1 /*
2  * Copyright (C) 2011 Red Hat, Inc.
3  *
4  * This file is released under the GPL.
5  */
6 
7 #include "dm-btree-internal.h"
8 #include "dm-space-map.h"
9 #include "dm-transaction-manager.h"
10 
11 #include <linux/export.h>
12 #include <linux/device-mapper.h>
13 
14 #define DM_MSG_PREFIX "btree"
15 
16 /*----------------------------------------------------------------
17  * Array manipulation
18  *--------------------------------------------------------------*/
19 static void memcpy_disk(void *dest, const void *src, size_t len)
20 	__dm_written_to_disk(src)
21 {
22 	memcpy(dest, src, len);
23 	__dm_unbless_for_disk(src);
24 }
25 
26 static void array_insert(void *base, size_t elt_size, unsigned nr_elts,
27 			 unsigned index, void *elt)
28 	__dm_written_to_disk(elt)
29 {
30 	if (index < nr_elts)
31 		memmove(base + (elt_size * (index + 1)),
32 			base + (elt_size * index),
33 			(nr_elts - index) * elt_size);
34 
35 	memcpy_disk(base + (elt_size * index), elt, elt_size);
36 }
37 
38 /*----------------------------------------------------------------*/
39 
40 /* makes the assumption that no two keys are the same. */
41 static int bsearch(struct btree_node *n, uint64_t key, int want_hi)
42 {
43 	int lo = -1, hi = le32_to_cpu(n->header.nr_entries);
44 
45 	while (hi - lo > 1) {
46 		int mid = lo + ((hi - lo) / 2);
47 		uint64_t mid_key = le64_to_cpu(n->keys[mid]);
48 
49 		if (mid_key == key)
50 			return mid;
51 
52 		if (mid_key < key)
53 			lo = mid;
54 		else
55 			hi = mid;
56 	}
57 
58 	return want_hi ? hi : lo;
59 }
60 
61 int lower_bound(struct btree_node *n, uint64_t key)
62 {
63 	return bsearch(n, key, 0);
64 }
65 
66 static int upper_bound(struct btree_node *n, uint64_t key)
67 {
68 	return bsearch(n, key, 1);
69 }
70 
71 void inc_children(struct dm_transaction_manager *tm, struct btree_node *n,
72 		  struct dm_btree_value_type *vt)
73 {
74 	uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
75 
76 	if (le32_to_cpu(n->header.flags) & INTERNAL_NODE)
77 		dm_tm_with_runs(tm, value_ptr(n, 0), nr_entries, dm_tm_inc_range);
78 
79 	else if (vt->inc)
80 		vt->inc(vt->context, value_ptr(n, 0), nr_entries);
81 }
82 
83 static int insert_at(size_t value_size, struct btree_node *node, unsigned index,
84 		      uint64_t key, void *value)
85 		      __dm_written_to_disk(value)
86 {
87 	uint32_t nr_entries = le32_to_cpu(node->header.nr_entries);
88 	__le64 key_le = cpu_to_le64(key);
89 
90 	if (index > nr_entries ||
91 	    index >= le32_to_cpu(node->header.max_entries)) {
92 		DMERR("too many entries in btree node for insert");
93 		__dm_unbless_for_disk(value);
94 		return -ENOMEM;
95 	}
96 
97 	__dm_bless_for_disk(&key_le);
98 
99 	array_insert(node->keys, sizeof(*node->keys), nr_entries, index, &key_le);
100 	array_insert(value_base(node), value_size, nr_entries, index, value);
101 	node->header.nr_entries = cpu_to_le32(nr_entries + 1);
102 
103 	return 0;
104 }
105 
106 /*----------------------------------------------------------------*/
107 
108 /*
109  * We want 3n entries (for some n).  This works more nicely for repeated
110  * insert remove loops than (2n + 1).
111  */
112 static uint32_t calc_max_entries(size_t value_size, size_t block_size)
113 {
114 	uint32_t total, n;
115 	size_t elt_size = sizeof(uint64_t) + value_size; /* key + value */
116 
117 	block_size -= sizeof(struct node_header);
118 	total = block_size / elt_size;
119 	n = total / 3;		/* rounds down */
120 
121 	return 3 * n;
122 }
123 
124 int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root)
125 {
126 	int r;
127 	struct dm_block *b;
128 	struct btree_node *n;
129 	size_t block_size;
130 	uint32_t max_entries;
131 
132 	r = new_block(info, &b);
133 	if (r < 0)
134 		return r;
135 
136 	block_size = dm_bm_block_size(dm_tm_get_bm(info->tm));
137 	max_entries = calc_max_entries(info->value_type.size, block_size);
138 
139 	n = dm_block_data(b);
140 	memset(n, 0, block_size);
141 	n->header.flags = cpu_to_le32(LEAF_NODE);
142 	n->header.nr_entries = cpu_to_le32(0);
143 	n->header.max_entries = cpu_to_le32(max_entries);
144 	n->header.value_size = cpu_to_le32(info->value_type.size);
145 
146 	*root = dm_block_location(b);
147 	unlock_block(info, b);
148 
149 	return 0;
150 }
151 EXPORT_SYMBOL_GPL(dm_btree_empty);
152 
153 /*----------------------------------------------------------------*/
154 
155 /*
156  * Deletion uses a recursive algorithm, since we have limited stack space
157  * we explicitly manage our own stack on the heap.
158  */
159 #define MAX_SPINE_DEPTH 64
160 struct frame {
161 	struct dm_block *b;
162 	struct btree_node *n;
163 	unsigned level;
164 	unsigned nr_children;
165 	unsigned current_child;
166 };
167 
168 struct del_stack {
169 	struct dm_btree_info *info;
170 	struct dm_transaction_manager *tm;
171 	int top;
172 	struct frame spine[MAX_SPINE_DEPTH];
173 };
174 
175 static int top_frame(struct del_stack *s, struct frame **f)
176 {
177 	if (s->top < 0) {
178 		DMERR("btree deletion stack empty");
179 		return -EINVAL;
180 	}
181 
182 	*f = s->spine + s->top;
183 
184 	return 0;
185 }
186 
187 static int unprocessed_frames(struct del_stack *s)
188 {
189 	return s->top >= 0;
190 }
191 
192 static void prefetch_children(struct del_stack *s, struct frame *f)
193 {
194 	unsigned i;
195 	struct dm_block_manager *bm = dm_tm_get_bm(s->tm);
196 
197 	for (i = 0; i < f->nr_children; i++)
198 		dm_bm_prefetch(bm, value64(f->n, i));
199 }
200 
201 static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
202 {
203 	return f->level < (info->levels - 1);
204 }
205 
206 static int push_frame(struct del_stack *s, dm_block_t b, unsigned level)
207 {
208 	int r;
209 	uint32_t ref_count;
210 
211 	if (s->top >= MAX_SPINE_DEPTH - 1) {
212 		DMERR("btree deletion stack out of memory");
213 		return -ENOMEM;
214 	}
215 
216 	r = dm_tm_ref(s->tm, b, &ref_count);
217 	if (r)
218 		return r;
219 
220 	if (ref_count > 1)
221 		/*
222 		 * This is a shared node, so we can just decrement it's
223 		 * reference counter and leave the children.
224 		 */
225 		dm_tm_dec(s->tm, b);
226 
227 	else {
228 		uint32_t flags;
229 		struct frame *f = s->spine + ++s->top;
230 
231 		r = dm_tm_read_lock(s->tm, b, &btree_node_validator, &f->b);
232 		if (r) {
233 			s->top--;
234 			return r;
235 		}
236 
237 		f->n = dm_block_data(f->b);
238 		f->level = level;
239 		f->nr_children = le32_to_cpu(f->n->header.nr_entries);
240 		f->current_child = 0;
241 
242 		flags = le32_to_cpu(f->n->header.flags);
243 		if (flags & INTERNAL_NODE || is_internal_level(s->info, f))
244 			prefetch_children(s, f);
245 	}
246 
247 	return 0;
248 }
249 
250 static void pop_frame(struct del_stack *s)
251 {
252 	struct frame *f = s->spine + s->top--;
253 
254 	dm_tm_dec(s->tm, dm_block_location(f->b));
255 	dm_tm_unlock(s->tm, f->b);
256 }
257 
258 static void unlock_all_frames(struct del_stack *s)
259 {
260 	struct frame *f;
261 
262 	while (unprocessed_frames(s)) {
263 		f = s->spine + s->top--;
264 		dm_tm_unlock(s->tm, f->b);
265 	}
266 }
267 
268 int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
269 {
270 	int r;
271 	struct del_stack *s;
272 
273 	/*
274 	 * dm_btree_del() is called via an ioctl, as such should be
275 	 * considered an FS op.  We can't recurse back into the FS, so we
276 	 * allocate GFP_NOFS.
277 	 */
278 	s = kmalloc(sizeof(*s), GFP_NOFS);
279 	if (!s)
280 		return -ENOMEM;
281 	s->info = info;
282 	s->tm = info->tm;
283 	s->top = -1;
284 
285 	r = push_frame(s, root, 0);
286 	if (r)
287 		goto out;
288 
289 	while (unprocessed_frames(s)) {
290 		uint32_t flags;
291 		struct frame *f;
292 		dm_block_t b;
293 
294 		r = top_frame(s, &f);
295 		if (r)
296 			goto out;
297 
298 		if (f->current_child >= f->nr_children) {
299 			pop_frame(s);
300 			continue;
301 		}
302 
303 		flags = le32_to_cpu(f->n->header.flags);
304 		if (flags & INTERNAL_NODE) {
305 			b = value64(f->n, f->current_child);
306 			f->current_child++;
307 			r = push_frame(s, b, f->level);
308 			if (r)
309 				goto out;
310 
311 		} else if (is_internal_level(info, f)) {
312 			b = value64(f->n, f->current_child);
313 			f->current_child++;
314 			r = push_frame(s, b, f->level + 1);
315 			if (r)
316 				goto out;
317 
318 		} else {
319 			if (info->value_type.dec)
320 				info->value_type.dec(info->value_type.context,
321 						     value_ptr(f->n, 0), f->nr_children);
322 			pop_frame(s);
323 		}
324 	}
325 out:
326 	if (r) {
327 		/* cleanup all frames of del_stack */
328 		unlock_all_frames(s);
329 	}
330 	kfree(s);
331 
332 	return r;
333 }
334 EXPORT_SYMBOL_GPL(dm_btree_del);
335 
336 /*----------------------------------------------------------------*/
337 
338 static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
339 			    int (*search_fn)(struct btree_node *, uint64_t),
340 			    uint64_t *result_key, void *v, size_t value_size)
341 {
342 	int i, r;
343 	uint32_t flags, nr_entries;
344 
345 	do {
346 		r = ro_step(s, block);
347 		if (r < 0)
348 			return r;
349 
350 		i = search_fn(ro_node(s), key);
351 
352 		flags = le32_to_cpu(ro_node(s)->header.flags);
353 		nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
354 		if (i < 0 || i >= nr_entries)
355 			return -ENODATA;
356 
357 		if (flags & INTERNAL_NODE)
358 			block = value64(ro_node(s), i);
359 
360 	} while (!(flags & LEAF_NODE));
361 
362 	*result_key = le64_to_cpu(ro_node(s)->keys[i]);
363 	if (v)
364 		memcpy(v, value_ptr(ro_node(s), i), value_size);
365 
366 	return 0;
367 }
368 
369 int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
370 		    uint64_t *keys, void *value_le)
371 {
372 	unsigned level, last_level = info->levels - 1;
373 	int r = -ENODATA;
374 	uint64_t rkey;
375 	__le64 internal_value_le;
376 	struct ro_spine spine;
377 
378 	init_ro_spine(&spine, info);
379 	for (level = 0; level < info->levels; level++) {
380 		size_t size;
381 		void *value_p;
382 
383 		if (level == last_level) {
384 			value_p = value_le;
385 			size = info->value_type.size;
386 
387 		} else {
388 			value_p = &internal_value_le;
389 			size = sizeof(uint64_t);
390 		}
391 
392 		r = btree_lookup_raw(&spine, root, keys[level],
393 				     lower_bound, &rkey,
394 				     value_p, size);
395 
396 		if (!r) {
397 			if (rkey != keys[level]) {
398 				exit_ro_spine(&spine);
399 				return -ENODATA;
400 			}
401 		} else {
402 			exit_ro_spine(&spine);
403 			return r;
404 		}
405 
406 		root = le64_to_cpu(internal_value_le);
407 	}
408 	exit_ro_spine(&spine);
409 
410 	return r;
411 }
412 EXPORT_SYMBOL_GPL(dm_btree_lookup);
413 
414 static int dm_btree_lookup_next_single(struct dm_btree_info *info, dm_block_t root,
415 				       uint64_t key, uint64_t *rkey, void *value_le)
416 {
417 	int r, i;
418 	uint32_t flags, nr_entries;
419 	struct dm_block *node;
420 	struct btree_node *n;
421 
422 	r = bn_read_lock(info, root, &node);
423 	if (r)
424 		return r;
425 
426 	n = dm_block_data(node);
427 	flags = le32_to_cpu(n->header.flags);
428 	nr_entries = le32_to_cpu(n->header.nr_entries);
429 
430 	if (flags & INTERNAL_NODE) {
431 		i = lower_bound(n, key);
432 		if (i < 0) {
433 			/*
434 			 * avoid early -ENODATA return when all entries are
435 			 * higher than the search @key.
436 			 */
437 			i = 0;
438 		}
439 		if (i >= nr_entries) {
440 			r = -ENODATA;
441 			goto out;
442 		}
443 
444 		r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
445 		if (r == -ENODATA && i < (nr_entries - 1)) {
446 			i++;
447 			r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
448 		}
449 
450 	} else {
451 		i = upper_bound(n, key);
452 		if (i < 0 || i >= nr_entries) {
453 			r = -ENODATA;
454 			goto out;
455 		}
456 
457 		*rkey = le64_to_cpu(n->keys[i]);
458 		memcpy(value_le, value_ptr(n, i), info->value_type.size);
459 	}
460 out:
461 	dm_tm_unlock(info->tm, node);
462 	return r;
463 }
464 
465 int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
466 			 uint64_t *keys, uint64_t *rkey, void *value_le)
467 {
468 	unsigned level;
469 	int r = -ENODATA;
470 	__le64 internal_value_le;
471 	struct ro_spine spine;
472 
473 	init_ro_spine(&spine, info);
474 	for (level = 0; level < info->levels - 1u; level++) {
475 		r = btree_lookup_raw(&spine, root, keys[level],
476 				     lower_bound, rkey,
477 				     &internal_value_le, sizeof(uint64_t));
478 		if (r)
479 			goto out;
480 
481 		if (*rkey != keys[level]) {
482 			r = -ENODATA;
483 			goto out;
484 		}
485 
486 		root = le64_to_cpu(internal_value_le);
487 	}
488 
489 	r = dm_btree_lookup_next_single(info, root, keys[level], rkey, value_le);
490 out:
491 	exit_ro_spine(&spine);
492 	return r;
493 }
494 
495 EXPORT_SYMBOL_GPL(dm_btree_lookup_next);
496 
497 /*----------------------------------------------------------------*/
498 
499 /*
500  * Copies entries from one region of a btree node to another.  The regions
501  * must not overlap.
502  */
503 static void copy_entries(struct btree_node *dest, unsigned dest_offset,
504 			 struct btree_node *src, unsigned src_offset,
505 			 unsigned count)
506 {
507 	size_t value_size = le32_to_cpu(dest->header.value_size);
508 	memcpy(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
509 	memcpy(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
510 }
511 
512 /*
513  * Moves entries from one region fo a btree node to another.  The regions
514  * may overlap.
515  */
516 static void move_entries(struct btree_node *dest, unsigned dest_offset,
517 			 struct btree_node *src, unsigned src_offset,
518 			 unsigned count)
519 {
520 	size_t value_size = le32_to_cpu(dest->header.value_size);
521 	memmove(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
522 	memmove(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
523 }
524 
525 /*
526  * Erases the first 'count' entries of a btree node, shifting following
527  * entries down into their place.
528  */
529 static void shift_down(struct btree_node *n, unsigned count)
530 {
531 	move_entries(n, 0, n, count, le32_to_cpu(n->header.nr_entries) - count);
532 }
533 
534 /*
535  * Moves entries in a btree node up 'count' places, making space for
536  * new entries at the start of the node.
537  */
538 static void shift_up(struct btree_node *n, unsigned count)
539 {
540 	move_entries(n, count, n, 0, le32_to_cpu(n->header.nr_entries));
541 }
542 
543 /*
544  * Redistributes entries between two btree nodes to make them
545  * have similar numbers of entries.
546  */
547 static void redistribute2(struct btree_node *left, struct btree_node *right)
548 {
549 	unsigned nr_left = le32_to_cpu(left->header.nr_entries);
550 	unsigned nr_right = le32_to_cpu(right->header.nr_entries);
551 	unsigned total = nr_left + nr_right;
552 	unsigned target_left = total / 2;
553 	unsigned target_right = total - target_left;
554 
555 	if (nr_left < target_left) {
556 		unsigned delta = target_left - nr_left;
557 		copy_entries(left, nr_left, right, 0, delta);
558 		shift_down(right, delta);
559 	} else if (nr_left > target_left) {
560 		unsigned delta = nr_left - target_left;
561 		if (nr_right)
562 			shift_up(right, delta);
563 		copy_entries(right, 0, left, target_left, delta);
564 	}
565 
566 	left->header.nr_entries = cpu_to_le32(target_left);
567 	right->header.nr_entries = cpu_to_le32(target_right);
568 }
569 
570 /*
571  * Redistribute entries between three nodes.  Assumes the central
572  * node is empty.
573  */
574 static void redistribute3(struct btree_node *left, struct btree_node *center,
575 			  struct btree_node *right)
576 {
577 	unsigned nr_left = le32_to_cpu(left->header.nr_entries);
578 	unsigned nr_center = le32_to_cpu(center->header.nr_entries);
579 	unsigned nr_right = le32_to_cpu(right->header.nr_entries);
580 	unsigned total, target_left, target_center, target_right;
581 
582 	BUG_ON(nr_center);
583 
584 	total = nr_left + nr_right;
585 	target_left = total / 3;
586 	target_center = (total - target_left) / 2;
587 	target_right = (total - target_left - target_center);
588 
589 	if (nr_left < target_left) {
590 		unsigned left_short = target_left - nr_left;
591 		copy_entries(left, nr_left, right, 0, left_short);
592 		copy_entries(center, 0, right, left_short, target_center);
593 		shift_down(right, nr_right - target_right);
594 
595 	} else if (nr_left < (target_left + target_center)) {
596 		unsigned left_to_center = nr_left - target_left;
597 		copy_entries(center, 0, left, target_left, left_to_center);
598 		copy_entries(center, left_to_center, right, 0, target_center - left_to_center);
599 		shift_down(right, nr_right - target_right);
600 
601 	} else {
602 		unsigned right_short = target_right - nr_right;
603 		shift_up(right, right_short);
604 		copy_entries(right, 0, left, nr_left - right_short, right_short);
605 		copy_entries(center, 0, left, target_left, nr_left - target_left);
606 	}
607 
608 	left->header.nr_entries = cpu_to_le32(target_left);
609 	center->header.nr_entries = cpu_to_le32(target_center);
610 	right->header.nr_entries = cpu_to_le32(target_right);
611 }
612 
613 /*
614  * Splits a node by creating a sibling node and shifting half the nodes
615  * contents across.  Assumes there is a parent node, and it has room for
616  * another child.
617  *
618  * Before:
619  *	  +--------+
620  *	  | Parent |
621  *	  +--------+
622  *	     |
623  *	     v
624  *	+----------+
625  *	| A ++++++ |
626  *	+----------+
627  *
628  *
629  * After:
630  *		+--------+
631  *		| Parent |
632  *		+--------+
633  *		  |	|
634  *		  v	+------+
635  *	    +---------+	       |
636  *	    | A* +++  |	       v
637  *	    +---------+	  +-------+
638  *			  | B +++ |
639  *			  +-------+
640  *
641  * Where A* is a shadow of A.
642  */
643 static int split_one_into_two(struct shadow_spine *s, unsigned parent_index,
644 			      struct dm_btree_value_type *vt, uint64_t key)
645 {
646 	int r;
647 	struct dm_block *left, *right, *parent;
648 	struct btree_node *ln, *rn, *pn;
649 	__le64 location;
650 
651 	left = shadow_current(s);
652 
653 	r = new_block(s->info, &right);
654 	if (r < 0)
655 		return r;
656 
657 	ln = dm_block_data(left);
658 	rn = dm_block_data(right);
659 
660 	rn->header.flags = ln->header.flags;
661 	rn->header.nr_entries = cpu_to_le32(0);
662 	rn->header.max_entries = ln->header.max_entries;
663 	rn->header.value_size = ln->header.value_size;
664 	redistribute2(ln, rn);
665 
666 	/* patch up the parent */
667 	parent = shadow_parent(s);
668 	pn = dm_block_data(parent);
669 
670 	location = cpu_to_le64(dm_block_location(right));
671 	__dm_bless_for_disk(&location);
672 	r = insert_at(sizeof(__le64), pn, parent_index + 1,
673 		      le64_to_cpu(rn->keys[0]), &location);
674 	if (r) {
675 		unlock_block(s->info, right);
676 		return r;
677 	}
678 
679 	/* patch up the spine */
680 	if (key < le64_to_cpu(rn->keys[0])) {
681 		unlock_block(s->info, right);
682 		s->nodes[1] = left;
683 	} else {
684 		unlock_block(s->info, left);
685 		s->nodes[1] = right;
686 	}
687 
688 	return 0;
689 }
690 
691 /*
692  * We often need to modify a sibling node.  This function shadows a particular
693  * child of the given parent node.  Making sure to update the parent to point
694  * to the new shadow.
695  */
696 static int shadow_child(struct dm_btree_info *info, struct dm_btree_value_type *vt,
697 			struct btree_node *parent, unsigned index,
698 			struct dm_block **result)
699 {
700 	int r, inc;
701 	dm_block_t root;
702 	struct btree_node *node;
703 
704 	root = value64(parent, index);
705 
706 	r = dm_tm_shadow_block(info->tm, root, &btree_node_validator,
707 			       result, &inc);
708 	if (r)
709 		return r;
710 
711 	node = dm_block_data(*result);
712 
713 	if (inc)
714 		inc_children(info->tm, node, vt);
715 
716 	*((__le64 *) value_ptr(parent, index)) =
717 		cpu_to_le64(dm_block_location(*result));
718 
719 	return 0;
720 }
721 
722 /*
723  * Splits two nodes into three.  This is more work, but results in fuller
724  * nodes, so saves metadata space.
725  */
726 static int split_two_into_three(struct shadow_spine *s, unsigned parent_index,
727                                 struct dm_btree_value_type *vt, uint64_t key)
728 {
729 	int r;
730 	unsigned middle_index;
731 	struct dm_block *left, *middle, *right, *parent;
732 	struct btree_node *ln, *rn, *mn, *pn;
733 	__le64 location;
734 
735 	parent = shadow_parent(s);
736 	pn = dm_block_data(parent);
737 
738 	if (parent_index == 0) {
739 		middle_index = 1;
740 		left = shadow_current(s);
741 		r = shadow_child(s->info, vt, pn, parent_index + 1, &right);
742 		if (r)
743 			return r;
744 	} else {
745 		middle_index = parent_index;
746 		right = shadow_current(s);
747 		r = shadow_child(s->info, vt, pn, parent_index - 1, &left);
748 		if (r)
749 			return r;
750 	}
751 
752 	r = new_block(s->info, &middle);
753 	if (r < 0)
754 		return r;
755 
756 	ln = dm_block_data(left);
757 	mn = dm_block_data(middle);
758 	rn = dm_block_data(right);
759 
760 	mn->header.nr_entries = cpu_to_le32(0);
761 	mn->header.flags = ln->header.flags;
762 	mn->header.max_entries = ln->header.max_entries;
763 	mn->header.value_size = ln->header.value_size;
764 
765 	redistribute3(ln, mn, rn);
766 
767 	/* patch up the parent */
768 	pn->keys[middle_index] = rn->keys[0];
769 	location = cpu_to_le64(dm_block_location(middle));
770 	__dm_bless_for_disk(&location);
771 	r = insert_at(sizeof(__le64), pn, middle_index,
772 		      le64_to_cpu(mn->keys[0]), &location);
773 	if (r) {
774 		if (shadow_current(s) != left)
775 			unlock_block(s->info, left);
776 
777 		unlock_block(s->info, middle);
778 
779 		if (shadow_current(s) != right)
780 			unlock_block(s->info, right);
781 
782 	        return r;
783 	}
784 
785 
786 	/* patch up the spine */
787 	if (key < le64_to_cpu(mn->keys[0])) {
788 		unlock_block(s->info, middle);
789 		unlock_block(s->info, right);
790 		s->nodes[1] = left;
791 	} else if (key < le64_to_cpu(rn->keys[0])) {
792 		unlock_block(s->info, left);
793 		unlock_block(s->info, right);
794 		s->nodes[1] = middle;
795 	} else {
796 		unlock_block(s->info, left);
797 		unlock_block(s->info, middle);
798 		s->nodes[1] = right;
799 	}
800 
801 	return 0;
802 }
803 
804 /*----------------------------------------------------------------*/
805 
806 /*
807  * Splits a node by creating two new children beneath the given node.
808  *
809  * Before:
810  *	  +----------+
811  *	  | A ++++++ |
812  *	  +----------+
813  *
814  *
815  * After:
816  *	+------------+
817  *	| A (shadow) |
818  *	+------------+
819  *	    |	|
820  *   +------+	+----+
821  *   |		     |
822  *   v		     v
823  * +-------+	 +-------+
824  * | B +++ |	 | C +++ |
825  * +-------+	 +-------+
826  */
827 static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
828 {
829 	int r;
830 	size_t size;
831 	unsigned nr_left, nr_right;
832 	struct dm_block *left, *right, *new_parent;
833 	struct btree_node *pn, *ln, *rn;
834 	__le64 val;
835 
836 	new_parent = shadow_current(s);
837 
838 	pn = dm_block_data(new_parent);
839 	size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
840 		sizeof(__le64) : s->info->value_type.size;
841 
842 	/* create & init the left block */
843 	r = new_block(s->info, &left);
844 	if (r < 0)
845 		return r;
846 
847 	ln = dm_block_data(left);
848 	nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
849 
850 	ln->header.flags = pn->header.flags;
851 	ln->header.nr_entries = cpu_to_le32(nr_left);
852 	ln->header.max_entries = pn->header.max_entries;
853 	ln->header.value_size = pn->header.value_size;
854 	memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
855 	memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
856 
857 	/* create & init the right block */
858 	r = new_block(s->info, &right);
859 	if (r < 0) {
860 		unlock_block(s->info, left);
861 		return r;
862 	}
863 
864 	rn = dm_block_data(right);
865 	nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
866 
867 	rn->header.flags = pn->header.flags;
868 	rn->header.nr_entries = cpu_to_le32(nr_right);
869 	rn->header.max_entries = pn->header.max_entries;
870 	rn->header.value_size = pn->header.value_size;
871 	memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
872 	memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
873 	       nr_right * size);
874 
875 	/* new_parent should just point to l and r now */
876 	pn->header.flags = cpu_to_le32(INTERNAL_NODE);
877 	pn->header.nr_entries = cpu_to_le32(2);
878 	pn->header.max_entries = cpu_to_le32(
879 		calc_max_entries(sizeof(__le64),
880 				 dm_bm_block_size(
881 					 dm_tm_get_bm(s->info->tm))));
882 	pn->header.value_size = cpu_to_le32(sizeof(__le64));
883 
884 	val = cpu_to_le64(dm_block_location(left));
885 	__dm_bless_for_disk(&val);
886 	pn->keys[0] = ln->keys[0];
887 	memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
888 
889 	val = cpu_to_le64(dm_block_location(right));
890 	__dm_bless_for_disk(&val);
891 	pn->keys[1] = rn->keys[0];
892 	memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
893 
894 	unlock_block(s->info, left);
895 	unlock_block(s->info, right);
896 	return 0;
897 }
898 
899 /*----------------------------------------------------------------*/
900 
901 /*
902  * Redistributes a node's entries with its left sibling.
903  */
904 static int rebalance_left(struct shadow_spine *s, struct dm_btree_value_type *vt,
905 			  unsigned parent_index, uint64_t key)
906 {
907 	int r;
908 	struct dm_block *sib;
909 	struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
910 
911 	r = shadow_child(s->info, vt, parent, parent_index - 1, &sib);
912 	if (r)
913 		return r;
914 
915 	left = dm_block_data(sib);
916 	right = dm_block_data(shadow_current(s));
917 	redistribute2(left, right);
918 	*key_ptr(parent, parent_index) = right->keys[0];
919 
920 	if (key < le64_to_cpu(right->keys[0])) {
921 		unlock_block(s->info, s->nodes[1]);
922 		s->nodes[1] = sib;
923 	} else {
924 		unlock_block(s->info, sib);
925 	}
926 
927 	return 0;
928 }
929 
930 /*
931  * Redistributes a nodes entries with its right sibling.
932  */
933 static int rebalance_right(struct shadow_spine *s, struct dm_btree_value_type *vt,
934 			   unsigned parent_index, uint64_t key)
935 {
936 	int r;
937 	struct dm_block *sib;
938 	struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
939 
940 	r = shadow_child(s->info, vt, parent, parent_index + 1, &sib);
941 	if (r)
942 		return r;
943 
944 	left = dm_block_data(shadow_current(s));
945 	right = dm_block_data(sib);
946 	redistribute2(left, right);
947 	*key_ptr(parent, parent_index + 1) = right->keys[0];
948 
949 	if (key < le64_to_cpu(right->keys[0])) {
950 		unlock_block(s->info, sib);
951 	} else {
952 		unlock_block(s->info, s->nodes[1]);
953 		s->nodes[1] = sib;
954 	}
955 
956 	return 0;
957 }
958 
959 /*
960  * Returns the number of spare entries in a node.
961  */
962 static int get_node_free_space(struct dm_btree_info *info, dm_block_t b, unsigned *space)
963 {
964 	int r;
965 	unsigned nr_entries;
966 	struct dm_block *block;
967 	struct btree_node *node;
968 
969 	r = bn_read_lock(info, b, &block);
970 	if (r)
971 		return r;
972 
973 	node = dm_block_data(block);
974 	nr_entries = le32_to_cpu(node->header.nr_entries);
975 	*space = le32_to_cpu(node->header.max_entries) - nr_entries;
976 
977 	unlock_block(info, block);
978 	return 0;
979 }
980 
981 /*
982  * Make space in a node, either by moving some entries to a sibling,
983  * or creating a new sibling node.  SPACE_THRESHOLD defines the minimum
984  * number of free entries that must be in the sibling to make the move
985  * worth while.  If the siblings are shared (eg, part of a snapshot),
986  * then they are not touched, since this break sharing and so consume
987  * more space than we save.
988  */
989 #define SPACE_THRESHOLD 8
990 static int rebalance_or_split(struct shadow_spine *s, struct dm_btree_value_type *vt,
991 			      unsigned parent_index, uint64_t key)
992 {
993 	int r;
994 	struct btree_node *parent = dm_block_data(shadow_parent(s));
995 	unsigned nr_parent = le32_to_cpu(parent->header.nr_entries);
996 	unsigned free_space;
997 	int left_shared = 0, right_shared = 0;
998 
999 	/* Should we move entries to the left sibling? */
1000 	if (parent_index > 0) {
1001 		dm_block_t left_b = value64(parent, parent_index - 1);
1002 		r = dm_tm_block_is_shared(s->info->tm, left_b, &left_shared);
1003 		if (r)
1004 			return r;
1005 
1006 		if (!left_shared) {
1007 			r = get_node_free_space(s->info, left_b, &free_space);
1008 			if (r)
1009 				return r;
1010 
1011 			if (free_space >= SPACE_THRESHOLD)
1012 				return rebalance_left(s, vt, parent_index, key);
1013 		}
1014 	}
1015 
1016 	/* Should we move entries to the right sibling? */
1017 	if (parent_index < (nr_parent - 1)) {
1018 		dm_block_t right_b = value64(parent, parent_index + 1);
1019 		r = dm_tm_block_is_shared(s->info->tm, right_b, &right_shared);
1020 		if (r)
1021 			return r;
1022 
1023 		if (!right_shared) {
1024 			r = get_node_free_space(s->info, right_b, &free_space);
1025 			if (r)
1026 				return r;
1027 
1028 			if (free_space >= SPACE_THRESHOLD)
1029 				return rebalance_right(s, vt, parent_index, key);
1030 		}
1031 	}
1032 
1033 	/*
1034 	 * We need to split the node, normally we split two nodes
1035 	 * into three.	But when inserting a sequence that is either
1036 	 * monotonically increasing or decreasing it's better to split
1037 	 * a single node into two.
1038 	 */
1039 	if (left_shared || right_shared || (nr_parent <= 2) ||
1040 	    (parent_index == 0) || (parent_index + 1 == nr_parent)) {
1041 		return split_one_into_two(s, parent_index, vt, key);
1042 	} else {
1043 		return split_two_into_three(s, parent_index, vt, key);
1044 	}
1045 }
1046 
1047 /*
1048  * Does the node contain a particular key?
1049  */
1050 static bool contains_key(struct btree_node *node, uint64_t key)
1051 {
1052 	int i = lower_bound(node, key);
1053 
1054 	if (i >= 0 && le64_to_cpu(node->keys[i]) == key)
1055 		return true;
1056 
1057 	return false;
1058 }
1059 
1060 /*
1061  * In general we preemptively make sure there's a free entry in every
1062  * node on the spine when doing an insert.  But we can avoid that with
1063  * leaf nodes if we know it's an overwrite.
1064  */
1065 static bool has_space_for_insert(struct btree_node *node, uint64_t key)
1066 {
1067 	if (node->header.nr_entries == node->header.max_entries) {
1068 		if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1069 			/* we don't need space if it's an overwrite */
1070 			return contains_key(node, key);
1071 		}
1072 
1073 		return false;
1074 	}
1075 
1076 	return true;
1077 }
1078 
1079 static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
1080 			    struct dm_btree_value_type *vt,
1081 			    uint64_t key, unsigned *index)
1082 {
1083 	int r, i = *index, top = 1;
1084 	struct btree_node *node;
1085 
1086 	for (;;) {
1087 		r = shadow_step(s, root, vt);
1088 		if (r < 0)
1089 			return r;
1090 
1091 		node = dm_block_data(shadow_current(s));
1092 
1093 		/*
1094 		 * We have to patch up the parent node, ugly, but I don't
1095 		 * see a way to do this automatically as part of the spine
1096 		 * op.
1097 		 */
1098 		if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
1099 			__le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1100 
1101 			__dm_bless_for_disk(&location);
1102 			memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1103 				    &location, sizeof(__le64));
1104 		}
1105 
1106 		node = dm_block_data(shadow_current(s));
1107 
1108 		if (!has_space_for_insert(node, key)) {
1109 			if (top)
1110 				r = btree_split_beneath(s, key);
1111 			else
1112 				r = rebalance_or_split(s, vt, i, key);
1113 
1114 			if (r < 0)
1115 				return r;
1116 
1117 			/* making space can cause the current node to change */
1118 			node = dm_block_data(shadow_current(s));
1119 		}
1120 
1121 		i = lower_bound(node, key);
1122 
1123 		if (le32_to_cpu(node->header.flags) & LEAF_NODE)
1124 			break;
1125 
1126 		if (i < 0) {
1127 			/* change the bounds on the lowest key */
1128 			node->keys[0] = cpu_to_le64(key);
1129 			i = 0;
1130 		}
1131 
1132 		root = value64(node, i);
1133 		top = 0;
1134 	}
1135 
1136 	if (i < 0 || le64_to_cpu(node->keys[i]) != key)
1137 		i++;
1138 
1139 	*index = i;
1140 	return 0;
1141 }
1142 
1143 static int __btree_get_overwrite_leaf(struct shadow_spine *s, dm_block_t root,
1144 				      uint64_t key, int *index)
1145 {
1146 	int r, i = -1;
1147 	struct btree_node *node;
1148 
1149 	*index = 0;
1150 	for (;;) {
1151 		r = shadow_step(s, root, &s->info->value_type);
1152 		if (r < 0)
1153 			return r;
1154 
1155 		node = dm_block_data(shadow_current(s));
1156 
1157 		/*
1158 		 * We have to patch up the parent node, ugly, but I don't
1159 		 * see a way to do this automatically as part of the spine
1160 		 * op.
1161 		 */
1162 		if (shadow_has_parent(s) && i >= 0) {
1163 			__le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1164 
1165 			__dm_bless_for_disk(&location);
1166 			memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1167 				    &location, sizeof(__le64));
1168 		}
1169 
1170 		node = dm_block_data(shadow_current(s));
1171 		i = lower_bound(node, key);
1172 
1173 		BUG_ON(i < 0);
1174 		BUG_ON(i >= le32_to_cpu(node->header.nr_entries));
1175 
1176 		if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1177 			if (key != le64_to_cpu(node->keys[i]))
1178 				return -EINVAL;
1179 			break;
1180 		}
1181 
1182 		root = value64(node, i);
1183 	}
1184 
1185 	*index = i;
1186 	return 0;
1187 }
1188 
1189 int btree_get_overwrite_leaf(struct dm_btree_info *info, dm_block_t root,
1190 			     uint64_t key, int *index,
1191 			     dm_block_t *new_root, struct dm_block **leaf)
1192 {
1193 	int r;
1194 	struct shadow_spine spine;
1195 
1196 	BUG_ON(info->levels > 1);
1197 	init_shadow_spine(&spine, info);
1198 	r = __btree_get_overwrite_leaf(&spine, root, key, index);
1199 	if (!r) {
1200 		*new_root = shadow_root(&spine);
1201 		*leaf = shadow_current(&spine);
1202 
1203 		/*
1204 		 * Decrement the count so exit_shadow_spine() doesn't
1205 		 * unlock the leaf.
1206 		 */
1207 		spine.count--;
1208 	}
1209 	exit_shadow_spine(&spine);
1210 
1211 	return r;
1212 }
1213 
1214 static bool need_insert(struct btree_node *node, uint64_t *keys,
1215 			unsigned level, unsigned index)
1216 {
1217         return ((index >= le32_to_cpu(node->header.nr_entries)) ||
1218 		(le64_to_cpu(node->keys[index]) != keys[level]));
1219 }
1220 
1221 static int insert(struct dm_btree_info *info, dm_block_t root,
1222 		  uint64_t *keys, void *value, dm_block_t *new_root,
1223 		  int *inserted)
1224 		  __dm_written_to_disk(value)
1225 {
1226 	int r;
1227 	unsigned level, index = -1, last_level = info->levels - 1;
1228 	dm_block_t block = root;
1229 	struct shadow_spine spine;
1230 	struct btree_node *n;
1231 	struct dm_btree_value_type le64_type;
1232 
1233 	init_le64_type(info->tm, &le64_type);
1234 	init_shadow_spine(&spine, info);
1235 
1236 	for (level = 0; level < (info->levels - 1); level++) {
1237 		r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
1238 		if (r < 0)
1239 			goto bad;
1240 
1241 		n = dm_block_data(shadow_current(&spine));
1242 
1243 		if (need_insert(n, keys, level, index)) {
1244 			dm_block_t new_tree;
1245 			__le64 new_le;
1246 
1247 			r = dm_btree_empty(info, &new_tree);
1248 			if (r < 0)
1249 				goto bad;
1250 
1251 			new_le = cpu_to_le64(new_tree);
1252 			__dm_bless_for_disk(&new_le);
1253 
1254 			r = insert_at(sizeof(uint64_t), n, index,
1255 				      keys[level], &new_le);
1256 			if (r)
1257 				goto bad;
1258 		}
1259 
1260 		if (level < last_level)
1261 			block = value64(n, index);
1262 	}
1263 
1264 	r = btree_insert_raw(&spine, block, &info->value_type,
1265 			     keys[level], &index);
1266 	if (r < 0)
1267 		goto bad;
1268 
1269 	n = dm_block_data(shadow_current(&spine));
1270 
1271 	if (need_insert(n, keys, level, index)) {
1272 		if (inserted)
1273 			*inserted = 1;
1274 
1275 		r = insert_at(info->value_type.size, n, index,
1276 			      keys[level], value);
1277 		if (r)
1278 			goto bad_unblessed;
1279 	} else {
1280 		if (inserted)
1281 			*inserted = 0;
1282 
1283 		if (info->value_type.dec &&
1284 		    (!info->value_type.equal ||
1285 		     !info->value_type.equal(
1286 			     info->value_type.context,
1287 			     value_ptr(n, index),
1288 			     value))) {
1289 			info->value_type.dec(info->value_type.context,
1290 					     value_ptr(n, index), 1);
1291 		}
1292 		memcpy_disk(value_ptr(n, index),
1293 			    value, info->value_type.size);
1294 	}
1295 
1296 	*new_root = shadow_root(&spine);
1297 	exit_shadow_spine(&spine);
1298 
1299 	return 0;
1300 
1301 bad:
1302 	__dm_unbless_for_disk(value);
1303 bad_unblessed:
1304 	exit_shadow_spine(&spine);
1305 	return r;
1306 }
1307 
1308 int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
1309 		    uint64_t *keys, void *value, dm_block_t *new_root)
1310 		    __dm_written_to_disk(value)
1311 {
1312 	return insert(info, root, keys, value, new_root, NULL);
1313 }
1314 EXPORT_SYMBOL_GPL(dm_btree_insert);
1315 
1316 int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
1317 			   uint64_t *keys, void *value, dm_block_t *new_root,
1318 			   int *inserted)
1319 			   __dm_written_to_disk(value)
1320 {
1321 	return insert(info, root, keys, value, new_root, inserted);
1322 }
1323 EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
1324 
1325 /*----------------------------------------------------------------*/
1326 
1327 static int find_key(struct ro_spine *s, dm_block_t block, bool find_highest,
1328 		    uint64_t *result_key, dm_block_t *next_block)
1329 {
1330 	int i, r;
1331 	uint32_t flags;
1332 
1333 	do {
1334 		r = ro_step(s, block);
1335 		if (r < 0)
1336 			return r;
1337 
1338 		flags = le32_to_cpu(ro_node(s)->header.flags);
1339 		i = le32_to_cpu(ro_node(s)->header.nr_entries);
1340 		if (!i)
1341 			return -ENODATA;
1342 		else
1343 			i--;
1344 
1345 		if (find_highest)
1346 			*result_key = le64_to_cpu(ro_node(s)->keys[i]);
1347 		else
1348 			*result_key = le64_to_cpu(ro_node(s)->keys[0]);
1349 
1350 		if (next_block || flags & INTERNAL_NODE) {
1351 			if (find_highest)
1352 				block = value64(ro_node(s), i);
1353 			else
1354 				block = value64(ro_node(s), 0);
1355 		}
1356 
1357 	} while (flags & INTERNAL_NODE);
1358 
1359 	if (next_block)
1360 		*next_block = block;
1361 	return 0;
1362 }
1363 
1364 static int dm_btree_find_key(struct dm_btree_info *info, dm_block_t root,
1365 			     bool find_highest, uint64_t *result_keys)
1366 {
1367 	int r = 0, count = 0, level;
1368 	struct ro_spine spine;
1369 
1370 	init_ro_spine(&spine, info);
1371 	for (level = 0; level < info->levels; level++) {
1372 		r = find_key(&spine, root, find_highest, result_keys + level,
1373 			     level == info->levels - 1 ? NULL : &root);
1374 		if (r == -ENODATA) {
1375 			r = 0;
1376 			break;
1377 
1378 		} else if (r)
1379 			break;
1380 
1381 		count++;
1382 	}
1383 	exit_ro_spine(&spine);
1384 
1385 	return r ? r : count;
1386 }
1387 
1388 int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
1389 			      uint64_t *result_keys)
1390 {
1391 	return dm_btree_find_key(info, root, true, result_keys);
1392 }
1393 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
1394 
1395 int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
1396 			     uint64_t *result_keys)
1397 {
1398 	return dm_btree_find_key(info, root, false, result_keys);
1399 }
1400 EXPORT_SYMBOL_GPL(dm_btree_find_lowest_key);
1401 
1402 /*----------------------------------------------------------------*/
1403 
1404 /*
1405  * FIXME: We shouldn't use a recursive algorithm when we have limited stack
1406  * space.  Also this only works for single level trees.
1407  */
1408 static int walk_node(struct dm_btree_info *info, dm_block_t block,
1409 		     int (*fn)(void *context, uint64_t *keys, void *leaf),
1410 		     void *context)
1411 {
1412 	int r;
1413 	unsigned i, nr;
1414 	struct dm_block *node;
1415 	struct btree_node *n;
1416 	uint64_t keys;
1417 
1418 	r = bn_read_lock(info, block, &node);
1419 	if (r)
1420 		return r;
1421 
1422 	n = dm_block_data(node);
1423 
1424 	nr = le32_to_cpu(n->header.nr_entries);
1425 	for (i = 0; i < nr; i++) {
1426 		if (le32_to_cpu(n->header.flags) & INTERNAL_NODE) {
1427 			r = walk_node(info, value64(n, i), fn, context);
1428 			if (r)
1429 				goto out;
1430 		} else {
1431 			keys = le64_to_cpu(*key_ptr(n, i));
1432 			r = fn(context, &keys, value_ptr(n, i));
1433 			if (r)
1434 				goto out;
1435 		}
1436 	}
1437 
1438 out:
1439 	dm_tm_unlock(info->tm, node);
1440 	return r;
1441 }
1442 
1443 int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
1444 		  int (*fn)(void *context, uint64_t *keys, void *leaf),
1445 		  void *context)
1446 {
1447 	BUG_ON(info->levels > 1);
1448 	return walk_node(info, root, fn, context);
1449 }
1450 EXPORT_SYMBOL_GPL(dm_btree_walk);
1451 
1452 /*----------------------------------------------------------------*/
1453 
1454 static void prefetch_values(struct dm_btree_cursor *c)
1455 {
1456 	unsigned i, nr;
1457 	__le64 value_le;
1458 	struct cursor_node *n = c->nodes + c->depth - 1;
1459 	struct btree_node *bn = dm_block_data(n->b);
1460 	struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
1461 
1462 	BUG_ON(c->info->value_type.size != sizeof(value_le));
1463 
1464 	nr = le32_to_cpu(bn->header.nr_entries);
1465 	for (i = 0; i < nr; i++) {
1466 		memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
1467 		dm_bm_prefetch(bm, le64_to_cpu(value_le));
1468 	}
1469 }
1470 
1471 static bool leaf_node(struct dm_btree_cursor *c)
1472 {
1473 	struct cursor_node *n = c->nodes + c->depth - 1;
1474 	struct btree_node *bn = dm_block_data(n->b);
1475 
1476 	return le32_to_cpu(bn->header.flags) & LEAF_NODE;
1477 }
1478 
1479 static int push_node(struct dm_btree_cursor *c, dm_block_t b)
1480 {
1481 	int r;
1482 	struct cursor_node *n = c->nodes + c->depth;
1483 
1484 	if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
1485 		DMERR("couldn't push cursor node, stack depth too high");
1486 		return -EINVAL;
1487 	}
1488 
1489 	r = bn_read_lock(c->info, b, &n->b);
1490 	if (r)
1491 		return r;
1492 
1493 	n->index = 0;
1494 	c->depth++;
1495 
1496 	if (c->prefetch_leaves || !leaf_node(c))
1497 		prefetch_values(c);
1498 
1499 	return 0;
1500 }
1501 
1502 static void pop_node(struct dm_btree_cursor *c)
1503 {
1504 	c->depth--;
1505 	unlock_block(c->info, c->nodes[c->depth].b);
1506 }
1507 
1508 static int inc_or_backtrack(struct dm_btree_cursor *c)
1509 {
1510 	struct cursor_node *n;
1511 	struct btree_node *bn;
1512 
1513 	for (;;) {
1514 		if (!c->depth)
1515 			return -ENODATA;
1516 
1517 		n = c->nodes + c->depth - 1;
1518 		bn = dm_block_data(n->b);
1519 
1520 		n->index++;
1521 		if (n->index < le32_to_cpu(bn->header.nr_entries))
1522 			break;
1523 
1524 		pop_node(c);
1525 	}
1526 
1527 	return 0;
1528 }
1529 
1530 static int find_leaf(struct dm_btree_cursor *c)
1531 {
1532 	int r = 0;
1533 	struct cursor_node *n;
1534 	struct btree_node *bn;
1535 	__le64 value_le;
1536 
1537 	for (;;) {
1538 		n = c->nodes + c->depth - 1;
1539 		bn = dm_block_data(n->b);
1540 
1541 		if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
1542 			break;
1543 
1544 		memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
1545 		r = push_node(c, le64_to_cpu(value_le));
1546 		if (r) {
1547 			DMERR("push_node failed");
1548 			break;
1549 		}
1550 	}
1551 
1552 	if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
1553 		return -ENODATA;
1554 
1555 	return r;
1556 }
1557 
1558 int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
1559 			  bool prefetch_leaves, struct dm_btree_cursor *c)
1560 {
1561 	int r;
1562 
1563 	c->info = info;
1564 	c->root = root;
1565 	c->depth = 0;
1566 	c->prefetch_leaves = prefetch_leaves;
1567 
1568 	r = push_node(c, root);
1569 	if (r)
1570 		return r;
1571 
1572 	return find_leaf(c);
1573 }
1574 EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
1575 
1576 void dm_btree_cursor_end(struct dm_btree_cursor *c)
1577 {
1578 	while (c->depth)
1579 		pop_node(c);
1580 }
1581 EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
1582 
1583 int dm_btree_cursor_next(struct dm_btree_cursor *c)
1584 {
1585 	int r = inc_or_backtrack(c);
1586 	if (!r) {
1587 		r = find_leaf(c);
1588 		if (r)
1589 			DMERR("find_leaf failed");
1590 	}
1591 
1592 	return r;
1593 }
1594 EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
1595 
1596 int dm_btree_cursor_skip(struct dm_btree_cursor *c, uint32_t count)
1597 {
1598 	int r = 0;
1599 
1600 	while (count-- && !r)
1601 		r = dm_btree_cursor_next(c);
1602 
1603 	return r;
1604 }
1605 EXPORT_SYMBOL_GPL(dm_btree_cursor_skip);
1606 
1607 int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
1608 {
1609 	if (c->depth) {
1610 		struct cursor_node *n = c->nodes + c->depth - 1;
1611 		struct btree_node *bn = dm_block_data(n->b);
1612 
1613 		if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
1614 			return -EINVAL;
1615 
1616 		*key = le64_to_cpu(*key_ptr(bn, n->index));
1617 		memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
1618 		return 0;
1619 
1620 	} else
1621 		return -ENODATA;
1622 }
1623 EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);
1624