xref: /titanic_51/usr/src/common/avl/avl.c (revision 70818f5837509317d1f5dac4d82d7b5a2d547c29)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 
29 /*
30  * AVL - generic AVL tree implementation for kernel use
31  *
32  * A complete description of AVL trees can be found in many CS textbooks.
33  *
34  * Here is a very brief overview. An AVL tree is a binary search tree that is
35  * almost perfectly balanced. By "almost" perfectly balanced, we mean that at
36  * any given node, the left and right subtrees are allowed to differ in height
37  * by at most 1 level.
38  *
39  * This relaxation from a perfectly balanced binary tree allows doing
40  * insertion and deletion relatively efficiently. Searching the tree is
41  * still a fast operation, roughly O(log(N)).
42  *
43  * The key to insertion and deletion is a set of tree maniuplations called
44  * rotations, which bring unbalanced subtrees back into the semi-balanced state.
45  *
46  * This implementation of AVL trees has the following peculiarities:
47  *
48  *	- The AVL specific data structures are physically embedded as fields
49  *	  in the "using" data structures.  To maintain generality the code
50  *	  must constantly translate between "avl_node_t *" and containing
51  *	  data structure "void *"s by adding/subracting the avl_offset.
52  *
53  *	- Since the AVL data is always embedded in other structures, there is
54  *	  no locking or memory allocation in the AVL routines. This must be
55  *	  provided for by the enclosing data structure's semantics. Typically,
56  *	  avl_insert()/_add()/_remove()/avl_insert_here() require some kind of
57  *	  exclusive write lock. Other operations require a read lock.
58  *
59  *      - The implementation uses iteration instead of explicit recursion,
60  *	  since it is intended to run on limited size kernel stacks. Since
61  *	  there is no recursion stack present to move "up" in the tree,
62  *	  there is an explicit "parent" link in the avl_node_t.
63  *
64  *      - The left/right children pointers of a node are in an array.
65  *	  In the code, variables (instead of constants) are used to represent
66  *	  left and right indices.  The implementation is written as if it only
67  *	  dealt with left handed manipulations.  By changing the value assigned
68  *	  to "left", the code also works for right handed trees.  The
69  *	  following variables/terms are frequently used:
70  *
71  *		int left;	// 0 when dealing with left children,
72  *				// 1 for dealing with right children
73  *
74  *		int left_heavy;	// -1 when left subtree is taller at some node,
75  *				// +1 when right subtree is taller
76  *
77  *		int right;	// will be the opposite of left (0 or 1)
78  *		int right_heavy;// will be the opposite of left_heavy (-1 or 1)
79  *
80  *		int direction;  // 0 for "<" (ie. left child); 1 for ">" (right)
81  *
82  *	  Though it is a little more confusing to read the code, the approach
83  *	  allows using half as much code (and hence cache footprint) for tree
84  *	  manipulations and eliminates many conditional branches.
85  *
86  *	- The avl_index_t is an opaque "cookie" used to find nodes at or
87  *	  adjacent to where a new value would be inserted in the tree. The value
88  *	  is a modified "avl_node_t *".  The bottom bit (normally 0 for a
89  *	  pointer) is set to indicate if that the new node has a value greater
90  *	  than the value of the indicated "avl_node_t *".
91  */
92 
93 #include <sys/types.h>
94 #include <sys/param.h>
95 #include <sys/debug.h>
96 #include <sys/avl.h>
97 #include <sys/cmn_err.h>
98 
99 /*
100  * Small arrays to translate between balance (or diff) values and child indeces.
101  *
102  * Code that deals with binary tree data structures will randomly use
103  * left and right children when examining a tree.  C "if()" statements
104  * which evaluate randomly suffer from very poor hardware branch prediction.
105  * In this code we avoid some of the branch mispredictions by using the
106  * following translation arrays. They replace random branches with an
107  * additional memory reference. Since the translation arrays are both very
108  * small the data should remain efficiently in cache.
109  */
110 static const int  avl_child2balance[2]	= {-1, 1};
111 static const int  avl_balance2child[]	= {0, 0, 1};
112 
113 
114 /*
115  * Walk from one node to the previous valued node (ie. an infix walk
116  * towards the left). At any given node we do one of 2 things:
117  *
118  * - If there is a left child, go to it, then to it's rightmost descendant.
119  *
120  * - otherwise we return thru parent nodes until we've come from a right child.
121  *
122  * Return Value:
123  * NULL - if at the end of the nodes
124  * otherwise next node
125  */
126 void *
127 avl_walk(avl_tree_t *tree, void	*oldnode, int left)
128 {
129 	size_t off = tree->avl_offset;
130 	avl_node_t *node = AVL_DATA2NODE(oldnode, off);
131 	int right = 1 - left;
132 	int was_child;
133 
134 
135 	/*
136 	 * nowhere to walk to if tree is empty
137 	 */
138 	if (node == NULL)
139 		return (NULL);
140 
141 	/*
142 	 * Visit the previous valued node. There are two possibilities:
143 	 *
144 	 * If this node has a left child, go down one left, then all
145 	 * the way right.
146 	 */
147 	if (node->avl_child[left] != NULL) {
148 		for (node = node->avl_child[left];
149 		    node->avl_child[right] != NULL;
150 		    node = node->avl_child[right])
151 			;
152 	/*
153 	 * Otherwise, return thru left children as far as we can.
154 	 */
155 	} else {
156 		for (;;) {
157 			was_child = AVL_XCHILD(node);
158 			node = AVL_XPARENT(node);
159 			if (node == NULL)
160 				return (NULL);
161 			if (was_child == right)
162 				break;
163 		}
164 	}
165 
166 	return (AVL_NODE2DATA(node, off));
167 }
168 
169 /*
170  * Return the lowest valued node in a tree or NULL.
171  * (leftmost child from root of tree)
172  */
173 void *
174 avl_first(avl_tree_t *tree)
175 {
176 	avl_node_t *node;
177 	avl_node_t *prev = NULL;
178 	size_t off = tree->avl_offset;
179 
180 	for (node = tree->avl_root; node != NULL; node = node->avl_child[0])
181 		prev = node;
182 
183 	if (prev != NULL)
184 		return (AVL_NODE2DATA(prev, off));
185 	return (NULL);
186 }
187 
188 /*
189  * Return the highest valued node in a tree or NULL.
190  * (rightmost child from root of tree)
191  */
192 void *
193 avl_last(avl_tree_t *tree)
194 {
195 	avl_node_t *node;
196 	avl_node_t *prev = NULL;
197 	size_t off = tree->avl_offset;
198 
199 	for (node = tree->avl_root; node != NULL; node = node->avl_child[1])
200 		prev = node;
201 
202 	if (prev != NULL)
203 		return (AVL_NODE2DATA(prev, off));
204 	return (NULL);
205 }
206 
207 /*
208  * Access the node immediately before or after an insertion point.
209  *
210  * "avl_index_t" is a (avl_node_t *) with the bottom bit indicating a child
211  *
212  * Return value:
213  *	NULL: no node in the given direction
214  *	"void *"  of the found tree node
215  */
216 void *
217 avl_nearest(avl_tree_t *tree, avl_index_t where, int direction)
218 {
219 	int child = AVL_INDEX2CHILD(where);
220 	avl_node_t *node = AVL_INDEX2NODE(where);
221 	void *data;
222 	size_t off = tree->avl_offset;
223 
224 	if (node == NULL) {
225 		ASSERT(tree->avl_root == NULL);
226 		return (NULL);
227 	}
228 	data = AVL_NODE2DATA(node, off);
229 	if (child != direction)
230 		return (data);
231 
232 	return (avl_walk(tree, data, direction));
233 }
234 
235 
236 /*
237  * Search for the node which contains "value".  The algorithm is a
238  * simple binary tree search.
239  *
240  * return value:
241  *	NULL: the value is not in the AVL tree
242  *		*where (if not NULL)  is set to indicate the insertion point
243  *	"void *"  of the found tree node
244  */
245 void *
246 avl_find(avl_tree_t *tree, void *value, avl_index_t *where)
247 {
248 	avl_node_t *node;
249 	avl_node_t *prev = NULL;
250 	int child = 0;
251 	int diff;
252 	size_t off = tree->avl_offset;
253 
254 	for (node = tree->avl_root; node != NULL;
255 	    node = node->avl_child[child]) {
256 
257 		prev = node;
258 
259 		diff = tree->avl_compar(value, AVL_NODE2DATA(node, off));
260 		ASSERT(-1 <= diff && diff <= 1);
261 		if (diff == 0) {
262 #ifdef DEBUG
263 			if (where != NULL)
264 				*where = 0;
265 #endif
266 			return (AVL_NODE2DATA(node, off));
267 		}
268 		child = avl_balance2child[1 + diff];
269 
270 	}
271 
272 	if (where != NULL)
273 		*where = AVL_MKINDEX(prev, child);
274 
275 	return (NULL);
276 }
277 
278 
279 /*
280  * Perform a rotation to restore balance at the subtree given by depth.
281  *
282  * This routine is used by both insertion and deletion. The return value
283  * indicates:
284  *	 0 : subtree did not change height
285  *	!0 : subtree was reduced in height
286  *
287  * The code is written as if handling left rotations, right rotations are
288  * symmetric and handled by swapping values of variables right/left[_heavy]
289  *
290  * On input balance is the "new" balance at "node". This value is either
291  * -2 or +2.
292  */
293 static int
294 avl_rotation(avl_tree_t *tree, avl_node_t *node, int balance)
295 {
296 	int left = !(balance < 0);	/* when balance = -2, left will be 0 */
297 	int right = 1 - left;
298 	int left_heavy = balance >> 1;
299 	int right_heavy = -left_heavy;
300 	avl_node_t *parent = AVL_XPARENT(node);
301 	avl_node_t *child = node->avl_child[left];
302 	avl_node_t *cright;
303 	avl_node_t *gchild;
304 	avl_node_t *gright;
305 	avl_node_t *gleft;
306 	int which_child = AVL_XCHILD(node);
307 	int child_bal = AVL_XBALANCE(child);
308 
309 	/* BEGIN CSTYLED */
310 	/*
311 	 * case 1 : node is overly left heavy, the left child is balanced or
312 	 * also left heavy. This requires the following rotation.
313 	 *
314 	 *                   (node bal:-2)
315 	 *                    /           \
316 	 *                   /             \
317 	 *              (child bal:0 or -1)
318 	 *              /    \
319 	 *             /      \
320 	 *                     cright
321 	 *
322 	 * becomes:
323 	 *
324 	 *              (child bal:1 or 0)
325 	 *              /        \
326 	 *             /          \
327 	 *                        (node bal:-1 or 0)
328 	 *                         /     \
329 	 *                        /       \
330 	 *                     cright
331 	 *
332 	 * we detect this situation by noting that child's balance is not
333 	 * right_heavy.
334 	 */
335 	/* END CSTYLED */
336 	if (child_bal != right_heavy) {
337 
338 		/*
339 		 * compute new balance of nodes
340 		 *
341 		 * If child used to be left heavy (now balanced) we reduced
342 		 * the height of this sub-tree -- used in "return...;" below
343 		 */
344 		child_bal += right_heavy; /* adjust towards right */
345 
346 		/*
347 		 * move "cright" to be node's left child
348 		 */
349 		cright = child->avl_child[right];
350 		node->avl_child[left] = cright;
351 		if (cright != NULL) {
352 			AVL_SETPARENT(cright, node);
353 			AVL_SETCHILD(cright, left);
354 		}
355 
356 		/*
357 		 * move node to be child's right child
358 		 */
359 		child->avl_child[right] = node;
360 		AVL_SETBALANCE(node, -child_bal);
361 		AVL_SETCHILD(node, right);
362 		AVL_SETPARENT(node, child);
363 
364 		/*
365 		 * update the pointer into this subtree
366 		 */
367 		AVL_SETBALANCE(child, child_bal);
368 		AVL_SETCHILD(child, which_child);
369 		AVL_SETPARENT(child, parent);
370 		if (parent != NULL)
371 			parent->avl_child[which_child] = child;
372 		else
373 			tree->avl_root = child;
374 
375 		return (child_bal == 0);
376 	}
377 
378 	/* BEGIN CSTYLED */
379 	/*
380 	 * case 2 : When node is left heavy, but child is right heavy we use
381 	 * a different rotation.
382 	 *
383 	 *                   (node b:-2)
384 	 *                    /   \
385 	 *                   /     \
386 	 *                  /       \
387 	 *             (child b:+1)
388 	 *              /     \
389 	 *             /       \
390 	 *                   (gchild b: != 0)
391 	 *                     /  \
392 	 *                    /    \
393 	 *                 gleft   gright
394 	 *
395 	 * becomes:
396 	 *
397 	 *              (gchild b:0)
398 	 *              /       \
399 	 *             /         \
400 	 *            /           \
401 	 *        (child b:?)   (node b:?)
402 	 *         /  \          /   \
403 	 *        /    \        /     \
404 	 *            gleft   gright
405 	 *
406 	 * computing the new balances is more complicated. As an example:
407 	 *	 if gchild was right_heavy, then child is now left heavy
408 	 *		else it is balanced
409 	 */
410 	/* END CSTYLED */
411 	gchild = child->avl_child[right];
412 	gleft = gchild->avl_child[left];
413 	gright = gchild->avl_child[right];
414 
415 	/*
416 	 * move gright to left child of node and
417 	 *
418 	 * move gleft to right child of node
419 	 */
420 	node->avl_child[left] = gright;
421 	if (gright != NULL) {
422 		AVL_SETPARENT(gright, node);
423 		AVL_SETCHILD(gright, left);
424 	}
425 
426 	child->avl_child[right] = gleft;
427 	if (gleft != NULL) {
428 		AVL_SETPARENT(gleft, child);
429 		AVL_SETCHILD(gleft, right);
430 	}
431 
432 	/*
433 	 * move child to left child of gchild and
434 	 *
435 	 * move node to right child of gchild and
436 	 *
437 	 * fixup parent of all this to point to gchild
438 	 */
439 	balance = AVL_XBALANCE(gchild);
440 	gchild->avl_child[left] = child;
441 	AVL_SETBALANCE(child, (balance == right_heavy ? left_heavy : 0));
442 	AVL_SETPARENT(child, gchild);
443 	AVL_SETCHILD(child, left);
444 
445 	gchild->avl_child[right] = node;
446 	AVL_SETBALANCE(node, (balance == left_heavy ? right_heavy : 0));
447 	AVL_SETPARENT(node, gchild);
448 	AVL_SETCHILD(node, right);
449 
450 	AVL_SETBALANCE(gchild, 0);
451 	AVL_SETPARENT(gchild, parent);
452 	AVL_SETCHILD(gchild, which_child);
453 	if (parent != NULL)
454 		parent->avl_child[which_child] = gchild;
455 	else
456 		tree->avl_root = gchild;
457 
458 	return (1);	/* the new tree is always shorter */
459 }
460 
461 
462 /*
463  * Insert a new node into an AVL tree at the specified (from avl_find()) place.
464  *
465  * Newly inserted nodes are always leaf nodes in the tree, since avl_find()
466  * searches out to the leaf positions.  The avl_index_t indicates the node
467  * which will be the parent of the new node.
468  *
469  * After the node is inserted, a single rotation further up the tree may
470  * be necessary to maintain an acceptable AVL balance.
471  */
472 void
473 avl_insert(avl_tree_t *tree, void *new_data, avl_index_t where)
474 {
475 	avl_node_t *node;
476 	avl_node_t *parent = AVL_INDEX2NODE(where);
477 	int old_balance;
478 	int new_balance;
479 	int which_child = AVL_INDEX2CHILD(where);
480 	size_t off = tree->avl_offset;
481 
482 	ASSERT(tree);
483 #ifdef _LP64
484 	ASSERT(((uintptr_t)new_data & 0x7) == 0);
485 #endif
486 
487 	node = AVL_DATA2NODE(new_data, off);
488 
489 	/*
490 	 * First, add the node to the tree at the indicated position.
491 	 */
492 	++tree->avl_numnodes;
493 
494 	node->avl_child[0] = NULL;
495 	node->avl_child[1] = NULL;
496 
497 	AVL_SETCHILD(node, which_child);
498 	AVL_SETBALANCE(node, 0);
499 	AVL_SETPARENT(node, parent);
500 	if (parent != NULL) {
501 		ASSERT(parent->avl_child[which_child] == NULL);
502 		parent->avl_child[which_child] = node;
503 	} else {
504 		ASSERT(tree->avl_root == NULL);
505 		tree->avl_root = node;
506 	}
507 	/*
508 	 * Now, back up the tree modifying the balance of all nodes above the
509 	 * insertion point. If we get to a highly unbalanced ancestor, we
510 	 * need to do a rotation.  If we back out of the tree we are done.
511 	 * If we brought any subtree into perfect balance (0), we are also done.
512 	 */
513 	for (;;) {
514 		node = parent;
515 		if (node == NULL)
516 			return;
517 
518 		/*
519 		 * Compute the new balance
520 		 */
521 		old_balance = AVL_XBALANCE(node);
522 		new_balance = old_balance + avl_child2balance[which_child];
523 
524 		/*
525 		 * If we introduced equal balance, then we are done immediately
526 		 */
527 		if (new_balance == 0) {
528 			AVL_SETBALANCE(node, 0);
529 			return;
530 		}
531 
532 		/*
533 		 * If both old and new are not zero we went
534 		 * from -1 to -2 balance, do a rotation.
535 		 */
536 		if (old_balance != 0)
537 			break;
538 
539 		AVL_SETBALANCE(node, new_balance);
540 		parent = AVL_XPARENT(node);
541 		which_child = AVL_XCHILD(node);
542 	}
543 
544 	/*
545 	 * perform a rotation to fix the tree and return
546 	 */
547 	(void) avl_rotation(tree, node, new_balance);
548 }
549 
550 /*
551  * Insert "new_data" in "tree" in the given "direction" either after or
552  * before (AVL_AFTER, AVL_BEFORE) the data "here".
553  *
554  * Insertions can only be done at empty leaf points in the tree, therefore
555  * if the given child of the node is already present we move to either
556  * the AVL_PREV or AVL_NEXT and reverse the insertion direction. Since
557  * every other node in the tree is a leaf, this always works.
558  *
559  * To help developers using this interface, we assert that the new node
560  * is correctly ordered at every step of the way in DEBUG kernels.
561  */
562 void
563 avl_insert_here(
564 	avl_tree_t *tree,
565 	void *new_data,
566 	void *here,
567 	int direction)
568 {
569 	avl_node_t *node;
570 	int child = direction;	/* rely on AVL_BEFORE == 0, AVL_AFTER == 1 */
571 #ifdef DEBUG
572 	int diff;
573 #endif
574 
575 	ASSERT(tree != NULL);
576 	ASSERT(new_data != NULL);
577 	ASSERT(here != NULL);
578 	ASSERT(direction == AVL_BEFORE || direction == AVL_AFTER);
579 
580 	/*
581 	 * If corresponding child of node is not NULL, go to the neighboring
582 	 * node and reverse the insertion direction.
583 	 */
584 	node = AVL_DATA2NODE(here, tree->avl_offset);
585 
586 #ifdef DEBUG
587 	diff = tree->avl_compar(new_data, here);
588 	ASSERT(-1 <= diff && diff <= 1);
589 	ASSERT(diff != 0);
590 	ASSERT(diff > 0 ? child == 1 : child == 0);
591 #endif
592 
593 	if (node->avl_child[child] != NULL) {
594 		node = node->avl_child[child];
595 		child = 1 - child;
596 		while (node->avl_child[child] != NULL) {
597 #ifdef DEBUG
598 			diff = tree->avl_compar(new_data,
599 			    AVL_NODE2DATA(node, tree->avl_offset));
600 			ASSERT(-1 <= diff && diff <= 1);
601 			ASSERT(diff != 0);
602 			ASSERT(diff > 0 ? child == 1 : child == 0);
603 #endif
604 			node = node->avl_child[child];
605 		}
606 #ifdef DEBUG
607 		diff = tree->avl_compar(new_data,
608 		    AVL_NODE2DATA(node, tree->avl_offset));
609 		ASSERT(-1 <= diff && diff <= 1);
610 		ASSERT(diff != 0);
611 		ASSERT(diff > 0 ? child == 1 : child == 0);
612 #endif
613 	}
614 	ASSERT(node->avl_child[child] == NULL);
615 
616 	avl_insert(tree, new_data, AVL_MKINDEX(node, child));
617 }
618 
619 /*
620  * Add a new node to an AVL tree.
621  */
622 void
623 avl_add(avl_tree_t *tree, void *new_node)
624 {
625 	avl_index_t where;
626 
627 	/*
628 	 * This is unfortunate.  We want to call panic() here, even for
629 	 * non-DEBUG kernels.  In userland, however, we can't depend on anything
630 	 * in libc or else the rtld build process gets confused.  So, all we can
631 	 * do in userland is resort to a normal ASSERT().
632 	 */
633 	if (avl_find(tree, new_node, &where) != NULL)
634 #ifdef _KERNEL
635 		panic("avl_find() succeeded inside avl_add()");
636 #else
637 		ASSERT(0);
638 #endif
639 	avl_insert(tree, new_node, where);
640 }
641 
642 /*
643  * Delete a node from the AVL tree.  Deletion is similar to insertion, but
644  * with 2 complications.
645  *
646  * First, we may be deleting an interior node. Consider the following subtree:
647  *
648  *     d           c            c
649  *    / \         / \          / \
650  *   b   e       b   e        b   e
651  *  / \	        / \          /
652  * a   c       a            a
653  *
654  * When we are deleting node (d), we find and bring up an adjacent valued leaf
655  * node, say (c), to take the interior node's place. In the code this is
656  * handled by temporarily swapping (d) and (c) in the tree and then using
657  * common code to delete (d) from the leaf position.
658  *
659  * Secondly, an interior deletion from a deep tree may require more than one
660  * rotation to fix the balance. This is handled by moving up the tree through
661  * parents and applying rotations as needed. The return value from
662  * avl_rotation() is used to detect when a subtree did not change overall
663  * height due to a rotation.
664  */
665 void
666 avl_remove(avl_tree_t *tree, void *data)
667 {
668 	avl_node_t *delete;
669 	avl_node_t *parent;
670 	avl_node_t *node;
671 	avl_node_t tmp;
672 	int old_balance;
673 	int new_balance;
674 	int left;
675 	int right;
676 	int which_child;
677 	size_t off = tree->avl_offset;
678 
679 	ASSERT(tree);
680 
681 	delete = AVL_DATA2NODE(data, off);
682 
683 	/*
684 	 * Deletion is easiest with a node that has at most 1 child.
685 	 * We swap a node with 2 children with a sequentially valued
686 	 * neighbor node. That node will have at most 1 child. Note this
687 	 * has no effect on the ordering of the remaining nodes.
688 	 *
689 	 * As an optimization, we choose the greater neighbor if the tree
690 	 * is right heavy, otherwise the left neighbor. This reduces the
691 	 * number of rotations needed.
692 	 */
693 	if (delete->avl_child[0] != NULL && delete->avl_child[1] != NULL) {
694 
695 		/*
696 		 * choose node to swap from whichever side is taller
697 		 */
698 		old_balance = AVL_XBALANCE(delete);
699 		left = avl_balance2child[old_balance + 1];
700 		right = 1 - left;
701 
702 		/*
703 		 * get to the previous value'd node
704 		 * (down 1 left, as far as possible right)
705 		 */
706 		for (node = delete->avl_child[left];
707 		    node->avl_child[right] != NULL;
708 		    node = node->avl_child[right])
709 			;
710 
711 		/*
712 		 * create a temp placeholder for 'node'
713 		 * move 'node' to delete's spot in the tree
714 		 */
715 		tmp = *node;
716 
717 		*node = *delete;
718 		if (node->avl_child[left] == node)
719 			node->avl_child[left] = &tmp;
720 
721 		parent = AVL_XPARENT(node);
722 		if (parent != NULL)
723 			parent->avl_child[AVL_XCHILD(node)] = node;
724 		else
725 			tree->avl_root = node;
726 		AVL_SETPARENT(node->avl_child[left], node);
727 		AVL_SETPARENT(node->avl_child[right], node);
728 
729 		/*
730 		 * Put tmp where node used to be (just temporary).
731 		 * It always has a parent and at most 1 child.
732 		 */
733 		delete = &tmp;
734 		parent = AVL_XPARENT(delete);
735 		parent->avl_child[AVL_XCHILD(delete)] = delete;
736 		which_child = (delete->avl_child[1] != 0);
737 		if (delete->avl_child[which_child] != NULL)
738 			AVL_SETPARENT(delete->avl_child[which_child], delete);
739 	}
740 
741 
742 	/*
743 	 * Here we know "delete" is at least partially a leaf node. It can
744 	 * be easily removed from the tree.
745 	 */
746 	ASSERT(tree->avl_numnodes > 0);
747 	--tree->avl_numnodes;
748 	parent = AVL_XPARENT(delete);
749 	which_child = AVL_XCHILD(delete);
750 	if (delete->avl_child[0] != NULL)
751 		node = delete->avl_child[0];
752 	else
753 		node = delete->avl_child[1];
754 
755 	/*
756 	 * Connect parent directly to node (leaving out delete).
757 	 */
758 	if (node != NULL) {
759 		AVL_SETPARENT(node, parent);
760 		AVL_SETCHILD(node, which_child);
761 	}
762 	if (parent == NULL) {
763 		tree->avl_root = node;
764 		return;
765 	}
766 	parent->avl_child[which_child] = node;
767 
768 
769 	/*
770 	 * Since the subtree is now shorter, begin adjusting parent balances
771 	 * and performing any needed rotations.
772 	 */
773 	do {
774 
775 		/*
776 		 * Move up the tree and adjust the balance
777 		 *
778 		 * Capture the parent and which_child values for the next
779 		 * iteration before any rotations occur.
780 		 */
781 		node = parent;
782 		old_balance = AVL_XBALANCE(node);
783 		new_balance = old_balance - avl_child2balance[which_child];
784 		parent = AVL_XPARENT(node);
785 		which_child = AVL_XCHILD(node);
786 
787 		/*
788 		 * If a node was in perfect balance but isn't anymore then
789 		 * we can stop, since the height didn't change above this point
790 		 * due to a deletion.
791 		 */
792 		if (old_balance == 0) {
793 			AVL_SETBALANCE(node, new_balance);
794 			break;
795 		}
796 
797 		/*
798 		 * If the new balance is zero, we don't need to rotate
799 		 * else
800 		 * need a rotation to fix the balance.
801 		 * If the rotation doesn't change the height
802 		 * of the sub-tree we have finished adjusting.
803 		 */
804 		if (new_balance == 0)
805 			AVL_SETBALANCE(node, new_balance);
806 		else if (!avl_rotation(tree, node, new_balance))
807 			break;
808 	} while (parent != NULL);
809 }
810 
811 #define	AVL_REINSERT(tree, obj)		\
812 	avl_remove((tree), (obj));	\
813 	avl_add((tree), (obj))
814 
815 boolean_t
816 avl_update_lt(avl_tree_t *t, void *obj)
817 {
818 	void *neighbor;
819 
820 	ASSERT(((neighbor = AVL_NEXT(t, obj)) == NULL) ||
821 	    (t->avl_compar(obj, neighbor) <= 0));
822 
823 	neighbor = AVL_PREV(t, obj);
824 	if ((neighbor != NULL) && (t->avl_compar(obj, neighbor) < 0)) {
825 		AVL_REINSERT(t, obj);
826 		return (B_TRUE);
827 	}
828 
829 	return (B_FALSE);
830 }
831 
832 boolean_t
833 avl_update_gt(avl_tree_t *t, void *obj)
834 {
835 	void *neighbor;
836 
837 	ASSERT(((neighbor = AVL_PREV(t, obj)) == NULL) ||
838 	    (t->avl_compar(obj, neighbor) >= 0));
839 
840 	neighbor = AVL_NEXT(t, obj);
841 	if ((neighbor != NULL) && (t->avl_compar(obj, neighbor) > 0)) {
842 		AVL_REINSERT(t, obj);
843 		return (B_TRUE);
844 	}
845 
846 	return (B_FALSE);
847 }
848 
849 boolean_t
850 avl_update(avl_tree_t *t, void *obj)
851 {
852 	void *neighbor;
853 
854 	neighbor = AVL_PREV(t, obj);
855 	if ((neighbor != NULL) && (t->avl_compar(obj, neighbor) < 0)) {
856 		AVL_REINSERT(t, obj);
857 		return (B_TRUE);
858 	}
859 
860 	neighbor = AVL_NEXT(t, obj);
861 	if ((neighbor != NULL) && (t->avl_compar(obj, neighbor) > 0)) {
862 		AVL_REINSERT(t, obj);
863 		return (B_TRUE);
864 	}
865 
866 	return (B_FALSE);
867 }
868 
869 /*
870  * initialize a new AVL tree
871  */
872 void
873 avl_create(avl_tree_t *tree, int (*compar) (const void *, const void *),
874     size_t size, size_t offset)
875 {
876 	ASSERT(tree);
877 	ASSERT(compar);
878 	ASSERT(size > 0);
879 	ASSERT(size >= offset + sizeof (avl_node_t));
880 #ifdef _LP64
881 	ASSERT((offset & 0x7) == 0);
882 #endif
883 
884 	tree->avl_compar = compar;
885 	tree->avl_root = NULL;
886 	tree->avl_numnodes = 0;
887 	tree->avl_size = size;
888 	tree->avl_offset = offset;
889 }
890 
891 /*
892  * Delete a tree.
893  */
894 /* ARGSUSED */
895 void
896 avl_destroy(avl_tree_t *tree)
897 {
898 	ASSERT(tree);
899 	ASSERT(tree->avl_numnodes == 0);
900 	ASSERT(tree->avl_root == NULL);
901 }
902 
903 
904 /*
905  * Return the number of nodes in an AVL tree.
906  */
907 ulong_t
908 avl_numnodes(avl_tree_t *tree)
909 {
910 	ASSERT(tree);
911 	return (tree->avl_numnodes);
912 }
913 
914 boolean_t
915 avl_is_empty(avl_tree_t *tree)
916 {
917 	ASSERT(tree);
918 	return (tree->avl_numnodes == 0);
919 }
920 
921 #define	CHILDBIT	(1L)
922 
923 /*
924  * Post-order tree walk used to visit all tree nodes and destroy the tree
925  * in post order. This is used for destroying a tree w/o paying any cost
926  * for rebalancing it.
927  *
928  * example:
929  *
930  *	void *cookie = NULL;
931  *	my_data_t *node;
932  *
933  *	while ((node = avl_destroy_nodes(tree, &cookie)) != NULL)
934  *		free(node);
935  *	avl_destroy(tree);
936  *
937  * The cookie is really an avl_node_t to the current node's parent and
938  * an indication of which child you looked at last.
939  *
940  * On input, a cookie value of CHILDBIT indicates the tree is done.
941  */
942 void *
943 avl_destroy_nodes(avl_tree_t *tree, void **cookie)
944 {
945 	avl_node_t	*node;
946 	avl_node_t	*parent;
947 	int		child;
948 	void		*first;
949 	size_t		off = tree->avl_offset;
950 
951 	/*
952 	 * Initial calls go to the first node or it's right descendant.
953 	 */
954 	if (*cookie == NULL) {
955 		first = avl_first(tree);
956 
957 		/*
958 		 * deal with an empty tree
959 		 */
960 		if (first == NULL) {
961 			*cookie = (void *)CHILDBIT;
962 			return (NULL);
963 		}
964 
965 		node = AVL_DATA2NODE(first, off);
966 		parent = AVL_XPARENT(node);
967 		goto check_right_side;
968 	}
969 
970 	/*
971 	 * If there is no parent to return to we are done.
972 	 */
973 	parent = (avl_node_t *)((uintptr_t)(*cookie) & ~CHILDBIT);
974 	if (parent == NULL) {
975 		if (tree->avl_root != NULL) {
976 			ASSERT(tree->avl_numnodes == 1);
977 			tree->avl_root = NULL;
978 			tree->avl_numnodes = 0;
979 		}
980 		return (NULL);
981 	}
982 
983 	/*
984 	 * Remove the child pointer we just visited from the parent and tree.
985 	 */
986 	child = (uintptr_t)(*cookie) & CHILDBIT;
987 	parent->avl_child[child] = NULL;
988 	ASSERT(tree->avl_numnodes > 1);
989 	--tree->avl_numnodes;
990 
991 	/*
992 	 * If we just did a right child or there isn't one, go up to parent.
993 	 */
994 	if (child == 1 || parent->avl_child[1] == NULL) {
995 		node = parent;
996 		parent = AVL_XPARENT(parent);
997 		goto done;
998 	}
999 
1000 	/*
1001 	 * Do parent's right child, then leftmost descendent.
1002 	 */
1003 	node = parent->avl_child[1];
1004 	while (node->avl_child[0] != NULL) {
1005 		parent = node;
1006 		node = node->avl_child[0];
1007 	}
1008 
1009 	/*
1010 	 * If here, we moved to a left child. It may have one
1011 	 * child on the right (when balance == +1).
1012 	 */
1013 check_right_side:
1014 	if (node->avl_child[1] != NULL) {
1015 		ASSERT(AVL_XBALANCE(node) == 1);
1016 		parent = node;
1017 		node = node->avl_child[1];
1018 		ASSERT(node->avl_child[0] == NULL &&
1019 		    node->avl_child[1] == NULL);
1020 	} else {
1021 		ASSERT(AVL_XBALANCE(node) <= 0);
1022 	}
1023 
1024 done:
1025 	if (parent == NULL) {
1026 		*cookie = (void *)CHILDBIT;
1027 		ASSERT(node == tree->avl_root);
1028 	} else {
1029 		*cookie = (void *)((uintptr_t)parent | AVL_XCHILD(node));
1030 	}
1031 
1032 	return (AVL_NODE2DATA(node, off));
1033 }
1034