ViewVC Help

View File | Revision Log | Show Annotations | Download File | Root Listing

root/jsr166/jsr166/src/jsr166y/LinkedTransferQueue.java

Comparing jsr166/src/jsr166y/LinkedTransferQueue.java (file contents):
Revision 1.66 by jsr166, Mon Nov 2 18:38:37 2009 UTC vs.
Revision 1.67 by dl, Sat Nov 14 20:27:18 2009 UTC

#	Line 206 | Line 206 | public class LinkedTransferQueue<E> exte
206		* additional GC bookkeeping ("write barriers") that are sometimes
207		* more costly than the writes themselves because of contention).
208		*
209	–	* Removal of interior nodes (due to timed out or interrupted
210	–	* waits, or calls to remove(x) or Iterator.remove) can use a
211	–	* scheme roughly similar to that described in Scherer, Lea, and
212	–	* Scott's SynchronousQueue. Given a predecessor, we can unsplice
213	–	* any node except the (actual) tail of the queue. To avoid
214	–	* build-up of cancelled trailing nodes, upon a request to remove
215	–	* a trailing node, it is placed in field "cleanMe" to be
216	–	* unspliced upon the next call to unsplice any other node.
217	–	* Situations needing such mechanics are not common but do occur
218	–	* in practice; for example when an unbounded series of short
219	–	* timed calls to poll repeatedly time out but never otherwise
220	–	* fall off the list because of an untimed call to take at the
221	–	* front of the queue. Note that maintaining field cleanMe does
222	–	* not otherwise much impact garbage retention even if never
223	–	* cleared by some other call because the held node will
224	–	* eventually either directly or indirectly lead to a self-link
225	–	* once off the list.
226	–	*
209		* *** Overview of implementation ***
210		*
211		* We use a threshold-based approach to updates, with a slack
#	Line 239 | Line 221 | public class LinkedTransferQueue<E> exte
221		* per-thread one available, but even ThreadLocalRandom is too
222		* heavy for these purposes.
223		*
224	<	* With such a small slack threshold value, it is rarely
225	<	* worthwhile to augment this with path short-circuiting; i.e.,
226	<	* unsplicing nodes between head and the first unmatched node, or
227	<	* 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).
224	>	* With such a small slack threshold value, it is not worthwhile
225	>	* to augment this with path short-circuiting (i.e., unsplicing
226	>	* interior nodes) except in the case of cancellation/removal (see
227	>	* below).
228		*
229		* We allow both the head and tail fields to be null before any
230		* nodes are enqueued; initializing upon first append.  This
#	Line 329 | Line 306 | public class LinkedTransferQueue<E> exte
306		*    versa) compared to their predecessors receive additional
307		*    chained spins, reflecting longer paths typically required to
308		*    unblock threads during phase changes.
309	+	*
310	+	*
311	+	* ** Unlinking removed interior nodes **
312	+	*
313	+	* In addition to minimizing garbage retention via self-linking
314	+	* described above, we also unlink removed interior nodes. These
315	+	* may arise due to timed out or interrupted waits, or calls to
316	+	* remove(x) or Iterator.remove.  Normally, given a node that was
317	+	* at one time known to be the predecessor of some node s that is
318	+	* to be removed, we can unsplice s by CASing the next field of
319	+	* its predecessor if it still points to s (otherwise s must
320	+	* already have been removed or is now offlist). But there are two
321	+	* situations in which we cannot guarantee to make node s
322	+	* unreachable in this way: (1) If s is the trailing node of list
323	+	* (i.e., with null next), then it is pinned as the target node
324	+	* for appends, so can only be removed later when other nodes are
325	+	* appended. (2) We cannot necessarily unlink s given a
326	+	* predecessor node that is matched (including the case of being
327	+	* cancelled): the predecessor may already be already unspliced,
328	+	* in which case some previous reachable node may still point to
329	+	* s.  (For further explanation see Herlihy & Shavit "The Art of
330	+	* Multiprocessor Programming" chapter 9).  Although, in both
331	+	* cases, we can rule out the need for further action if either s
332	+	* or its predecessor are (or can be made to be) at, or fall off
333	+	* from, the head of list.
334	+	*
335	+	* Without taking these into account, it would be possible for an
336	+	* unbounded number of supposedly removed nodes to remain
337	+	* reachable.  Situations leading to such buildup are uncommon but
338	+	* can occur in practice; for example when a series of short timed
339	+	* calls to poll repeatedly time out but never otherwise fall off
340	+	* the list because of an untimed call to take at the front of the
341	+	* queue.
342	+	*
343	+	* When these cases arise, rather than always retraversing the
344	+	* entire list to find an actual predecessor to unlink (which
345	+	* won't help for case (1) anyway), we record a conservative
346	+	* estimate of possible unsplice failures (in "sweepVotes).  We
347	+	* trigger a full sweep when the estimate exceeds a threshold
348	+	* indicating the maximum number of estimated removal failures to
349	+	* tolerate before sweeping through, unlinking cancelled nodes
350	+	* that were not unlinked upon initial removal. We perform sweeps
351	+	* by the thread hitting threshold (rather than background threads
352	+	* or by spreading work to other threads) because in the main
353	+	* contexts in which removal occurs, the caller is already
354	+	* timed-out, cancelled, or performing a potentially O(n)
355	+	* operation (i.e., remove(x)), none of which are time-critical
356	+	* enough to warrant the overhead that alternatives would impose
357	+	* on other threads.
358	+	*
359	+	* Because the sweepVotes estimate is conservative, and because
360	+	* nodes become unlinked "naturally" as they fall off the head of
361	+	* the queue, and because we allow votes to accumulate even while
362	+	* sweeps are in progress, there are typically signficantly fewer
363	+	* such nodes than estimated.  Choice of a threshold value
364	+	* balances the likelihood of wasted effort and contention, versus
365	+	* providing a worst-case bound on retention of interior nodes in
366	+	* quiescent queues. The value defined below was chosen
367	+	* empirically to balance these under various timeout scenarios.
368	+	*
369	+	* Note that we cannot self-link unlinked interior nodes during
370	+	* sweeps. However, the associated garbage chains terminate when
371	+	* some successor ultimately falls off the head of the list and is
372	+	* self-linked.
373		*/
374
375		/** True if on multiprocessor */
#	Line 355 | Line 396 | public class LinkedTransferQueue<E> exte
396		private static final int CHAINED_SPINS = FRONT_SPINS >>> 1;
397
398		/**
399	+	* The maximum number of estimated removal failures (sweepVotes)
400	+	* to tolerate before sweeping through the queue unlinking
401	+	* cancelled nodes that were not unlinked upon initial
402	+	* removal. See above for explanation. The value must be at least
403	+	* two to avoid useless sweeps when removing trailing nodes.
404	+	*/
405	+	static final int SWEEP_THRESHOLD = 32;
406	+
407	+	/**
408		* Queue nodes. Uses Object, not E, for items to allow forgetting
409		* them after use.  Relies heavily on Unsafe mechanics to minimize
410	<	* unnecessary ordering constraints: Writes that intrinsically
411	<	* precede or follow CASes use simple relaxed forms.  Other
362	<	* cleanups use releasing/lazy writes.
410	>	* unnecessary ordering constraints: Writes that are intrinsically
411	>	* ordered wrt other accesses or CASes use simple relaxed forms.
412		*/
413		static final class Node {
414		final boolean isData;   // false if this is a request node
#	Line 395 | Line 444 | public class LinkedTransferQueue<E> exte
444		}
445
446		/**
447	<	* Sets item to self (using a releasing/lazy write) and waiter
448	<	* to null, to avoid garbage retention after extracting or
449	<	* cancelling.
447	>	* Sets item to self and waiter to null, to avoid garbage
448	>	* retention after matching or cancelling. Uses relaxed writes
449	>	* bacause order is already constrained in the only calling
450	>	* contexts: item is forgotten only after volatile/atomic
451	>	* mechanics that extract items.  Similarly, clearing waiter
452	>	* follows either CAS or return from park (if ever parked;
453	>	* else we don't care).
454		*/
455		final void forgetContents() {
456	<	UNSAFE.putOrderedObject(this, itemOffset, this);
457	<	UNSAFE.putOrderedObject(this, waiterOffset, null);
456	>	UNSAFE.putObject(this, itemOffset, this);
457	>	UNSAFE.putObject(this, waiterOffset, null);
458		}
459
460		/**
#	Line 459 | Line 512 | public class LinkedTransferQueue<E> exte
512		/** head of the queue; null until first enqueue */
513		transient volatile Node head;
514
462	–	/** predecessor of dangling unspliceable node */
463	–	private transient volatile Node cleanMe; // decl here reduces contention
464	–
515		/** tail of the queue; null until first append */
516		private transient volatile Node tail;
517
518	+	/** The number of apparent failures to unsplice removed nodes */
519	+	private transient volatile int sweepVotes;
520	+
521		// CAS methods for fields
522		private boolean casTail(Node cmp, Node val) {
523		return UNSAFE.compareAndSwapObject(this, tailOffset, cmp, val);
#	Line 474 | Line 527 | public class LinkedTransferQueue<E> exte
527		return UNSAFE.compareAndSwapObject(this, headOffset, cmp, val);
528		}
529
530	<	private boolean casCleanMe(Node cmp, Node val) {
531	<	return UNSAFE.compareAndSwapObject(this, cleanMeOffset, cmp, val);
530	>	private boolean casSweepVotes(int cmp, int val) {
531	>	return UNSAFE.compareAndSwapInt(this, sweepVotesOffset, cmp, val);
532		}
533
534		/*
#	Line 517 | Line 570 | public class LinkedTransferQueue<E> exte
570		break;
571		if (p.casItem(item, e)) { // match
572		for (Node q = p; q != h;) {
573	<	Node n = q.next;  // update head by 2
574	<	if (n != null)    // unless singleton
522	<	q = n;
523	<	if (head == h && casHead(h, q)) {
573	>	Node n = q.next;  // update by 2 unless singleton
574	>	if (head == h && casHead(h, n == null? q : n)) {
575		h.forgetNext();
576		break;
577		}                 // advance and retry
#	Line 610 | Line 661 | public class LinkedTransferQueue<E> exte
661		return this.<E>cast(item);
662		}
663		if ((w.isInterrupted() || (timed && nanos <= 0)) &&
664	<	s.casItem(e, s)) {       // cancel
664	>	s.casItem(e, s)) {        // cancel
665		unsplice(pred, s);
666		return e;
667		}
#	Line 620 | Line 671 | public class LinkedTransferQueue<E> exte
671		randomYields = ThreadLocalRandom.current();
672		}
673		else if (spins > 0) {             // spin
674	<	if (--spins == 0)
675	<	shortenHeadPath();        // reduce slack before blocking
625	<	else if (randomYields.nextInt(CHAINED_SPINS) == 0)
674	>	--spins;
675	>	if (randomYields.nextInt(CHAINED_SPINS) == 0)
676		Thread.yield();           // occasionally yield
677		}
678		else if (s.waiter == null) {
#	Line 636 | Line 686 | public class LinkedTransferQueue<E> exte
686		}
687		else {
688		LockSupport.park(this);
639	–	s.waiter = null;
640	–	spins = -1;                   // spin if front upon wakeup
689		}
690		}
691		}
#	Line 658 | Line 706 | public class LinkedTransferQueue<E> exte
706		return 0;
707		}
708
661	–	/**
662	–	* Tries (once) to unsplice nodes between head and first unmatched
663	–	* or trailing node; failing on contention.
664	–	*/
665	–	private void shortenHeadPath() {
666	–	Node h, hn, p, q;
667	–	if ((p = h = head) != null && h.isMatched() &&
668	–	(q = hn = h.next) != null) {
669	–	Node n;
670	–	while ((n = q.next) != q) {
671	–	if (n == null || !q.isMatched()) {
672	–	if (hn != q && h.next == hn)
673	–	h.casNext(hn, q);
674	–	break;
675	–	}
676	–	p = q;
677	–	q = n;
678	–	}
679	–	}
680	–	}
681	–
709		/* -------------- Traversal methods -------------- */
710
711		/**
#	Line 791 | Line 818 | public class LinkedTransferQueue<E> exte
818		public final void remove() {
819		Node p = lastRet;
820		if (p == null) throw new IllegalStateException();
821	<	findAndRemoveDataNode(lastPred, p);
821	>	if (p.tryMatchData())
822	>	unsplice(lastPred, p);
823		}
824		}
825
#	Line 801 | Line 829 | public class LinkedTransferQueue<E> exte
829		* Unsplices (now or later) the given deleted/cancelled node with
830		* the given predecessor.
831		*
832	<	* @param pred predecessor of node to be unspliced
832	>	* @param pred a node that was at one time known to be the
833	>	* predecessor of s, or null or s itself if s is/was at head
834		* @param s the node to be unspliced
835		*/
836	<	private void unsplice(Node pred, Node s) {
837	<	s.forgetContents(); // clear unneeded fields
836	>	final void unsplice(Node pred, Node s) {
837	>	s.forgetContents(); // forget unneeded fields
838		/*
839	<	* At any given time, exactly one node on list cannot be
840	<	* unlinked -- the last inserted node. To accommodate this, if
841	<	* we cannot unlink s, we save its predecessor as "cleanMe",
842	<	* processing the previously saved version first. Because only
843	<	* one node in the list can have a null next, at least one of
815	<	* node s or the node previously saved can always be
816	<	* processed, so this always terminates.
839	>	* See above for rationale. Briefly: if pred still points to
840	>	* s, try to unlink s.  If s cannot be unlinked, because it is
841	>	* trailing node or pred might be unlinked, and neither pred
842	>	* nor s are head or offlist, add to sweepVotes, and if enough
843	>	* votes have accumulated, sweep.
844		*/
845	<	if (pred != null && pred != s) {
846	<	while (pred.next == s) {
847	<	Node oldpred = (cleanMe == null) ? null : reclean();
848	<	Node n = s.next;
849	<	if (n != null) {
850	<	if (n != s)
851	<	pred.casNext(s, n);
852	<	break;
845	>	if (pred != null && pred != s && pred.next == s) {
846	>	Node n = s.next;
847	>	if (n == null ||
848	>	(n != s && pred.casNext(s, n) && pred.isMatched())) {
849	>	for (;;) {               // check if at, or could be, head
850	>	Node h = head;
851	>	if (h == pred || h == s || h == null)
852	>	return;          // at head or list empty
853	>	if (!h.isMatched())
854	>	break;
855	>	Node hn = h.next;
856	>	if (hn == null)
857	>	return;          // now empty
858	>	if (hn != h && casHead(h, hn))
859	>	h.forgetNext();  // advance head
860		}
861	<	if (oldpred == pred ||      // Already saved
862	<	((oldpred == null || oldpred.next == s) &&
863	<	casCleanMe(oldpred, pred))) {
864	<	break;
861	>	if (pred.next != pred && s.next != s) { // recheck if offlist
862	>	for (;;) {           // sweep now if enough votes
863	>	int v = sweepVotes;
864	>	if (v < SWEEP_THRESHOLD) {
865	>	if (casSweepVotes(v, v + 1))
866	>	break;
867	>	}
868	>	else if (casSweepVotes(v, 0)) {
869	>	sweep();
870	>	break;
871	>	}
872	>	}
873		}
874		}
875		}
876		}
877
878		/**
879	<	* Tries to unsplice the deleted/cancelled node held in cleanMe
838	<	* that was previously uncleanable because it was at tail.
839	<	*
840	<	* @return current cleanMe node (or null)
879	>	* Unlink matched nodes encountered in a traversal from head
880		*/
881	<	private Node reclean() {
882	<	/*
883	<	* cleanMe is, or at one time was, predecessor of a cancelled
884	<	* node s that was the tail so could not be unspliced.  If it
885	<	* is no longer the tail, try to unsplice if necessary and
886	<	* make cleanMe slot available.  This differs from similar
887	<	* code in unsplice() because we must check that pred still
849	<	* points to a matched node that can be unspliced -- if not,
850	<	* we can (must) clear cleanMe without unsplicing.  This can
851	<	* loop only due to contention.
852	<	*/
853	<	Node pred;
854	<	while ((pred = cleanMe) != null) {
855	<	Node s = pred.next;
856	<	Node n;
857	<	if (s == null || s == pred || !s.isMatched())
858	<	casCleanMe(pred, null); // already gone
859	<	else if ((n = s.next) != null) {
860	<	if (n != s)
861	<	pred.casNext(s, n);
862	<	casCleanMe(pred, null);
863	<	}
881	>	private void sweep() {
882	>	Node p = head, s, n;
883	>	while (p != null && (s = p.next) != null && (n = s.next) != null) {
884	>	if (p == s || s == n)
885	>	p = head; // stale
886	>	else if (s.isMatched())
887	>	p.casNext(s, n);
888		else
889	<	break;
866	<	}
867	<	return pred;
868	<	}
869	<
870	<	/**
871	<	* Main implementation of Iterator.remove(). Finds
872	<	* and unsplices the given data node.
873	<	*
874	<	* @param possiblePred possible predecessor of s
875	<	* @param s the node to remove
876	<	*/
877	<	final void findAndRemoveDataNode(Node possiblePred, Node s) {
878	<	assert s.isData;
879	<	if (s.tryMatchData()) {
880	<	if (possiblePred != null && possiblePred.next == s)
881	<	unsplice(possiblePred, s); // was actual predecessor
882	<	else {
883	<	for (Node pred = null, p = head; p != null; ) {
884	<	if (p == s) {
885	<	unsplice(pred, p);
886	<	break;
887	<	}
888	<	if (p.isUnmatchedRequest())
889	<	break;
890	<	pred = p;
891	<	if ((p = p.next) == pred) { // stale
892	<	pred = null;
893	<	p = head;
894	<	}
895	<	}
896	<	}
889	>	p = s;
890		}
891		}
892
#	Line 1225 | Line 1218 | public class LinkedTransferQueue<E> exte
1218		objectFieldOffset(UNSAFE, "head", LinkedTransferQueue.class);
1219		private static final long tailOffset =
1220		objectFieldOffset(UNSAFE, "tail", LinkedTransferQueue.class);
1221	<	private static final long cleanMeOffset =
1222	<	objectFieldOffset(UNSAFE, "cleanMe", LinkedTransferQueue.class);
1221	>	private static final long sweepVotesOffset =
1222	>	objectFieldOffset(UNSAFE, "sweepVotes", LinkedTransferQueue.class);
1223
1224		static long objectFieldOffset(sun.misc.Unsafe UNSAFE,
1225		String field, Class<?> klazz) {

Diff Legend

–	Removed lines
+	Added lines
<	Changed lines
>	Changed lines