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 static int upper_bound(struct btree_node *n, uint64_t key)
  67 {
  68         return bsearch(n, key, 1);
  69 }
  70
  71 void inc_children(struct dm_transaction_manager *tm, struct btree_node *n,
  72                   struct dm_btree_value_type *vt)
  73 {
  74         unsigned i;
  75         uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
  76
  77         if (le32_to_cpu(n->header.flags) & INTERNAL_NODE)
  78                 for (i = 0; i < nr_entries; i++)
  79                         dm_tm_inc(tm, value64(n, i));
  80         else if (vt->inc)
  81                 for (i = 0; i < nr_entries; i++)
  82                         vt->inc(vt->context, value_ptr(n, i));
  83 }
  84
  85 static int insert_at(size_t value_size, struct btree_node *node, unsigned index,
  86                      uint64_t key, void *value)
  87         __dm_written_to_disk(value)
  88 {
  89         uint32_t nr_entries = le32_to_cpu(node->header.nr_entries);
  90         uint32_t max_entries = le32_to_cpu(node->header.max_entries);
  91         __le64 key_le = cpu_to_le64(key);
  92
  93         if (index > nr_entries ||
  94             index >= max_entries ||
  95             nr_entries >= max_entries) {
  96                 DMERR("too many entries in btree node for insert");
  97                 __dm_unbless_for_disk(value);
  98                 return -ENOMEM;
  99         }
 100
 101         __dm_bless_for_disk(&key_le);
 102
 103         array_insert(node->keys, sizeof(*node->keys), nr_entries, index, &key_le);
 104         array_insert(value_base(node), value_size, nr_entries, index, value);
 105         node->header.nr_entries = cpu_to_le32(nr_entries + 1);
 106
 107         return 0;
 108 }
 109
 110 /*----------------------------------------------------------------*/
 111
 112 /*
 113  * We want 3n entries (for some n).  This works more nicely for repeated
 114  * insert remove loops than (2n + 1).
 115  */
 116 static uint32_t calc_max_entries(size_t value_size, size_t block_size)
 117 {
 118         uint32_t total, n;
 119         size_t elt_size = sizeof(uint64_t) + value_size; /* key + value */
 120
 121         block_size -= sizeof(struct node_header);
 122         total = block_size / elt_size;
 123         n = total / 3;          /* rounds down */
 124
 125         return 3 * n;
 126 }
 127
 128 int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root)
 129 {
 130         int r;
 131         struct dm_block *b;
 132         struct btree_node *n;
 133         size_t block_size;
 134         uint32_t max_entries;
 135
 136         r = new_block(info, &b);
 137         if (r < 0)
 138                 return r;
 139
 140         block_size = dm_bm_block_size(dm_tm_get_bm(info->tm));
 141         max_entries = calc_max_entries(info->value_type.size, block_size);
 142
 143         n = dm_block_data(b);
 144         memset(n, 0, block_size);
 145         n->header.flags = cpu_to_le32(LEAF_NODE);
 146         n->header.nr_entries = cpu_to_le32(0);
 147         n->header.max_entries = cpu_to_le32(max_entries);
 148         n->header.value_size = cpu_to_le32(info->value_type.size);
 149
 150         *root = dm_block_location(b);
 151         unlock_block(info, b);
 152
 153         return 0;
 154 }
 155 EXPORT_SYMBOL_GPL(dm_btree_empty);
 156
 157 /*----------------------------------------------------------------*/
 158
 159 /*
 160  * Deletion uses a recursive algorithm, since we have limited stack space
 161  * we explicitly manage our own stack on the heap.
 162  */
 163 #define MAX_SPINE_DEPTH 64
 164 struct frame {
 165         struct dm_block *b;
 166         struct btree_node *n;
 167         unsigned level;
 168         unsigned nr_children;
 169         unsigned current_child;
 170 };
 171
 172 struct del_stack {
 173         struct dm_btree_info *info;
 174         struct dm_transaction_manager *tm;
 175         int top;
 176         struct frame spine[MAX_SPINE_DEPTH];
 177 };
 178
 179 static int top_frame(struct del_stack *s, struct frame **f)
 180 {
 181         if (s->top < 0) {
 182                 DMERR("btree deletion stack empty");
 183                 return -EINVAL;
 184         }
 185
 186         *f = s->spine + s->top;
 187
 188         return 0;
 189 }
 190
 191 static int unprocessed_frames(struct del_stack *s)
 192 {
 193         return s->top >= 0;
 194 }
 195
 196 static void prefetch_children(struct del_stack *s, struct frame *f)
 197 {
 198         unsigned i;
 199         struct dm_block_manager *bm = dm_tm_get_bm(s->tm);
 200
 201         for (i = 0; i < f->nr_children; i++)
 202                 dm_bm_prefetch(bm, value64(f->n, i));
 203 }
 204
 205 static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
 206 {
 207         return f->level < (info->levels - 1);
 208 }
 209
 210 static int push_frame(struct del_stack *s, dm_block_t b, unsigned level)
 211 {
 212         int r;
 213         uint32_t ref_count;
 214
 215         if (s->top >= MAX_SPINE_DEPTH - 1) {
 216                 DMERR("btree deletion stack out of memory");
 217                 return -ENOMEM;
 218         }
 219
 220         r = dm_tm_ref(s->tm, b, &ref_count);
 221         if (r)
 222                 return r;
 223
 224         if (ref_count > 1)
 225                 /*
 226                  * This is a shared node, so we can just decrement it's
 227                  * reference counter and leave the children.
 228                  */
 229                 dm_tm_dec(s->tm, b);
 230
 231         else {
 232                 uint32_t flags;
 233                 struct frame *f = s->spine + ++s->top;
 234
 235                 r = dm_tm_read_lock(s->tm, b, &btree_node_validator, &f->b);
 236                 if (r) {
 237                         s->top--;
 238                         return r;
 239                 }
 240
 241                 f->n = dm_block_data(f->b);
 242                 f->level = level;
 243                 f->nr_children = le32_to_cpu(f->n->header.nr_entries);
 244                 f->current_child = 0;
 245
 246                 flags = le32_to_cpu(f->n->header.flags);
 247                 if (flags & INTERNAL_NODE || is_internal_level(s->info, f))
 248                         prefetch_children(s, f);
 249         }
 250
 251         return 0;
 252 }
 253
 254 static void pop_frame(struct del_stack *s)
 255 {
 256         struct frame *f = s->spine + s->top--;
 257
 258         dm_tm_dec(s->tm, dm_block_location(f->b));
 259         dm_tm_unlock(s->tm, f->b);
 260 }
 261
 262 static void unlock_all_frames(struct del_stack *s)
 263 {
 264         struct frame *f;
 265
 266         while (unprocessed_frames(s)) {
 267                 f = s->spine + s->top--;
 268                 dm_tm_unlock(s->tm, f->b);
 269         }
 270 }
 271
 272 int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
 273 {
 274         int r;
 275         struct del_stack *s;
 276
 277         /*
 278          * dm_btree_del() is called via an ioctl, as such should be
 279          * considered an FS op.  We can't recurse back into the FS, so we
 280          * allocate GFP_NOFS.
 281          */
 282         s = kmalloc(sizeof(*s), GFP_NOFS);
 283         if (!s)
 284                 return -ENOMEM;
 285         s->info = info;
 286         s->tm = info->tm;
 287         s->top = -1;
 288
 289         r = push_frame(s, root, 0);
 290         if (r)
 291                 goto out;
 292
 293         while (unprocessed_frames(s)) {
 294                 uint32_t flags;
 295                 struct frame *f;
 296                 dm_block_t b;
 297
 298                 r = top_frame(s, &f);
 299                 if (r)
 300                         goto out;
 301
 302                 if (f->current_child >= f->nr_children) {
 303                         pop_frame(s);
 304                         continue;
 305                 }
 306
 307                 flags = le32_to_cpu(f->n->header.flags);
 308                 if (flags & INTERNAL_NODE) {
 309                         b = value64(f->n, f->current_child);
 310                         f->current_child++;
 311                         r = push_frame(s, b, f->level);
 312                         if (r)
 313                                 goto out;
 314
 315                 } else if (is_internal_level(info, f)) {
 316                         b = value64(f->n, f->current_child);
 317                         f->current_child++;
 318                         r = push_frame(s, b, f->level + 1);
 319                         if (r)
 320                                 goto out;
 321
 322                 } else {
 323                         if (info->value_type.dec) {
 324                                 unsigned i;
 325
 326                                 for (i = 0; i < f->nr_children; i++)
 327                                         info->value_type.dec(info->value_type.context,
 328                                                              value_ptr(f->n, i));
 329                         }
 330                         pop_frame(s);
 331                 }
 332         }
 333 out:
 334         if (r) {
 335                 /* cleanup all frames of del_stack */
 336                 unlock_all_frames(s);
 337         }
 338         kfree(s);
 339
 340         return r;
 341 }
 342 EXPORT_SYMBOL_GPL(dm_btree_del);
 343
 344 /*----------------------------------------------------------------*/
 345
 346 static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
 347                             int (*search_fn)(struct btree_node *, uint64_t),
 348                             uint64_t *result_key, void *v, size_t value_size)
 349 {
 350         int i, r;
 351         uint32_t flags, nr_entries;
 352
 353         do {
 354                 r = ro_step(s, block);
 355                 if (r < 0)
 356                         return r;
 357
 358                 i = search_fn(ro_node(s), key);
 359
 360                 flags = le32_to_cpu(ro_node(s)->header.flags);
 361                 nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
 362                 if (i < 0 || i >= nr_entries)
 363                         return -ENODATA;
 364
 365                 if (flags & INTERNAL_NODE)
 366                         block = value64(ro_node(s), i);
 367
 368         } while (!(flags & LEAF_NODE));
 369
 370         *result_key = le64_to_cpu(ro_node(s)->keys[i]);
 371         memcpy(v, value_ptr(ro_node(s), i), value_size);
 372
 373         return 0;
 374 }
 375
 376 int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
 377                     uint64_t *keys, void *value_le)
 378 {
 379         unsigned level, last_level = info->levels - 1;
 380         int r = -ENODATA;
 381         uint64_t rkey;
 382         __le64 internal_value_le;
 383         struct ro_spine spine;
 384
 385         init_ro_spine(&spine, info);
 386         for (level = 0; level < info->levels; level++) {
 387                 size_t size;
 388                 void *value_p;
 389
 390                 if (level == last_level) {
 391                         value_p = value_le;
 392                         size = info->value_type.size;
 393
 394                 } else {
 395                         value_p = &internal_value_le;
 396                         size = sizeof(uint64_t);
 397                 }
 398
 399                 r = btree_lookup_raw(&spine, root, keys[level],
 400                                      lower_bound, &rkey,
 401                                      value_p, size);
 402
 403                 if (!r) {
 404                         if (rkey != keys[level]) {
 405                                 exit_ro_spine(&spine);
 406                                 return -ENODATA;
 407                         }
 408                 } else {
 409                         exit_ro_spine(&spine);
 410                         return r;
 411                 }
 412
 413                 root = le64_to_cpu(internal_value_le);
 414         }
 415         exit_ro_spine(&spine);
 416
 417         return r;
 418 }
 419 EXPORT_SYMBOL_GPL(dm_btree_lookup);
 420
 421 static int dm_btree_lookup_next_single(struct dm_btree_info *info, dm_block_t root,
 422                                        uint64_t key, uint64_t *rkey, void *value_le)
 423 {
 424         int r, i;
 425         uint32_t flags, nr_entries;
 426         struct dm_block *node;
 427         struct btree_node *n;
 428
 429         r = bn_read_lock(info, root, &node);
 430         if (r)
 431                 return r;
 432
 433         n = dm_block_data(node);
 434         flags = le32_to_cpu(n->header.flags);
 435         nr_entries = le32_to_cpu(n->header.nr_entries);
 436
 437         if (flags & INTERNAL_NODE) {
 438                 i = lower_bound(n, key);
 439                 if (i < 0) {
 440                         /*
 441                          * avoid early -ENODATA return when all entries are
 442                          * higher than the search @key.
 443                          */
 444                         i = 0;
 445                 }
 446                 if (i >= nr_entries) {
 447                         r = -ENODATA;
 448                         goto out;
 449                 }
 450
 451                 r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
 452                 if (r == -ENODATA && i < (nr_entries - 1)) {
 453                         i++;
 454                         r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
 455                 }
 456
 457         } else {
 458                 i = upper_bound(n, key);
 459                 if (i < 0 || i >= nr_entries) {
 460                         r = -ENODATA;
 461                         goto out;
 462                 }
 463
 464                 *rkey = le64_to_cpu(n->keys[i]);
 465                 memcpy(value_le, value_ptr(n, i), info->value_type.size);
 466         }
 467 out:
 468         dm_tm_unlock(info->tm, node);
 469         return r;
 470 }
 471
 472 int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
 473                          uint64_t *keys, uint64_t *rkey, void *value_le)
 474 {
 475         unsigned level;
 476         int r = -ENODATA;
 477         __le64 internal_value_le;
 478         struct ro_spine spine;
 479
 480         init_ro_spine(&spine, info);
 481         for (level = 0; level < info->levels - 1u; level++) {
 482                 r = btree_lookup_raw(&spine, root, keys[level],
 483                                      lower_bound, rkey,
 484                                      &internal_value_le, sizeof(uint64_t));
 485                 if (r)
 486                         goto out;
 487
 488                 if (*rkey != keys[level]) {
 489                         r = -ENODATA;
 490                         goto out;
 491                 }
 492
 493                 root = le64_to_cpu(internal_value_le);
 494         }
 495
 496         r = dm_btree_lookup_next_single(info, root, keys[level], rkey, value_le);
 497 out:
 498         exit_ro_spine(&spine);
 499         return r;
 500 }
 501
 502 EXPORT_SYMBOL_GPL(dm_btree_lookup_next);
 503
 504 /*
 505  * Splits a node by creating a sibling node and shifting half the nodes
 506  * contents across.  Assumes there is a parent node, and it has room for
 507  * another child.
 508  *
 509  * Before:
 510  *        +--------+
 511  *        | Parent |
 512  *        +--------+
 513  *           |
 514  *           v
 515  *      +----------+
 516  *      | A ++++++ |
 517  *      +----------+
 518  *
 519  *
 520  * After:
 521  *              +--------+
 522  *              | Parent |
 523  *              +--------+
 524  *                |     |
 525  *                v     +------+
 526  *          +---------+        |
 527  *          | A* +++  |        v
 528  *          +---------+   +-------+
 529  *                        | B +++ |
 530  *                        +-------+
 531  *
 532  * Where A* is a shadow of A.
 533  */
 534 static int btree_split_sibling(struct shadow_spine *s, unsigned parent_index,
 535                                uint64_t key)
 536 {
 537         int r;
 538         size_t size;
 539         unsigned nr_left, nr_right;
 540         struct dm_block *left, *right, *parent;
 541         struct btree_node *ln, *rn, *pn;
 542         __le64 location;
 543
 544         left = shadow_current(s);
 545
 546         r = new_block(s->info, &right);
 547         if (r < 0)
 548                 return r;
 549
 550         ln = dm_block_data(left);
 551         rn = dm_block_data(right);
 552
 553         nr_left = le32_to_cpu(ln->header.nr_entries) / 2;
 554         nr_right = le32_to_cpu(ln->header.nr_entries) - nr_left;
 555
 556         ln->header.nr_entries = cpu_to_le32(nr_left);
 557
 558         rn->header.flags = ln->header.flags;
 559         rn->header.nr_entries = cpu_to_le32(nr_right);
 560         rn->header.max_entries = ln->header.max_entries;
 561         rn->header.value_size = ln->header.value_size;
 562         memcpy(rn->keys, ln->keys + nr_left, nr_right * sizeof(rn->keys[0]));
 563
 564         size = le32_to_cpu(ln->header.flags) & INTERNAL_NODE ?
 565                 sizeof(uint64_t) : s->info->value_type.size;
 566         memcpy(value_ptr(rn, 0), value_ptr(ln, nr_left),
 567                size * nr_right);
 568
 569         /*
 570          * Patch up the parent
 571          */
 572         parent = shadow_parent(s);
 573
 574         pn = dm_block_data(parent);
 575         location = cpu_to_le64(dm_block_location(left));
 576         __dm_bless_for_disk(&location);
 577         memcpy_disk(value_ptr(pn, parent_index),
 578                     &location, sizeof(__le64));
 579
 580         location = cpu_to_le64(dm_block_location(right));
 581         __dm_bless_for_disk(&location);
 582
 583         r = insert_at(sizeof(__le64), pn, parent_index + 1,
 584                       le64_to_cpu(rn->keys[0]), &location);
 585         if (r) {
 586                 unlock_block(s->info, right);
 587                 return r;
 588         }
 589
 590         if (key < le64_to_cpu(rn->keys[0])) {
 591                 unlock_block(s->info, right);
 592                 s->nodes[1] = left;
 593         } else {
 594                 unlock_block(s->info, left);
 595                 s->nodes[1] = right;
 596         }
 597
 598         return 0;
 599 }
 600
 601 /*
 602  * Splits a node by creating two new children beneath the given node.
 603  *
 604  * Before:
 605  *        +----------+
 606  *        | A ++++++ |
 607  *        +----------+
 608  *
 609  *
 610  * After:
 611  *      +------------+
 612  *      | A (shadow) |
 613  *      +------------+
 614  *          |   |
 615  *   +------+   +----+
 616  *   |               |
 617  *   v               v
 618  * +-------+     +-------+
 619  * | B +++ |     | C +++ |
 620  * +-------+     +-------+
 621  */
 622 static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
 623 {
 624         int r;
 625         size_t size;
 626         unsigned nr_left, nr_right;
 627         struct dm_block *left, *right, *new_parent;
 628         struct btree_node *pn, *ln, *rn;
 629         __le64 val;
 630
 631         new_parent = shadow_current(s);
 632
 633         pn = dm_block_data(new_parent);
 634         size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
 635                 sizeof(__le64) : s->info->value_type.size;
 636
 637         /* create & init the left block */
 638         r = new_block(s->info, &left);
 639         if (r < 0)
 640                 return r;
 641
 642         ln = dm_block_data(left);
 643         nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
 644
 645         ln->header.flags = pn->header.flags;
 646         ln->header.nr_entries = cpu_to_le32(nr_left);
 647         ln->header.max_entries = pn->header.max_entries;
 648         ln->header.value_size = pn->header.value_size;
 649         memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
 650         memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
 651
 652         /* create & init the right block */
 653         r = new_block(s->info, &right);
 654         if (r < 0) {
 655                 unlock_block(s->info, left);
 656                 return r;
 657         }
 658
 659         rn = dm_block_data(right);
 660         nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
 661
 662         rn->header.flags = pn->header.flags;
 663         rn->header.nr_entries = cpu_to_le32(nr_right);
 664         rn->header.max_entries = pn->header.max_entries;
 665         rn->header.value_size = pn->header.value_size;
 666         memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
 667         memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
 668                nr_right * size);
 669
 670         /* new_parent should just point to l and r now */
 671         pn->header.flags = cpu_to_le32(INTERNAL_NODE);
 672         pn->header.nr_entries = cpu_to_le32(2);
 673         pn->header.max_entries = cpu_to_le32(
 674                 calc_max_entries(sizeof(__le64),
 675                                  dm_bm_block_size(
 676                                          dm_tm_get_bm(s->info->tm))));
 677         pn->header.value_size = cpu_to_le32(sizeof(__le64));
 678
 679         val = cpu_to_le64(dm_block_location(left));
 680         __dm_bless_for_disk(&val);
 681         pn->keys[0] = ln->keys[0];
 682         memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
 683
 684         val = cpu_to_le64(dm_block_location(right));
 685         __dm_bless_for_disk(&val);
 686         pn->keys[1] = rn->keys[0];
 687         memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
 688
 689         unlock_block(s->info, left);
 690         unlock_block(s->info, right);
 691         return 0;
 692 }
 693
 694 static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
 695                             struct dm_btree_value_type *vt,
 696                             uint64_t key, unsigned *index)
 697 {
 698         int r, i = *index, top = 1;
 699         struct btree_node *node;
 700
 701         for (;;) {
 702                 r = shadow_step(s, root, vt);
 703                 if (r < 0)
 704                         return r;
 705
 706                 node = dm_block_data(shadow_current(s));
 707
 708                 /*
 709                  * We have to patch up the parent node, ugly, but I don't
 710                  * see a way to do this automatically as part of the spine
 711                  * op.
 712                  */
 713                 if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
 714                         __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
 715
 716                         __dm_bless_for_disk(&location);
 717                         memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
 718                                     &location, sizeof(__le64));
 719                 }
 720
 721                 node = dm_block_data(shadow_current(s));
 722
 723                 if (node->header.nr_entries == node->header.max_entries) {
 724                         if (top)
 725                                 r = btree_split_beneath(s, key);
 726                         else
 727                                 r = btree_split_sibling(s, i, key);
 728
 729                         if (r < 0)
 730                                 return r;
 731                 }
 732
 733                 node = dm_block_data(shadow_current(s));
 734
 735                 i = lower_bound(node, key);
 736
 737                 if (le32_to_cpu(node->header.flags) & LEAF_NODE)
 738                         break;
 739
 740                 if (i < 0) {
 741                         /* change the bounds on the lowest key */
 742                         node->keys[0] = cpu_to_le64(key);
 743                         i = 0;
 744                 }
 745
 746                 root = value64(node, i);
 747                 top = 0;
 748         }
 749
 750         if (i < 0 || le64_to_cpu(node->keys[i]) != key)
 751                 i++;
 752
 753         *index = i;
 754         return 0;
 755 }
 756
 757 static bool need_insert(struct btree_node *node, uint64_t *keys,
 758                         unsigned level, unsigned index)
 759 {
 760         return ((index >= le32_to_cpu(node->header.nr_entries)) ||
 761                 (le64_to_cpu(node->keys[index]) != keys[level]));
 762 }
 763
 764 static int insert(struct dm_btree_info *info, dm_block_t root,
 765                   uint64_t *keys, void *value, dm_block_t *new_root,
 766                   int *inserted)
 767                   __dm_written_to_disk(value)
 768 {
 769         int r;
 770         unsigned level, index = -1, last_level = info->levels - 1;
 771         dm_block_t block = root;
 772         struct shadow_spine spine;
 773         struct btree_node *n;
 774         struct dm_btree_value_type le64_type;
 775
 776         init_le64_type(info->tm, &le64_type);
 777         init_shadow_spine(&spine, info);
 778
 779         for (level = 0; level < (info->levels - 1); level++) {
 780                 r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
 781                 if (r < 0)
 782                         goto bad;
 783
 784                 n = dm_block_data(shadow_current(&spine));
 785
 786                 if (need_insert(n, keys, level, index)) {
 787                         dm_block_t new_tree;
 788                         __le64 new_le;
 789
 790                         r = dm_btree_empty(info, &new_tree);
 791                         if (r < 0)
 792                                 goto bad;
 793
 794                         new_le = cpu_to_le64(new_tree);
 795                         __dm_bless_for_disk(&new_le);
 796
 797                         r = insert_at(sizeof(uint64_t), n, index,
 798                                       keys[level], &new_le);
 799                         if (r)
 800                                 goto bad;
 801                 }
 802
 803                 if (level < last_level)
 804                         block = value64(n, index);
 805         }
 806
 807         r = btree_insert_raw(&spine, block, &info->value_type,
 808                              keys[level], &index);
 809         if (r < 0)
 810                 goto bad;
 811
 812         n = dm_block_data(shadow_current(&spine));
 813
 814         if (need_insert(n, keys, level, index)) {
 815                 if (inserted)
 816                         *inserted = 1;
 817
 818                 r = insert_at(info->value_type.size, n, index,
 819                               keys[level], value);
 820                 if (r)
 821                         goto bad_unblessed;
 822         } else {
 823                 if (inserted)
 824                         *inserted = 0;
 825
 826                 if (info->value_type.dec &&
 827                     (!info->value_type.equal ||
 828                      !info->value_type.equal(
 829                              info->value_type.context,
 830                              value_ptr(n, index),
 831                              value))) {
 832                         info->value_type.dec(info->value_type.context,
 833                                              value_ptr(n, index));
 834                 }
 835                 memcpy_disk(value_ptr(n, index),
 836                             value, info->value_type.size);
 837         }
 838
 839         *new_root = shadow_root(&spine);
 840         exit_shadow_spine(&spine);
 841
 842         return 0;
 843
 844 bad:
 845         __dm_unbless_for_disk(value);
 846 bad_unblessed:
 847         exit_shadow_spine(&spine);
 848         return r;
 849 }
 850
 851 int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
 852                     uint64_t *keys, void *value, dm_block_t *new_root)
 853                     __dm_written_to_disk(value)
 854 {
 855         return insert(info, root, keys, value, new_root, NULL);
 856 }
 857 EXPORT_SYMBOL_GPL(dm_btree_insert);
 858
 859 int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
 860                            uint64_t *keys, void *value, dm_block_t *new_root,
 861                            int *inserted)
 862                            __dm_written_to_disk(value)
 863 {
 864         return insert(info, root, keys, value, new_root, inserted);
 865 }
 866 EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
 867
 868 /*----------------------------------------------------------------*/
 869
 870 static int find_key(struct ro_spine *s, dm_block_t block, bool find_highest,
 871                     uint64_t *result_key, dm_block_t *next_block)
 872 {
 873         int i, r;
 874         uint32_t flags;
 875
 876         do {
 877                 r = ro_step(s, block);
 878                 if (r < 0)
 879                         return r;
 880
 881                 flags = le32_to_cpu(ro_node(s)->header.flags);
 882                 i = le32_to_cpu(ro_node(s)->header.nr_entries);
 883                 if (!i)
 884                         return -ENODATA;
 885                 else
 886                         i--;
 887
 888                 if (find_highest)
 889                         *result_key = le64_to_cpu(ro_node(s)->keys[i]);
 890                 else
 891                         *result_key = le64_to_cpu(ro_node(s)->keys[0]);
 892
 893                 if (next_block || flags & INTERNAL_NODE) {
 894                         if (find_highest)
 895                                 block = value64(ro_node(s), i);
 896                         else
 897                                 block = value64(ro_node(s), 0);
 898                 }
 899
 900         } while (flags & INTERNAL_NODE);
 901
 902         if (next_block)
 903                 *next_block = block;
 904         return 0;
 905 }
 906
 907 static int dm_btree_find_key(struct dm_btree_info *info, dm_block_t root,
 908                              bool find_highest, uint64_t *result_keys)
 909 {
 910         int r = 0, count = 0, level;
 911         struct ro_spine spine;
 912
 913         init_ro_spine(&spine, info);
 914         for (level = 0; level < info->levels; level++) {
 915                 r = find_key(&spine, root, find_highest, result_keys + level,
 916                              level == info->levels - 1 ? NULL : &root);
 917                 if (r == -ENODATA) {
 918                         r = 0;
 919                         break;
 920
 921                 } else if (r)
 922                         break;
 923
 924                 count++;
 925         }
 926         exit_ro_spine(&spine);
 927
 928         return r ? r : count;
 929 }
 930
 931 int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
 932                               uint64_t *result_keys)
 933 {
 934         return dm_btree_find_key(info, root, true, result_keys);
 935 }
 936 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
 937
 938 int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
 939                              uint64_t *result_keys)
 940 {
 941         return dm_btree_find_key(info, root, false, result_keys);
 942 }
 943 EXPORT_SYMBOL_GPL(dm_btree_find_lowest_key);
 944
 945 /*----------------------------------------------------------------*/
 946
 947 /*
 948  * FIXME: We shouldn't use a recursive algorithm when we have limited stack
 949  * space.  Also this only works for single level trees.
 950  */
 951 static int walk_node(struct dm_btree_info *info, dm_block_t block,
 952                      int (*fn)(void *context, uint64_t *keys, void *leaf),
 953                      void *context)
 954 {
 955         int r;
 956         unsigned i, nr;
 957         struct dm_block *node;
 958         struct btree_node *n;
 959         uint64_t keys;
 960
 961         r = bn_read_lock(info, block, &node);
 962         if (r)
 963                 return r;
 964
 965         n = dm_block_data(node);
 966
 967         nr = le32_to_cpu(n->header.nr_entries);
 968         for (i = 0; i < nr; i++) {
 969                 if (le32_to_cpu(n->header.flags) & INTERNAL_NODE) {
 970                         r = walk_node(info, value64(n, i), fn, context);
 971                         if (r)
 972                                 goto out;
 973                 } else {
 974                         keys = le64_to_cpu(*key_ptr(n, i));
 975                         r = fn(context, &keys, value_ptr(n, i));
 976                         if (r)
 977                                 goto out;
 978                 }
 979         }
 980
 981 out:
 982         dm_tm_unlock(info->tm, node);
 983         return r;
 984 }
 985
 986 int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
 987                   int (*fn)(void *context, uint64_t *keys, void *leaf),
 988                   void *context)
 989 {
 990         BUG_ON(info->levels > 1);
 991         return walk_node(info, root, fn, context);
 992 }
 993 EXPORT_SYMBOL_GPL(dm_btree_walk);
 994
 995 /*----------------------------------------------------------------*/
 996
 997 static void prefetch_values(struct dm_btree_cursor *c)
 998 {
 999         unsigned i, nr;
1000         __le64 value_le;
1001         struct cursor_node *n = c->nodes + c->depth - 1;
1002         struct btree_node *bn = dm_block_data(n->b);
1003         struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
1004
1005         BUG_ON(c->info->value_type.size != sizeof(value_le));
1006
1007         nr = le32_to_cpu(bn->header.nr_entries);
1008         for (i = 0; i < nr; i++) {
1009                 memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
1010                 dm_bm_prefetch(bm, le64_to_cpu(value_le));
1011         }
1012 }
1013
1014 static bool leaf_node(struct dm_btree_cursor *c)
1015 {
1016         struct cursor_node *n = c->nodes + c->depth - 1;
1017         struct btree_node *bn = dm_block_data(n->b);
1018
1019         return le32_to_cpu(bn->header.flags) & LEAF_NODE;
1020 }
1021
1022 static int push_node(struct dm_btree_cursor *c, dm_block_t b)
1023 {
1024         int r;
1025         struct cursor_node *n = c->nodes + c->depth;
1026
1027         if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
1028                 DMERR("couldn't push cursor node, stack depth too high");
1029                 return -EINVAL;
1030         }
1031
1032         r = bn_read_lock(c->info, b, &n->b);
1033         if (r)
1034                 return r;
1035
1036         n->index = 0;
1037         c->depth++;
1038
1039         if (c->prefetch_leaves || !leaf_node(c))
1040                 prefetch_values(c);
1041
1042         return 0;
1043 }
1044
1045 static void pop_node(struct dm_btree_cursor *c)
1046 {
1047         c->depth--;
1048         unlock_block(c->info, c->nodes[c->depth].b);
1049 }
1050
1051 static int inc_or_backtrack(struct dm_btree_cursor *c)
1052 {
1053         struct cursor_node *n;
1054         struct btree_node *bn;
1055
1056         for (;;) {
1057                 if (!c->depth)
1058                         return -ENODATA;
1059
1060                 n = c->nodes + c->depth - 1;
1061                 bn = dm_block_data(n->b);
1062
1063                 n->index++;
1064                 if (n->index < le32_to_cpu(bn->header.nr_entries))
1065                         break;
1066
1067                 pop_node(c);
1068         }
1069
1070         return 0;
1071 }
1072
1073 static int find_leaf(struct dm_btree_cursor *c)
1074 {
1075         int r = 0;
1076         struct cursor_node *n;
1077         struct btree_node *bn;
1078         __le64 value_le;
1079
1080         for (;;) {
1081                 n = c->nodes + c->depth - 1;
1082                 bn = dm_block_data(n->b);
1083
1084                 if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
1085                         break;
1086
1087                 memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
1088                 r = push_node(c, le64_to_cpu(value_le));
1089                 if (r) {
1090                         DMERR("push_node failed");
1091                         break;
1092                 }
1093         }
1094
1095         if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
1096                 return -ENODATA;
1097
1098         return r;
1099 }
1100
1101 int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
1102                           bool prefetch_leaves, struct dm_btree_cursor *c)
1103 {
1104         int r;
1105
1106         c->info = info;
1107         c->root = root;
1108         c->depth = 0;
1109         c->prefetch_leaves = prefetch_leaves;
1110
1111         r = push_node(c, root);
1112         if (r)
1113                 return r;
1114
1115         return find_leaf(c);
1116 }
1117 EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
1118
1119 void dm_btree_cursor_end(struct dm_btree_cursor *c)
1120 {
1121         while (c->depth)
1122                 pop_node(c);
1123 }
1124 EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
1125
1126 int dm_btree_cursor_next(struct dm_btree_cursor *c)
1127 {
1128         int r = inc_or_backtrack(c);
1129         if (!r) {
1130                 r = find_leaf(c);
1131                 if (r)
1132                         DMERR("find_leaf failed");
1133         }
1134
1135         return r;
1136 }
1137 EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
1138
1139 int dm_btree_cursor_skip(struct dm_btree_cursor *c, uint32_t count)
1140 {
1141         int r = 0;
1142
1143         while (count-- && !r)
1144                 r = dm_btree_cursor_next(c);
1145
1146         return r;
1147 }
1148 EXPORT_SYMBOL_GPL(dm_btree_cursor_skip);
1149
1150 int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
1151 {
1152         if (c->depth) {
1153                 struct cursor_node *n = c->nodes + c->depth - 1;
1154                 struct btree_node *bn = dm_block_data(n->b);
1155
1156                 if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
1157                         return -EINVAL;
1158
1159                 *key = le64_to_cpu(*key_ptr(bn, n->index));
1160                 memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
1161                 return 0;
1162
1163         } else
1164                 return -ENODATA;
1165 }
1166 EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);