xref: /linux/drivers/md/persistent-data/dm-btree.c (revision 26fbb4c8c7c3ee9a4c3b4de555a8587b5a19154e)
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 	unsigned i;
75 	uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
76 
77 	if (le32_to_cpu(n->header.flags) & INTERNAL_NODE)
78 		for (i = 0; i < nr_entries; i++)
79 			dm_tm_inc(tm, value64(n, i));
80 	else if (vt->inc)
81 		for (i = 0; i < nr_entries; i++)
82 			vt->inc(vt->context, value_ptr(n, i));
83 }
84 
85 static int insert_at(size_t value_size, struct btree_node *node, unsigned index,
86 		      uint64_t key, void *value)
87 		      __dm_written_to_disk(value)
88 {
89 	uint32_t nr_entries = le32_to_cpu(node->header.nr_entries);
90 	__le64 key_le = cpu_to_le64(key);
91 
92 	if (index > nr_entries ||
93 	    index >= le32_to_cpu(node->header.max_entries)) {
94 		DMERR("too many entries in btree node for insert");
95 		__dm_unbless_for_disk(value);
96 		return -ENOMEM;
97 	}
98 
99 	__dm_bless_for_disk(&key_le);
100 
101 	array_insert(node->keys, sizeof(*node->keys), nr_entries, index, &key_le);
102 	array_insert(value_base(node), value_size, nr_entries, index, value);
103 	node->header.nr_entries = cpu_to_le32(nr_entries + 1);
104 
105 	return 0;
106 }
107 
108 /*----------------------------------------------------------------*/
109 
110 /*
111  * We want 3n entries (for some n).  This works more nicely for repeated
112  * insert remove loops than (2n + 1).
113  */
114 static uint32_t calc_max_entries(size_t value_size, size_t block_size)
115 {
116 	uint32_t total, n;
117 	size_t elt_size = sizeof(uint64_t) + value_size; /* key + value */
118 
119 	block_size -= sizeof(struct node_header);
120 	total = block_size / elt_size;
121 	n = total / 3;		/* rounds down */
122 
123 	return 3 * n;
124 }
125 
126 int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root)
127 {
128 	int r;
129 	struct dm_block *b;
130 	struct btree_node *n;
131 	size_t block_size;
132 	uint32_t max_entries;
133 
134 	r = new_block(info, &b);
135 	if (r < 0)
136 		return r;
137 
138 	block_size = dm_bm_block_size(dm_tm_get_bm(info->tm));
139 	max_entries = calc_max_entries(info->value_type.size, block_size);
140 
141 	n = dm_block_data(b);
142 	memset(n, 0, block_size);
143 	n->header.flags = cpu_to_le32(LEAF_NODE);
144 	n->header.nr_entries = cpu_to_le32(0);
145 	n->header.max_entries = cpu_to_le32(max_entries);
146 	n->header.value_size = cpu_to_le32(info->value_type.size);
147 
148 	*root = dm_block_location(b);
149 	unlock_block(info, b);
150 
151 	return 0;
152 }
153 EXPORT_SYMBOL_GPL(dm_btree_empty);
154 
155 /*----------------------------------------------------------------*/
156 
157 /*
158  * Deletion uses a recursive algorithm, since we have limited stack space
159  * we explicitly manage our own stack on the heap.
160  */
161 #define MAX_SPINE_DEPTH 64
162 struct frame {
163 	struct dm_block *b;
164 	struct btree_node *n;
165 	unsigned level;
166 	unsigned nr_children;
167 	unsigned current_child;
168 };
169 
170 struct del_stack {
171 	struct dm_btree_info *info;
172 	struct dm_transaction_manager *tm;
173 	int top;
174 	struct frame spine[MAX_SPINE_DEPTH];
175 };
176 
177 static int top_frame(struct del_stack *s, struct frame **f)
178 {
179 	if (s->top < 0) {
180 		DMERR("btree deletion stack empty");
181 		return -EINVAL;
182 	}
183 
184 	*f = s->spine + s->top;
185 
186 	return 0;
187 }
188 
189 static int unprocessed_frames(struct del_stack *s)
190 {
191 	return s->top >= 0;
192 }
193 
194 static void prefetch_children(struct del_stack *s, struct frame *f)
195 {
196 	unsigned i;
197 	struct dm_block_manager *bm = dm_tm_get_bm(s->tm);
198 
199 	for (i = 0; i < f->nr_children; i++)
200 		dm_bm_prefetch(bm, value64(f->n, i));
201 }
202 
203 static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
204 {
205 	return f->level < (info->levels - 1);
206 }
207 
208 static int push_frame(struct del_stack *s, dm_block_t b, unsigned level)
209 {
210 	int r;
211 	uint32_t ref_count;
212 
213 	if (s->top >= MAX_SPINE_DEPTH - 1) {
214 		DMERR("btree deletion stack out of memory");
215 		return -ENOMEM;
216 	}
217 
218 	r = dm_tm_ref(s->tm, b, &ref_count);
219 	if (r)
220 		return r;
221 
222 	if (ref_count > 1)
223 		/*
224 		 * This is a shared node, so we can just decrement it's
225 		 * reference counter and leave the children.
226 		 */
227 		dm_tm_dec(s->tm, b);
228 
229 	else {
230 		uint32_t flags;
231 		struct frame *f = s->spine + ++s->top;
232 
233 		r = dm_tm_read_lock(s->tm, b, &btree_node_validator, &f->b);
234 		if (r) {
235 			s->top--;
236 			return r;
237 		}
238 
239 		f->n = dm_block_data(f->b);
240 		f->level = level;
241 		f->nr_children = le32_to_cpu(f->n->header.nr_entries);
242 		f->current_child = 0;
243 
244 		flags = le32_to_cpu(f->n->header.flags);
245 		if (flags & INTERNAL_NODE || is_internal_level(s->info, f))
246 			prefetch_children(s, f);
247 	}
248 
249 	return 0;
250 }
251 
252 static void pop_frame(struct del_stack *s)
253 {
254 	struct frame *f = s->spine + s->top--;
255 
256 	dm_tm_dec(s->tm, dm_block_location(f->b));
257 	dm_tm_unlock(s->tm, f->b);
258 }
259 
260 static void unlock_all_frames(struct del_stack *s)
261 {
262 	struct frame *f;
263 
264 	while (unprocessed_frames(s)) {
265 		f = s->spine + s->top--;
266 		dm_tm_unlock(s->tm, f->b);
267 	}
268 }
269 
270 int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
271 {
272 	int r;
273 	struct del_stack *s;
274 
275 	/*
276 	 * dm_btree_del() is called via an ioctl, as such should be
277 	 * considered an FS op.  We can't recurse back into the FS, so we
278 	 * allocate GFP_NOFS.
279 	 */
280 	s = kmalloc(sizeof(*s), GFP_NOFS);
281 	if (!s)
282 		return -ENOMEM;
283 	s->info = info;
284 	s->tm = info->tm;
285 	s->top = -1;
286 
287 	r = push_frame(s, root, 0);
288 	if (r)
289 		goto out;
290 
291 	while (unprocessed_frames(s)) {
292 		uint32_t flags;
293 		struct frame *f;
294 		dm_block_t b;
295 
296 		r = top_frame(s, &f);
297 		if (r)
298 			goto out;
299 
300 		if (f->current_child >= f->nr_children) {
301 			pop_frame(s);
302 			continue;
303 		}
304 
305 		flags = le32_to_cpu(f->n->header.flags);
306 		if (flags & INTERNAL_NODE) {
307 			b = value64(f->n, f->current_child);
308 			f->current_child++;
309 			r = push_frame(s, b, f->level);
310 			if (r)
311 				goto out;
312 
313 		} else if (is_internal_level(info, f)) {
314 			b = value64(f->n, f->current_child);
315 			f->current_child++;
316 			r = push_frame(s, b, f->level + 1);
317 			if (r)
318 				goto out;
319 
320 		} else {
321 			if (info->value_type.dec) {
322 				unsigned i;
323 
324 				for (i = 0; i < f->nr_children; i++)
325 					info->value_type.dec(info->value_type.context,
326 							     value_ptr(f->n, i));
327 			}
328 			pop_frame(s);
329 		}
330 	}
331 out:
332 	if (r) {
333 		/* cleanup all frames of del_stack */
334 		unlock_all_frames(s);
335 	}
336 	kfree(s);
337 
338 	return r;
339 }
340 EXPORT_SYMBOL_GPL(dm_btree_del);
341 
342 /*----------------------------------------------------------------*/
343 
344 static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
345 			    int (*search_fn)(struct btree_node *, uint64_t),
346 			    uint64_t *result_key, void *v, size_t value_size)
347 {
348 	int i, r;
349 	uint32_t flags, nr_entries;
350 
351 	do {
352 		r = ro_step(s, block);
353 		if (r < 0)
354 			return r;
355 
356 		i = search_fn(ro_node(s), key);
357 
358 		flags = le32_to_cpu(ro_node(s)->header.flags);
359 		nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
360 		if (i < 0 || i >= nr_entries)
361 			return -ENODATA;
362 
363 		if (flags & INTERNAL_NODE)
364 			block = value64(ro_node(s), i);
365 
366 	} while (!(flags & LEAF_NODE));
367 
368 	*result_key = le64_to_cpu(ro_node(s)->keys[i]);
369 	if (v)
370 		memcpy(v, value_ptr(ro_node(s), i), value_size);
371 
372 	return 0;
373 }
374 
375 int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
376 		    uint64_t *keys, void *value_le)
377 {
378 	unsigned level, last_level = info->levels - 1;
379 	int r = -ENODATA;
380 	uint64_t rkey;
381 	__le64 internal_value_le;
382 	struct ro_spine spine;
383 
384 	init_ro_spine(&spine, info);
385 	for (level = 0; level < info->levels; level++) {
386 		size_t size;
387 		void *value_p;
388 
389 		if (level == last_level) {
390 			value_p = value_le;
391 			size = info->value_type.size;
392 
393 		} else {
394 			value_p = &internal_value_le;
395 			size = sizeof(uint64_t);
396 		}
397 
398 		r = btree_lookup_raw(&spine, root, keys[level],
399 				     lower_bound, &rkey,
400 				     value_p, size);
401 
402 		if (!r) {
403 			if (rkey != keys[level]) {
404 				exit_ro_spine(&spine);
405 				return -ENODATA;
406 			}
407 		} else {
408 			exit_ro_spine(&spine);
409 			return r;
410 		}
411 
412 		root = le64_to_cpu(internal_value_le);
413 	}
414 	exit_ro_spine(&spine);
415 
416 	return r;
417 }
418 EXPORT_SYMBOL_GPL(dm_btree_lookup);
419 
420 static int dm_btree_lookup_next_single(struct dm_btree_info *info, dm_block_t root,
421 				       uint64_t key, uint64_t *rkey, void *value_le)
422 {
423 	int r, i;
424 	uint32_t flags, nr_entries;
425 	struct dm_block *node;
426 	struct btree_node *n;
427 
428 	r = bn_read_lock(info, root, &node);
429 	if (r)
430 		return r;
431 
432 	n = dm_block_data(node);
433 	flags = le32_to_cpu(n->header.flags);
434 	nr_entries = le32_to_cpu(n->header.nr_entries);
435 
436 	if (flags & INTERNAL_NODE) {
437 		i = lower_bound(n, key);
438 		if (i < 0) {
439 			/*
440 			 * avoid early -ENODATA return when all entries are
441 			 * higher than the search @key.
442 			 */
443 			i = 0;
444 		}
445 		if (i >= nr_entries) {
446 			r = -ENODATA;
447 			goto out;
448 		}
449 
450 		r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
451 		if (r == -ENODATA && i < (nr_entries - 1)) {
452 			i++;
453 			r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
454 		}
455 
456 	} else {
457 		i = upper_bound(n, key);
458 		if (i < 0 || i >= nr_entries) {
459 			r = -ENODATA;
460 			goto out;
461 		}
462 
463 		*rkey = le64_to_cpu(n->keys[i]);
464 		memcpy(value_le, value_ptr(n, i), info->value_type.size);
465 	}
466 out:
467 	dm_tm_unlock(info->tm, node);
468 	return r;
469 }
470 
471 int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
472 			 uint64_t *keys, uint64_t *rkey, void *value_le)
473 {
474 	unsigned level;
475 	int r = -ENODATA;
476 	__le64 internal_value_le;
477 	struct ro_spine spine;
478 
479 	init_ro_spine(&spine, info);
480 	for (level = 0; level < info->levels - 1u; level++) {
481 		r = btree_lookup_raw(&spine, root, keys[level],
482 				     lower_bound, rkey,
483 				     &internal_value_le, sizeof(uint64_t));
484 		if (r)
485 			goto out;
486 
487 		if (*rkey != keys[level]) {
488 			r = -ENODATA;
489 			goto out;
490 		}
491 
492 		root = le64_to_cpu(internal_value_le);
493 	}
494 
495 	r = dm_btree_lookup_next_single(info, root, keys[level], rkey, value_le);
496 out:
497 	exit_ro_spine(&spine);
498 	return r;
499 }
500 
501 EXPORT_SYMBOL_GPL(dm_btree_lookup_next);
502 
503 /*
504  * Splits a node by creating a sibling node and shifting half the nodes
505  * contents across.  Assumes there is a parent node, and it has room for
506  * another child.
507  *
508  * Before:
509  *	  +--------+
510  *	  | Parent |
511  *	  +--------+
512  *	     |
513  *	     v
514  *	+----------+
515  *	| A ++++++ |
516  *	+----------+
517  *
518  *
519  * After:
520  *		+--------+
521  *		| Parent |
522  *		+--------+
523  *		  |	|
524  *		  v	+------+
525  *	    +---------+	       |
526  *	    | A* +++  |	       v
527  *	    +---------+	  +-------+
528  *			  | B +++ |
529  *			  +-------+
530  *
531  * Where A* is a shadow of A.
532  */
533 static int btree_split_sibling(struct shadow_spine *s, unsigned parent_index,
534 			       uint64_t key)
535 {
536 	int r;
537 	size_t size;
538 	unsigned nr_left, nr_right;
539 	struct dm_block *left, *right, *parent;
540 	struct btree_node *ln, *rn, *pn;
541 	__le64 location;
542 
543 	left = shadow_current(s);
544 
545 	r = new_block(s->info, &right);
546 	if (r < 0)
547 		return r;
548 
549 	ln = dm_block_data(left);
550 	rn = dm_block_data(right);
551 
552 	nr_left = le32_to_cpu(ln->header.nr_entries) / 2;
553 	nr_right = le32_to_cpu(ln->header.nr_entries) - nr_left;
554 
555 	ln->header.nr_entries = cpu_to_le32(nr_left);
556 
557 	rn->header.flags = ln->header.flags;
558 	rn->header.nr_entries = cpu_to_le32(nr_right);
559 	rn->header.max_entries = ln->header.max_entries;
560 	rn->header.value_size = ln->header.value_size;
561 	memcpy(rn->keys, ln->keys + nr_left, nr_right * sizeof(rn->keys[0]));
562 
563 	size = le32_to_cpu(ln->header.flags) & INTERNAL_NODE ?
564 		sizeof(uint64_t) : s->info->value_type.size;
565 	memcpy(value_ptr(rn, 0), value_ptr(ln, nr_left),
566 	       size * nr_right);
567 
568 	/*
569 	 * Patch up the parent
570 	 */
571 	parent = shadow_parent(s);
572 
573 	pn = dm_block_data(parent);
574 	location = cpu_to_le64(dm_block_location(left));
575 	__dm_bless_for_disk(&location);
576 	memcpy_disk(value_ptr(pn, parent_index),
577 		    &location, sizeof(__le64));
578 
579 	location = cpu_to_le64(dm_block_location(right));
580 	__dm_bless_for_disk(&location);
581 
582 	r = insert_at(sizeof(__le64), pn, parent_index + 1,
583 		      le64_to_cpu(rn->keys[0]), &location);
584 	if (r) {
585 		unlock_block(s->info, right);
586 		return r;
587 	}
588 
589 	if (key < le64_to_cpu(rn->keys[0])) {
590 		unlock_block(s->info, right);
591 		s->nodes[1] = left;
592 	} else {
593 		unlock_block(s->info, left);
594 		s->nodes[1] = right;
595 	}
596 
597 	return 0;
598 }
599 
600 /*
601  * Splits a node by creating two new children beneath the given node.
602  *
603  * Before:
604  *	  +----------+
605  *	  | A ++++++ |
606  *	  +----------+
607  *
608  *
609  * After:
610  *	+------------+
611  *	| A (shadow) |
612  *	+------------+
613  *	    |	|
614  *   +------+	+----+
615  *   |		     |
616  *   v		     v
617  * +-------+	 +-------+
618  * | B +++ |	 | C +++ |
619  * +-------+	 +-------+
620  */
621 static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
622 {
623 	int r;
624 	size_t size;
625 	unsigned nr_left, nr_right;
626 	struct dm_block *left, *right, *new_parent;
627 	struct btree_node *pn, *ln, *rn;
628 	__le64 val;
629 
630 	new_parent = shadow_current(s);
631 
632 	pn = dm_block_data(new_parent);
633 	size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
634 		sizeof(__le64) : s->info->value_type.size;
635 
636 	/* create & init the left block */
637 	r = new_block(s->info, &left);
638 	if (r < 0)
639 		return r;
640 
641 	ln = dm_block_data(left);
642 	nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
643 
644 	ln->header.flags = pn->header.flags;
645 	ln->header.nr_entries = cpu_to_le32(nr_left);
646 	ln->header.max_entries = pn->header.max_entries;
647 	ln->header.value_size = pn->header.value_size;
648 	memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
649 	memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
650 
651 	/* create & init the right block */
652 	r = new_block(s->info, &right);
653 	if (r < 0) {
654 		unlock_block(s->info, left);
655 		return r;
656 	}
657 
658 	rn = dm_block_data(right);
659 	nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
660 
661 	rn->header.flags = pn->header.flags;
662 	rn->header.nr_entries = cpu_to_le32(nr_right);
663 	rn->header.max_entries = pn->header.max_entries;
664 	rn->header.value_size = pn->header.value_size;
665 	memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
666 	memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
667 	       nr_right * size);
668 
669 	/* new_parent should just point to l and r now */
670 	pn->header.flags = cpu_to_le32(INTERNAL_NODE);
671 	pn->header.nr_entries = cpu_to_le32(2);
672 	pn->header.max_entries = cpu_to_le32(
673 		calc_max_entries(sizeof(__le64),
674 				 dm_bm_block_size(
675 					 dm_tm_get_bm(s->info->tm))));
676 	pn->header.value_size = cpu_to_le32(sizeof(__le64));
677 
678 	val = cpu_to_le64(dm_block_location(left));
679 	__dm_bless_for_disk(&val);
680 	pn->keys[0] = ln->keys[0];
681 	memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
682 
683 	val = cpu_to_le64(dm_block_location(right));
684 	__dm_bless_for_disk(&val);
685 	pn->keys[1] = rn->keys[0];
686 	memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
687 
688 	unlock_block(s->info, left);
689 	unlock_block(s->info, right);
690 	return 0;
691 }
692 
693 static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
694 			    struct dm_btree_value_type *vt,
695 			    uint64_t key, unsigned *index)
696 {
697 	int r, i = *index, top = 1;
698 	struct btree_node *node;
699 
700 	for (;;) {
701 		r = shadow_step(s, root, vt);
702 		if (r < 0)
703 			return r;
704 
705 		node = dm_block_data(shadow_current(s));
706 
707 		/*
708 		 * We have to patch up the parent node, ugly, but I don't
709 		 * see a way to do this automatically as part of the spine
710 		 * op.
711 		 */
712 		if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
713 			__le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
714 
715 			__dm_bless_for_disk(&location);
716 			memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
717 				    &location, sizeof(__le64));
718 		}
719 
720 		node = dm_block_data(shadow_current(s));
721 
722 		if (node->header.nr_entries == node->header.max_entries) {
723 			if (top)
724 				r = btree_split_beneath(s, key);
725 			else
726 				r = btree_split_sibling(s, i, key);
727 
728 			if (r < 0)
729 				return r;
730 		}
731 
732 		node = dm_block_data(shadow_current(s));
733 
734 		i = lower_bound(node, key);
735 
736 		if (le32_to_cpu(node->header.flags) & LEAF_NODE)
737 			break;
738 
739 		if (i < 0) {
740 			/* change the bounds on the lowest key */
741 			node->keys[0] = cpu_to_le64(key);
742 			i = 0;
743 		}
744 
745 		root = value64(node, i);
746 		top = 0;
747 	}
748 
749 	if (i < 0 || le64_to_cpu(node->keys[i]) != key)
750 		i++;
751 
752 	*index = i;
753 	return 0;
754 }
755 
756 static bool need_insert(struct btree_node *node, uint64_t *keys,
757 			unsigned level, unsigned index)
758 {
759         return ((index >= le32_to_cpu(node->header.nr_entries)) ||
760 		(le64_to_cpu(node->keys[index]) != keys[level]));
761 }
762 
763 static int insert(struct dm_btree_info *info, dm_block_t root,
764 		  uint64_t *keys, void *value, dm_block_t *new_root,
765 		  int *inserted)
766 		  __dm_written_to_disk(value)
767 {
768 	int r;
769 	unsigned level, index = -1, last_level = info->levels - 1;
770 	dm_block_t block = root;
771 	struct shadow_spine spine;
772 	struct btree_node *n;
773 	struct dm_btree_value_type le64_type;
774 
775 	init_le64_type(info->tm, &le64_type);
776 	init_shadow_spine(&spine, info);
777 
778 	for (level = 0; level < (info->levels - 1); level++) {
779 		r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
780 		if (r < 0)
781 			goto bad;
782 
783 		n = dm_block_data(shadow_current(&spine));
784 
785 		if (need_insert(n, keys, level, index)) {
786 			dm_block_t new_tree;
787 			__le64 new_le;
788 
789 			r = dm_btree_empty(info, &new_tree);
790 			if (r < 0)
791 				goto bad;
792 
793 			new_le = cpu_to_le64(new_tree);
794 			__dm_bless_for_disk(&new_le);
795 
796 			r = insert_at(sizeof(uint64_t), n, index,
797 				      keys[level], &new_le);
798 			if (r)
799 				goto bad;
800 		}
801 
802 		if (level < last_level)
803 			block = value64(n, index);
804 	}
805 
806 	r = btree_insert_raw(&spine, block, &info->value_type,
807 			     keys[level], &index);
808 	if (r < 0)
809 		goto bad;
810 
811 	n = dm_block_data(shadow_current(&spine));
812 
813 	if (need_insert(n, keys, level, index)) {
814 		if (inserted)
815 			*inserted = 1;
816 
817 		r = insert_at(info->value_type.size, n, index,
818 			      keys[level], value);
819 		if (r)
820 			goto bad_unblessed;
821 	} else {
822 		if (inserted)
823 			*inserted = 0;
824 
825 		if (info->value_type.dec &&
826 		    (!info->value_type.equal ||
827 		     !info->value_type.equal(
828 			     info->value_type.context,
829 			     value_ptr(n, index),
830 			     value))) {
831 			info->value_type.dec(info->value_type.context,
832 					     value_ptr(n, index));
833 		}
834 		memcpy_disk(value_ptr(n, index),
835 			    value, info->value_type.size);
836 	}
837 
838 	*new_root = shadow_root(&spine);
839 	exit_shadow_spine(&spine);
840 
841 	return 0;
842 
843 bad:
844 	__dm_unbless_for_disk(value);
845 bad_unblessed:
846 	exit_shadow_spine(&spine);
847 	return r;
848 }
849 
850 int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
851 		    uint64_t *keys, void *value, dm_block_t *new_root)
852 		    __dm_written_to_disk(value)
853 {
854 	return insert(info, root, keys, value, new_root, NULL);
855 }
856 EXPORT_SYMBOL_GPL(dm_btree_insert);
857 
858 int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
859 			   uint64_t *keys, void *value, dm_block_t *new_root,
860 			   int *inserted)
861 			   __dm_written_to_disk(value)
862 {
863 	return insert(info, root, keys, value, new_root, inserted);
864 }
865 EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
866 
867 /*----------------------------------------------------------------*/
868 
869 static int find_key(struct ro_spine *s, dm_block_t block, bool find_highest,
870 		    uint64_t *result_key, dm_block_t *next_block)
871 {
872 	int i, r;
873 	uint32_t flags;
874 
875 	do {
876 		r = ro_step(s, block);
877 		if (r < 0)
878 			return r;
879 
880 		flags = le32_to_cpu(ro_node(s)->header.flags);
881 		i = le32_to_cpu(ro_node(s)->header.nr_entries);
882 		if (!i)
883 			return -ENODATA;
884 		else
885 			i--;
886 
887 		if (find_highest)
888 			*result_key = le64_to_cpu(ro_node(s)->keys[i]);
889 		else
890 			*result_key = le64_to_cpu(ro_node(s)->keys[0]);
891 
892 		if (next_block || flags & INTERNAL_NODE) {
893 			if (find_highest)
894 				block = value64(ro_node(s), i);
895 			else
896 				block = value64(ro_node(s), 0);
897 		}
898 
899 	} while (flags & INTERNAL_NODE);
900 
901 	if (next_block)
902 		*next_block = block;
903 	return 0;
904 }
905 
906 static int dm_btree_find_key(struct dm_btree_info *info, dm_block_t root,
907 			     bool find_highest, uint64_t *result_keys)
908 {
909 	int r = 0, count = 0, level;
910 	struct ro_spine spine;
911 
912 	init_ro_spine(&spine, info);
913 	for (level = 0; level < info->levels; level++) {
914 		r = find_key(&spine, root, find_highest, result_keys + level,
915 			     level == info->levels - 1 ? NULL : &root);
916 		if (r == -ENODATA) {
917 			r = 0;
918 			break;
919 
920 		} else if (r)
921 			break;
922 
923 		count++;
924 	}
925 	exit_ro_spine(&spine);
926 
927 	return r ? r : count;
928 }
929 
930 int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
931 			      uint64_t *result_keys)
932 {
933 	return dm_btree_find_key(info, root, true, result_keys);
934 }
935 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
936 
937 int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
938 			     uint64_t *result_keys)
939 {
940 	return dm_btree_find_key(info, root, false, result_keys);
941 }
942 EXPORT_SYMBOL_GPL(dm_btree_find_lowest_key);
943 
944 /*----------------------------------------------------------------*/
945 
946 /*
947  * FIXME: We shouldn't use a recursive algorithm when we have limited stack
948  * space.  Also this only works for single level trees.
949  */
950 static int walk_node(struct dm_btree_info *info, dm_block_t block,
951 		     int (*fn)(void *context, uint64_t *keys, void *leaf),
952 		     void *context)
953 {
954 	int r;
955 	unsigned i, nr;
956 	struct dm_block *node;
957 	struct btree_node *n;
958 	uint64_t keys;
959 
960 	r = bn_read_lock(info, block, &node);
961 	if (r)
962 		return r;
963 
964 	n = dm_block_data(node);
965 
966 	nr = le32_to_cpu(n->header.nr_entries);
967 	for (i = 0; i < nr; i++) {
968 		if (le32_to_cpu(n->header.flags) & INTERNAL_NODE) {
969 			r = walk_node(info, value64(n, i), fn, context);
970 			if (r)
971 				goto out;
972 		} else {
973 			keys = le64_to_cpu(*key_ptr(n, i));
974 			r = fn(context, &keys, value_ptr(n, i));
975 			if (r)
976 				goto out;
977 		}
978 	}
979 
980 out:
981 	dm_tm_unlock(info->tm, node);
982 	return r;
983 }
984 
985 int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
986 		  int (*fn)(void *context, uint64_t *keys, void *leaf),
987 		  void *context)
988 {
989 	BUG_ON(info->levels > 1);
990 	return walk_node(info, root, fn, context);
991 }
992 EXPORT_SYMBOL_GPL(dm_btree_walk);
993 
994 /*----------------------------------------------------------------*/
995 
996 static void prefetch_values(struct dm_btree_cursor *c)
997 {
998 	unsigned i, nr;
999 	__le64 value_le;
1000 	struct cursor_node *n = c->nodes + c->depth - 1;
1001 	struct btree_node *bn = dm_block_data(n->b);
1002 	struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
1003 
1004 	BUG_ON(c->info->value_type.size != sizeof(value_le));
1005 
1006 	nr = le32_to_cpu(bn->header.nr_entries);
1007 	for (i = 0; i < nr; i++) {
1008 		memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
1009 		dm_bm_prefetch(bm, le64_to_cpu(value_le));
1010 	}
1011 }
1012 
1013 static bool leaf_node(struct dm_btree_cursor *c)
1014 {
1015 	struct cursor_node *n = c->nodes + c->depth - 1;
1016 	struct btree_node *bn = dm_block_data(n->b);
1017 
1018 	return le32_to_cpu(bn->header.flags) & LEAF_NODE;
1019 }
1020 
1021 static int push_node(struct dm_btree_cursor *c, dm_block_t b)
1022 {
1023 	int r;
1024 	struct cursor_node *n = c->nodes + c->depth;
1025 
1026 	if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
1027 		DMERR("couldn't push cursor node, stack depth too high");
1028 		return -EINVAL;
1029 	}
1030 
1031 	r = bn_read_lock(c->info, b, &n->b);
1032 	if (r)
1033 		return r;
1034 
1035 	n->index = 0;
1036 	c->depth++;
1037 
1038 	if (c->prefetch_leaves || !leaf_node(c))
1039 		prefetch_values(c);
1040 
1041 	return 0;
1042 }
1043 
1044 static void pop_node(struct dm_btree_cursor *c)
1045 {
1046 	c->depth--;
1047 	unlock_block(c->info, c->nodes[c->depth].b);
1048 }
1049 
1050 static int inc_or_backtrack(struct dm_btree_cursor *c)
1051 {
1052 	struct cursor_node *n;
1053 	struct btree_node *bn;
1054 
1055 	for (;;) {
1056 		if (!c->depth)
1057 			return -ENODATA;
1058 
1059 		n = c->nodes + c->depth - 1;
1060 		bn = dm_block_data(n->b);
1061 
1062 		n->index++;
1063 		if (n->index < le32_to_cpu(bn->header.nr_entries))
1064 			break;
1065 
1066 		pop_node(c);
1067 	}
1068 
1069 	return 0;
1070 }
1071 
1072 static int find_leaf(struct dm_btree_cursor *c)
1073 {
1074 	int r = 0;
1075 	struct cursor_node *n;
1076 	struct btree_node *bn;
1077 	__le64 value_le;
1078 
1079 	for (;;) {
1080 		n = c->nodes + c->depth - 1;
1081 		bn = dm_block_data(n->b);
1082 
1083 		if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
1084 			break;
1085 
1086 		memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
1087 		r = push_node(c, le64_to_cpu(value_le));
1088 		if (r) {
1089 			DMERR("push_node failed");
1090 			break;
1091 		}
1092 	}
1093 
1094 	if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
1095 		return -ENODATA;
1096 
1097 	return r;
1098 }
1099 
1100 int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
1101 			  bool prefetch_leaves, struct dm_btree_cursor *c)
1102 {
1103 	int r;
1104 
1105 	c->info = info;
1106 	c->root = root;
1107 	c->depth = 0;
1108 	c->prefetch_leaves = prefetch_leaves;
1109 
1110 	r = push_node(c, root);
1111 	if (r)
1112 		return r;
1113 
1114 	return find_leaf(c);
1115 }
1116 EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
1117 
1118 void dm_btree_cursor_end(struct dm_btree_cursor *c)
1119 {
1120 	while (c->depth)
1121 		pop_node(c);
1122 }
1123 EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
1124 
1125 int dm_btree_cursor_next(struct dm_btree_cursor *c)
1126 {
1127 	int r = inc_or_backtrack(c);
1128 	if (!r) {
1129 		r = find_leaf(c);
1130 		if (r)
1131 			DMERR("find_leaf failed");
1132 	}
1133 
1134 	return r;
1135 }
1136 EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
1137 
1138 int dm_btree_cursor_skip(struct dm_btree_cursor *c, uint32_t count)
1139 {
1140 	int r = 0;
1141 
1142 	while (count-- && !r)
1143 		r = dm_btree_cursor_next(c);
1144 
1145 	return r;
1146 }
1147 EXPORT_SYMBOL_GPL(dm_btree_cursor_skip);
1148 
1149 int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
1150 {
1151 	if (c->depth) {
1152 		struct cursor_node *n = c->nodes + c->depth - 1;
1153 		struct btree_node *bn = dm_block_data(n->b);
1154 
1155 		if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
1156 			return -EINVAL;
1157 
1158 		*key = le64_to_cpu(*key_ptr(bn, n->index));
1159 		memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
1160 		return 0;
1161 
1162 	} else
1163 		return -ENODATA;
1164 }
1165 EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);
1166