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 embedded elements, such as |
108 |
> |
* support deletion of interior elements, such as |
109 |
|
* j.u.c.ConcurrentLinkedQueue.) |
110 |
|
* |
111 |
< |
* Once a node is matched, its item can never again change. We |
112 |
< |
* may thus arrange that the linked list of them contains a prefix |
113 |
< |
* of zero or more matched nodes, followed by a suffix of zero or |
114 |
< |
* more unmatched nodes. (Note that we allow both the prefix and |
115 |
< |
* suffix to be zero length, which in turn means that we do not |
116 |
< |
* use a dummy header.) If we were not concerned with either time |
117 |
< |
* or space efficiency, we could correctly perform enqueue and |
118 |
< |
* dequeue operations by traversing from a pointer to the initial |
119 |
< |
* node; CASing the item of the first unmatched node on match and |
120 |
< |
* CASing the next field of the trailing node on appends. While |
121 |
< |
* this would be a terrible idea in itself, it does have the |
122 |
< |
* benefit of not requiring ANY atomic updates on head/tail |
123 |
< |
* fields. |
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 |
< |
* that reflects the tradeoff of sometimes requiring extra traversal |
129 |
< |
* steps to locate the first and/or last unmatched nodes, versus |
130 |
< |
* the reduced overhead and contention of fewer updates to queue |
131 |
< |
* pointers. For example, a possible snapshot of a queue is: |
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 |
|
* | | |
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. |
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 |
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 releasing another thread) to be read-only, |
165 |
< |
* thus not introducing any further contention. As described |
166 |
< |
* below, we implement this by performing slack maintenance |
167 |
< |
* retries only after these points. |
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. During traversals, threads may sometimes |
172 |
< |
* shortcut the "next" link path from the current "head" node to |
173 |
< |
* be closer to the currently known first unmatched node. Again, |
174 |
< |
* this may be triggered with using thresholds or randomization. |
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" |
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 |
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 internal nodes (due to timed out or interrupted |
210 |
< |
* waits, or calls to remove or Iterator.remove) uses a scheme |
211 |
< |
* roughly similar to that in Scherer, Lea, and Scott |
212 |
< |
* SynchronousQueue. Given a predecessor, we can unsplice any node |
213 |
< |
* except the (actual) tail of the queue. To avoid build-up of |
214 |
< |
* cancelled trailing nodes, upon a request to remove a trailing |
215 |
< |
* node, it is placed in field "cleanMe" to be unspliced later. |
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 target |
230 |
< |
* slack of two. The slack value is hard-wired: a path greater |
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 max slack at one. Avoiding tracking |
236 |
< |
* explicit counts across situations slightly simplifies an |
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 value, path short-circuiting is rarely |
243 |
< |
* worthwhile. However, it is used (in awaitMatch) immediately |
244 |
< |
* before a waiting thread starts to block, as a final bit of |
245 |
< |
* helping at a point when contention with others is extremely |
246 |
< |
* unlikely (since if other threads that could release it are |
247 |
< |
* operating, then the current thread wouldn't be blocking). |
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 nine separate methods. |
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 |
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 retry loops until the slack is at most two. Traversals |
280 |
< |
* also check if the initial head is now off-list, in which |
281 |
< |
* case they start at the new head. |
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, try to append a new node |
292 |
< |
* to the list (or if head was null, establish the first |
293 |
< |
* node). Nodes can be appended only if their predecessors are |
294 |
< |
* either already matched or are of the same mode. If we detect |
295 |
< |
* otherwise, then a new node with opposite mode must have been |
296 |
< |
* appended during traversal, so must restart at phase 1. The |
297 |
< |
* traversal and update steps are otherwise similar to phase 1: |
298 |
< |
* Retrying upon CAS misses and checking for staleness. In |
299 |
< |
* particular, if a self-link is encountered, then we can |
300 |
< |
* safely jump to a node on the list by continuing the |
301 |
< |
* traversal at current head. |
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 |
< |
* current thread was interrupted or the wait timed out. On |
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 |
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 much smaller (1/4) spins for nodes |
324 |
< |
* that are not known to be front but whose predecessors have |
325 |
< |
* not blocked -- these "chained" spins avoid artifacts of |
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 |
< |
* spins, reflecting the longer code path lengths necessary to |
331 |
< |
* release them under contention. |
330 |
> |
* chained spins, reflecting longer paths typically required to |
331 |
> |
* unblock threads during phase changes. |
332 |
|
*/ |
333 |
|
|
334 |
|
/** True if on multiprocessor */ |
336 |
|
Runtime.getRuntime().availableProcessors() > 1; |
337 |
|
|
338 |
|
/** |
339 |
< |
* The number of times to spin (with on average one randomly |
340 |
< |
* interspersed call to Thread.yield) on multiprocessor before |
341 |
< |
* blocking when a node is apparently the first waiter in the |
342 |
< |
* queue. See above for explanation. Must be a power of two. The |
343 |
< |
* value is empirically derived -- it works pretty well across a |
344 |
< |
* variety of processors, numbers of CPUs, and OSes. |
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. |
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 >>> 2; |
355 |
> |
private static final int CHAINED_SPINS = FRONT_SPINS >>> 1; |
356 |
|
|
357 |
|
/** |
358 |
|
* Queue nodes. Uses Object, not E, for items to allow forgetting |
361 |
|
* precede or follow CASes use simple relaxed forms. Other |
362 |
|
* cleanups use releasing/lazy writes. |
363 |
|
*/ |
364 |
< |
static final class Node { |
364 |
> |
static final class Node<E> { |
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; |
367 |
> |
volatile Node<E> next; |
368 |
|
volatile Thread waiter; // null until waiting |
369 |
|
|
370 |
|
// CAS methods for fields |
371 |
< |
final boolean casNext(Node cmp, Node val) { |
371 |
> |
final boolean casNext(Node<E> cmp, Node<E> val) { |
372 |
|
return UNSAFE.compareAndSwapObject(this, nextOffset, cmp, val); |
373 |
|
} |
374 |
|
|
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 |
|
|
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) { |
384 |
> |
Node(E item, boolean isData) { |
385 |
|
UNSAFE.putObject(this, itemOffset, item); // relaxed write |
386 |
|
this.isData = isData; |
387 |
|
} |
410 |
|
*/ |
411 |
|
final boolean isMatched() { |
412 |
|
Object x = item; |
413 |
< |
return x == this || (x != null) != isData; |
413 |
> |
return (x == this) || ((x == null) == isData); |
414 |
> |
} |
415 |
> |
|
416 |
> |
/** |
417 |
> |
* Returns true if this is an unmatched request node. |
418 |
> |
*/ |
419 |
> |
final boolean isUnmatchedRequest() { |
420 |
> |
return !isData && item == null; |
421 |
|
} |
422 |
|
|
423 |
|
/** |
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); |
457 |
|
} |
458 |
|
|
459 |
|
/** head of the queue; null until first enqueue */ |
460 |
< |
private transient volatile Node head; |
460 |
> |
transient volatile Node<E> head; |
461 |
|
|
462 |
|
/** predecessor of dangling unspliceable node */ |
463 |
< |
private transient volatile Node cleanMe; // decl here to reduce contention |
463 |
> |
private transient volatile Node<E> cleanMe; // decl here reduces contention |
464 |
|
|
465 |
|
/** tail of the queue; null until first append */ |
466 |
< |
private transient volatile Node tail; |
466 |
> |
private transient volatile Node<E> tail; |
467 |
|
|
468 |
|
// CAS methods for fields |
469 |
< |
private boolean casTail(Node cmp, Node val) { |
469 |
> |
private boolean casTail(Node<E> cmp, Node<E> val) { |
470 |
|
return UNSAFE.compareAndSwapObject(this, tailOffset, cmp, val); |
471 |
|
} |
472 |
|
|
473 |
< |
private boolean casHead(Node cmp, Node val) { |
473 |
> |
private boolean casHead(Node<E> cmp, Node<E> val) { |
474 |
|
return UNSAFE.compareAndSwapObject(this, headOffset, cmp, val); |
475 |
|
} |
476 |
|
|
477 |
< |
private boolean casCleanMe(Node cmp, Node val) { |
477 |
> |
private boolean casCleanMe(Node<E> cmp, Node<E> val) { |
478 |
|
return UNSAFE.compareAndSwapObject(this, cleanMeOffset, cmp, val); |
479 |
|
} |
480 |
|
|
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 |
|
* Implements all queuing methods. See above for explanation. |
498 |
|
* |
503 |
|
* @return an item if matched, else e |
504 |
|
* @throws NullPointerException if haveData mode but e is null |
505 |
|
*/ |
506 |
< |
private Object xfer(Object e, boolean haveData, int how, long nanos) { |
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 |
509 |
> |
Node<E> s = null; // the node to append, if needed |
510 |
|
|
511 |
|
retry: for (;;) { // restart on append race |
512 |
|
|
513 |
< |
for (Node h = head, p = h; p != null;) { // find & match first node |
513 |
> |
for (Node<E> h = head, p = h; p != null;) { |
514 |
> |
// find & match first node |
515 |
|
boolean isData = p.isData; |
516 |
|
Object item = p.item; |
517 |
|
if (item != p && (item != null) == isData) { // unmatched |
518 |
|
if (isData == haveData) // can't match |
519 |
|
break; |
520 |
|
if (p.casItem(item, e)) { // match |
521 |
< |
Thread w = p.waiter; |
522 |
< |
while (p != h) { // update head |
523 |
< |
Node n = p.next; // by 2 unless singleton |
524 |
< |
if (n != null) |
525 |
< |
p = n; |
477 |
< |
if (head == h && casHead(h, p)) { |
521 |
> |
for (Node<E> q = p; q != h;) { |
522 |
> |
Node<E> n = q.next; // update head by 2 |
523 |
> |
if (n != null) // unless singleton |
524 |
> |
q = n; |
525 |
> |
if (head == h && casHead(h, q)) { |
526 |
|
h.forgetNext(); |
527 |
|
break; |
528 |
|
} // advance and retry |
529 |
|
if ((h = head) == null || |
530 |
< |
(p = h.next) == null || !p.isMatched()) |
530 |
> |
(q = h.next) == null || !q.isMatched()) |
531 |
|
break; // unless slack < 2 |
532 |
|
} |
533 |
< |
LockSupport.unpark(w); |
534 |
< |
return item; |
533 |
> |
LockSupport.unpark(p.waiter); |
534 |
> |
return this.<E>cast(item); |
535 |
|
} |
536 |
|
} |
537 |
< |
Node n = p.next; |
537 |
> |
Node<E> n = p.next; |
538 |
|
p = (p != n) ? n : (h = head); // Use head if p offlist |
539 |
|
} |
540 |
|
|
541 |
|
if (how >= ASYNC) { // No matches available |
542 |
|
if (s == null) |
543 |
< |
s = new Node(e, haveData); |
544 |
< |
Node pred = tryAppend(s, haveData); |
543 |
> |
s = new Node<E>(e, haveData); |
544 |
> |
Node<E> pred = tryAppend(s, haveData); |
545 |
|
if (pred == null) |
546 |
|
continue retry; // lost race vs opposite mode |
547 |
|
if (how >= SYNC) |
548 |
< |
return awaitMatch(pred, s, e, how, nanos); |
548 |
> |
return awaitMatch(s, pred, e, how, nanos); |
549 |
|
} |
550 |
|
return e; // not waiting |
551 |
|
} |
554 |
|
/** |
555 |
|
* Tries to append node s as tail. |
556 |
|
* |
509 |
– |
* @param haveData true if appending in data mode |
557 |
|
* @param s the node to append |
558 |
+ |
* @param haveData true if appending in data mode |
559 |
|
* @return null on failure due to losing race with append in |
560 |
|
* different mode, else s's predecessor, or s itself if no |
561 |
|
* predecessor |
562 |
|
*/ |
563 |
< |
private Node tryAppend(Node s, boolean haveData) { |
564 |
< |
for (Node t = tail, p = t;;) { // move p to actual tail and append |
565 |
< |
Node n, u; // temps for reads of next & tail |
563 |
> |
private Node<E> tryAppend(Node<E> s, boolean haveData) { |
564 |
> |
for (Node<E> t = tail, p = t;;) { // move p to last node and append |
565 |
> |
Node<E> n, u; // temps for reads of next & tail |
566 |
|
if (p == null && (p = head) == null) { |
567 |
|
if (casHead(null, s)) |
568 |
|
return s; // initialize |
569 |
|
} |
570 |
|
else if (p.cannotPrecede(haveData)) |
571 |
|
return null; // lost race vs opposite mode |
572 |
< |
else if ((n = p.next) != null) // Not tail; keep traversing |
572 |
> |
else if ((n = p.next) != null) // not last; keep traversing |
573 |
|
p = p != t && t != (u = tail) ? (t = u) : // stale tail |
574 |
|
(p != n) ? n : null; // restart if off list |
575 |
|
else if (!p.casNext(null, s)) |
576 |
|
p = p.next; // re-read on CAS failure |
577 |
|
else { |
578 |
< |
if (p != t) { // Update if slack now >= 2 |
578 |
> |
if (p != t) { // update if slack now >= 2 |
579 |
|
while ((tail != t || !casTail(t, s)) && |
580 |
|
(t = tail) != null && |
581 |
|
(s = t.next) != null && // advance and retry |
589 |
|
/** |
590 |
|
* Spins/yields/blocks until node s is matched or caller gives up. |
591 |
|
* |
544 |
– |
* @param pred the predecessor of s or s or null if none |
592 |
|
* @param s the waiting node |
593 |
+ |
* @param pred the predecessor of s, or s itself if it has no |
594 |
+ |
* predecessor, or null if unknown (the null case does not occur |
595 |
+ |
* in any current calls but may in possible future extensions) |
596 |
|
* @param e the comparison value for checking match |
597 |
|
* @param how either SYNC or TIMEOUT |
598 |
|
* @param nanos timeout value |
599 |
|
* @return matched item, or e if unmatched on interrupt or timeout |
600 |
|
*/ |
601 |
< |
private Object awaitMatch(Node pred, Node s, Object e, |
552 |
< |
int how, long nanos) { |
601 |
> |
private E awaitMatch(Node<E> s, Node<E> pred, E e, int how, long nanos) { |
602 |
|
long lastTime = (how == TIMEOUT) ? System.nanoTime() : 0L; |
603 |
|
Thread w = Thread.currentThread(); |
604 |
|
int spins = -1; // initialized after first item and cancel checks |
607 |
|
for (;;) { |
608 |
|
Object item = s.item; |
609 |
|
if (item != e) { // matched |
610 |
+ |
assert item != s; |
611 |
|
s.forgetContents(); // avoid garbage |
612 |
< |
return item; |
612 |
> |
return this.<E>cast(item); |
613 |
|
} |
614 |
|
if ((w.isInterrupted() || (how == TIMEOUT && nanos <= 0)) && |
615 |
< |
s.casItem(e, s)) { // cancel |
615 |
> |
s.casItem(e, s)) { // cancel |
616 |
|
unsplice(pred, s); |
617 |
|
return e; |
618 |
|
} |
621 |
|
if ((spins = spinsFor(pred, s.isData)) > 0) |
622 |
|
randomYields = ThreadLocalRandom.current(); |
623 |
|
} |
624 |
< |
else if (spins > 0) { // spin, occasionally yield |
625 |
< |
if (randomYields.nextInt(FRONT_SPINS) == 0) |
626 |
< |
Thread.yield(); |
627 |
< |
--spins; |
624 |
> |
else if (spins > 0) { // spin |
625 |
> |
if (--spins == 0) |
626 |
> |
shortenHeadPath(); // reduce slack before blocking |
627 |
> |
else if (randomYields.nextInt(CHAINED_SPINS) == 0) |
628 |
> |
Thread.yield(); // occasionally yield |
629 |
|
} |
630 |
|
else if (s.waiter == null) { |
631 |
< |
shortenHeadPath(); // reduce slack before blocking |
581 |
< |
s.waiter = w; // request unpark |
631 |
> |
s.waiter = w; // request unpark then recheck |
632 |
|
} |
633 |
|
else if (how == TIMEOUT) { |
634 |
|
long now = System.nanoTime(); |
638 |
|
} |
639 |
|
else { |
640 |
|
LockSupport.park(this); |
641 |
+ |
s.waiter = null; |
642 |
|
spins = -1; // spin if front upon wakeup |
643 |
|
} |
644 |
|
} |
648 |
|
* Returns spin/yield value for a node with given predecessor and |
649 |
|
* data mode. See above for explanation. |
650 |
|
*/ |
651 |
< |
private static int spinsFor(Node pred, boolean haveData) { |
651 |
> |
private static int spinsFor(Node<?> pred, boolean haveData) { |
652 |
|
if (MP && pred != null) { |
653 |
< |
boolean predData = pred.isData; |
654 |
< |
if (predData != haveData) // front and phase change |
655 |
< |
return FRONT_SPINS + (FRONT_SPINS >>> 1); |
605 |
< |
if (predData != (pred.item != null)) // probably at front |
653 |
> |
if (pred.isData != haveData) // phase change |
654 |
> |
return FRONT_SPINS + CHAINED_SPINS; |
655 |
> |
if (pred.isMatched()) // probably at front |
656 |
|
return FRONT_SPINS; |
657 |
|
if (pred.waiter == null) // pred apparently spinning |
658 |
|
return CHAINED_SPINS; |
665 |
|
* or trailing node; failing on contention. |
666 |
|
*/ |
667 |
|
private void shortenHeadPath() { |
668 |
< |
Node h, hn, p, q; |
668 |
> |
Node<E> h, hn, p, q; |
669 |
|
if ((p = h = head) != null && h.isMatched() && |
670 |
|
(q = hn = h.next) != null) { |
671 |
< |
Node n; |
671 |
> |
Node<E> n; |
672 |
|
while ((n = q.next) != q) { |
673 |
|
if (n == null || !q.isMatched()) { |
674 |
|
if (hn != q && h.next == hn) |
687 |
|
* Returns the first unmatched node of the given mode, or null if |
688 |
|
* none. Used by methods isEmpty, hasWaitingConsumer. |
689 |
|
*/ |
690 |
< |
private Node firstOfMode(boolean data) { |
691 |
< |
for (Node p = head; p != null; ) { |
690 |
> |
private Node<E> firstOfMode(boolean data) { |
691 |
> |
for (Node<E> p = head; p != null; ) { |
692 |
|
if (!p.isMatched()) |
693 |
|
return (p.isData == data) ? p : null; |
694 |
< |
Node n = p.next; |
694 |
> |
Node<E> n = p.next; |
695 |
|
p = (n != p) ? n : head; |
696 |
|
} |
697 |
|
return null; |
699 |
|
|
700 |
|
/** |
701 |
|
* Returns the item in the first unmatched node with isData; or |
702 |
< |
* null if none. Used by peek. |
702 |
> |
* null if none. Used by peek. |
703 |
|
*/ |
704 |
< |
private Object firstDataItem() { |
705 |
< |
for (Node p = head; p != null; ) { |
704 |
> |
private E firstDataItem() { |
705 |
> |
for (Node<E> p = head; p != null; ) { |
706 |
|
boolean isData = p.isData; |
707 |
|
Object item = p.item; |
708 |
|
if (item != p && (item != null) == isData) |
709 |
< |
return isData ? item : null; |
710 |
< |
Node n = p.next; |
709 |
> |
return isData ? this.<E>cast(item) : null; |
710 |
> |
Node<E> n = p.next; |
711 |
|
p = (n != p) ? n : head; |
712 |
|
} |
713 |
|
return null; |
719 |
|
*/ |
720 |
|
private int countOfMode(boolean data) { |
721 |
|
int count = 0; |
722 |
< |
for (Node p = head; p != null; ) { |
722 |
> |
for (Node<E> p = head; p != null; ) { |
723 |
|
if (!p.isMatched()) { |
724 |
|
if (p.isData != data) |
725 |
|
return 0; |
726 |
|
if (++count == Integer.MAX_VALUE) // saturated |
727 |
|
break; |
728 |
|
} |
729 |
< |
Node n = p.next; |
729 |
> |
Node<E> n = p.next; |
730 |
|
if (n != p) |
731 |
|
p = n; |
732 |
|
else { |
738 |
|
} |
739 |
|
|
740 |
|
final class Itr implements Iterator<E> { |
741 |
< |
private Node nextNode; // next node to return item for |
742 |
< |
private Object nextItem; // the corresponding item |
743 |
< |
private Node lastRet; // last returned node, to support remove |
741 |
> |
private Node<E> nextNode; // next node to return item for |
742 |
> |
private E nextItem; // the corresponding item |
743 |
> |
private Node<E> lastRet; // last returned node, to support remove |
744 |
|
|
745 |
|
/** |
746 |
|
* Moves to next node after prev, or first node if prev null. |
747 |
|
*/ |
748 |
< |
private void advance(Node prev) { |
748 |
> |
private void advance(Node<E> prev) { |
749 |
|
lastRet = prev; |
750 |
< |
Node p; |
750 |
> |
Node<E> p; |
751 |
|
if (prev == null || (p = prev.next) == prev) |
752 |
|
p = head; |
753 |
|
while (p != null) { |
754 |
|
Object item = p.item; |
755 |
|
if (p.isData) { |
756 |
|
if (item != null && item != p) { |
757 |
< |
nextItem = item; |
757 |
> |
nextItem = LinkedTransferQueue.this.<E>cast(item); |
758 |
|
nextNode = p; |
759 |
|
return; |
760 |
|
} |
761 |
|
} |
762 |
|
else if (item == null) |
763 |
|
break; |
764 |
< |
Node n = p.next; |
764 |
> |
Node<E> n = p.next; |
765 |
|
p = (n != p) ? n : head; |
766 |
|
} |
767 |
|
nextNode = null; |
776 |
|
} |
777 |
|
|
778 |
|
public final E next() { |
779 |
< |
Node p = nextNode; |
779 |
> |
Node<E> p = nextNode; |
780 |
|
if (p == null) throw new NoSuchElementException(); |
781 |
< |
Object e = nextItem; |
781 |
> |
E e = nextItem; |
782 |
|
advance(p); |
783 |
< |
return (E) e; |
783 |
> |
return e; |
784 |
|
} |
785 |
|
|
786 |
|
public final void remove() { |
787 |
< |
Node p = lastRet; |
787 |
> |
Node<E> p = lastRet; |
788 |
|
if (p == null) throw new IllegalStateException(); |
789 |
|
lastRet = null; |
790 |
< |
findAndRemoveNode(p); |
790 |
> |
findAndRemoveDataNode(p); |
791 |
|
} |
792 |
|
} |
793 |
|
|
800 |
|
* @param pred predecessor of node to be unspliced |
801 |
|
* @param s the node to be unspliced |
802 |
|
*/ |
803 |
< |
private void unsplice(Node pred, Node s) { |
803 |
> |
private void unsplice(Node<E> pred, Node<E> s) { |
804 |
|
s.forgetContents(); // clear unneeded fields |
805 |
|
/* |
806 |
|
* At any given time, exactly one node on list cannot be |
807 |
< |
* deleted -- the last inserted node. To accommodate this, if |
808 |
< |
* we cannot delete s, we save its predecessor as "cleanMe", |
807 |
> |
* unlinked -- the last inserted node. To accommodate this, if |
808 |
> |
* we cannot unlink s, we save its predecessor as "cleanMe", |
809 |
|
* processing the previously saved version first. Because only |
810 |
|
* one node in the list can have a null next, at least one of |
811 |
|
* node s or the node previously saved can always be |
813 |
|
*/ |
814 |
|
if (pred != null && pred != s) { |
815 |
|
while (pred.next == s) { |
816 |
< |
Node oldpred = (cleanMe == null) ? null : reclean(); |
817 |
< |
Node n = s.next; |
816 |
> |
Node<E> oldpred = (cleanMe == null) ? null : reclean(); |
817 |
> |
Node<E> n = s.next; |
818 |
|
if (n != null) { |
819 |
|
if (n != s) |
820 |
|
pred.casNext(s, n); |
833 |
|
* |
834 |
|
* @return current cleanMe node (or null) |
835 |
|
*/ |
836 |
< |
private Node reclean() { |
836 |
> |
private Node<E> reclean() { |
837 |
|
/* |
838 |
|
* cleanMe is, or at one time was, predecessor of a cancelled |
839 |
|
* node s that was the tail so could not be unspliced. If it |
844 |
|
* we can (must) clear cleanMe without unsplicing. This can |
845 |
|
* loop only due to contention. |
846 |
|
*/ |
847 |
< |
Node pred; |
847 |
> |
Node<E> pred; |
848 |
|
while ((pred = cleanMe) != null) { |
849 |
< |
Node s = pred.next; |
850 |
< |
Node n; |
849 |
> |
Node<E> s = pred.next; |
850 |
> |
Node<E> n; |
851 |
|
if (s == null || s == pred || !s.isMatched()) |
852 |
|
casCleanMe(pred, null); // already gone |
853 |
|
else if ((n = s.next) != null) { |
863 |
|
|
864 |
|
/** |
865 |
|
* Main implementation of Iterator.remove(). Find |
866 |
< |
* and unsplice the given node. |
866 |
> |
* and unsplice the given data node. |
867 |
|
*/ |
868 |
< |
final void findAndRemoveNode(Node s) { |
868 |
> |
final void findAndRemoveDataNode(Node<E> s) { |
869 |
> |
assert s.isData; |
870 |
|
if (s.tryMatchData()) { |
871 |
< |
Node pred = null; |
821 |
< |
Node p = head; |
822 |
< |
while (p != null) { |
871 |
> |
for (Node<E> pred = null, p = head; p != null; ) { |
872 |
|
if (p == s) { |
873 |
|
unsplice(pred, p); |
874 |
|
break; |
875 |
|
} |
876 |
< |
if (!p.isData && !p.isMatched()) |
876 |
> |
if (p.isUnmatchedRequest()) |
877 |
|
break; |
878 |
|
pred = p; |
879 |
|
if ((p = p.next) == pred) { // stale |
889 |
|
*/ |
890 |
|
private boolean findAndRemove(Object e) { |
891 |
|
if (e != null) { |
892 |
< |
Node pred = null; |
844 |
< |
Node p = head; |
845 |
< |
while (p != null) { |
892 |
> |
for (Node<E> pred = null, p = head; p != null; ) { |
893 |
|
Object item = p.item; |
894 |
|
if (p.isData) { |
895 |
|
if (item != null && item != p && e.equals(item) && |
901 |
|
else if (item == null) |
902 |
|
break; |
903 |
|
pred = p; |
904 |
< |
if ((p = p.next) == pred) { |
904 |
> |
if ((p = p.next) == pred) { // stale |
905 |
|
pred = null; |
906 |
|
p = head; |
907 |
|
} |
1037 |
|
} |
1038 |
|
|
1039 |
|
public E take() throws InterruptedException { |
1040 |
< |
Object e = xfer(null, false, SYNC, 0); |
1040 |
> |
E e = xfer(null, false, SYNC, 0); |
1041 |
|
if (e != null) |
1042 |
< |
return (E)e; |
1042 |
> |
return e; |
1043 |
|
Thread.interrupted(); |
1044 |
|
throw new InterruptedException(); |
1045 |
|
} |
1046 |
|
|
1047 |
|
public E poll(long timeout, TimeUnit unit) throws InterruptedException { |
1048 |
< |
Object e = xfer(null, false, TIMEOUT, unit.toNanos(timeout)); |
1048 |
> |
E e = xfer(null, false, TIMEOUT, unit.toNanos(timeout)); |
1049 |
|
if (e != null || !Thread.interrupted()) |
1050 |
< |
return (E)e; |
1050 |
> |
return e; |
1051 |
|
throw new InterruptedException(); |
1052 |
|
} |
1053 |
|
|
1054 |
|
public E poll() { |
1055 |
< |
return (E)xfer(null, false, NOW, 0); |
1055 |
> |
return xfer(null, false, NOW, 0); |
1056 |
|
} |
1057 |
|
|
1058 |
|
/** |
1109 |
|
} |
1110 |
|
|
1111 |
|
public E peek() { |
1112 |
< |
return (E) firstDataItem(); |
1112 |
> |
return firstDataItem(); |
1113 |
|
} |
1114 |
|
|
1115 |
|
/** |
1205 |
|
} |
1206 |
|
} |
1207 |
|
|
1161 |
– |
|
1208 |
|
// Unsafe mechanics |
1209 |
|
|
1210 |
|
private static final sun.misc.Unsafe UNSAFE = getUnsafe(); |
1227 |
|
} |
1228 |
|
} |
1229 |
|
|
1230 |
< |
private static sun.misc.Unsafe getUnsafe() { |
1230 |
> |
/** |
1231 |
> |
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
1232 |
> |
* Replace with a simple call to Unsafe.getUnsafe when integrating |
1233 |
> |
* into a jdk. |
1234 |
> |
* |
1235 |
> |
* @return a sun.misc.Unsafe |
1236 |
> |
*/ |
1237 |
> |
static sun.misc.Unsafe getUnsafe() { |
1238 |
|
try { |
1239 |
|
return sun.misc.Unsafe.getUnsafe(); |
1240 |
|
} catch (SecurityException se) { |