Linux Audio

Check our new training course

Loading...
v6.2
 
   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 *--------------------------------------------------------------*/
 
  19static 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
  26static 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. */
  41static 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
  61int lower_bound(struct btree_node *n, uint64_t key)
  62{
  63	return bsearch(n, key, 0);
  64}
  65
  66static int upper_bound(struct btree_node *n, uint64_t key)
  67{
  68	return bsearch(n, key, 1);
  69}
  70
  71void 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
  83static 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	uint32_t max_entries = le32_to_cpu(node->header.max_entries);
  89	__le64 key_le = cpu_to_le64(key);
  90
  91	if (index > nr_entries ||
  92	    index >= max_entries ||
  93	    nr_entries >= 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 */
 114static 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
 126int 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}
 153EXPORT_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
 162struct frame {
 163	struct dm_block *b;
 164	struct btree_node *n;
 165	unsigned level;
 166	unsigned nr_children;
 167	unsigned current_child;
 168};
 169
 170struct 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
 177static 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
 189static int unprocessed_frames(struct del_stack *s)
 190{
 191	return s->top >= 0;
 192}
 193
 194static 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
 203static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
 204{
 205	return f->level < (info->levels - 1);
 206}
 207
 208static 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
 252static 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
 260static 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
 270int 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				info->value_type.dec(info->value_type.context,
 323						     value_ptr(f->n, 0), f->nr_children);
 324			pop_frame(s);
 325		}
 326	}
 327out:
 328	if (r) {
 329		/* cleanup all frames of del_stack */
 330		unlock_all_frames(s);
 331	}
 332	kfree(s);
 333
 334	return r;
 335}
 336EXPORT_SYMBOL_GPL(dm_btree_del);
 337
 338/*----------------------------------------------------------------*/
 339
 340static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
 341			    int (*search_fn)(struct btree_node *, uint64_t),
 342			    uint64_t *result_key, void *v, size_t value_size)
 343{
 344	int i, r;
 345	uint32_t flags, nr_entries;
 346
 347	do {
 348		r = ro_step(s, block);
 349		if (r < 0)
 350			return r;
 351
 352		i = search_fn(ro_node(s), key);
 353
 354		flags = le32_to_cpu(ro_node(s)->header.flags);
 355		nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
 356		if (i < 0 || i >= nr_entries)
 357			return -ENODATA;
 358
 359		if (flags & INTERNAL_NODE)
 360			block = value64(ro_node(s), i);
 361
 362	} while (!(flags & LEAF_NODE));
 363
 364	*result_key = le64_to_cpu(ro_node(s)->keys[i]);
 365	if (v)
 366		memcpy(v, value_ptr(ro_node(s), i), value_size);
 367
 368	return 0;
 369}
 370
 371int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
 372		    uint64_t *keys, void *value_le)
 373{
 374	unsigned level, last_level = info->levels - 1;
 375	int r = -ENODATA;
 376	uint64_t rkey;
 377	__le64 internal_value_le;
 378	struct ro_spine spine;
 379
 380	init_ro_spine(&spine, info);
 381	for (level = 0; level < info->levels; level++) {
 382		size_t size;
 383		void *value_p;
 384
 385		if (level == last_level) {
 386			value_p = value_le;
 387			size = info->value_type.size;
 388
 389		} else {
 390			value_p = &internal_value_le;
 391			size = sizeof(uint64_t);
 392		}
 393
 394		r = btree_lookup_raw(&spine, root, keys[level],
 395				     lower_bound, &rkey,
 396				     value_p, size);
 397
 398		if (!r) {
 399			if (rkey != keys[level]) {
 400				exit_ro_spine(&spine);
 401				return -ENODATA;
 402			}
 403		} else {
 404			exit_ro_spine(&spine);
 405			return r;
 406		}
 407
 408		root = le64_to_cpu(internal_value_le);
 409	}
 410	exit_ro_spine(&spine);
 411
 412	return r;
 413}
 414EXPORT_SYMBOL_GPL(dm_btree_lookup);
 415
 416static int dm_btree_lookup_next_single(struct dm_btree_info *info, dm_block_t root,
 417				       uint64_t key, uint64_t *rkey, void *value_le)
 418{
 419	int r, i;
 420	uint32_t flags, nr_entries;
 421	struct dm_block *node;
 422	struct btree_node *n;
 423
 424	r = bn_read_lock(info, root, &node);
 425	if (r)
 426		return r;
 427
 428	n = dm_block_data(node);
 429	flags = le32_to_cpu(n->header.flags);
 430	nr_entries = le32_to_cpu(n->header.nr_entries);
 431
 432	if (flags & INTERNAL_NODE) {
 433		i = lower_bound(n, key);
 434		if (i < 0) {
 435			/*
 436			 * avoid early -ENODATA return when all entries are
 437			 * higher than the search @key.
 438			 */
 439			i = 0;
 440		}
 441		if (i >= nr_entries) {
 442			r = -ENODATA;
 443			goto out;
 444		}
 445
 446		r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
 447		if (r == -ENODATA && i < (nr_entries - 1)) {
 448			i++;
 449			r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
 450		}
 451
 452	} else {
 453		i = upper_bound(n, key);
 454		if (i < 0 || i >= nr_entries) {
 455			r = -ENODATA;
 456			goto out;
 457		}
 458
 459		*rkey = le64_to_cpu(n->keys[i]);
 460		memcpy(value_le, value_ptr(n, i), info->value_type.size);
 461	}
 462out:
 463	dm_tm_unlock(info->tm, node);
 464	return r;
 465}
 466
 467int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
 468			 uint64_t *keys, uint64_t *rkey, void *value_le)
 469{
 470	unsigned level;
 471	int r = -ENODATA;
 472	__le64 internal_value_le;
 473	struct ro_spine spine;
 474
 475	init_ro_spine(&spine, info);
 476	for (level = 0; level < info->levels - 1u; level++) {
 477		r = btree_lookup_raw(&spine, root, keys[level],
 478				     lower_bound, rkey,
 479				     &internal_value_le, sizeof(uint64_t));
 480		if (r)
 481			goto out;
 482
 483		if (*rkey != keys[level]) {
 484			r = -ENODATA;
 485			goto out;
 486		}
 487
 488		root = le64_to_cpu(internal_value_le);
 489	}
 490
 491	r = dm_btree_lookup_next_single(info, root, keys[level], rkey, value_le);
 492out:
 493	exit_ro_spine(&spine);
 494	return r;
 495}
 496
 497EXPORT_SYMBOL_GPL(dm_btree_lookup_next);
 498
 499/*----------------------------------------------------------------*/
 500
 501/*
 502 * Copies entries from one region of a btree node to another.  The regions
 503 * must not overlap.
 504 */
 505static void copy_entries(struct btree_node *dest, unsigned dest_offset,
 506			 struct btree_node *src, unsigned src_offset,
 507			 unsigned count)
 508{
 509	size_t value_size = le32_to_cpu(dest->header.value_size);
 
 510	memcpy(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
 511	memcpy(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
 512}
 513
 514/*
 515 * Moves entries from one region fo a btree node to another.  The regions
 516 * may overlap.
 517 */
 518static void move_entries(struct btree_node *dest, unsigned dest_offset,
 519			 struct btree_node *src, unsigned src_offset,
 520			 unsigned count)
 521{
 522	size_t value_size = le32_to_cpu(dest->header.value_size);
 
 523	memmove(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
 524	memmove(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
 525}
 526
 527/*
 528 * Erases the first 'count' entries of a btree node, shifting following
 529 * entries down into their place.
 530 */
 531static void shift_down(struct btree_node *n, unsigned count)
 532{
 533	move_entries(n, 0, n, count, le32_to_cpu(n->header.nr_entries) - count);
 534}
 535
 536/*
 537 * Moves entries in a btree node up 'count' places, making space for
 538 * new entries at the start of the node.
 539 */
 540static void shift_up(struct btree_node *n, unsigned count)
 541{
 542	move_entries(n, count, n, 0, le32_to_cpu(n->header.nr_entries));
 543}
 544
 545/*
 546 * Redistributes entries between two btree nodes to make them
 547 * have similar numbers of entries.
 548 */
 549static void redistribute2(struct btree_node *left, struct btree_node *right)
 550{
 551	unsigned nr_left = le32_to_cpu(left->header.nr_entries);
 552	unsigned nr_right = le32_to_cpu(right->header.nr_entries);
 553	unsigned total = nr_left + nr_right;
 554	unsigned target_left = total / 2;
 555	unsigned target_right = total - target_left;
 556
 557	if (nr_left < target_left) {
 558		unsigned delta = target_left - nr_left;
 
 559		copy_entries(left, nr_left, right, 0, delta);
 560		shift_down(right, delta);
 561	} else if (nr_left > target_left) {
 562		unsigned delta = nr_left - target_left;
 
 563		if (nr_right)
 564			shift_up(right, delta);
 565		copy_entries(right, 0, left, target_left, delta);
 566	}
 567
 568	left->header.nr_entries = cpu_to_le32(target_left);
 569	right->header.nr_entries = cpu_to_le32(target_right);
 570}
 571
 572/*
 573 * Redistribute entries between three nodes.  Assumes the central
 574 * node is empty.
 575 */
 576static void redistribute3(struct btree_node *left, struct btree_node *center,
 577			  struct btree_node *right)
 578{
 579	unsigned nr_left = le32_to_cpu(left->header.nr_entries);
 580	unsigned nr_center = le32_to_cpu(center->header.nr_entries);
 581	unsigned nr_right = le32_to_cpu(right->header.nr_entries);
 582	unsigned total, target_left, target_center, target_right;
 583
 584	BUG_ON(nr_center);
 585
 586	total = nr_left + nr_right;
 587	target_left = total / 3;
 588	target_center = (total - target_left) / 2;
 589	target_right = (total - target_left - target_center);
 590
 591	if (nr_left < target_left) {
 592		unsigned left_short = target_left - nr_left;
 
 593		copy_entries(left, nr_left, right, 0, left_short);
 594		copy_entries(center, 0, right, left_short, target_center);
 595		shift_down(right, nr_right - target_right);
 596
 597	} else if (nr_left < (target_left + target_center)) {
 598		unsigned left_to_center = nr_left - target_left;
 
 599		copy_entries(center, 0, left, target_left, left_to_center);
 600		copy_entries(center, left_to_center, right, 0, target_center - left_to_center);
 601		shift_down(right, nr_right - target_right);
 602
 603	} else {
 604		unsigned right_short = target_right - nr_right;
 
 605		shift_up(right, right_short);
 606		copy_entries(right, 0, left, nr_left - right_short, right_short);
 607		copy_entries(center, 0, left, target_left, nr_left - target_left);
 608	}
 609
 610	left->header.nr_entries = cpu_to_le32(target_left);
 611	center->header.nr_entries = cpu_to_le32(target_center);
 612	right->header.nr_entries = cpu_to_le32(target_right);
 613}
 614
 615/*
 616 * Splits a node by creating a sibling node and shifting half the nodes
 617 * contents across.  Assumes there is a parent node, and it has room for
 618 * another child.
 619 *
 620 * Before:
 621 *	  +--------+
 622 *	  | Parent |
 623 *	  +--------+
 624 *	     |
 625 *	     v
 626 *	+----------+
 627 *	| A ++++++ |
 628 *	+----------+
 629 *
 630 *
 631 * After:
 632 *		+--------+
 633 *		| Parent |
 634 *		+--------+
 635 *		  |	|
 636 *		  v	+------+
 637 *	    +---------+	       |
 638 *	    | A* +++  |	       v
 639 *	    +---------+	  +-------+
 640 *			  | B +++ |
 641 *			  +-------+
 642 *
 643 * Where A* is a shadow of A.
 644 */
 645static int split_one_into_two(struct shadow_spine *s, unsigned parent_index,
 646			      struct dm_btree_value_type *vt, uint64_t key)
 647{
 648	int r;
 649	struct dm_block *left, *right, *parent;
 650	struct btree_node *ln, *rn, *pn;
 651	__le64 location;
 652
 653	left = shadow_current(s);
 654
 655	r = new_block(s->info, &right);
 656	if (r < 0)
 657		return r;
 658
 659	ln = dm_block_data(left);
 660	rn = dm_block_data(right);
 661
 662	rn->header.flags = ln->header.flags;
 663	rn->header.nr_entries = cpu_to_le32(0);
 664	rn->header.max_entries = ln->header.max_entries;
 665	rn->header.value_size = ln->header.value_size;
 666	redistribute2(ln, rn);
 667
 668	/* patch up the parent */
 669	parent = shadow_parent(s);
 670	pn = dm_block_data(parent);
 671
 672	location = cpu_to_le64(dm_block_location(right));
 673	__dm_bless_for_disk(&location);
 674	r = insert_at(sizeof(__le64), pn, parent_index + 1,
 675		      le64_to_cpu(rn->keys[0]), &location);
 676	if (r) {
 677		unlock_block(s->info, right);
 678		return r;
 679	}
 680
 681	/* patch up the spine */
 682	if (key < le64_to_cpu(rn->keys[0])) {
 683		unlock_block(s->info, right);
 684		s->nodes[1] = left;
 685	} else {
 686		unlock_block(s->info, left);
 687		s->nodes[1] = right;
 688	}
 689
 690	return 0;
 691}
 692
 693/*
 694 * We often need to modify a sibling node.  This function shadows a particular
 695 * child of the given parent node.  Making sure to update the parent to point
 696 * to the new shadow.
 697 */
 698static int shadow_child(struct dm_btree_info *info, struct dm_btree_value_type *vt,
 699			struct btree_node *parent, unsigned index,
 700			struct dm_block **result)
 701{
 702	int r, inc;
 703	dm_block_t root;
 704	struct btree_node *node;
 705
 706	root = value64(parent, index);
 707
 708	r = dm_tm_shadow_block(info->tm, root, &btree_node_validator,
 709			       result, &inc);
 710	if (r)
 711		return r;
 712
 713	node = dm_block_data(*result);
 714
 715	if (inc)
 716		inc_children(info->tm, node, vt);
 717
 718	*((__le64 *) value_ptr(parent, index)) =
 719		cpu_to_le64(dm_block_location(*result));
 720
 721	return 0;
 722}
 723
 724/*
 725 * Splits two nodes into three.  This is more work, but results in fuller
 726 * nodes, so saves metadata space.
 727 */
 728static int split_two_into_three(struct shadow_spine *s, unsigned parent_index,
 729                                struct dm_btree_value_type *vt, uint64_t key)
 730{
 731	int r;
 732	unsigned middle_index;
 733	struct dm_block *left, *middle, *right, *parent;
 734	struct btree_node *ln, *rn, *mn, *pn;
 735	__le64 location;
 736
 737	parent = shadow_parent(s);
 738	pn = dm_block_data(parent);
 739
 740	if (parent_index == 0) {
 741		middle_index = 1;
 742		left = shadow_current(s);
 743		r = shadow_child(s->info, vt, pn, parent_index + 1, &right);
 744		if (r)
 745			return r;
 746	} else {
 747		middle_index = parent_index;
 748		right = shadow_current(s);
 749		r = shadow_child(s->info, vt, pn, parent_index - 1, &left);
 750		if (r)
 751			return r;
 752	}
 753
 754	r = new_block(s->info, &middle);
 755	if (r < 0)
 756		return r;
 757
 758	ln = dm_block_data(left);
 759	mn = dm_block_data(middle);
 760	rn = dm_block_data(right);
 761
 762	mn->header.nr_entries = cpu_to_le32(0);
 763	mn->header.flags = ln->header.flags;
 764	mn->header.max_entries = ln->header.max_entries;
 765	mn->header.value_size = ln->header.value_size;
 766
 767	redistribute3(ln, mn, rn);
 768
 769	/* patch up the parent */
 770	pn->keys[middle_index] = rn->keys[0];
 771	location = cpu_to_le64(dm_block_location(middle));
 772	__dm_bless_for_disk(&location);
 773	r = insert_at(sizeof(__le64), pn, middle_index,
 774		      le64_to_cpu(mn->keys[0]), &location);
 775	if (r) {
 776		if (shadow_current(s) != left)
 777			unlock_block(s->info, left);
 778
 779		unlock_block(s->info, middle);
 780
 781		if (shadow_current(s) != right)
 782			unlock_block(s->info, right);
 783
 784	        return r;
 785	}
 786
 787
 788	/* patch up the spine */
 789	if (key < le64_to_cpu(mn->keys[0])) {
 790		unlock_block(s->info, middle);
 791		unlock_block(s->info, right);
 792		s->nodes[1] = left;
 793	} else if (key < le64_to_cpu(rn->keys[0])) {
 794		unlock_block(s->info, left);
 795		unlock_block(s->info, right);
 796		s->nodes[1] = middle;
 797	} else {
 798		unlock_block(s->info, left);
 799		unlock_block(s->info, middle);
 800		s->nodes[1] = right;
 801	}
 802
 803	return 0;
 804}
 805
 806/*----------------------------------------------------------------*/
 807
 808/*
 809 * Splits a node by creating two new children beneath the given node.
 810 *
 811 * Before:
 812 *	  +----------+
 813 *	  | A ++++++ |
 814 *	  +----------+
 815 *
 816 *
 817 * After:
 818 *	+------------+
 819 *	| A (shadow) |
 820 *	+------------+
 821 *	    |	|
 822 *   +------+	+----+
 823 *   |		     |
 824 *   v		     v
 825 * +-------+	 +-------+
 826 * | B +++ |	 | C +++ |
 827 * +-------+	 +-------+
 828 */
 829static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
 830{
 831	int r;
 832	size_t size;
 833	unsigned nr_left, nr_right;
 834	struct dm_block *left, *right, *new_parent;
 835	struct btree_node *pn, *ln, *rn;
 836	__le64 val;
 837
 838	new_parent = shadow_current(s);
 839
 840	pn = dm_block_data(new_parent);
 841	size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
 842		sizeof(__le64) : s->info->value_type.size;
 843
 844	/* create & init the left block */
 845	r = new_block(s->info, &left);
 846	if (r < 0)
 847		return r;
 848
 849	ln = dm_block_data(left);
 850	nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
 851
 852	ln->header.flags = pn->header.flags;
 853	ln->header.nr_entries = cpu_to_le32(nr_left);
 854	ln->header.max_entries = pn->header.max_entries;
 855	ln->header.value_size = pn->header.value_size;
 856	memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
 857	memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
 858
 859	/* create & init the right block */
 860	r = new_block(s->info, &right);
 861	if (r < 0) {
 862		unlock_block(s->info, left);
 863		return r;
 864	}
 865
 866	rn = dm_block_data(right);
 867	nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
 868
 869	rn->header.flags = pn->header.flags;
 870	rn->header.nr_entries = cpu_to_le32(nr_right);
 871	rn->header.max_entries = pn->header.max_entries;
 872	rn->header.value_size = pn->header.value_size;
 873	memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
 874	memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
 875	       nr_right * size);
 876
 877	/* new_parent should just point to l and r now */
 878	pn->header.flags = cpu_to_le32(INTERNAL_NODE);
 879	pn->header.nr_entries = cpu_to_le32(2);
 880	pn->header.max_entries = cpu_to_le32(
 881		calc_max_entries(sizeof(__le64),
 882				 dm_bm_block_size(
 883					 dm_tm_get_bm(s->info->tm))));
 884	pn->header.value_size = cpu_to_le32(sizeof(__le64));
 885
 886	val = cpu_to_le64(dm_block_location(left));
 887	__dm_bless_for_disk(&val);
 888	pn->keys[0] = ln->keys[0];
 889	memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
 890
 891	val = cpu_to_le64(dm_block_location(right));
 892	__dm_bless_for_disk(&val);
 893	pn->keys[1] = rn->keys[0];
 894	memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
 895
 896	unlock_block(s->info, left);
 897	unlock_block(s->info, right);
 898	return 0;
 899}
 900
 901/*----------------------------------------------------------------*/
 902
 903/*
 904 * Redistributes a node's entries with its left sibling.
 905 */
 906static int rebalance_left(struct shadow_spine *s, struct dm_btree_value_type *vt,
 907			  unsigned parent_index, uint64_t key)
 908{
 909	int r;
 910	struct dm_block *sib;
 911	struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
 912
 913	r = shadow_child(s->info, vt, parent, parent_index - 1, &sib);
 914	if (r)
 915		return r;
 916
 917	left = dm_block_data(sib);
 918	right = dm_block_data(shadow_current(s));
 919	redistribute2(left, right);
 920	*key_ptr(parent, parent_index) = right->keys[0];
 921
 922	if (key < le64_to_cpu(right->keys[0])) {
 923		unlock_block(s->info, s->nodes[1]);
 924		s->nodes[1] = sib;
 925	} else {
 926		unlock_block(s->info, sib);
 927	}
 928
 929	return 0;
 930}
 931
 932/*
 933 * Redistributes a nodes entries with its right sibling.
 934 */
 935static int rebalance_right(struct shadow_spine *s, struct dm_btree_value_type *vt,
 936			   unsigned parent_index, uint64_t key)
 937{
 938	int r;
 939	struct dm_block *sib;
 940	struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
 941
 942	r = shadow_child(s->info, vt, parent, parent_index + 1, &sib);
 943	if (r)
 944		return r;
 945
 946	left = dm_block_data(shadow_current(s));
 947	right = dm_block_data(sib);
 948	redistribute2(left, right);
 949	*key_ptr(parent, parent_index + 1) = right->keys[0];
 950
 951	if (key < le64_to_cpu(right->keys[0])) {
 952		unlock_block(s->info, sib);
 953	} else {
 954		unlock_block(s->info, s->nodes[1]);
 955		s->nodes[1] = sib;
 956	}
 957
 958	return 0;
 959}
 960
 961/*
 962 * Returns the number of spare entries in a node.
 963 */
 964static int get_node_free_space(struct dm_btree_info *info, dm_block_t b, unsigned *space)
 965{
 966	int r;
 967	unsigned nr_entries;
 968	struct dm_block *block;
 969	struct btree_node *node;
 970
 971	r = bn_read_lock(info, b, &block);
 972	if (r)
 973		return r;
 974
 975	node = dm_block_data(block);
 976	nr_entries = le32_to_cpu(node->header.nr_entries);
 977	*space = le32_to_cpu(node->header.max_entries) - nr_entries;
 978
 979	unlock_block(info, block);
 980	return 0;
 981}
 982
 983/*
 984 * Make space in a node, either by moving some entries to a sibling,
 985 * or creating a new sibling node.  SPACE_THRESHOLD defines the minimum
 986 * number of free entries that must be in the sibling to make the move
 987 * worth while.  If the siblings are shared (eg, part of a snapshot),
 988 * then they are not touched, since this break sharing and so consume
 989 * more space than we save.
 990 */
 991#define SPACE_THRESHOLD 8
 992static int rebalance_or_split(struct shadow_spine *s, struct dm_btree_value_type *vt,
 993			      unsigned parent_index, uint64_t key)
 994{
 995	int r;
 996	struct btree_node *parent = dm_block_data(shadow_parent(s));
 997	unsigned nr_parent = le32_to_cpu(parent->header.nr_entries);
 998	unsigned free_space;
 999	int left_shared = 0, right_shared = 0;
1000
1001	/* Should we move entries to the left sibling? */
1002	if (parent_index > 0) {
1003		dm_block_t left_b = value64(parent, parent_index - 1);
 
1004		r = dm_tm_block_is_shared(s->info->tm, left_b, &left_shared);
1005		if (r)
1006			return r;
1007
1008		if (!left_shared) {
1009			r = get_node_free_space(s->info, left_b, &free_space);
1010			if (r)
1011				return r;
1012
1013			if (free_space >= SPACE_THRESHOLD)
1014				return rebalance_left(s, vt, parent_index, key);
1015		}
1016	}
1017
1018	/* Should we move entries to the right sibling? */
1019	if (parent_index < (nr_parent - 1)) {
1020		dm_block_t right_b = value64(parent, parent_index + 1);
 
1021		r = dm_tm_block_is_shared(s->info->tm, right_b, &right_shared);
1022		if (r)
1023			return r;
1024
1025		if (!right_shared) {
1026			r = get_node_free_space(s->info, right_b, &free_space);
1027			if (r)
1028				return r;
1029
1030			if (free_space >= SPACE_THRESHOLD)
1031				return rebalance_right(s, vt, parent_index, key);
1032		}
1033	}
1034
1035	/*
1036	 * We need to split the node, normally we split two nodes
1037	 * into three.	But when inserting a sequence that is either
1038	 * monotonically increasing or decreasing it's better to split
1039	 * a single node into two.
1040	 */
1041	if (left_shared || right_shared || (nr_parent <= 2) ||
1042	    (parent_index == 0) || (parent_index + 1 == nr_parent)) {
1043		return split_one_into_two(s, parent_index, vt, key);
1044	} else {
1045		return split_two_into_three(s, parent_index, vt, key);
1046	}
1047}
1048
1049/*
1050 * Does the node contain a particular key?
1051 */
1052static bool contains_key(struct btree_node *node, uint64_t key)
1053{
1054	int i = lower_bound(node, key);
1055
1056	if (i >= 0 && le64_to_cpu(node->keys[i]) == key)
1057		return true;
1058
1059	return false;
1060}
1061
1062/*
1063 * In general we preemptively make sure there's a free entry in every
1064 * node on the spine when doing an insert.  But we can avoid that with
1065 * leaf nodes if we know it's an overwrite.
1066 */
1067static bool has_space_for_insert(struct btree_node *node, uint64_t key)
1068{
1069	if (node->header.nr_entries == node->header.max_entries) {
1070		if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1071			/* we don't need space if it's an overwrite */
1072			return contains_key(node, key);
1073		}
1074
1075		return false;
1076	}
1077
1078	return true;
1079}
1080
1081static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
1082			    struct dm_btree_value_type *vt,
1083			    uint64_t key, unsigned *index)
1084{
1085	int r, i = *index, top = 1;
1086	struct btree_node *node;
1087
1088	for (;;) {
1089		r = shadow_step(s, root, vt);
1090		if (r < 0)
1091			return r;
1092
1093		node = dm_block_data(shadow_current(s));
1094
1095		/*
1096		 * We have to patch up the parent node, ugly, but I don't
1097		 * see a way to do this automatically as part of the spine
1098		 * op.
1099		 */
1100		if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
1101			__le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1102
1103			__dm_bless_for_disk(&location);
1104			memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1105				    &location, sizeof(__le64));
1106		}
1107
1108		node = dm_block_data(shadow_current(s));
1109
1110		if (!has_space_for_insert(node, key)) {
1111			if (top)
1112				r = btree_split_beneath(s, key);
1113			else
1114				r = rebalance_or_split(s, vt, i, key);
1115
1116			if (r < 0)
1117				return r;
1118
1119			/* making space can cause the current node to change */
1120			node = dm_block_data(shadow_current(s));
1121		}
1122
1123		i = lower_bound(node, key);
1124
1125		if (le32_to_cpu(node->header.flags) & LEAF_NODE)
1126			break;
1127
1128		if (i < 0) {
1129			/* change the bounds on the lowest key */
1130			node->keys[0] = cpu_to_le64(key);
1131			i = 0;
1132		}
1133
1134		root = value64(node, i);
1135		top = 0;
1136	}
1137
1138	if (i < 0 || le64_to_cpu(node->keys[i]) != key)
1139		i++;
1140
1141	*index = i;
1142	return 0;
1143}
1144
1145static int __btree_get_overwrite_leaf(struct shadow_spine *s, dm_block_t root,
1146				      uint64_t key, int *index)
1147{
1148	int r, i = -1;
1149	struct btree_node *node;
1150
1151	*index = 0;
1152	for (;;) {
1153		r = shadow_step(s, root, &s->info->value_type);
1154		if (r < 0)
1155			return r;
1156
1157		node = dm_block_data(shadow_current(s));
1158
1159		/*
1160		 * We have to patch up the parent node, ugly, but I don't
1161		 * see a way to do this automatically as part of the spine
1162		 * op.
1163		 */
1164		if (shadow_has_parent(s) && i >= 0) {
1165			__le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1166
1167			__dm_bless_for_disk(&location);
1168			memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1169				    &location, sizeof(__le64));
1170		}
1171
1172		node = dm_block_data(shadow_current(s));
1173		i = lower_bound(node, key);
1174
1175		BUG_ON(i < 0);
1176		BUG_ON(i >= le32_to_cpu(node->header.nr_entries));
1177
1178		if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1179			if (key != le64_to_cpu(node->keys[i]))
1180				return -EINVAL;
1181			break;
1182		}
1183
1184		root = value64(node, i);
1185	}
1186
1187	*index = i;
1188	return 0;
1189}
1190
1191int btree_get_overwrite_leaf(struct dm_btree_info *info, dm_block_t root,
1192			     uint64_t key, int *index,
1193			     dm_block_t *new_root, struct dm_block **leaf)
1194{
1195	int r;
1196	struct shadow_spine spine;
1197
1198	BUG_ON(info->levels > 1);
1199	init_shadow_spine(&spine, info);
1200	r = __btree_get_overwrite_leaf(&spine, root, key, index);
1201	if (!r) {
1202		*new_root = shadow_root(&spine);
1203		*leaf = shadow_current(&spine);
1204
1205		/*
1206		 * Decrement the count so exit_shadow_spine() doesn't
1207		 * unlock the leaf.
1208		 */
1209		spine.count--;
1210	}
1211	exit_shadow_spine(&spine);
1212
1213	return r;
1214}
1215
1216static bool need_insert(struct btree_node *node, uint64_t *keys,
1217			unsigned level, unsigned index)
1218{
1219        return ((index >= le32_to_cpu(node->header.nr_entries)) ||
1220		(le64_to_cpu(node->keys[index]) != keys[level]));
1221}
1222
1223static int insert(struct dm_btree_info *info, dm_block_t root,
1224		  uint64_t *keys, void *value, dm_block_t *new_root,
1225		  int *inserted)
1226		  __dm_written_to_disk(value)
1227{
1228	int r;
1229	unsigned level, index = -1, last_level = info->levels - 1;
1230	dm_block_t block = root;
1231	struct shadow_spine spine;
1232	struct btree_node *n;
1233	struct dm_btree_value_type le64_type;
1234
1235	init_le64_type(info->tm, &le64_type);
1236	init_shadow_spine(&spine, info);
1237
1238	for (level = 0; level < (info->levels - 1); level++) {
1239		r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
1240		if (r < 0)
1241			goto bad;
1242
1243		n = dm_block_data(shadow_current(&spine));
1244
1245		if (need_insert(n, keys, level, index)) {
1246			dm_block_t new_tree;
1247			__le64 new_le;
1248
1249			r = dm_btree_empty(info, &new_tree);
1250			if (r < 0)
1251				goto bad;
1252
1253			new_le = cpu_to_le64(new_tree);
1254			__dm_bless_for_disk(&new_le);
1255
1256			r = insert_at(sizeof(uint64_t), n, index,
1257				      keys[level], &new_le);
1258			if (r)
1259				goto bad;
1260		}
1261
1262		if (level < last_level)
1263			block = value64(n, index);
1264	}
1265
1266	r = btree_insert_raw(&spine, block, &info->value_type,
1267			     keys[level], &index);
1268	if (r < 0)
1269		goto bad;
1270
1271	n = dm_block_data(shadow_current(&spine));
1272
1273	if (need_insert(n, keys, level, index)) {
1274		if (inserted)
1275			*inserted = 1;
1276
1277		r = insert_at(info->value_type.size, n, index,
1278			      keys[level], value);
1279		if (r)
1280			goto bad_unblessed;
1281	} else {
1282		if (inserted)
1283			*inserted = 0;
1284
1285		if (info->value_type.dec &&
1286		    (!info->value_type.equal ||
1287		     !info->value_type.equal(
1288			     info->value_type.context,
1289			     value_ptr(n, index),
1290			     value))) {
1291			info->value_type.dec(info->value_type.context,
1292					     value_ptr(n, index), 1);
1293		}
1294		memcpy_disk(value_ptr(n, index),
1295			    value, info->value_type.size);
1296	}
1297
1298	*new_root = shadow_root(&spine);
1299	exit_shadow_spine(&spine);
1300
1301	return 0;
1302
1303bad:
1304	__dm_unbless_for_disk(value);
1305bad_unblessed:
1306	exit_shadow_spine(&spine);
1307	return r;
1308}
1309
1310int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
1311		    uint64_t *keys, void *value, dm_block_t *new_root)
1312		    __dm_written_to_disk(value)
1313{
1314	return insert(info, root, keys, value, new_root, NULL);
1315}
1316EXPORT_SYMBOL_GPL(dm_btree_insert);
1317
1318int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
1319			   uint64_t *keys, void *value, dm_block_t *new_root,
1320			   int *inserted)
1321			   __dm_written_to_disk(value)
1322{
1323	return insert(info, root, keys, value, new_root, inserted);
1324}
1325EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
1326
1327/*----------------------------------------------------------------*/
1328
1329static int find_key(struct ro_spine *s, dm_block_t block, bool find_highest,
1330		    uint64_t *result_key, dm_block_t *next_block)
1331{
1332	int i, r;
1333	uint32_t flags;
1334
1335	do {
1336		r = ro_step(s, block);
1337		if (r < 0)
1338			return r;
1339
1340		flags = le32_to_cpu(ro_node(s)->header.flags);
1341		i = le32_to_cpu(ro_node(s)->header.nr_entries);
1342		if (!i)
1343			return -ENODATA;
1344		else
1345			i--;
1346
1347		if (find_highest)
1348			*result_key = le64_to_cpu(ro_node(s)->keys[i]);
1349		else
1350			*result_key = le64_to_cpu(ro_node(s)->keys[0]);
1351
1352		if (next_block || flags & INTERNAL_NODE) {
1353			if (find_highest)
1354				block = value64(ro_node(s), i);
1355			else
1356				block = value64(ro_node(s), 0);
1357		}
1358
1359	} while (flags & INTERNAL_NODE);
1360
1361	if (next_block)
1362		*next_block = block;
1363	return 0;
1364}
1365
1366static int dm_btree_find_key(struct dm_btree_info *info, dm_block_t root,
1367			     bool find_highest, uint64_t *result_keys)
1368{
1369	int r = 0, count = 0, level;
1370	struct ro_spine spine;
1371
1372	init_ro_spine(&spine, info);
1373	for (level = 0; level < info->levels; level++) {
1374		r = find_key(&spine, root, find_highest, result_keys + level,
1375			     level == info->levels - 1 ? NULL : &root);
1376		if (r == -ENODATA) {
1377			r = 0;
1378			break;
1379
1380		} else if (r)
1381			break;
1382
1383		count++;
1384	}
1385	exit_ro_spine(&spine);
1386
1387	return r ? r : count;
1388}
1389
1390int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
1391			      uint64_t *result_keys)
1392{
1393	return dm_btree_find_key(info, root, true, result_keys);
1394}
1395EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
1396
1397int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
1398			     uint64_t *result_keys)
1399{
1400	return dm_btree_find_key(info, root, false, result_keys);
1401}
1402EXPORT_SYMBOL_GPL(dm_btree_find_lowest_key);
1403
1404/*----------------------------------------------------------------*/
1405
1406/*
1407 * FIXME: We shouldn't use a recursive algorithm when we have limited stack
1408 * space.  Also this only works for single level trees.
1409 */
1410static int walk_node(struct dm_btree_info *info, dm_block_t block,
1411		     int (*fn)(void *context, uint64_t *keys, void *leaf),
1412		     void *context)
1413{
1414	int r;
1415	unsigned i, nr;
1416	struct dm_block *node;
1417	struct btree_node *n;
1418	uint64_t keys;
1419
1420	r = bn_read_lock(info, block, &node);
1421	if (r)
1422		return r;
1423
1424	n = dm_block_data(node);
1425
1426	nr = le32_to_cpu(n->header.nr_entries);
1427	for (i = 0; i < nr; i++) {
1428		if (le32_to_cpu(n->header.flags) & INTERNAL_NODE) {
1429			r = walk_node(info, value64(n, i), fn, context);
1430			if (r)
1431				goto out;
1432		} else {
1433			keys = le64_to_cpu(*key_ptr(n, i));
1434			r = fn(context, &keys, value_ptr(n, i));
1435			if (r)
1436				goto out;
1437		}
1438	}
1439
1440out:
1441	dm_tm_unlock(info->tm, node);
1442	return r;
1443}
1444
1445int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
1446		  int (*fn)(void *context, uint64_t *keys, void *leaf),
1447		  void *context)
1448{
1449	BUG_ON(info->levels > 1);
1450	return walk_node(info, root, fn, context);
1451}
1452EXPORT_SYMBOL_GPL(dm_btree_walk);
1453
1454/*----------------------------------------------------------------*/
1455
1456static void prefetch_values(struct dm_btree_cursor *c)
1457{
1458	unsigned i, nr;
1459	__le64 value_le;
1460	struct cursor_node *n = c->nodes + c->depth - 1;
1461	struct btree_node *bn = dm_block_data(n->b);
1462	struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
1463
1464	BUG_ON(c->info->value_type.size != sizeof(value_le));
1465
1466	nr = le32_to_cpu(bn->header.nr_entries);
1467	for (i = 0; i < nr; i++) {
1468		memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
1469		dm_bm_prefetch(bm, le64_to_cpu(value_le));
1470	}
1471}
1472
1473static bool leaf_node(struct dm_btree_cursor *c)
1474{
1475	struct cursor_node *n = c->nodes + c->depth - 1;
1476	struct btree_node *bn = dm_block_data(n->b);
1477
1478	return le32_to_cpu(bn->header.flags) & LEAF_NODE;
1479}
1480
1481static int push_node(struct dm_btree_cursor *c, dm_block_t b)
1482{
1483	int r;
1484	struct cursor_node *n = c->nodes + c->depth;
1485
1486	if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
1487		DMERR("couldn't push cursor node, stack depth too high");
1488		return -EINVAL;
1489	}
1490
1491	r = bn_read_lock(c->info, b, &n->b);
1492	if (r)
1493		return r;
1494
1495	n->index = 0;
1496	c->depth++;
1497
1498	if (c->prefetch_leaves || !leaf_node(c))
1499		prefetch_values(c);
1500
1501	return 0;
1502}
1503
1504static void pop_node(struct dm_btree_cursor *c)
1505{
1506	c->depth--;
1507	unlock_block(c->info, c->nodes[c->depth].b);
1508}
1509
1510static int inc_or_backtrack(struct dm_btree_cursor *c)
1511{
1512	struct cursor_node *n;
1513	struct btree_node *bn;
1514
1515	for (;;) {
1516		if (!c->depth)
1517			return -ENODATA;
1518
1519		n = c->nodes + c->depth - 1;
1520		bn = dm_block_data(n->b);
1521
1522		n->index++;
1523		if (n->index < le32_to_cpu(bn->header.nr_entries))
1524			break;
1525
1526		pop_node(c);
1527	}
1528
1529	return 0;
1530}
1531
1532static int find_leaf(struct dm_btree_cursor *c)
1533{
1534	int r = 0;
1535	struct cursor_node *n;
1536	struct btree_node *bn;
1537	__le64 value_le;
1538
1539	for (;;) {
1540		n = c->nodes + c->depth - 1;
1541		bn = dm_block_data(n->b);
1542
1543		if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
1544			break;
1545
1546		memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
1547		r = push_node(c, le64_to_cpu(value_le));
1548		if (r) {
1549			DMERR("push_node failed");
1550			break;
1551		}
1552	}
1553
1554	if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
1555		return -ENODATA;
1556
1557	return r;
1558}
1559
1560int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
1561			  bool prefetch_leaves, struct dm_btree_cursor *c)
1562{
1563	int r;
1564
1565	c->info = info;
1566	c->root = root;
1567	c->depth = 0;
1568	c->prefetch_leaves = prefetch_leaves;
1569
1570	r = push_node(c, root);
1571	if (r)
1572		return r;
1573
1574	return find_leaf(c);
1575}
1576EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
1577
1578void dm_btree_cursor_end(struct dm_btree_cursor *c)
1579{
1580	while (c->depth)
1581		pop_node(c);
1582}
1583EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
1584
1585int dm_btree_cursor_next(struct dm_btree_cursor *c)
1586{
1587	int r = inc_or_backtrack(c);
 
1588	if (!r) {
1589		r = find_leaf(c);
1590		if (r)
1591			DMERR("find_leaf failed");
1592	}
1593
1594	return r;
1595}
1596EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
1597
1598int dm_btree_cursor_skip(struct dm_btree_cursor *c, uint32_t count)
1599{
1600	int r = 0;
1601
1602	while (count-- && !r)
1603		r = dm_btree_cursor_next(c);
1604
1605	return r;
1606}
1607EXPORT_SYMBOL_GPL(dm_btree_cursor_skip);
1608
1609int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
1610{
1611	if (c->depth) {
1612		struct cursor_node *n = c->nodes + c->depth - 1;
1613		struct btree_node *bn = dm_block_data(n->b);
1614
1615		if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
1616			return -EINVAL;
1617
1618		*key = le64_to_cpu(*key_ptr(bn, n->index));
1619		memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
1620		return 0;
1621
1622	} else
1623		return -ENODATA;
1624}
1625EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);
v6.13.7
   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 */
  22static 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
  29static 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. */
  44static 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
  64int lower_bound(struct btree_node *n, uint64_t key)
  65{
  66	return bsearch(n, key, 0);
  67}
  68
  69static int upper_bound(struct btree_node *n, uint64_t key)
  70{
  71	return bsearch(n, key, 1);
  72}
  73
  74void 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
  86static 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 */
 117static 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
 129int 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}
 156EXPORT_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
 165struct 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
 173struct 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
 180static 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
 192static int unprocessed_frames(struct del_stack *s)
 193{
 194	return s->top >= 0;
 195}
 196
 197static 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
 206static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
 207{
 208	return f->level < (info->levels - 1);
 209}
 210
 211static 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
 255static 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
 263static 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
 273int 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	}
 330out:
 331	if (r) {
 332		/* cleanup all frames of del_stack */
 333		unlock_all_frames(s);
 334	}
 335	kfree(s);
 336
 337	return r;
 338}
 339EXPORT_SYMBOL_GPL(dm_btree_del);
 340
 341/*----------------------------------------------------------------*/
 342
 343static 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
 374int 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}
 417EXPORT_SYMBOL_GPL(dm_btree_lookup);
 418
 419static 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	}
 465out:
 466	dm_tm_unlock(info->tm, node);
 467	return r;
 468}
 469
 470int 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);
 495out:
 496	exit_ro_spine(&spine);
 497	return r;
 498}
 
 499EXPORT_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 */
 507static 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 */
 521static 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 */
 535static 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 */
 544static 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 */
 553static 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 */
 582static 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 */
 654static 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 */
 707static 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 */
 737static 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 */
 838static 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 */
 915static 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 */
 944static 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 */
 973static 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
1001static 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 */
1063static 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 */
1078static 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
1092static 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
1156static 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
1202int 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
1227static 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
1234static 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
1314bad:
1315	__dm_unbless_for_disk(value);
1316bad_unblessed:
1317	exit_shadow_spine(&spine);
1318	return r;
1319}
1320
1321int 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}
1327EXPORT_SYMBOL_GPL(dm_btree_insert);
1328
1329int 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}
1336EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
1337
1338/*----------------------------------------------------------------*/
1339
1340static 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
1377static 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
1401int 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}
1406EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
1407
1408int 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}
1413EXPORT_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 */
1421static 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
1451out:
1452	dm_tm_unlock(info->tm, node);
1453	return r;
1454}
1455
1456int 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}
1463EXPORT_SYMBOL_GPL(dm_btree_walk);
1464
1465/*----------------------------------------------------------------*/
1466
1467static 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
1484static 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
1492static 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
1515static void pop_node(struct dm_btree_cursor *c)
1516{
1517	c->depth--;
1518	unlock_block(c->info, c->nodes[c->depth].b);
1519}
1520
1521static 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
1543static 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
1571int 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}
1587EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
1588
1589void dm_btree_cursor_end(struct dm_btree_cursor *c)
1590{
1591	while (c->depth)
1592		pop_node(c);
1593}
1594EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
1595
1596int 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}
1608EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
1609
1610int 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}
1619EXPORT_SYMBOL_GPL(dm_btree_cursor_skip);
1620
1621int 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}
1637EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);