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 int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
 235 {
 236         int r;
 237         struct del_stack *s;
 238
 239         s = kmalloc(sizeof(*s), GFP_NOIO);
 240         if (!s)
 241                 return -ENOMEM;
 242         s->tm = info->tm;
 243         s->top = -1;
 244
 245         r = push_frame(s, root, 1);
 246         if (r)
 247                 goto out;
 248
 249         while (unprocessed_frames(s)) {
 250                 uint32_t flags;
 251                 struct frame *f;
 252                 dm_block_t b;
 253
 254                 r = top_frame(s, &f);
 255                 if (r)
 256                         goto out;
 257
 258                 if (f->current_child >= f->nr_children) {
 259                         pop_frame(s);
 260                         continue;
 261                 }
 262
 263                 flags = le32_to_cpu(f->n->header.flags);
 264                 if (flags & INTERNAL_NODE) {
 265                         b = value64(f->n, f->current_child);
 266                         f->current_child++;
 267                         r = push_frame(s, b, f->level);
 268                         if (r)
 269                                 goto out;
 270
 271                 } else if (f->level != (info->levels - 1)) {
 272                         b = value64(f->n, f->current_child);
 273                         f->current_child++;
 274                         r = push_frame(s, b, f->level + 1);
 275                         if (r)
 276                                 goto out;
 277
 278                 } else {
 279                         if (info->value_type.dec) {
 280                                 unsigned i;
 281
 282                                 for (i = 0; i < f->nr_children; i++)
 283                                         info->value_type.dec(info->value_type.context,
 284                                                              value_ptr(f->n, i, info->value_type.size));
 285                         }
 286                         f->current_child = f->nr_children;
 287                 }
 288         }
 289
 290 out:
 291         kfree(s);
 292         return r;
 293 }
 294 EXPORT_SYMBOL_GPL(dm_btree_del);
 295
 296 /*----------------------------------------------------------------*/
 297
 298 static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
 299                             int (*search_fn)(struct btree_node *, uint64_t),
 300                             uint64_t *result_key, void *v, size_t value_size)
 301 {
 302         int i, r;
 303         uint32_t flags, nr_entries;
 304
 305         do {
 306                 r = ro_step(s, block);
 307                 if (r < 0)
 308                         return r;
 309
 310                 i = search_fn(ro_node(s), key);
 311
 312                 flags = le32_to_cpu(ro_node(s)->header.flags);
 313                 nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
 314                 if (i < 0 || i >= nr_entries)
 315                         return -ENODATA;
 316
 317                 if (flags & INTERNAL_NODE)
 318                         block = value64(ro_node(s), i);
 319
 320         } while (!(flags & LEAF_NODE));
 321
 322         *result_key = le64_to_cpu(ro_node(s)->keys[i]);
 323         memcpy(v, value_ptr(ro_node(s), i, value_size), value_size);
 324
 325         return 0;
 326 }
 327
 328 int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
 329                     uint64_t *keys, void *value_le)
 330 {
 331         unsigned level, last_level = info->levels - 1;
 332         int r = -ENODATA;
 333         uint64_t rkey;
 334         __le64 internal_value_le;
 335         struct ro_spine spine;
 336
 337         init_ro_spine(&spine, info);
 338         for (level = 0; level < info->levels; level++) {
 339                 size_t size;
 340                 void *value_p;
 341
 342                 if (level == last_level) {
 343                         value_p = value_le;
 344                         size = info->value_type.size;
 345
 346                 } else {
 347                         value_p = &internal_value_le;
 348                         size = sizeof(uint64_t);
 349                 }
 350
 351                 r = btree_lookup_raw(&spine, root, keys[level],
 352                                      lower_bound, &rkey,
 353                                      value_p, size);
 354
 355                 if (!r) {
 356                         if (rkey != keys[level]) {
 357                                 exit_ro_spine(&spine);
 358                                 return -ENODATA;
 359                         }
 360                 } else {
 361                         exit_ro_spine(&spine);
 362                         return r;
 363                 }
 364
 365                 root = le64_to_cpu(internal_value_le);
 366         }
 367         exit_ro_spine(&spine);
 368
 369         return r;
 370 }
 371 EXPORT_SYMBOL_GPL(dm_btree_lookup);
 372
 373 /*
 374  * Splits a node by creating a sibling node and shifting half the nodes
 375  * contents across.  Assumes there is a parent node, and it has room for
 376  * another child.
 377  *
 378  * Before:
 379  *        +--------+
 380  *        | Parent |
 381  *        +--------+
 382  *           |
 383  *           v
 384  *      +----------+
 385  *      | A ++++++ |
 386  *      +----------+
 387  *
 388  *
 389  * After:
 390  *              +--------+
 391  *              | Parent |
 392  *              +--------+
 393  *                |     |
 394  *                v     +------+
 395  *          +---------+        |
 396  *          | A* +++  |        v
 397  *          +---------+   +-------+
 398  *                        | B +++ |
 399  *                        +-------+
 400  *
 401  * Where A* is a shadow of A.
 402  */
 403 static int btree_split_sibling(struct shadow_spine *s, dm_block_t root,
 404                                unsigned parent_index, uint64_t key)
 405 {
 406         int r;
 407         size_t size;
 408         unsigned nr_left, nr_right;
 409         struct dm_block *left, *right, *parent;
 410         struct btree_node *ln, *rn, *pn;
 411         __le64 location;
 412
 413         left = shadow_current(s);
 414
 415         r = new_block(s->info, &right);
 416         if (r < 0)
 417                 return r;
 418
 419         ln = dm_block_data(left);
 420         rn = dm_block_data(right);
 421
 422         nr_left = le32_to_cpu(ln->header.nr_entries) / 2;
 423         nr_right = le32_to_cpu(ln->header.nr_entries) - nr_left;
 424
 425         ln->header.nr_entries = cpu_to_le32(nr_left);
 426
 427         rn->header.flags = ln->header.flags;
 428         rn->header.nr_entries = cpu_to_le32(nr_right);
 429         rn->header.max_entries = ln->header.max_entries;
 430         rn->header.value_size = ln->header.value_size;
 431         memcpy(rn->keys, ln->keys + nr_left, nr_right * sizeof(rn->keys[0]));
 432
 433         size = le32_to_cpu(ln->header.flags) & INTERNAL_NODE ?
 434                 sizeof(uint64_t) : s->info->value_type.size;
 435         memcpy(value_ptr(rn, 0, size), value_ptr(ln, nr_left, size),
 436                size * nr_right);
 437
 438         /*
 439          * Patch up the parent
 440          */
 441         parent = shadow_parent(s);
 442
 443         pn = dm_block_data(parent);
 444         location = cpu_to_le64(dm_block_location(left));
 445         __dm_bless_for_disk(&location);
 446         memcpy_disk(value_ptr(pn, parent_index, sizeof(__le64)),
 447                     &location, sizeof(__le64));
 448
 449         location = cpu_to_le64(dm_block_location(right));
 450         __dm_bless_for_disk(&location);
 451
 452         r = insert_at(sizeof(__le64), pn, parent_index + 1,
 453                       le64_to_cpu(rn->keys[0]), &location);
 454         if (r) {
 455                 unlock_block(s->info, right);
 456                 return r;
 457         }
 458
 459         if (key < le64_to_cpu(rn->keys[0])) {
 460                 unlock_block(s->info, right);
 461                 s->nodes[1] = left;
 462         } else {
 463                 unlock_block(s->info, left);
 464                 s->nodes[1] = right;
 465         }
 466
 467         return 0;
 468 }
 469
 470 /*
 471  * Splits a node by creating two new children beneath the given node.
 472  *
 473  * Before:
 474  *        +----------+
 475  *        | A ++++++ |
 476  *        +----------+
 477  *
 478  *
 479  * After:
 480  *      +------------+
 481  *      | A (shadow) |
 482  *      +------------+
 483  *          |   |
 484  *   +------+   +----+
 485  *   |               |
 486  *   v               v
 487  * +-------+     +-------+
 488  * | B +++ |     | C +++ |
 489  * +-------+     +-------+
 490  */
 491 static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
 492 {
 493         int r;
 494         size_t size;
 495         unsigned nr_left, nr_right;
 496         struct dm_block *left, *right, *new_parent;
 497         struct btree_node *pn, *ln, *rn;
 498         __le64 val;
 499
 500         new_parent = shadow_current(s);
 501
 502         r = new_block(s->info, &left);
 503         if (r < 0)
 504                 return r;
 505
 506         r = new_block(s->info, &right);
 507         if (r < 0) {
 508                 unlock_block(s->info, left);
 509                 return r;
 510         }
 511
 512         pn = dm_block_data(new_parent);
 513         ln = dm_block_data(left);
 514         rn = dm_block_data(right);
 515
 516         nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
 517         nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
 518
 519         ln->header.flags = pn->header.flags;
 520         ln->header.nr_entries = cpu_to_le32(nr_left);
 521         ln->header.max_entries = pn->header.max_entries;
 522         ln->header.value_size = pn->header.value_size;
 523
 524         rn->header.flags = pn->header.flags;
 525         rn->header.nr_entries = cpu_to_le32(nr_right);
 526         rn->header.max_entries = pn->header.max_entries;
 527         rn->header.value_size = pn->header.value_size;
 528
 529         memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
 530         memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
 531
 532         size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
 533                 sizeof(__le64) : s->info->value_type.size;
 534         memcpy(value_ptr(ln, 0, size), value_ptr(pn, 0, size), nr_left * size);
 535         memcpy(value_ptr(rn, 0, size), value_ptr(pn, nr_left, size),
 536                nr_right * size);
 537
 538         /* new_parent should just point to l and r now */
 539         pn->header.flags = cpu_to_le32(INTERNAL_NODE);
 540         pn->header.nr_entries = cpu_to_le32(2);
 541         pn->header.max_entries = cpu_to_le32(
 542                 calc_max_entries(sizeof(__le64),
 543                                  dm_bm_block_size(
 544                                          dm_tm_get_bm(s->info->tm))));
 545         pn->header.value_size = cpu_to_le32(sizeof(__le64));
 546
 547         val = cpu_to_le64(dm_block_location(left));
 548         __dm_bless_for_disk(&val);
 549         pn->keys[0] = ln->keys[0];
 550         memcpy_disk(value_ptr(pn, 0, sizeof(__le64)), &val, sizeof(__le64));
 551
 552         val = cpu_to_le64(dm_block_location(right));
 553         __dm_bless_for_disk(&val);
 554         pn->keys[1] = rn->keys[0];
 555         memcpy_disk(value_ptr(pn, 1, sizeof(__le64)), &val, sizeof(__le64));
 556
 557         /*
 558          * rejig the spine.  This is ugly, since it knows too
 559          * much about the spine
 560          */
 561         if (s->nodes[0] != new_parent) {
 562                 unlock_block(s->info, s->nodes[0]);
 563                 s->nodes[0] = new_parent;
 564         }
 565         if (key < le64_to_cpu(rn->keys[0])) {
 566                 unlock_block(s->info, right);
 567                 s->nodes[1] = left;
 568         } else {
 569                 unlock_block(s->info, left);
 570                 s->nodes[1] = right;
 571         }
 572         s->count = 2;
 573
 574         return 0;
 575 }
 576
 577 static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
 578                             struct dm_btree_value_type *vt,
 579                             uint64_t key, unsigned *index)
 580 {
 581         int r, i = *index, top = 1;
 582         struct btree_node *node;
 583
 584         for (;;) {
 585                 r = shadow_step(s, root, vt);
 586                 if (r < 0)
 587                         return r;
 588
 589                 node = dm_block_data(shadow_current(s));
 590
 591                 /*
 592                  * We have to patch up the parent node, ugly, but I don't
 593                  * see a way to do this automatically as part of the spine
 594                  * op.
 595                  */
 596                 if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
 597                         __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
 598
 599                         __dm_bless_for_disk(&location);
 600                         memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i, sizeof(uint64_t)),
 601                                     &location, sizeof(__le64));
 602                 }
 603
 604                 node = dm_block_data(shadow_current(s));
 605
 606                 if (node->header.nr_entries == node->header.max_entries) {
 607                         if (top)
 608                                 r = btree_split_beneath(s, key);
 609                         else
 610                                 r = btree_split_sibling(s, root, i, key);
 611
 612                         if (r < 0)
 613                                 return r;
 614                 }
 615
 616                 node = dm_block_data(shadow_current(s));
 617
 618                 i = lower_bound(node, key);
 619
 620                 if (le32_to_cpu(node->header.flags) & LEAF_NODE)
 621                         break;
 622
 623                 if (i < 0) {
 624                         /* change the bounds on the lowest key */
 625                         node->keys[0] = cpu_to_le64(key);
 626                         i = 0;
 627                 }
 628
 629                 root = value64(node, i);
 630                 top = 0;
 631         }
 632
 633         if (i < 0 || le64_to_cpu(node->keys[i]) != key)
 634                 i++;
 635
 636         *index = i;
 637         return 0;
 638 }
 639
 640 static int insert(struct dm_btree_info *info, dm_block_t root,
 641                   uint64_t *keys, void *value, dm_block_t *new_root,
 642                   int *inserted)
 643                   __dm_written_to_disk(value)
 644 {
 645         int r, need_insert;
 646         unsigned level, index = -1, last_level = info->levels - 1;
 647         dm_block_t block = root;
 648         struct shadow_spine spine;
 649         struct btree_node *n;
 650         struct dm_btree_value_type le64_type;
 651
 652         init_le64_type(info->tm, &le64_type);
 653         init_shadow_spine(&spine, info);
 654
 655         for (level = 0; level < (info->levels - 1); level++) {
 656                 r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
 657                 if (r < 0)
 658                         goto bad;
 659
 660                 n = dm_block_data(shadow_current(&spine));
 661                 need_insert = ((index >= le32_to_cpu(n->header.nr_entries)) ||
 662                                (le64_to_cpu(n->keys[index]) != keys[level]));
 663
 664                 if (need_insert) {
 665                         dm_block_t new_tree;
 666                         __le64 new_le;
 667
 668                         r = dm_btree_empty(info, &new_tree);
 669                         if (r < 0)
 670                                 goto bad;
 671
 672                         new_le = cpu_to_le64(new_tree);
 673                         __dm_bless_for_disk(&new_le);
 674
 675                         r = insert_at(sizeof(uint64_t), n, index,
 676                                       keys[level], &new_le);
 677                         if (r)
 678                                 goto bad;
 679                 }
 680
 681                 if (level < last_level)
 682                         block = value64(n, index);
 683         }
 684
 685         r = btree_insert_raw(&spine, block, &info->value_type,
 686                              keys[level], &index);
 687         if (r < 0)
 688                 goto bad;
 689
 690         n = dm_block_data(shadow_current(&spine));
 691         need_insert = ((index >= le32_to_cpu(n->header.nr_entries)) ||
 692                        (le64_to_cpu(n->keys[index]) != keys[level]));
 693
 694         if (need_insert) {
 695                 if (inserted)
 696                         *inserted = 1;
 697
 698                 r = insert_at(info->value_type.size, n, index,
 699                               keys[level], value);
 700                 if (r)
 701                         goto bad_unblessed;
 702         } else {
 703                 if (inserted)
 704                         *inserted = 0;
 705
 706                 if (info->value_type.dec &&
 707                     (!info->value_type.equal ||
 708                      !info->value_type.equal(
 709                              info->value_type.context,
 710                              value_ptr(n, index, info->value_type.size),
 711                              value))) {
 712                         info->value_type.dec(info->value_type.context,
 713                                              value_ptr(n, index, info->value_type.size));
 714                 }
 715                 memcpy_disk(value_ptr(n, index, info->value_type.size),
 716                             value, info->value_type.size);
 717         }
 718
 719         *new_root = shadow_root(&spine);
 720         exit_shadow_spine(&spine);
 721
 722         return 0;
 723
 724 bad:
 725         __dm_unbless_for_disk(value);
 726 bad_unblessed:
 727         exit_shadow_spine(&spine);
 728         return r;
 729 }
 730
 731 int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
 732                     uint64_t *keys, void *value, dm_block_t *new_root)
 733                     __dm_written_to_disk(value)
 734 {
 735         return insert(info, root, keys, value, new_root, NULL);
 736 }
 737 EXPORT_SYMBOL_GPL(dm_btree_insert);
 738
 739 int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
 740                            uint64_t *keys, void *value, dm_block_t *new_root,
 741                            int *inserted)
 742                            __dm_written_to_disk(value)
 743 {
 744         return insert(info, root, keys, value, new_root, inserted);
 745 }
 746 EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
 747
 748 /*----------------------------------------------------------------*/
 749
 750 static int find_highest_key(struct ro_spine *s, dm_block_t block,
 751                             uint64_t *result_key, dm_block_t *next_block)
 752 {
 753         int i, r;
 754         uint32_t flags;
 755
 756         do {
 757                 r = ro_step(s, block);
 758                 if (r < 0)
 759                         return r;
 760
 761                 flags = le32_to_cpu(ro_node(s)->header.flags);
 762                 i = le32_to_cpu(ro_node(s)->header.nr_entries);
 763                 if (!i)
 764                         return -ENODATA;
 765                 else
 766                         i--;
 767
 768                 *result_key = le64_to_cpu(ro_node(s)->keys[i]);
 769                 if (next_block || flags & INTERNAL_NODE)
 770                         block = value64(ro_node(s), i);
 771
 772         } while (flags & INTERNAL_NODE);
 773
 774         if (next_block)
 775                 *next_block = block;
 776         return 0;
 777 }
 778
 779 int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
 780                               uint64_t *result_keys)
 781 {
 782         int r = 0, count = 0, level;
 783         struct ro_spine spine;
 784
 785         init_ro_spine(&spine, info);
 786         for (level = 0; level < info->levels; level++) {
 787                 r = find_highest_key(&spine, root, result_keys + level,
 788                                      level == info->levels - 1 ? NULL : &root);
 789                 if (r == -ENODATA) {
 790                         r = 0;
 791                         break;
 792
 793                 } else if (r)
 794                         break;
 795
 796                 count++;
 797         }
 798         exit_ro_spine(&spine);
 799
 800         return r ? r : count;
 801 }
 802 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);