      }
1140  
1141      /**
# Line 808 | Line 1151 | public class LinkedTransferQueue<E> exte
1151       * @return the number of elements in this queue
1152       */
1153      public int size() {
1154 <        for (;;) {
812 <            int count = 0;
813 <            Node<E> pred = traversalHead();
814 <            for (;;) {
815 <                Node<E> q = pred.next;
816 <                if (q == pred) // restart
817 <                    break;
818 <                if (q == null || !q.isData)
819 <                    return count;
820 <                Object x = q.get();
821 <                if (x != null && x != q) {
822 <                    if (++count == Integer.MAX_VALUE) // saturated
823 <                        return count;
824 <                }
825 <                pred = q;
826 <            }
827 <        }
1154 >        return countOfMode(true);
1155      }
1156  
1157      public int getWaitingConsumerCount() {
1158 <        // converse of size -- count valid non-data nodes
832 <        for (;;) {
833 <            int count = 0;
834 <            Node<E> pred = traversalHead();
835 <            for (;;) {
836 <                Node<E> q = pred.next;
837 <                if (q == pred) // restart
838 <                    break;
839 <                if (q == null || q.isData)
840 <                    return count;
841 <                Object x = q.get();
842 <                if (x == null) {
843 <                    if (++count == Integer.MAX_VALUE) // saturated
844 <                        return count;
845 <                }
846 <                pred = q;
847 <            }
848 <        }
1158 >        return countOfMode(false);
1159      }
1160  
1161      /**
# Line 860 | Line 1170 | public class LinkedTransferQueue<E> exte
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)
864 <            return false;
865 <        for (;;) {
866 <            Node<E> pred = traversalHead();
867 <            for (;;) {
868 <                Node<E> q = pred.next;
869 <                if (q == pred) // restart
870 <                    break;
871 <                if (q == null || !q.isData)
872 <                    return false;
873 <                Object x = q.get();
874 <                if (x != null && x != q && o.equals(x) &&
875 <                    q.compareAndSet(x, q)) {
876 <                    clean(pred, q);
877 <                    return true;
878 <                }
879 <                pred = q;
880 <            }
881 <        }
1173 >        return findAndRemove(o);
1174      }
1175  
1176      /**
# Line 893 | 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 909 | 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();
920        resetHeadAndTail();
1212          for (;;) {
1213              @SuppressWarnings("unchecked") E item = (E) s.readObject();
1214              if (item == null)
# Line 927 | Line 1218 | public class LinkedTransferQueue<E> exte
1218          }
1219      }
1220  
930    // Support for resetting head/tail while deserializing
931    private void resetHeadAndTail() {
932        Node<E> dummy = new Node<E>(null, false);
933        UNSAFE.putObjectVolatile(this, headOffset,
934                                 new PaddedAtomicReference<Node<E>>(dummy));
935        UNSAFE.putObjectVolatile(this, tailOffset,
936                                 new PaddedAtomicReference<Node<E>>(dummy));
937        UNSAFE.putObjectVolatile(this, cleanMeOffset,
938                                 new PaddedAtomicReference<Node<E>>(null));
939    }
940
1221      // Unsafe mechanics
1222  
1223      private static final sun.misc.Unsafe UNSAFE = getUnsafe();
# Line 945 | Line 1225 | public class LinkedTransferQueue<E> exte
1225          objectFieldOffset(UNSAFE, "head", LinkedTransferQueue.class);
1226      private static final long tailOffset =
1227          objectFieldOffset(UNSAFE, "tail", LinkedTransferQueue.class);
1228 <    private static final long cleanMeOffset =
1229 <        objectFieldOffset(UNSAFE, "cleanMe", LinkedTransferQueue.class);
950 <
1228 >    private static final long sweepVotesOffset =
1229 >        objectFieldOffset(UNSAFE, "sweepVotes", LinkedTransferQueue.class);
1230  
1231      static long objectFieldOffset(sun.misc.Unsafe UNSAFE,
1232                                    String field, Class<?> klazz) {
# Line 968 | 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 988 | Line 1267 | public class LinkedTransferQueue<E> exte
1267              }
1268          }
1269      }
1270 +
1271   }

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines