drivers/md/persistent-data/dm-btree.c

   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 void inc_children(struct dm_transaction_manager *tm, struct btree_node *n,
  67                   struct dm_btree_value_type *vt)
  68 {
  69         unsigned i;
  70         uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
  71
  72         if (le32_to_cpu(n->header.flags) & INTERNAL_NODE)
  73                 for (i = 0; i < nr_entries; i++)
  74                         dm_tm_inc(tm, value64(n, i));
  75         else if (vt->inc)
  76                 for (i = 0; i < nr_entries; i++)
  77                         vt->inc(vt->context,
  78                                 value_ptr(n, i, vt->size));
  79 }
  80
  81 static int insert_at(size_t value_size, struct btree_node *node, unsigned index,
  82                       uint64_t key, void *value)
  83                       __dm_written_to_disk(value)
  84 {
  85         uint32_t nr_entries = le32_to_cpu(node->header.nr_entries);
  86         __le64 key_le = cpu_to_le64(key);
  87
  88         if (index > nr_entries ||
  89             index >= le32_to_cpu(node->header.max_entries)) {
  90                 DMERR("too many entries in btree node for insert");
  91                 __dm_unbless_for_disk(value);
  92                 return -ENOMEM;
  93         }
  94
  95         __dm_bless_for_disk(&key_le);
  96
  97         array_insert(node->keys, sizeof(*node->keys), nr_entries, index, &key_le);
  98         array_insert(value_base(node), value_size, nr_entries, index, value);
  99         node->header.nr_entries = cpu_to_le32(nr_entries + 1);
 100
 101         return 0;
 102 }
 103
 104 /*----------------------------------------------------------------*/
 105
 106 /*
 107  * We want 3n entries (for some n).  This works more nicely for repeated
 108  * insert remove loops than (2n + 1).
 109  */
 110 static uint32_t calc_max_entries(size_t value_size, size_t block_size)
 111 {
 112         uint32_t total, n;
 113         size_t elt_size = sizeof(uint64_t) + value_size; /* key + value */
 114
 115         block_size -= sizeof(struct node_header);
 116         total = block_size / elt_size;
 117         n = total / 3;          /* rounds down */
 118
 119         return 3 * n;
 120 }
 121
 122 int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root)
 123 {
 124         int r;
 125         struct dm_block *b;
 126         struct btree_node *n;
 127         size_t block_size;
 128         uint32_t max_entries;
 129
 130         r = new_block(info, &b);
 131         if (r < 0)
 132                 return r;
 133
 134         block_size = dm_bm_block_size(dm_tm_get_bm(info->tm));
 135         max_entries = calc_max_entries(info->value_type.size, block_size);
 136
 137         n = dm_block_data(b);
 138         memset(n, 0, block_size);
 139         n->header.flags = cpu_to_le32(LEAF_NODE);
 140         n->header.nr_entries = cpu_to_le32(0);
 141         n->header.max_entries = cpu_to_le32(max_entries);
 142         n->header.value_size = cpu_to_le32(info->value_type.size);
 143
 144         *root = dm_block_location(b);
 145         return unlock_block(info, b);
 146 }
 147 EXPORT_SYMBOL_GPL(dm_btree_empty);
 148
 149 /*----------------------------------------------------------------*/
 150
 151 /*
 152  * Deletion uses a recursive algorithm, since we have limited stack space
 153  * we explicitly manage our own stack on the heap.
 154  */
 155 #define MAX_SPINE_DEPTH 64
 156 struct frame {
 157         struct dm_block *b;
 158         struct btree_node *n;
 159         unsigned level;
 160         unsigned nr_children;
 161         unsigned current_child;
 162 };
 163
 164 struct del_stack {
 165         struct dm_transaction_manager *tm;
 166         int top;
 167         struct frame spine[MAX_SPINE_DEPTH];
 168 };
 169
 170 static int top_frame(struct del_stack *s, struct frame **f)
 171 {
 172         if (s->top < 0) {
 173                 DMERR("btree deletion stack empty");
 174                 return -EINVAL;
 175         }
 176
 177         *f = s->spine + s->top;
 178
 179         return 0;
 180 }
 181
 182 static int unprocessed_frames(struct del_stack *s)
 183 {
 184         return s->top >= 0;
 185 }
 186
 187 static int push_frame(struct del_stack *s, dm_block_t b, unsigned level)
 188 {
 189         int r;
 190         uint32_t ref_count;
 191
 192         if (s->top >= MAX_SPINE_DEPTH - 1) {
 193                 DMERR("btree deletion stack out of memory");
 194                 return -ENOMEM;
 195         }
 196
 197         r = dm_tm_ref(s->tm, b, &ref_count);
 198         if (r)
 199                 return r;
 200
 201         if (ref_count > 1)
 202                 /*
 203                  * This is a shared node, so we can just decrement it's
 204                  * reference counter and leave the children.
 205                  */
 206                 dm_tm_dec(s->tm, b);
 207
 208         else {
 209                 struct frame *f = s->spine + ++s->top;
 210
 211                 r = dm_tm_read_lock(s->tm, b, &btree_node_validator, &f->b);
 212                 if (r) {
 213                         s->top--;
 214                         return r;
 215                 }
 216
 217                 f->n = dm_block_data(f->b);
 218                 f->level = level;
 219                 f->nr_children = le32_to_cpu(f->n->header.nr_entries);
 220                 f->current_child = 0;
 221         }
 222
 223         return 0;
 224 }
 225
 226 static void pop_frame(struct del_stack *s)
 227 {
 228         struct frame *f = s->spine + s->top--;
 229
 230         dm_tm_dec(s->tm, dm_block_location(f->b));
 231         dm_tm_unlock(s->tm, f->b);
 232 }
 233
 234 static void unlock_all_frames(struct del_stack *s)
 235 {
 236         struct frame *f;
 237
 238         while (unprocessed_frames(s)) {
 239                 f = s->spine + s->top--;
 240                 dm_tm_unlock(s->tm, f->b);
 241         }
 242 }
 243
 244 int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
 245 {
 246         int r;
 247         struct del_stack *s;
 248
 249         s = kmalloc(sizeof(*s), GFP_NOIO);
 250         if (!s)
 251                 return -ENOMEM;
 252         s->tm = info->tm;
 253         s->top = -1;
 254
 255         r = push_frame(s, root, 1);
 256         if (r)
 257                 goto out;
 258
 259         while (unprocessed_frames(s)) {
 260                 uint32_t flags;
 261                 struct frame *f;
 262                 dm_block_t b;
 263
 264                 r = top_frame(s, &f);
 265                 if (r)
 266                         goto out;
 267
 268                 if (f->current_child >= f->nr_children) {
 269                         pop_frame(s);
 270                         continue;
 271                 }
 272
 273                 flags = le32_to_cpu(f->n->header.flags);
 274                 if (flags & INTERNAL_NODE) {
 275                         b = value64(f->n, f->current_child);
 276                         f->current_child++;
 277                         r = push_frame(s, b, f->level);
 278                         if (r)
 279                                 goto out;
 280
 281                 } else if (f->level != (info->levels - 1)) {
 282                         b = value64(f->n, f->current_child);
 283                         f->current_child++;
 284                         r = push_frame(s, b, f->level + 1);
 285                         if (r)
 286                                 goto out;
 287
 288                 } else {
 289                         if (info->value_type.dec) {
 290                                 unsigned i;
 291
 292                                 for (i = 0; i < f->nr_children; i++)
 293                                         info->value_type.dec(info->value_type.context,
 294                                                              value_ptr(f->n, i, info->value_type.size));
 295                         }
 296                         f->current_child = f->nr_children;
 297                 }
 298         }
 299 out:
 300         if (r) {
 301                 /* cleanup all frames of del_stack */
 302                 unlock_all_frames(s);
 303         }
 304         kfree(s);
 305
 306         return r;
 307 }
 308 EXPORT_SYMBOL_GPL(dm_btree_del);
 309
 310 /*----------------------------------------------------------------*/
 311
 312 static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
 313                             int (*search_fn)(struct btree_node *, uint64_t),
 314                             uint64_t *result_key, void *v, size_t value_size)
 315 {
 316         int i, r;
 317         uint32_t flags, nr_entries;
 318
 319         do {
 320                 r = ro_step(s, block);
 321                 if (r < 0)
 322                         return r;
 323
 324                 i = search_fn(ro_node(s), key);
 325
 326                 flags = le32_to_cpu(ro_node(s)->header.flags);
 327                 nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
 328                 if (i < 0 || i >= nr_entries)
 329                         return -ENODATA;
 330
 331                 if (flags & INTERNAL_NODE)
 332                         block = value64(ro_node(s), i);
 333
 334         } while (!(flags & LEAF_NODE));
 335
 336         *result_key = le64_to_cpu(ro_node(s)->keys[i]);
 337         memcpy(v, value_ptr(ro_node(s), i, value_size), value_size);
 338
 339         return 0;
 340 }
 341
 342 int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
 343                     uint64_t *keys, void *value_le)
 344 {
 345         unsigned level, last_level = info->levels - 1;
 346         int r = -ENODATA;
 347         uint64_t rkey;
 348         __le64 internal_value_le;
 349         struct ro_spine spine;
 350
 351         init_ro_spine(&spine, info);
 352         for (level = 0; level < info->levels; level++) {
 353                 size_t size;
 354                 void *value_p;
 355
 356                 if (level == last_level) {
 357                         value_p = value_le;
 358                         size = info->value_type.size;
 359
 360                 } else {
 361                         value_p = &internal_value_le;
 362                         size = sizeof(uint64_t);
 363                 }
 364
 365                 r = btree_lookup_raw(&spine, root, keys[level],
 366                                      lower_bound, &rkey,
 367                                      value_p, size);
 368
 369                 if (!r) {
 370                         if (rkey != keys[level]) {
 371                                 exit_ro_spine(&spine);
 372                                 return -ENODATA;
 373                         }
 374                 } else {
 375                         exit_ro_spine(&spine);
 376                         return r;
 377                 }
 378
 379                 root = le64_to_cpu(internal_value_le);
 380         }
 381         exit_ro_spine(&spine);
 382
 383         return r;
 384 }
 385 EXPORT_SYMBOL_GPL(dm_btree_lookup);
 386
 387 /*
 388  * Splits a node by creating a sibling node and shifting half the nodes
 389  * contents across.  Assumes there is a parent node, and it has room for
 390  * another child.
 391  *
 392  * Before:
 393  *        +--------+
 394  *        | Parent |
 395  *        +--------+
 396  *           |
 397  *           v
 398  *      +----------+
 399  *      | A ++++++ |
 400  *      +----------+
 401  *
 402  *
 403  * After:
 404  *              +--------+
 405  *              | Parent |
 406  *              +--------+
 407  *                |     |
 408  *                v     +------+
 409  *          +---------+        |
 410  *          | A* +++  |        v
 411  *          +---------+   +-------+
 412  *                        | B +++ |
 413  *                        +-------+
 414  *
 415  * Where A* is a shadow of A.
 416  */
 417 static int btree_split_sibling(struct shadow_spine *s, dm_block_t root,
 418                                unsigned parent_index, uint64_t key)
 419 {
 420         int r;
 421         size_t size;
 422         unsigned nr_left, nr_right;
 423         struct dm_block *left, *right, *parent;
 424         struct btree_node *ln, *rn, *pn;
 425         __le64 location;
 426
 427         left = shadow_current(s);
 428
 429         r = new_block(s->info, &right);
 430         if (r < 0)
 431                 return r;
 432
 433         ln = dm_block_data(left);
 434         rn = dm_block_data(right);
 435
 436         nr_left = le32_to_cpu(ln->header.nr_entries) / 2;
 437         nr_right = le32_to_cpu(ln->header.nr_entries) - nr_left;
 438
 439         ln->header.nr_entries = cpu_to_le32(nr_left);
 440
 441         rn->header.flags = ln->header.flags;
 442         rn->header.nr_entries = cpu_to_le32(nr_right);
 443         rn->header.max_entries = ln->header.max_entries;
 444         rn->header.value_size = ln->header.value_size;
 445         memcpy(rn->keys, ln->keys + nr_left, nr_right * sizeof(rn->keys[0]));
 446
 447         size = le32_to_cpu(ln->header.flags) & INTERNAL_NODE ?
 448                 sizeof(uint64_t) : s->info->value_type.size;
 449         memcpy(value_ptr(rn, 0, size), value_ptr(ln, nr_left, size),
 450                size * nr_right);
 451
 452         /*
 453          * Patch up the parent
 454          */
 455         parent = shadow_parent(s);
 456
 457         pn = dm_block_data(parent);
 458         location = cpu_to_le64(dm_block_location(left));
 459         __dm_bless_for_disk(&location);
 460         memcpy_disk(value_ptr(pn, parent_index, sizeof(__le64)),
 461                     &location, sizeof(__le64));
 462
 463         location = cpu_to_le64(dm_block_location(right));
 464         __dm_bless_for_disk(&location);
 465
 466         r = insert_at(sizeof(__le64), pn, parent_index + 1,
 467                       le64_to_cpu(rn->keys[0]), &location);
 468         if (r) {
 469                 unlock_block(s->info, right);
 470                 return r;
 471         }
 472
 473         if (key < le64_to_cpu(rn->keys[0])) {
 474                 unlock_block(s->info, right);
 475                 s->nodes[1] = left;
 476         } else {
 477                 unlock_block(s->info, left);
 478                 s->nodes[1] = right;
 479         }
 480
 481         return 0;
 482 }
 483
 484 /*
 485  * Splits a node by creating two new children beneath the given node.
 486  *
 487  * Before:
 488  *        +----------+
 489  *        | A ++++++ |
 490  *        +----------+
 491  *
 492  *
 493  * After:
 494  *      +------------+
 495  *      | A (shadow) |
 496  *      +------------+
 497  *          |   |
 498  *   +------+   +----+
 499  *   |               |
 500  *   v               v
 501  * +-------+     +-------+
 502  * | B +++ |     | C +++ |
 503  * +-------+     +-------+
 504  */
 505 static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
 506 {
 507         int r;
 508         size_t size;
 509         unsigned nr_left, nr_right;
 510         struct dm_block *left, *right, *new_parent;
 511         struct btree_node *pn, *ln, *rn;
 512         __le64 val;
 513
 514         new_parent = shadow_current(s);
 515
 516         r = new_block(s->info, &left);
 517         if (r < 0)
 518                 return r;
 519
 520         r = new_block(s->info, &right);
 521         if (r < 0) {
 522                 unlock_block(s->info, left);
 523                 return r;
 524         }
 525
 526         pn = dm_block_data(new_parent);
 527         ln = dm_block_data(left);
 528         rn = dm_block_data(right);
 529
 530         nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
 531         nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
 532
 533         ln->header.flags = pn->header.flags;
 534         ln->header.nr_entries = cpu_to_le32(nr_left);
 535         ln->header.max_entries = pn->header.max_entries;
 536         ln->header.value_size = pn->header.value_size;
 537
 538         rn->header.flags = pn->header.flags;
 539         rn->header.nr_entries = cpu_to_le32(nr_right);
 540         rn->header.max_entries = pn->header.max_entries;
 541         rn->header.value_size = pn->header.value_size;
 542
 543         memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
 544         memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
 545
 546         size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
 547                 sizeof(__le64) : s->info->value_type.size;
 548         memcpy(value_ptr(ln, 0, size), value_ptr(pn, 0, size), nr_left * size);
 549         memcpy(value_ptr(rn, 0, size), value_ptr(pn, nr_left, size),
 550                nr_right * size);
 551
 552         /* new_parent should just point to l and r now */
 553         pn->header.flags = cpu_to_le32(INTERNAL_NODE);
 554         pn->header.nr_entries = cpu_to_le32(2);
 555         pn->header.max_entries = cpu_to_le32(
 556                 calc_max_entries(sizeof(__le64),
 557                                  dm_bm_block_size(
 558                                          dm_tm_get_bm(s->info->tm))));
 559         pn->header.value_size = cpu_to_le32(sizeof(__le64));
 560
 561         val = cpu_to_le64(dm_block_location(left));
 562         __dm_bless_for_disk(&val);
 563         pn->keys[0] = ln->keys[0];
 564         memcpy_disk(value_ptr(pn, 0, sizeof(__le64)), &val, sizeof(__le64));
 565
 566         val = cpu_to_le64(dm_block_location(right));
 567         __dm_bless_for_disk(&val);
 568         pn->keys[1] = rn->keys[0];
 569         memcpy_disk(value_ptr(pn, 1, sizeof(__le64)), &val, sizeof(__le64));
 570
 571         unlock_block(s->info, left);
 572         unlock_block(s->info, right);
 573         return 0;
 574 }
 575
 576 static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
 577                             struct dm_btree_value_type *vt,
 578                             uint64_t key, unsigned *index)
 579 {
 580         int r, i = *index, top = 1;
 581         struct btree_node *node;
 582
 583         for (;;) {
 584                 r = shadow_step(s, root, vt);
 585                 if (r < 0)
 586                         return r;
 587
 588                 node = dm_block_data(shadow_current(s));
 589
 590                 /*
 591                  * We have to patch up the parent node, ugly, but I don't
 592                  * see a way to do this automatically as part of the spine
 593                  * op.
 594                  */
 595                 if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
 596                         __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
 597
 598                         __dm_bless_for_disk(&location);
 599                         memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i, sizeof(uint64_t)),
 600                                     &location, sizeof(__le64));
 601                 }
 602
 603                 node = dm_block_data(shadow_current(s));
 604
 605                 if (node->header.nr_entries == node->header.max_entries) {
 606                         if (top)
 607                                 r = btree_split_beneath(s, key);
 608                         else
 609                                 r = btree_split_sibling(s, root, i, key);
 610
 611                         if (r < 0)
 612                                 return r;
 613                 }
 614
 615                 node = dm_block_data(shadow_current(s));
 616
 617                 i = lower_bound(node, key);
 618
 619                 if (le32_to_cpu(node->header.flags) & LEAF_NODE)
 620                         break;
 621
 622                 if (i < 0) {
 623                         /* change the bounds on the lowest key */
 624                         node->keys[0] = cpu_to_le64(key);
 625                         i = 0;
 626                 }
 627
 628                 root = value64(node, i);
 629                 top = 0;
 630         }
 631
 632         if (i < 0 || le64_to_cpu(node->keys[i]) != key)
 633                 i++;
 634
 635         *index = i;
 636         return 0;
 637 }
 638
 639 static int insert(struct dm_btree_info *info, dm_block_t root,
 640                   uint64_t *keys, void *value, dm_block_t *new_root,
 641                   int *inserted)
 642                   __dm_written_to_disk(value)
 643 {
 644         int r, need_insert;
 645         unsigned level, index = -1, last_level = info->levels - 1;
 646         dm_block_t block = root;
 647         struct shadow_spine spine;
 648         struct btree_node *n;
 649         struct dm_btree_value_type le64_type;
 650
 651         init_le64_type(info->tm, &le64_type);
 652         init_shadow_spine(&spine, info);
 653
 654         for (level = 0; level < (info->levels - 1); level++) {
 655                 r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
 656                 if (r < 0)
 657                         goto bad;
 658
 659                 n = dm_block_data(shadow_current(&spine));
 660                 need_insert = ((index >= le32_to_cpu(n->header.nr_entries)) ||
 661                                (le64_to_cpu(n->keys[index]) != keys[level]));
 662
 663                 if (need_insert) {
 664                         dm_block_t new_tree;
 665                         __le64 new_le;
 666
 667                         r = dm_btree_empty(info, &new_tree);
 668                         if (r < 0)
 669                                 goto bad;
 670
 671                         new_le = cpu_to_le64(new_tree);
 672                         __dm_bless_for_disk(&new_le);
 673
 674                         r = insert_at(sizeof(uint64_t), n, index,
 675                                       keys[level], &new_le);
 676                         if (r)
 677                                 goto bad;
 678                 }
 679
 680                 if (level < last_level)
 681                         block = value64(n, index);
 682         }
 683
 684         r = btree_insert_raw(&spine, block, &info->value_type,
 685                              keys[level], &index);
 686         if (r < 0)
 687                 goto bad;
 688
 689         n = dm_block_data(shadow_current(&spine));
 690         need_insert = ((index >= le32_to_cpu(n->header.nr_entries)) ||
 691                        (le64_to_cpu(n->keys[index]) != keys[level]));
 692
 693         if (need_insert) {
 694                 if (inserted)
 695                         *inserted = 1;
 696
 697                 r = insert_at(info->value_type.size, n, index,
 698                               keys[level], value);
 699                 if (r)
 700                         goto bad_unblessed;
 701         } else {
 702                 if (inserted)
 703                         *inserted = 0;
 704
 705                 if (info->value_type.dec &&
 706                     (!info->value_type.equal ||
 707                      !info->value_type.equal(
 708                              info->value_type.context,
 709                              value_ptr(n, index, info->value_type.size),
 710                              value))) {
 711                         info->value_type.dec(info->value_type.context,
 712                                              value_ptr(n, index, info->value_type.size));
 713                 }
 714                 memcpy_disk(value_ptr(n, index, info->value_type.size),
 715                             value, info->value_type.size);
 716         }
 717
 718         *new_root = shadow_root(&spine);
 719         exit_shadow_spine(&spine);
 720
 721         return 0;
 722
 723 bad:
 724         __dm_unbless_for_disk(value);
 725 bad_unblessed:
 726         exit_shadow_spine(&spine);
 727         return r;
 728 }
 729
 730 int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
 731                     uint64_t *keys, void *value, dm_block_t *new_root)
 732                     __dm_written_to_disk(value)
 733 {
 734         return insert(info, root, keys, value, new_root, NULL);
 735 }
 736 EXPORT_SYMBOL_GPL(dm_btree_insert);
 737
 738 int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
 739                            uint64_t *keys, void *value, dm_block_t *new_root,
 740                            int *inserted)
 741                            __dm_written_to_disk(value)
 742 {
 743         return insert(info, root, keys, value, new_root, inserted);
 744 }
 745 EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
 746
 747 /*----------------------------------------------------------------*/
 748
 749 static int find_highest_key(struct ro_spine *s, dm_block_t block,
 750                             uint64_t *result_key, dm_block_t *next_block)
 751 {
 752         int i, r;
 753         uint32_t flags;
 754
 755         do {
 756                 r = ro_step(s, block);
 757                 if (r < 0)
 758                         return r;
 759
 760                 flags = le32_to_cpu(ro_node(s)->header.flags);
 761                 i = le32_to_cpu(ro_node(s)->header.nr_entries);
 762                 if (!i)
 763                         return -ENODATA;
 764                 else
 765                         i--;
 766
 767                 *result_key = le64_to_cpu(ro_node(s)->keys[i]);
 768                 if (next_block || flags & INTERNAL_NODE)
 769                         block = value64(ro_node(s), i);
 770
 771         } while (flags & INTERNAL_NODE);
 772
 773         if (next_block)
 774                 *next_block = block;
 775         return 0;
 776 }
 777
 778 int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
 779                               uint64_t *result_keys)
 780 {
 781         int r = 0, count = 0, level;
 782         struct ro_spine spine;
 783
 784         init_ro_spine(&spine, info);
 785         for (level = 0; level < info->levels; level++) {
 786                 r = find_highest_key(&spine, root, result_keys + level,
 787                                      level == info->levels - 1 ? NULL : &root);
 788                 if (r == -ENODATA) {
 789                         r = 0;
 790                         break;
 791
 792                 } else if (r)
 793                         break;
 794
 795                 count++;
 796         }
 797         exit_ro_spine(&spine);
 798
 799         return r ? r : count;
 800 }
 801 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);