1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2011 Red Hat, Inc.
4  *
5  * This file is released under the GPL.
6  */
7 
8 #include "dm-btree-internal.h"
9 #include "dm-space-map.h"
10 #include "dm-transaction-manager.h"
11 
12 #include <linux/export.h>
13 #include <linux/device-mapper.h>
14 
15 #define DM_MSG_PREFIX "btree"
16 
17 /*
18  *--------------------------------------------------------------
19  * Array manipulation
20  *--------------------------------------------------------------
21  */
22 static void memcpy_disk(void *dest, const void *src, size_t len)
23 	__dm_written_to_disk(src)
24 {
25 	memcpy(dest, src, len);
26 	__dm_unbless_for_disk(src);
27 }
28 
29 static void array_insert(void *base, size_t elt_size, unsigned int nr_elts,
30 			 unsigned int index, void *elt)
31 	__dm_written_to_disk(elt)
32 {
33 	if (index < nr_elts)
34 		memmove(base + (elt_size * (index + 1)),
35 			base + (elt_size * index),
36 			(nr_elts - index) * elt_size);
37 
38 	memcpy_disk(base + (elt_size * index), elt, elt_size);
39 }
40 
41 /*----------------------------------------------------------------*/
42 
43 /* makes the assumption that no two keys are the same. */
44 static int bsearch(struct btree_node *n, uint64_t key, int want_hi)
45 {
46 	int lo = -1, hi = le32_to_cpu(n->header.nr_entries);
47 
48 	while (hi - lo > 1) {
49 		int mid = lo + ((hi - lo) / 2);
50 		uint64_t mid_key = le64_to_cpu(n->keys[mid]);
51 
52 		if (mid_key == key)
53 			return mid;
54 
55 		if (mid_key < key)
56 			lo = mid;
57 		else
58 			hi = mid;
59 	}
60 
61 	return want_hi ? hi : lo;
62 }
63 
64 int lower_bound(struct btree_node *n, uint64_t key)
65 {
66 	return bsearch(n, key, 0);
67 }
68 
69 static int upper_bound(struct btree_node *n, uint64_t key)
70 {
71 	return bsearch(n, key, 1);
72 }
73 
74 void inc_children(struct dm_transaction_manager *tm, struct btree_node *n,
75 		  struct dm_btree_value_type *vt)
76 {
77 	uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
78 
79 	if (le32_to_cpu(n->header.flags) & INTERNAL_NODE)
80 		dm_tm_with_runs(tm, value_ptr(n, 0), nr_entries, dm_tm_inc_range);
81 
82 	else if (vt->inc)
83 		vt->inc(vt->context, value_ptr(n, 0), nr_entries);
84 }
85 
86 static int insert_at(size_t value_size, struct btree_node *node, unsigned int index,
87 		     uint64_t key, void *value)
88 	__dm_written_to_disk(value)
89 {
90 	uint32_t nr_entries = le32_to_cpu(node->header.nr_entries);
91 	uint32_t max_entries = le32_to_cpu(node->header.max_entries);
92 	__le64 key_le = cpu_to_le64(key);
93 
94 	if (index > nr_entries ||
95 	    index >= max_entries ||
96 	    nr_entries >= max_entries) {
97 		DMERR("too many entries in btree node for insert");
98 		__dm_unbless_for_disk(value);
99 		return -ENOMEM;
100 	}
101 
102 	__dm_bless_for_disk(&key_le);
103 
104 	array_insert(node->keys, sizeof(*node->keys), nr_entries, index, &key_le);
105 	array_insert(value_base(node), value_size, nr_entries, index, value);
106 	node->header.nr_entries = cpu_to_le32(nr_entries + 1);
107 
108 	return 0;
109 }
110 
111 /*----------------------------------------------------------------*/
112 
113 /*
114  * We want 3n entries (for some n).  This works more nicely for repeated
115  * insert remove loops than (2n + 1).
116  */
117 static uint32_t calc_max_entries(size_t value_size, size_t block_size)
118 {
119 	uint32_t total, n;
120 	size_t elt_size = sizeof(uint64_t) + value_size; /* key + value */
121 
122 	block_size -= sizeof(struct node_header);
123 	total = block_size / elt_size;
124 	n = total / 3;		/* rounds down */
125 
126 	return 3 * n;
127 }
128 
129 int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root)
130 {
131 	int r;
132 	struct dm_block *b;
133 	struct btree_node *n;
134 	size_t block_size;
135 	uint32_t max_entries;
136 
137 	r = new_block(info, &b);
138 	if (r < 0)
139 		return r;
140 
141 	block_size = dm_bm_block_size(dm_tm_get_bm(info->tm));
142 	max_entries = calc_max_entries(info->value_type.size, block_size);
143 
144 	n = dm_block_data(b);
145 	memset(n, 0, block_size);
146 	n->header.flags = cpu_to_le32(LEAF_NODE);
147 	n->header.nr_entries = cpu_to_le32(0);
148 	n->header.max_entries = cpu_to_le32(max_entries);
149 	n->header.value_size = cpu_to_le32(info->value_type.size);
150 
151 	*root = dm_block_location(b);
152 	unlock_block(info, b);
153 
154 	return 0;
155 }
156 EXPORT_SYMBOL_GPL(dm_btree_empty);
157 
158 /*----------------------------------------------------------------*/
159 
160 /*
161  * Deletion uses a recursive algorithm, since we have limited stack space
162  * we explicitly manage our own stack on the heap.
163  */
164 #define MAX_SPINE_DEPTH 64
165 struct frame {
166 	struct dm_block *b;
167 	struct btree_node *n;
168 	unsigned int level;
169 	unsigned int nr_children;
170 	unsigned int current_child;
171 };
172 
173 struct del_stack {
174 	struct dm_btree_info *info;
175 	struct dm_transaction_manager *tm;
176 	int top;
177 	struct frame spine[MAX_SPINE_DEPTH];
178 };
179 
180 static int top_frame(struct del_stack *s, struct frame **f)
181 {
182 	if (s->top < 0) {
183 		DMERR("btree deletion stack empty");
184 		return -EINVAL;
185 	}
186 
187 	*f = s->spine + s->top;
188 
189 	return 0;
190 }
191 
192 static int unprocessed_frames(struct del_stack *s)
193 {
194 	return s->top >= 0;
195 }
196 
197 static void prefetch_children(struct del_stack *s, struct frame *f)
198 {
199 	unsigned int i;
200 	struct dm_block_manager *bm = dm_tm_get_bm(s->tm);
201 
202 	for (i = 0; i < f->nr_children; i++)
203 		dm_bm_prefetch(bm, value64(f->n, i));
204 }
205 
206 static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
207 {
208 	return f->level < (info->levels - 1);
209 }
210 
211 static int push_frame(struct del_stack *s, dm_block_t b, unsigned int level)
212 {
213 	int r;
214 	uint32_t ref_count;
215 
216 	if (s->top >= MAX_SPINE_DEPTH - 1) {
217 		DMERR("btree deletion stack out of memory");
218 		return -ENOMEM;
219 	}
220 
221 	r = dm_tm_ref(s->tm, b, &ref_count);
222 	if (r)
223 		return r;
224 
225 	if (ref_count > 1)
226 		/*
227 		 * This is a shared node, so we can just decrement it's
228 		 * reference counter and leave the children.
229 		 */
230 		dm_tm_dec(s->tm, b);
231 
232 	else {
233 		uint32_t flags;
234 		struct frame *f = s->spine + ++s->top;
235 
236 		r = dm_tm_read_lock(s->tm, b, &btree_node_validator, &f->b);
237 		if (r) {
238 			s->top--;
239 			return r;
240 		}
241 
242 		f->n = dm_block_data(f->b);
243 		f->level = level;
244 		f->nr_children = le32_to_cpu(f->n->header.nr_entries);
245 		f->current_child = 0;
246 
247 		flags = le32_to_cpu(f->n->header.flags);
248 		if (flags & INTERNAL_NODE || is_internal_level(s->info, f))
249 			prefetch_children(s, f);
250 	}
251 
252 	return 0;
253 }
254 
255 static void pop_frame(struct del_stack *s)
256 {
257 	struct frame *f = s->spine + s->top--;
258 
259 	dm_tm_dec(s->tm, dm_block_location(f->b));
260 	dm_tm_unlock(s->tm, f->b);
261 }
262 
263 static void unlock_all_frames(struct del_stack *s)
264 {
265 	struct frame *f;
266 
267 	while (unprocessed_frames(s)) {
268 		f = s->spine + s->top--;
269 		dm_tm_unlock(s->tm, f->b);
270 	}
271 }
272 
273 int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
274 {
275 	int r;
276 	struct del_stack *s;
277 
278 	/*
279 	 * dm_btree_del() is called via an ioctl, as such should be
280 	 * considered an FS op.  We can't recurse back into the FS, so we
281 	 * allocate GFP_NOFS.
282 	 */
283 	s = kmalloc(sizeof(*s), GFP_NOFS);
284 	if (!s)
285 		return -ENOMEM;
286 	s->info = info;
287 	s->tm = info->tm;
288 	s->top = -1;
289 
290 	r = push_frame(s, root, 0);
291 	if (r)
292 		goto out;
293 
294 	while (unprocessed_frames(s)) {
295 		uint32_t flags;
296 		struct frame *f;
297 		dm_block_t b;
298 
299 		r = top_frame(s, &f);
300 		if (r)
301 			goto out;
302 
303 		if (f->current_child >= f->nr_children) {
304 			pop_frame(s);
305 			continue;
306 		}
307 
308 		flags = le32_to_cpu(f->n->header.flags);
309 		if (flags & INTERNAL_NODE) {
310 			b = value64(f->n, f->current_child);
311 			f->current_child++;
312 			r = push_frame(s, b, f->level);
313 			if (r)
314 				goto out;
315 
316 		} else if (is_internal_level(info, f)) {
317 			b = value64(f->n, f->current_child);
318 			f->current_child++;
319 			r = push_frame(s, b, f->level + 1);
320 			if (r)
321 				goto out;
322 
323 		} else {
324 			if (info->value_type.dec)
325 				info->value_type.dec(info->value_type.context,
326 						     value_ptr(f->n, 0), f->nr_children);
327 			pop_frame(s);
328 		}
329 	}
330 out:
331 	if (r) {
332 		/* cleanup all frames of del_stack */
333 		unlock_all_frames(s);
334 	}
335 	kfree(s);
336 
337 	return r;
338 }
339 EXPORT_SYMBOL_GPL(dm_btree_del);
340 
341 /*----------------------------------------------------------------*/
342 
343 static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
344 			    int (*search_fn)(struct btree_node *, uint64_t),
345 			    uint64_t *result_key, void *v, size_t value_size)
346 {
347 	int i, r;
348 	uint32_t flags, nr_entries;
349 
350 	do {
351 		r = ro_step(s, block);
352 		if (r < 0)
353 			return r;
354 
355 		i = search_fn(ro_node(s), key);
356 
357 		flags = le32_to_cpu(ro_node(s)->header.flags);
358 		nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
359 		if (i < 0 || i >= nr_entries)
360 			return -ENODATA;
361 
362 		if (flags & INTERNAL_NODE)
363 			block = value64(ro_node(s), i);
364 
365 	} while (!(flags & LEAF_NODE));
366 
367 	*result_key = le64_to_cpu(ro_node(s)->keys[i]);
368 	if (v)
369 		memcpy(v, value_ptr(ro_node(s), i), value_size);
370 
371 	return 0;
372 }
373 
374 int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
375 		    uint64_t *keys, void *value_le)
376 {
377 	unsigned int level, last_level = info->levels - 1;
378 	int r = -ENODATA;
379 	uint64_t rkey;
380 	__le64 internal_value_le;
381 	struct ro_spine spine;
382 
383 	init_ro_spine(&spine, info);
384 	for (level = 0; level < info->levels; level++) {
385 		size_t size;
386 		void *value_p;
387 
388 		if (level == last_level) {
389 			value_p = value_le;
390 			size = info->value_type.size;
391 
392 		} else {
393 			value_p = &internal_value_le;
394 			size = sizeof(uint64_t);
395 		}
396 
397 		r = btree_lookup_raw(&spine, root, keys[level],
398 				     lower_bound, &rkey,
399 				     value_p, size);
400 
401 		if (!r) {
402 			if (rkey != keys[level]) {
403 				exit_ro_spine(&spine);
404 				return -ENODATA;
405 			}
406 		} else {
407 			exit_ro_spine(&spine);
408 			return r;
409 		}
410 
411 		root = le64_to_cpu(internal_value_le);
412 	}
413 	exit_ro_spine(&spine);
414 
415 	return r;
416 }
417 EXPORT_SYMBOL_GPL(dm_btree_lookup);
418 
419 static int dm_btree_lookup_next_single(struct dm_btree_info *info, dm_block_t root,
420 				       uint64_t key, uint64_t *rkey, void *value_le)
421 {
422 	int r, i;
423 	uint32_t flags, nr_entries;
424 	struct dm_block *node;
425 	struct btree_node *n;
426 
427 	r = bn_read_lock(info, root, &node);
428 	if (r)
429 		return r;
430 
431 	n = dm_block_data(node);
432 	flags = le32_to_cpu(n->header.flags);
433 	nr_entries = le32_to_cpu(n->header.nr_entries);
434 
435 	if (flags & INTERNAL_NODE) {
436 		i = lower_bound(n, key);
437 		if (i < 0) {
438 			/*
439 			 * avoid early -ENODATA return when all entries are
440 			 * higher than the search @key.
441 			 */
442 			i = 0;
443 		}
444 		if (i >= nr_entries) {
445 			r = -ENODATA;
446 			goto out;
447 		}
448 
449 		r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
450 		if (r == -ENODATA && i < (nr_entries - 1)) {
451 			i++;
452 			r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
453 		}
454 
455 	} else {
456 		i = upper_bound(n, key);
457 		if (i < 0 || i >= nr_entries) {
458 			r = -ENODATA;
459 			goto out;
460 		}
461 
462 		*rkey = le64_to_cpu(n->keys[i]);
463 		memcpy(value_le, value_ptr(n, i), info->value_type.size);
464 	}
465 out:
466 	dm_tm_unlock(info->tm, node);
467 	return r;
468 }
469 
470 int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
471 			 uint64_t *keys, uint64_t *rkey, void *value_le)
472 {
473 	unsigned int level;
474 	int r = -ENODATA;
475 	__le64 internal_value_le;
476 	struct ro_spine spine;
477 
478 	init_ro_spine(&spine, info);
479 	for (level = 0; level < info->levels - 1u; level++) {
480 		r = btree_lookup_raw(&spine, root, keys[level],
481 				     lower_bound, rkey,
482 				     &internal_value_le, sizeof(uint64_t));
483 		if (r)
484 			goto out;
485 
486 		if (*rkey != keys[level]) {
487 			r = -ENODATA;
488 			goto out;
489 		}
490 
491 		root = le64_to_cpu(internal_value_le);
492 	}
493 
494 	r = dm_btree_lookup_next_single(info, root, keys[level], rkey, value_le);
495 out:
496 	exit_ro_spine(&spine);
497 	return r;
498 }
499 EXPORT_SYMBOL_GPL(dm_btree_lookup_next);
500 
501 /*----------------------------------------------------------------*/
502 
503 /*
504  * Copies entries from one region of a btree node to another.  The regions
505  * must not overlap.
506  */
507 static void copy_entries(struct btree_node *dest, unsigned int dest_offset,
508 			 struct btree_node *src, unsigned int src_offset,
509 			 unsigned int count)
510 {
511 	size_t value_size = le32_to_cpu(dest->header.value_size);
512 
513 	memcpy(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
514 	memcpy(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
515 }
516 
517 /*
518  * Moves entries from one region fo a btree node to another.  The regions
519  * may overlap.
520  */
521 static void move_entries(struct btree_node *dest, unsigned int dest_offset,
522 			 struct btree_node *src, unsigned int src_offset,
523 			 unsigned int count)
524 {
525 	size_t value_size = le32_to_cpu(dest->header.value_size);
526 
527 	memmove(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
528 	memmove(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
529 }
530 
531 /*
532  * Erases the first 'count' entries of a btree node, shifting following
533  * entries down into their place.
534  */
535 static void shift_down(struct btree_node *n, unsigned int count)
536 {
537 	move_entries(n, 0, n, count, le32_to_cpu(n->header.nr_entries) - count);
538 }
539 
540 /*
541  * Moves entries in a btree node up 'count' places, making space for
542  * new entries at the start of the node.
543  */
544 static void shift_up(struct btree_node *n, unsigned int count)
545 {
546 	move_entries(n, count, n, 0, le32_to_cpu(n->header.nr_entries));
547 }
548 
549 /*
550  * Redistributes entries between two btree nodes to make them
551  * have similar numbers of entries.
552  */
553 static void redistribute2(struct btree_node *left, struct btree_node *right)
554 {
555 	unsigned int nr_left = le32_to_cpu(left->header.nr_entries);
556 	unsigned int nr_right = le32_to_cpu(right->header.nr_entries);
557 	unsigned int total = nr_left + nr_right;
558 	unsigned int target_left = total / 2;
559 	unsigned int target_right = total - target_left;
560 
561 	if (nr_left < target_left) {
562 		unsigned int delta = target_left - nr_left;
563 
564 		copy_entries(left, nr_left, right, 0, delta);
565 		shift_down(right, delta);
566 	} else if (nr_left > target_left) {
567 		unsigned int delta = nr_left - target_left;
568 
569 		if (nr_right)
570 			shift_up(right, delta);
571 		copy_entries(right, 0, left, target_left, delta);
572 	}
573 
574 	left->header.nr_entries = cpu_to_le32(target_left);
575 	right->header.nr_entries = cpu_to_le32(target_right);
576 }
577 
578 /*
579  * Redistribute entries between three nodes.  Assumes the central
580  * node is empty.
581  */
582 static void redistribute3(struct btree_node *left, struct btree_node *center,
583 			  struct btree_node *right)
584 {
585 	unsigned int nr_left = le32_to_cpu(left->header.nr_entries);
586 	unsigned int nr_center = le32_to_cpu(center->header.nr_entries);
587 	unsigned int nr_right = le32_to_cpu(right->header.nr_entries);
588 	unsigned int total, target_left, target_center, target_right;
589 
590 	BUG_ON(nr_center);
591 
592 	total = nr_left + nr_right;
593 	target_left = total / 3;
594 	target_center = (total - target_left) / 2;
595 	target_right = (total - target_left - target_center);
596 
597 	if (nr_left < target_left) {
598 		unsigned int left_short = target_left - nr_left;
599 
600 		copy_entries(left, nr_left, right, 0, left_short);
601 		copy_entries(center, 0, right, left_short, target_center);
602 		shift_down(right, nr_right - target_right);
603 
604 	} else if (nr_left < (target_left + target_center)) {
605 		unsigned int left_to_center = nr_left - target_left;
606 
607 		copy_entries(center, 0, left, target_left, left_to_center);
608 		copy_entries(center, left_to_center, right, 0, target_center - left_to_center);
609 		shift_down(right, nr_right - target_right);
610 
611 	} else {
612 		unsigned int right_short = target_right - nr_right;
613 
614 		shift_up(right, right_short);
615 		copy_entries(right, 0, left, nr_left - right_short, right_short);
616 		copy_entries(center, 0, left, target_left, nr_left - target_left);
617 	}
618 
619 	left->header.nr_entries = cpu_to_le32(target_left);
620 	center->header.nr_entries = cpu_to_le32(target_center);
621 	right->header.nr_entries = cpu_to_le32(target_right);
622 }
623 
624 /*
625  * Splits a node by creating a sibling node and shifting half the nodes
626  * contents across.  Assumes there is a parent node, and it has room for
627  * another child.
628  *
629  * Before:
630  *	  +--------+
631  *	  | Parent |
632  *	  +--------+
633  *	     |
634  *	     v
635  *	+----------+
636  *	| A ++++++ |
637  *	+----------+
638  *
639  *
640  * After:
641  *		+--------+
642  *		| Parent |
643  *		+--------+
644  *		  |	|
645  *		  v	+------+
646  *	    +---------+	       |
647  *	    | A* +++  |	       v
648  *	    +---------+	  +-------+
649  *			  | B +++ |
650  *			  +-------+
651  *
652  * Where A* is a shadow of A.
653  */
654 static int split_one_into_two(struct shadow_spine *s, unsigned int parent_index,
655 			      struct dm_btree_value_type *vt, uint64_t key)
656 {
657 	int r;
658 	struct dm_block *left, *right, *parent;
659 	struct btree_node *ln, *rn, *pn;
660 	__le64 location;
661 
662 	left = shadow_current(s);
663 
664 	r = new_block(s->info, &right);
665 	if (r < 0)
666 		return r;
667 
668 	ln = dm_block_data(left);
669 	rn = dm_block_data(right);
670 
671 	rn->header.flags = ln->header.flags;
672 	rn->header.nr_entries = cpu_to_le32(0);
673 	rn->header.max_entries = ln->header.max_entries;
674 	rn->header.value_size = ln->header.value_size;
675 	redistribute2(ln, rn);
676 
677 	/* patch up the parent */
678 	parent = shadow_parent(s);
679 	pn = dm_block_data(parent);
680 
681 	location = cpu_to_le64(dm_block_location(right));
682 	__dm_bless_for_disk(&location);
683 	r = insert_at(sizeof(__le64), pn, parent_index + 1,
684 		      le64_to_cpu(rn->keys[0]), &location);
685 	if (r) {
686 		unlock_block(s->info, right);
687 		return r;
688 	}
689 
690 	/* patch up the spine */
691 	if (key < le64_to_cpu(rn->keys[0])) {
692 		unlock_block(s->info, right);
693 		s->nodes[1] = left;
694 	} else {
695 		unlock_block(s->info, left);
696 		s->nodes[1] = right;
697 	}
698 
699 	return 0;
700 }
701 
702 /*
703  * We often need to modify a sibling node.  This function shadows a particular
704  * child of the given parent node.  Making sure to update the parent to point
705  * to the new shadow.
706  */
707 static int shadow_child(struct dm_btree_info *info, struct dm_btree_value_type *vt,
708 			struct btree_node *parent, unsigned int index,
709 			struct dm_block **result)
710 {
711 	int r, inc;
712 	dm_block_t root;
713 	struct btree_node *node;
714 
715 	root = value64(parent, index);
716 
717 	r = dm_tm_shadow_block(info->tm, root, &btree_node_validator,
718 			       result, &inc);
719 	if (r)
720 		return r;
721 
722 	node = dm_block_data(*result);
723 
724 	if (inc)
725 		inc_children(info->tm, node, vt);
726 
727 	*((__le64 *) value_ptr(parent, index)) =
728 		cpu_to_le64(dm_block_location(*result));
729 
730 	return 0;
731 }
732 
733 /*
734  * Splits two nodes into three.  This is more work, but results in fuller
735  * nodes, so saves metadata space.
736  */
737 static int split_two_into_three(struct shadow_spine *s, unsigned int parent_index,
738 				struct dm_btree_value_type *vt, uint64_t key)
739 {
740 	int r;
741 	unsigned int middle_index;
742 	struct dm_block *left, *middle, *right, *parent;
743 	struct btree_node *ln, *rn, *mn, *pn;
744 	__le64 location;
745 
746 	parent = shadow_parent(s);
747 	pn = dm_block_data(parent);
748 
749 	if (parent_index == 0) {
750 		middle_index = 1;
751 		left = shadow_current(s);
752 		r = shadow_child(s->info, vt, pn, parent_index + 1, &right);
753 		if (r)
754 			return r;
755 	} else {
756 		middle_index = parent_index;
757 		right = shadow_current(s);
758 		r = shadow_child(s->info, vt, pn, parent_index - 1, &left);
759 		if (r)
760 			return r;
761 	}
762 
763 	r = new_block(s->info, &middle);
764 	if (r < 0)
765 		return r;
766 
767 	ln = dm_block_data(left);
768 	mn = dm_block_data(middle);
769 	rn = dm_block_data(right);
770 
771 	mn->header.nr_entries = cpu_to_le32(0);
772 	mn->header.flags = ln->header.flags;
773 	mn->header.max_entries = ln->header.max_entries;
774 	mn->header.value_size = ln->header.value_size;
775 
776 	redistribute3(ln, mn, rn);
777 
778 	/* patch up the parent */
779 	pn->keys[middle_index] = rn->keys[0];
780 	location = cpu_to_le64(dm_block_location(middle));
781 	__dm_bless_for_disk(&location);
782 	r = insert_at(sizeof(__le64), pn, middle_index,
783 		      le64_to_cpu(mn->keys[0]), &location);
784 	if (r) {
785 		if (shadow_current(s) != left)
786 			unlock_block(s->info, left);
787 
788 		unlock_block(s->info, middle);
789 
790 		if (shadow_current(s) != right)
791 			unlock_block(s->info, right);
792 
793 		return r;
794 	}
795 
796 
797 	/* patch up the spine */
798 	if (key < le64_to_cpu(mn->keys[0])) {
799 		unlock_block(s->info, middle);
800 		unlock_block(s->info, right);
801 		s->nodes[1] = left;
802 	} else if (key < le64_to_cpu(rn->keys[0])) {
803 		unlock_block(s->info, left);
804 		unlock_block(s->info, right);
805 		s->nodes[1] = middle;
806 	} else {
807 		unlock_block(s->info, left);
808 		unlock_block(s->info, middle);
809 		s->nodes[1] = right;
810 	}
811 
812 	return 0;
813 }
814 
815 /*----------------------------------------------------------------*/
816 
817 /*
818  * Splits a node by creating two new children beneath the given node.
819  *
820  * Before:
821  *	  +----------+
822  *	  | A ++++++ |
823  *	  +----------+
824  *
825  *
826  * After:
827  *	+------------+
828  *	| A (shadow) |
829  *	+------------+
830  *	    |	|
831  *   +------+	+----+
832  *   |		     |
833  *   v		     v
834  * +-------+	 +-------+
835  * | B +++ |	 | C +++ |
836  * +-------+	 +-------+
837  */
838 static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
839 {
840 	int r;
841 	size_t size;
842 	unsigned int nr_left, nr_right;
843 	struct dm_block *left, *right, *new_parent;
844 	struct btree_node *pn, *ln, *rn;
845 	__le64 val;
846 
847 	new_parent = shadow_current(s);
848 
849 	pn = dm_block_data(new_parent);
850 	size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
851 		sizeof(__le64) : s->info->value_type.size;
852 
853 	/* create & init the left block */
854 	r = new_block(s->info, &left);
855 	if (r < 0)
856 		return r;
857 
858 	ln = dm_block_data(left);
859 	nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
860 
861 	ln->header.flags = pn->header.flags;
862 	ln->header.nr_entries = cpu_to_le32(nr_left);
863 	ln->header.max_entries = pn->header.max_entries;
864 	ln->header.value_size = pn->header.value_size;
865 	memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
866 	memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
867 
868 	/* create & init the right block */
869 	r = new_block(s->info, &right);
870 	if (r < 0) {
871 		unlock_block(s->info, left);
872 		return r;
873 	}
874 
875 	rn = dm_block_data(right);
876 	nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
877 
878 	rn->header.flags = pn->header.flags;
879 	rn->header.nr_entries = cpu_to_le32(nr_right);
880 	rn->header.max_entries = pn->header.max_entries;
881 	rn->header.value_size = pn->header.value_size;
882 	memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
883 	memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
884 	       nr_right * size);
885 
886 	/* new_parent should just point to l and r now */
887 	pn->header.flags = cpu_to_le32(INTERNAL_NODE);
888 	pn->header.nr_entries = cpu_to_le32(2);
889 	pn->header.max_entries = cpu_to_le32(
890 		calc_max_entries(sizeof(__le64),
891 				 dm_bm_block_size(
892 					 dm_tm_get_bm(s->info->tm))));
893 	pn->header.value_size = cpu_to_le32(sizeof(__le64));
894 
895 	val = cpu_to_le64(dm_block_location(left));
896 	__dm_bless_for_disk(&val);
897 	pn->keys[0] = ln->keys[0];
898 	memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
899 
900 	val = cpu_to_le64(dm_block_location(right));
901 	__dm_bless_for_disk(&val);
902 	pn->keys[1] = rn->keys[0];
903 	memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
904 
905 	unlock_block(s->info, left);
906 	unlock_block(s->info, right);
907 	return 0;
908 }
909 
910 /*----------------------------------------------------------------*/
911 
912 /*
913  * Redistributes a node's entries with its left sibling.
914  */
915 static int rebalance_left(struct shadow_spine *s, struct dm_btree_value_type *vt,
916 			  unsigned int parent_index, uint64_t key)
917 {
918 	int r;
919 	struct dm_block *sib;
920 	struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
921 
922 	r = shadow_child(s->info, vt, parent, parent_index - 1, &sib);
923 	if (r)
924 		return r;
925 
926 	left = dm_block_data(sib);
927 	right = dm_block_data(shadow_current(s));
928 	redistribute2(left, right);
929 	*key_ptr(parent, parent_index) = right->keys[0];
930 
931 	if (key < le64_to_cpu(right->keys[0])) {
932 		unlock_block(s->info, s->nodes[1]);
933 		s->nodes[1] = sib;
934 	} else {
935 		unlock_block(s->info, sib);
936 	}
937 
938 	return 0;
939 }
940 
941 /*
942  * Redistributes a nodes entries with its right sibling.
943  */
944 static int rebalance_right(struct shadow_spine *s, struct dm_btree_value_type *vt,
945 			   unsigned int parent_index, uint64_t key)
946 {
947 	int r;
948 	struct dm_block *sib;
949 	struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
950 
951 	r = shadow_child(s->info, vt, parent, parent_index + 1, &sib);
952 	if (r)
953 		return r;
954 
955 	left = dm_block_data(shadow_current(s));
956 	right = dm_block_data(sib);
957 	redistribute2(left, right);
958 	*key_ptr(parent, parent_index + 1) = right->keys[0];
959 
960 	if (key < le64_to_cpu(right->keys[0])) {
961 		unlock_block(s->info, sib);
962 	} else {
963 		unlock_block(s->info, s->nodes[1]);
964 		s->nodes[1] = sib;
965 	}
966 
967 	return 0;
968 }
969 
970 /*
971  * Returns the number of spare entries in a node.
972  */
973 static int get_node_free_space(struct dm_btree_info *info, dm_block_t b, unsigned int *space)
974 {
975 	int r;
976 	unsigned int nr_entries;
977 	struct dm_block *block;
978 	struct btree_node *node;
979 
980 	r = bn_read_lock(info, b, &block);
981 	if (r)
982 		return r;
983 
984 	node = dm_block_data(block);
985 	nr_entries = le32_to_cpu(node->header.nr_entries);
986 	*space = le32_to_cpu(node->header.max_entries) - nr_entries;
987 
988 	unlock_block(info, block);
989 	return 0;
990 }
991 
992 /*
993  * Make space in a node, either by moving some entries to a sibling,
994  * or creating a new sibling node.  SPACE_THRESHOLD defines the minimum
995  * number of free entries that must be in the sibling to make the move
996  * worth while.  If the siblings are shared (eg, part of a snapshot),
997  * then they are not touched, since this break sharing and so consume
998  * more space than we save.
999  */
1000 #define SPACE_THRESHOLD 8
1001 static int rebalance_or_split(struct shadow_spine *s, struct dm_btree_value_type *vt,
1002 			      unsigned int parent_index, uint64_t key)
1003 {
1004 	int r;
1005 	struct btree_node *parent = dm_block_data(shadow_parent(s));
1006 	unsigned int nr_parent = le32_to_cpu(parent->header.nr_entries);
1007 	unsigned int free_space;
1008 	int left_shared = 0, right_shared = 0;
1009 
1010 	/* Should we move entries to the left sibling? */
1011 	if (parent_index > 0) {
1012 		dm_block_t left_b = value64(parent, parent_index - 1);
1013 
1014 		r = dm_tm_block_is_shared(s->info->tm, left_b, &left_shared);
1015 		if (r)
1016 			return r;
1017 
1018 		if (!left_shared) {
1019 			r = get_node_free_space(s->info, left_b, &free_space);
1020 			if (r)
1021 				return r;
1022 
1023 			if (free_space >= SPACE_THRESHOLD)
1024 				return rebalance_left(s, vt, parent_index, key);
1025 		}
1026 	}
1027 
1028 	/* Should we move entries to the right sibling? */
1029 	if (parent_index < (nr_parent - 1)) {
1030 		dm_block_t right_b = value64(parent, parent_index + 1);
1031 
1032 		r = dm_tm_block_is_shared(s->info->tm, right_b, &right_shared);
1033 		if (r)
1034 			return r;
1035 
1036 		if (!right_shared) {
1037 			r = get_node_free_space(s->info, right_b, &free_space);
1038 			if (r)
1039 				return r;
1040 
1041 			if (free_space >= SPACE_THRESHOLD)
1042 				return rebalance_right(s, vt, parent_index, key);
1043 		}
1044 	}
1045 
1046 	/*
1047 	 * We need to split the node, normally we split two nodes
1048 	 * into three.	But when inserting a sequence that is either
1049 	 * monotonically increasing or decreasing it's better to split
1050 	 * a single node into two.
1051 	 */
1052 	if (left_shared || right_shared || (nr_parent <= 2) ||
1053 	    (parent_index == 0) || (parent_index + 1 == nr_parent)) {
1054 		return split_one_into_two(s, parent_index, vt, key);
1055 	} else {
1056 		return split_two_into_three(s, parent_index, vt, key);
1057 	}
1058 }
1059 
1060 /*
1061  * Does the node contain a particular key?
1062  */
1063 static bool contains_key(struct btree_node *node, uint64_t key)
1064 {
1065 	int i = lower_bound(node, key);
1066 
1067 	if (i >= 0 && le64_to_cpu(node->keys[i]) == key)
1068 		return true;
1069 
1070 	return false;
1071 }
1072 
1073 /*
1074  * In general we preemptively make sure there's a free entry in every
1075  * node on the spine when doing an insert.  But we can avoid that with
1076  * leaf nodes if we know it's an overwrite.
1077  */
1078 static bool has_space_for_insert(struct btree_node *node, uint64_t key)
1079 {
1080 	if (node->header.nr_entries == node->header.max_entries) {
1081 		if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1082 			/* we don't need space if it's an overwrite */
1083 			return contains_key(node, key);
1084 		}
1085 
1086 		return false;
1087 	}
1088 
1089 	return true;
1090 }
1091 
1092 static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
1093 			    struct dm_btree_value_type *vt,
1094 			    uint64_t key, unsigned int *index)
1095 {
1096 	int r, i = *index, top = 1;
1097 	struct btree_node *node;
1098 
1099 	for (;;) {
1100 		r = shadow_step(s, root, vt);
1101 		if (r < 0)
1102 			return r;
1103 
1104 		node = dm_block_data(shadow_current(s));
1105 
1106 		/*
1107 		 * We have to patch up the parent node, ugly, but I don't
1108 		 * see a way to do this automatically as part of the spine
1109 		 * op.
1110 		 */
1111 		if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
1112 			__le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1113 
1114 			__dm_bless_for_disk(&location);
1115 			memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1116 				    &location, sizeof(__le64));
1117 		}
1118 
1119 		node = dm_block_data(shadow_current(s));
1120 
1121 		if (!has_space_for_insert(node, key)) {
1122 			if (top)
1123 				r = btree_split_beneath(s, key);
1124 			else
1125 				r = rebalance_or_split(s, vt, i, key);
1126 
1127 			if (r < 0)
1128 				return r;
1129 
1130 			/* making space can cause the current node to change */
1131 			node = dm_block_data(shadow_current(s));
1132 		}
1133 
1134 		i = lower_bound(node, key);
1135 
1136 		if (le32_to_cpu(node->header.flags) & LEAF_NODE)
1137 			break;
1138 
1139 		if (i < 0) {
1140 			/* change the bounds on the lowest key */
1141 			node->keys[0] = cpu_to_le64(key);
1142 			i = 0;
1143 		}
1144 
1145 		root = value64(node, i);
1146 		top = 0;
1147 	}
1148 
1149 	if (i < 0 || le64_to_cpu(node->keys[i]) != key)
1150 		i++;
1151 
1152 	*index = i;
1153 	return 0;
1154 }
1155 
1156 static int __btree_get_overwrite_leaf(struct shadow_spine *s, dm_block_t root,
1157 				      uint64_t key, int *index)
1158 {
1159 	int r, i = -1;
1160 	struct btree_node *node;
1161 
1162 	*index = 0;
1163 	for (;;) {
1164 		r = shadow_step(s, root, &s->info->value_type);
1165 		if (r < 0)
1166 			return r;
1167 
1168 		node = dm_block_data(shadow_current(s));
1169 
1170 		/*
1171 		 * We have to patch up the parent node, ugly, but I don't
1172 		 * see a way to do this automatically as part of the spine
1173 		 * op.
1174 		 */
1175 		if (shadow_has_parent(s) && i >= 0) {
1176 			__le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1177 
1178 			__dm_bless_for_disk(&location);
1179 			memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1180 				    &location, sizeof(__le64));
1181 		}
1182 
1183 		node = dm_block_data(shadow_current(s));
1184 		i = lower_bound(node, key);
1185 
1186 		BUG_ON(i < 0);
1187 		BUG_ON(i >= le32_to_cpu(node->header.nr_entries));
1188 
1189 		if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1190 			if (key != le64_to_cpu(node->keys[i]))
1191 				return -EINVAL;
1192 			break;
1193 		}
1194 
1195 		root = value64(node, i);
1196 	}
1197 
1198 	*index = i;
1199 	return 0;
1200 }
1201 
1202 int btree_get_overwrite_leaf(struct dm_btree_info *info, dm_block_t root,
1203 			     uint64_t key, int *index,
1204 			     dm_block_t *new_root, struct dm_block **leaf)
1205 {
1206 	int r;
1207 	struct shadow_spine spine;
1208 
1209 	BUG_ON(info->levels > 1);
1210 	init_shadow_spine(&spine, info);
1211 	r = __btree_get_overwrite_leaf(&spine, root, key, index);
1212 	if (!r) {
1213 		*new_root = shadow_root(&spine);
1214 		*leaf = shadow_current(&spine);
1215 
1216 		/*
1217 		 * Decrement the count so exit_shadow_spine() doesn't
1218 		 * unlock the leaf.
1219 		 */
1220 		spine.count--;
1221 	}
1222 	exit_shadow_spine(&spine);
1223 
1224 	return r;
1225 }
1226 
1227 static bool need_insert(struct btree_node *node, uint64_t *keys,
1228 			unsigned int level, unsigned int index)
1229 {
1230 	return ((index >= le32_to_cpu(node->header.nr_entries)) ||
1231 		(le64_to_cpu(node->keys[index]) != keys[level]));
1232 }
1233 
1234 static int insert(struct dm_btree_info *info, dm_block_t root,
1235 		  uint64_t *keys, void *value, dm_block_t *new_root,
1236 		  int *inserted)
1237 		  __dm_written_to_disk(value)
1238 {
1239 	int r;
1240 	unsigned int level, index = -1, last_level = info->levels - 1;
1241 	dm_block_t block = root;
1242 	struct shadow_spine spine;
1243 	struct btree_node *n;
1244 	struct dm_btree_value_type le64_type;
1245 
1246 	init_le64_type(info->tm, &le64_type);
1247 	init_shadow_spine(&spine, info);
1248 
1249 	for (level = 0; level < (info->levels - 1); level++) {
1250 		r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
1251 		if (r < 0)
1252 			goto bad;
1253 
1254 		n = dm_block_data(shadow_current(&spine));
1255 
1256 		if (need_insert(n, keys, level, index)) {
1257 			dm_block_t new_tree;
1258 			__le64 new_le;
1259 
1260 			r = dm_btree_empty(info, &new_tree);
1261 			if (r < 0)
1262 				goto bad;
1263 
1264 			new_le = cpu_to_le64(new_tree);
1265 			__dm_bless_for_disk(&new_le);
1266 
1267 			r = insert_at(sizeof(uint64_t), n, index,
1268 				      keys[level], &new_le);
1269 			if (r)
1270 				goto bad;
1271 		}
1272 
1273 		if (level < last_level)
1274 			block = value64(n, index);
1275 	}
1276 
1277 	r = btree_insert_raw(&spine, block, &info->value_type,
1278 			     keys[level], &index);
1279 	if (r < 0)
1280 		goto bad;
1281 
1282 	n = dm_block_data(shadow_current(&spine));
1283 
1284 	if (need_insert(n, keys, level, index)) {
1285 		if (inserted)
1286 			*inserted = 1;
1287 
1288 		r = insert_at(info->value_type.size, n, index,
1289 			      keys[level], value);
1290 		if (r)
1291 			goto bad_unblessed;
1292 	} else {
1293 		if (inserted)
1294 			*inserted = 0;
1295 
1296 		if (info->value_type.dec &&
1297 		    (!info->value_type.equal ||
1298 		     !info->value_type.equal(
1299 			     info->value_type.context,
1300 			     value_ptr(n, index),
1301 			     value))) {
1302 			info->value_type.dec(info->value_type.context,
1303 					     value_ptr(n, index), 1);
1304 		}
1305 		memcpy_disk(value_ptr(n, index),
1306 			    value, info->value_type.size);
1307 	}
1308 
1309 	*new_root = shadow_root(&spine);
1310 	exit_shadow_spine(&spine);
1311 
1312 	return 0;
1313 
1314 bad:
1315 	__dm_unbless_for_disk(value);
1316 bad_unblessed:
1317 	exit_shadow_spine(&spine);
1318 	return r;
1319 }
1320 
1321 int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
1322 		    uint64_t *keys, void *value, dm_block_t *new_root)
1323 	__dm_written_to_disk(value)
1324 {
1325 	return insert(info, root, keys, value, new_root, NULL);
1326 }
1327 EXPORT_SYMBOL_GPL(dm_btree_insert);
1328 
1329 int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
1330 			   uint64_t *keys, void *value, dm_block_t *new_root,
1331 			   int *inserted)
1332 	__dm_written_to_disk(value)
1333 {
1334 	return insert(info, root, keys, value, new_root, inserted);
1335 }
1336 EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
1337 
1338 /*----------------------------------------------------------------*/
1339 
1340 static int find_key(struct ro_spine *s, dm_block_t block, bool find_highest,
1341 		    uint64_t *result_key, dm_block_t *next_block)
1342 {
1343 	int i, r;
1344 	uint32_t flags;
1345 
1346 	do {
1347 		r = ro_step(s, block);
1348 		if (r < 0)
1349 			return r;
1350 
1351 		flags = le32_to_cpu(ro_node(s)->header.flags);
1352 		i = le32_to_cpu(ro_node(s)->header.nr_entries);
1353 		if (!i)
1354 			return -ENODATA;
1355 
1356 		i--;
1357 
1358 		if (find_highest)
1359 			*result_key = le64_to_cpu(ro_node(s)->keys[i]);
1360 		else
1361 			*result_key = le64_to_cpu(ro_node(s)->keys[0]);
1362 
1363 		if (next_block || flags & INTERNAL_NODE) {
1364 			if (find_highest)
1365 				block = value64(ro_node(s), i);
1366 			else
1367 				block = value64(ro_node(s), 0);
1368 		}
1369 
1370 	} while (flags & INTERNAL_NODE);
1371 
1372 	if (next_block)
1373 		*next_block = block;
1374 	return 0;
1375 }
1376 
1377 static int dm_btree_find_key(struct dm_btree_info *info, dm_block_t root,
1378 			     bool find_highest, uint64_t *result_keys)
1379 {
1380 	int r = 0, count = 0, level;
1381 	struct ro_spine spine;
1382 
1383 	init_ro_spine(&spine, info);
1384 	for (level = 0; level < info->levels; level++) {
1385 		r = find_key(&spine, root, find_highest, result_keys + level,
1386 			     level == info->levels - 1 ? NULL : &root);
1387 		if (r == -ENODATA) {
1388 			r = 0;
1389 			break;
1390 
1391 		} else if (r)
1392 			break;
1393 
1394 		count++;
1395 	}
1396 	exit_ro_spine(&spine);
1397 
1398 	return r ? r : count;
1399 }
1400 
1401 int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
1402 			      uint64_t *result_keys)
1403 {
1404 	return dm_btree_find_key(info, root, true, result_keys);
1405 }
1406 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
1407 
1408 int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
1409 			     uint64_t *result_keys)
1410 {
1411 	return dm_btree_find_key(info, root, false, result_keys);
1412 }
1413 EXPORT_SYMBOL_GPL(dm_btree_find_lowest_key);
1414 
1415 /*----------------------------------------------------------------*/
1416 
1417 /*
1418  * FIXME: We shouldn't use a recursive algorithm when we have limited stack
1419  * space.  Also this only works for single level trees.
1420  */
1421 static int walk_node(struct dm_btree_info *info, dm_block_t block,
1422 		     int (*fn)(void *context, uint64_t *keys, void *leaf),
1423 		     void *context)
1424 {
1425 	int r;
1426 	unsigned int i, nr;
1427 	struct dm_block *node;
1428 	struct btree_node *n;
1429 	uint64_t keys;
1430 
1431 	r = bn_read_lock(info, block, &node);
1432 	if (r)
1433 		return r;
1434 
1435 	n = dm_block_data(node);
1436 
1437 	nr = le32_to_cpu(n->header.nr_entries);
1438 	for (i = 0; i < nr; i++) {
1439 		if (le32_to_cpu(n->header.flags) & INTERNAL_NODE) {
1440 			r = walk_node(info, value64(n, i), fn, context);
1441 			if (r)
1442 				goto out;
1443 		} else {
1444 			keys = le64_to_cpu(*key_ptr(n, i));
1445 			r = fn(context, &keys, value_ptr(n, i));
1446 			if (r)
1447 				goto out;
1448 		}
1449 	}
1450 
1451 out:
1452 	dm_tm_unlock(info->tm, node);
1453 	return r;
1454 }
1455 
1456 int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
1457 		  int (*fn)(void *context, uint64_t *keys, void *leaf),
1458 		  void *context)
1459 {
1460 	BUG_ON(info->levels > 1);
1461 	return walk_node(info, root, fn, context);
1462 }
1463 EXPORT_SYMBOL_GPL(dm_btree_walk);
1464 
1465 /*----------------------------------------------------------------*/
1466 
1467 static void prefetch_values(struct dm_btree_cursor *c)
1468 {
1469 	unsigned int i, nr;
1470 	__le64 value_le;
1471 	struct cursor_node *n = c->nodes + c->depth - 1;
1472 	struct btree_node *bn = dm_block_data(n->b);
1473 	struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
1474 
1475 	BUG_ON(c->info->value_type.size != sizeof(value_le));
1476 
1477 	nr = le32_to_cpu(bn->header.nr_entries);
1478 	for (i = 0; i < nr; i++) {
1479 		memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
1480 		dm_bm_prefetch(bm, le64_to_cpu(value_le));
1481 	}
1482 }
1483 
1484 static bool leaf_node(struct dm_btree_cursor *c)
1485 {
1486 	struct cursor_node *n = c->nodes + c->depth - 1;
1487 	struct btree_node *bn = dm_block_data(n->b);
1488 
1489 	return le32_to_cpu(bn->header.flags) & LEAF_NODE;
1490 }
1491 
1492 static int push_node(struct dm_btree_cursor *c, dm_block_t b)
1493 {
1494 	int r;
1495 	struct cursor_node *n = c->nodes + c->depth;
1496 
1497 	if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
1498 		DMERR("couldn't push cursor node, stack depth too high");
1499 		return -EINVAL;
1500 	}
1501 
1502 	r = bn_read_lock(c->info, b, &n->b);
1503 	if (r)
1504 		return r;
1505 
1506 	n->index = 0;
1507 	c->depth++;
1508 
1509 	if (c->prefetch_leaves || !leaf_node(c))
1510 		prefetch_values(c);
1511 
1512 	return 0;
1513 }
1514 
1515 static void pop_node(struct dm_btree_cursor *c)
1516 {
1517 	c->depth--;
1518 	unlock_block(c->info, c->nodes[c->depth].b);
1519 }
1520 
1521 static int inc_or_backtrack(struct dm_btree_cursor *c)
1522 {
1523 	struct cursor_node *n;
1524 	struct btree_node *bn;
1525 
1526 	for (;;) {
1527 		if (!c->depth)
1528 			return -ENODATA;
1529 
1530 		n = c->nodes + c->depth - 1;
1531 		bn = dm_block_data(n->b);
1532 
1533 		n->index++;
1534 		if (n->index < le32_to_cpu(bn->header.nr_entries))
1535 			break;
1536 
1537 		pop_node(c);
1538 	}
1539 
1540 	return 0;
1541 }
1542 
1543 static int find_leaf(struct dm_btree_cursor *c)
1544 {
1545 	int r = 0;
1546 	struct cursor_node *n;
1547 	struct btree_node *bn;
1548 	__le64 value_le;
1549 
1550 	for (;;) {
1551 		n = c->nodes + c->depth - 1;
1552 		bn = dm_block_data(n->b);
1553 
1554 		if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
1555 			break;
1556 
1557 		memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
1558 		r = push_node(c, le64_to_cpu(value_le));
1559 		if (r) {
1560 			DMERR("push_node failed");
1561 			break;
1562 		}
1563 	}
1564 
1565 	if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
1566 		return -ENODATA;
1567 
1568 	return r;
1569 }
1570 
1571 int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
1572 			  bool prefetch_leaves, struct dm_btree_cursor *c)
1573 {
1574 	int r;
1575 
1576 	c->info = info;
1577 	c->root = root;
1578 	c->depth = 0;
1579 	c->prefetch_leaves = prefetch_leaves;
1580 
1581 	r = push_node(c, root);
1582 	if (r)
1583 		return r;
1584 
1585 	return find_leaf(c);
1586 }
1587 EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
1588 
1589 void dm_btree_cursor_end(struct dm_btree_cursor *c)
1590 {
1591 	while (c->depth)
1592 		pop_node(c);
1593 }
1594 EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
1595 
1596 int dm_btree_cursor_next(struct dm_btree_cursor *c)
1597 {
1598 	int r = inc_or_backtrack(c);
1599 
1600 	if (!r) {
1601 		r = find_leaf(c);
1602 		if (r)
1603 			DMERR("find_leaf failed");
1604 	}
1605 
1606 	return r;
1607 }
1608 EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
1609 
1610 int dm_btree_cursor_skip(struct dm_btree_cursor *c, uint32_t count)
1611 {
1612 	int r = 0;
1613 
1614 	while (count-- && !r)
1615 		r = dm_btree_cursor_next(c);
1616 
1617 	return r;
1618 }
1619 EXPORT_SYMBOL_GPL(dm_btree_cursor_skip);
1620 
1621 int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
1622 {
1623 	if (c->depth) {
1624 		struct cursor_node *n = c->nodes + c->depth - 1;
1625 		struct btree_node *bn = dm_block_data(n->b);
1626 
1627 		if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
1628 			return -EINVAL;
1629 
1630 		*key = le64_to_cpu(*key_ptr(bn, n->index));
1631 		memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
1632 		return 0;
1633 
1634 	} else
1635 		return -ENODATA;
1636 }
1637 EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);
1638