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         uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
  75
  76         if (le32_to_cpu(n->header.flags) & INTERNAL_NODE)
  77                 dm_tm_with_runs(tm, value_ptr(n, 0), nr_entries, dm_tm_inc_range);
  78
  79         else if (vt->inc)
  80                 vt->inc(vt->context, value_ptr(n, 0), nr_entries);
  81 }
  82
  83 static int insert_at(size_t value_size, struct btree_node *node, unsigned index,
  84                       uint64_t key, void *value)
  85                       __dm_written_to_disk(value)
  86 {
  87         uint32_t nr_entries = le32_to_cpu(node->header.nr_entries);
  88         __le64 key_le = cpu_to_le64(key);
  89
  90         if (index > nr_entries ||
  91             index >= le32_to_cpu(node->header.max_entries)) {
  92                 DMERR("too many entries in btree node for insert");
  93                 __dm_unbless_for_disk(value);
  94                 return -ENOMEM;
  95         }
  96
  97         __dm_bless_for_disk(&key_le);
  98
  99         array_insert(node->keys, sizeof(*node->keys), nr_entries, index, &key_le);
 100         array_insert(value_base(node), value_size, nr_entries, index, value);
 101         node->header.nr_entries = cpu_to_le32(nr_entries + 1);
 102
 103         return 0;
 104 }
 105
 106 /*----------------------------------------------------------------*/
 107
 108 /*
 109  * We want 3n entries (for some n).  This works more nicely for repeated
 110  * insert remove loops than (2n + 1).
 111  */
 112 static uint32_t calc_max_entries(size_t value_size, size_t block_size)
 113 {
 114         uint32_t total, n;
 115         size_t elt_size = sizeof(uint64_t) + value_size; /* key + value */
 116
 117         block_size -= sizeof(struct node_header);
 118         total = block_size / elt_size;
 119         n = total / 3;          /* rounds down */
 120
 121         return 3 * n;
 122 }
 123
 124 int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root)
 125 {
 126         int r;
 127         struct dm_block *b;
 128         struct btree_node *n;
 129         size_t block_size;
 130         uint32_t max_entries;
 131
 132         r = new_block(info, &b);
 133         if (r < 0)
 134                 return r;
 135
 136         block_size = dm_bm_block_size(dm_tm_get_bm(info->tm));
 137         max_entries = calc_max_entries(info->value_type.size, block_size);
 138
 139         n = dm_block_data(b);
 140         memset(n, 0, block_size);
 141         n->header.flags = cpu_to_le32(LEAF_NODE);
 142         n->header.nr_entries = cpu_to_le32(0);
 143         n->header.max_entries = cpu_to_le32(max_entries);
 144         n->header.value_size = cpu_to_le32(info->value_type.size);
 145
 146         *root = dm_block_location(b);
 147         unlock_block(info, b);
 148
 149         return 0;
 150 }
 151 EXPORT_SYMBOL_GPL(dm_btree_empty);
 152
 153 /*----------------------------------------------------------------*/
 154
 155 /*
 156  * Deletion uses a recursive algorithm, since we have limited stack space
 157  * we explicitly manage our own stack on the heap.
 158  */
 159 #define MAX_SPINE_DEPTH 64
 160 struct frame {
 161         struct dm_block *b;
 162         struct btree_node *n;
 163         unsigned level;
 164         unsigned nr_children;
 165         unsigned current_child;
 166 };
 167
 168 struct del_stack {
 169         struct dm_btree_info *info;
 170         struct dm_transaction_manager *tm;
 171         int top;
 172         struct frame spine[MAX_SPINE_DEPTH];
 173 };
 174
 175 static int top_frame(struct del_stack *s, struct frame **f)
 176 {
 177         if (s->top < 0) {
 178                 DMERR("btree deletion stack empty");
 179                 return -EINVAL;
 180         }
 181
 182         *f = s->spine + s->top;
 183
 184         return 0;
 185 }
 186
 187 static int unprocessed_frames(struct del_stack *s)
 188 {
 189         return s->top >= 0;
 190 }
 191
 192 static void prefetch_children(struct del_stack *s, struct frame *f)
 193 {
 194         unsigned i;
 195         struct dm_block_manager *bm = dm_tm_get_bm(s->tm);
 196
 197         for (i = 0; i < f->nr_children; i++)
 198                 dm_bm_prefetch(bm, value64(f->n, i));
 199 }
 200
 201 static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
 202 {
 203         return f->level < (info->levels - 1);
 204 }
 205
 206 static int push_frame(struct del_stack *s, dm_block_t b, unsigned level)
 207 {
 208         int r;
 209         uint32_t ref_count;
 210
 211         if (s->top >= MAX_SPINE_DEPTH - 1) {
 212                 DMERR("btree deletion stack out of memory");
 213                 return -ENOMEM;
 214         }
 215
 216         r = dm_tm_ref(s->tm, b, &ref_count);
 217         if (r)
 218                 return r;
 219
 220         if (ref_count > 1)
 221                 /*
 222                  * This is a shared node, so we can just decrement it's
 223                  * reference counter and leave the children.
 224                  */
 225                 dm_tm_dec(s->tm, b);
 226
 227         else {
 228                 uint32_t flags;
 229                 struct frame *f = s->spine + ++s->top;
 230
 231                 r = dm_tm_read_lock(s->tm, b, &btree_node_validator, &f->b);
 232                 if (r) {
 233                         s->top--;
 234                         return r;
 235                 }
 236
 237                 f->n = dm_block_data(f->b);
 238                 f->level = level;
 239                 f->nr_children = le32_to_cpu(f->n->header.nr_entries);
 240                 f->current_child = 0;
 241
 242                 flags = le32_to_cpu(f->n->header.flags);
 243                 if (flags & INTERNAL_NODE || is_internal_level(s->info, f))
 244                         prefetch_children(s, f);
 245         }
 246
 247         return 0;
 248 }
 249
 250 static void pop_frame(struct del_stack *s)
 251 {
 252         struct frame *f = s->spine + s->top--;
 253
 254         dm_tm_dec(s->tm, dm_block_location(f->b));
 255         dm_tm_unlock(s->tm, f->b);
 256 }
 257
 258 static void unlock_all_frames(struct del_stack *s)
 259 {
 260         struct frame *f;
 261
 262         while (unprocessed_frames(s)) {
 263                 f = s->spine + s->top--;
 264                 dm_tm_unlock(s->tm, f->b);
 265         }
 266 }
 267
 268 int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
 269 {
 270         int r;
 271         struct del_stack *s;
 272
 273         /*
 274          * dm_btree_del() is called via an ioctl, as such should be
 275          * considered an FS op.  We can't recurse back into the FS, so we
 276          * allocate GFP_NOFS.
 277          */
 278         s = kmalloc(sizeof(*s), GFP_NOFS);
 279         if (!s)
 280                 return -ENOMEM;
 281         s->info = info;
 282         s->tm = info->tm;
 283         s->top = -1;
 284
 285         r = push_frame(s, root, 0);
 286         if (r)
 287                 goto out;
 288
 289         while (unprocessed_frames(s)) {
 290                 uint32_t flags;
 291                 struct frame *f;
 292                 dm_block_t b;
 293
 294                 r = top_frame(s, &f);
 295                 if (r)
 296                         goto out;
 297
 298                 if (f->current_child >= f->nr_children) {
 299                         pop_frame(s);
 300                         continue;
 301                 }
 302
 303                 flags = le32_to_cpu(f->n->header.flags);
 304                 if (flags & INTERNAL_NODE) {
 305                         b = value64(f->n, f->current_child);
 306                         f->current_child++;
 307                         r = push_frame(s, b, f->level);
 308                         if (r)
 309                                 goto out;
 310
 311                 } else if (is_internal_level(info, f)) {
 312                         b = value64(f->n, f->current_child);
 313                         f->current_child++;
 314                         r = push_frame(s, b, f->level + 1);
 315                         if (r)
 316                                 goto out;
 317
 318                 } else {
 319                         if (info->value_type.dec)
 320                                 info->value_type.dec(info->value_type.context,
 321                                                      value_ptr(f->n, 0), f->nr_children);
 322                         pop_frame(s);
 323                 }
 324         }
 325 out:
 326         if (r) {
 327                 /* cleanup all frames of del_stack */
 328                 unlock_all_frames(s);
 329         }
 330         kfree(s);
 331
 332         return r;
 333 }
 334 EXPORT_SYMBOL_GPL(dm_btree_del);
 335
 336 /*----------------------------------------------------------------*/
 337
 338 static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
 339                             int (*search_fn)(struct btree_node *, uint64_t),
 340                             uint64_t *result_key, void *v, size_t value_size)
 341 {
 342         int i, r;
 343         uint32_t flags, nr_entries;
 344
 345         do {
 346                 r = ro_step(s, block);
 347                 if (r < 0)
 348                         return r;
 349
 350                 i = search_fn(ro_node(s), key);
 351
 352                 flags = le32_to_cpu(ro_node(s)->header.flags);
 353                 nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
 354                 if (i < 0 || i >= nr_entries)
 355                         return -ENODATA;
 356
 357                 if (flags & INTERNAL_NODE)
 358                         block = value64(ro_node(s), i);
 359
 360         } while (!(flags & LEAF_NODE));
 361
 362         *result_key = le64_to_cpu(ro_node(s)->keys[i]);
 363         if (v)
 364                 memcpy(v, value_ptr(ro_node(s), i), value_size);
 365
 366         return 0;
 367 }
 368
 369 int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
 370                     uint64_t *keys, void *value_le)
 371 {
 372         unsigned level, last_level = info->levels - 1;
 373         int r = -ENODATA;
 374         uint64_t rkey;
 375         __le64 internal_value_le;
 376         struct ro_spine spine;
 377
 378         init_ro_spine(&spine, info);
 379         for (level = 0; level < info->levels; level++) {
 380                 size_t size;
 381                 void *value_p;
 382
 383                 if (level == last_level) {
 384                         value_p = value_le;
 385                         size = info->value_type.size;
 386
 387                 } else {
 388                         value_p = &internal_value_le;
 389                         size = sizeof(uint64_t);
 390                 }
 391
 392                 r = btree_lookup_raw(&spine, root, keys[level],
 393                                      lower_bound, &rkey,
 394                                      value_p, size);
 395
 396                 if (!r) {
 397                         if (rkey != keys[level]) {
 398                                 exit_ro_spine(&spine);
 399                                 return -ENODATA;
 400                         }
 401                 } else {
 402                         exit_ro_spine(&spine);
 403                         return r;
 404                 }
 405
 406                 root = le64_to_cpu(internal_value_le);
 407         }
 408         exit_ro_spine(&spine);
 409
 410         return r;
 411 }
 412 EXPORT_SYMBOL_GPL(dm_btree_lookup);
 413
 414 static int dm_btree_lookup_next_single(struct dm_btree_info *info, dm_block_t root,
 415                                        uint64_t key, uint64_t *rkey, void *value_le)
 416 {
 417         int r, i;
 418         uint32_t flags, nr_entries;
 419         struct dm_block *node;
 420         struct btree_node *n;
 421
 422         r = bn_read_lock(info, root, &node);
 423         if (r)
 424                 return r;
 425
 426         n = dm_block_data(node);
 427         flags = le32_to_cpu(n->header.flags);
 428         nr_entries = le32_to_cpu(n->header.nr_entries);
 429
 430         if (flags & INTERNAL_NODE) {
 431                 i = lower_bound(n, key);
 432                 if (i < 0) {
 433                         /*
 434                          * avoid early -ENODATA return when all entries are
 435                          * higher than the search @key.
 436                          */
 437                         i = 0;
 438                 }
 439                 if (i >= nr_entries) {
 440                         r = -ENODATA;
 441                         goto out;
 442                 }
 443
 444                 r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
 445                 if (r == -ENODATA && i < (nr_entries - 1)) {
 446                         i++;
 447                         r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
 448                 }
 449
 450         } else {
 451                 i = upper_bound(n, key);
 452                 if (i < 0 || i >= nr_entries) {
 453                         r = -ENODATA;
 454                         goto out;
 455                 }
 456
 457                 *rkey = le64_to_cpu(n->keys[i]);
 458                 memcpy(value_le, value_ptr(n, i), info->value_type.size);
 459         }
 460 out:
 461         dm_tm_unlock(info->tm, node);
 462         return r;
 463 }
 464
 465 int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
 466                          uint64_t *keys, uint64_t *rkey, void *value_le)
 467 {
 468         unsigned level;
 469         int r = -ENODATA;
 470         __le64 internal_value_le;
 471         struct ro_spine spine;
 472
 473         init_ro_spine(&spine, info);
 474         for (level = 0; level < info->levels - 1u; level++) {
 475                 r = btree_lookup_raw(&spine, root, keys[level],
 476                                      lower_bound, rkey,
 477                                      &internal_value_le, sizeof(uint64_t));
 478                 if (r)
 479                         goto out;
 480
 481                 if (*rkey != keys[level]) {
 482                         r = -ENODATA;
 483                         goto out;
 484                 }
 485
 486                 root = le64_to_cpu(internal_value_le);
 487         }
 488
 489         r = dm_btree_lookup_next_single(info, root, keys[level], rkey, value_le);
 490 out:
 491         exit_ro_spine(&spine);
 492         return r;
 493 }
 494
 495 EXPORT_SYMBOL_GPL(dm_btree_lookup_next);
 496
 497 /*----------------------------------------------------------------*/
 498
 499 /*
 500  * Copies entries from one region of a btree node to another.  The regions
 501  * must not overlap.
 502  */
 503 static void copy_entries(struct btree_node *dest, unsigned dest_offset,
 504                          struct btree_node *src, unsigned src_offset,
 505                          unsigned count)
 506 {
 507         size_t value_size = le32_to_cpu(dest->header.value_size);
 508         memcpy(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
 509         memcpy(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
 510 }
 511
 512 /*
 513  * Moves entries from one region fo a btree node to another.  The regions
 514  * may overlap.
 515  */
 516 static void move_entries(struct btree_node *dest, unsigned dest_offset,
 517                          struct btree_node *src, unsigned src_offset,
 518                          unsigned count)
 519 {
 520         size_t value_size = le32_to_cpu(dest->header.value_size);
 521         memmove(dest->keys + dest_offset, src->keys + src_offset, count * sizeof(uint64_t));
 522         memmove(value_ptr(dest, dest_offset), value_ptr(src, src_offset), count * value_size);
 523 }
 524
 525 /*
 526  * Erases the first 'count' entries of a btree node, shifting following
 527  * entries down into their place.
 528  */
 529 static void shift_down(struct btree_node *n, unsigned count)
 530 {
 531         move_entries(n, 0, n, count, le32_to_cpu(n->header.nr_entries) - count);
 532 }
 533
 534 /*
 535  * Moves entries in a btree node up 'count' places, making space for
 536  * new entries at the start of the node.
 537  */
 538 static void shift_up(struct btree_node *n, unsigned count)
 539 {
 540         move_entries(n, count, n, 0, le32_to_cpu(n->header.nr_entries));
 541 }
 542
 543 /*
 544  * Redistributes entries between two btree nodes to make them
 545  * have similar numbers of entries.
 546  */
 547 static void redistribute2(struct btree_node *left, struct btree_node *right)
 548 {
 549         unsigned nr_left = le32_to_cpu(left->header.nr_entries);
 550         unsigned nr_right = le32_to_cpu(right->header.nr_entries);
 551         unsigned total = nr_left + nr_right;
 552         unsigned target_left = total / 2;
 553         unsigned target_right = total - target_left;
 554
 555         if (nr_left < target_left) {
 556                 unsigned delta = target_left - nr_left;
 557                 copy_entries(left, nr_left, right, 0, delta);
 558                 shift_down(right, delta);
 559         } else if (nr_left > target_left) {
 560                 unsigned delta = nr_left - target_left;
 561                 if (nr_right)
 562                         shift_up(right, delta);
 563                 copy_entries(right, 0, left, target_left, delta);
 564         }
 565
 566         left->header.nr_entries = cpu_to_le32(target_left);
 567         right->header.nr_entries = cpu_to_le32(target_right);
 568 }
 569
 570 /*
 571  * Redistribute entries between three nodes.  Assumes the central
 572  * node is empty.
 573  */
 574 static void redistribute3(struct btree_node *left, struct btree_node *center,
 575                           struct btree_node *right)
 576 {
 577         unsigned nr_left = le32_to_cpu(left->header.nr_entries);
 578         unsigned nr_center = le32_to_cpu(center->header.nr_entries);
 579         unsigned nr_right = le32_to_cpu(right->header.nr_entries);
 580         unsigned total, target_left, target_center, target_right;
 581
 582         BUG_ON(nr_center);
 583
 584         total = nr_left + nr_right;
 585         target_left = total / 3;
 586         target_center = (total - target_left) / 2;
 587         target_right = (total - target_left - target_center);
 588
 589         if (nr_left < target_left) {
 590                 unsigned left_short = target_left - nr_left;
 591                 copy_entries(left, nr_left, right, 0, left_short);
 592                 copy_entries(center, 0, right, left_short, target_center);
 593                 shift_down(right, nr_right - target_right);
 594
 595         } else if (nr_left < (target_left + target_center)) {
 596                 unsigned left_to_center = nr_left - target_left;
 597                 copy_entries(center, 0, left, target_left, left_to_center);
 598                 copy_entries(center, left_to_center, right, 0, target_center - left_to_center);
 599                 shift_down(right, nr_right - target_right);
 600
 601         } else {
 602                 unsigned right_short = target_right - nr_right;
 603                 shift_up(right, right_short);
 604                 copy_entries(right, 0, left, nr_left - right_short, right_short);
 605                 copy_entries(center, 0, left, target_left, nr_left - target_left);
 606         }
 607
 608         left->header.nr_entries = cpu_to_le32(target_left);
 609         center->header.nr_entries = cpu_to_le32(target_center);
 610         right->header.nr_entries = cpu_to_le32(target_right);
 611 }
 612
 613 /*
 614  * Splits a node by creating a sibling node and shifting half the nodes
 615  * contents across.  Assumes there is a parent node, and it has room for
 616  * another child.
 617  *
 618  * Before:
 619  *        +--------+
 620  *        | Parent |
 621  *        +--------+
 622  *           |
 623  *           v
 624  *      +----------+
 625  *      | A ++++++ |
 626  *      +----------+
 627  *
 628  *
 629  * After:
 630  *              +--------+
 631  *              | Parent |
 632  *              +--------+
 633  *                |     |
 634  *                v     +------+
 635  *          +---------+        |
 636  *          | A* +++  |        v
 637  *          +---------+   +-------+
 638  *                        | B +++ |
 639  *                        +-------+
 640  *
 641  * Where A* is a shadow of A.
 642  */
 643 static int split_one_into_two(struct shadow_spine *s, unsigned parent_index,
 644                               struct dm_btree_value_type *vt, uint64_t key)
 645 {
 646         int r;
 647         struct dm_block *left, *right, *parent;
 648         struct btree_node *ln, *rn, *pn;
 649         __le64 location;
 650
 651         left = shadow_current(s);
 652
 653         r = new_block(s->info, &right);
 654         if (r < 0)
 655                 return r;
 656
 657         ln = dm_block_data(left);
 658         rn = dm_block_data(right);
 659
 660         rn->header.flags = ln->header.flags;
 661         rn->header.nr_entries = cpu_to_le32(0);
 662         rn->header.max_entries = ln->header.max_entries;
 663         rn->header.value_size = ln->header.value_size;
 664         redistribute2(ln, rn);
 665
 666         /* patch up the parent */
 667         parent = shadow_parent(s);
 668         pn = dm_block_data(parent);
 669
 670         location = cpu_to_le64(dm_block_location(right));
 671         __dm_bless_for_disk(&location);
 672         r = insert_at(sizeof(__le64), pn, parent_index + 1,
 673                       le64_to_cpu(rn->keys[0]), &location);
 674         if (r) {
 675                 unlock_block(s->info, right);
 676                 return r;
 677         }
 678
 679         /* patch up the spine */
 680         if (key < le64_to_cpu(rn->keys[0])) {
 681                 unlock_block(s->info, right);
 682                 s->nodes[1] = left;
 683         } else {
 684                 unlock_block(s->info, left);
 685                 s->nodes[1] = right;
 686         }
 687
 688         return 0;
 689 }
 690
 691 /*
 692  * We often need to modify a sibling node.  This function shadows a particular
 693  * child of the given parent node.  Making sure to update the parent to point
 694  * to the new shadow.
 695  */
 696 static int shadow_child(struct dm_btree_info *info, struct dm_btree_value_type *vt,
 697                         struct btree_node *parent, unsigned index,
 698                         struct dm_block **result)
 699 {
 700         int r, inc;
 701         dm_block_t root;
 702         struct btree_node *node;
 703
 704         root = value64(parent, index);
 705
 706         r = dm_tm_shadow_block(info->tm, root, &btree_node_validator,
 707                                result, &inc);
 708         if (r)
 709                 return r;
 710
 711         node = dm_block_data(*result);
 712
 713         if (inc)
 714                 inc_children(info->tm, node, vt);
 715
 716         *((__le64 *) value_ptr(parent, index)) =
 717                 cpu_to_le64(dm_block_location(*result));
 718
 719         return 0;
 720 }
 721
 722 /*
 723  * Splits two nodes into three.  This is more work, but results in fuller
 724  * nodes, so saves metadata space.
 725  */
 726 static int split_two_into_three(struct shadow_spine *s, unsigned parent_index,
 727                                 struct dm_btree_value_type *vt, uint64_t key)
 728 {
 729         int r;
 730         unsigned middle_index;
 731         struct dm_block *left, *middle, *right, *parent;
 732         struct btree_node *ln, *rn, *mn, *pn;
 733         __le64 location;
 734
 735         parent = shadow_parent(s);
 736         pn = dm_block_data(parent);
 737
 738         if (parent_index == 0) {
 739                 middle_index = 1;
 740                 left = shadow_current(s);
 741                 r = shadow_child(s->info, vt, pn, parent_index + 1, &right);
 742                 if (r)
 743                         return r;
 744         } else {
 745                 middle_index = parent_index;
 746                 right = shadow_current(s);
 747                 r = shadow_child(s->info, vt, pn, parent_index - 1, &left);
 748                 if (r)
 749                         return r;
 750         }
 751
 752         r = new_block(s->info, &middle);
 753         if (r < 0)
 754                 return r;
 755
 756         ln = dm_block_data(left);
 757         mn = dm_block_data(middle);
 758         rn = dm_block_data(right);
 759
 760         mn->header.nr_entries = cpu_to_le32(0);
 761         mn->header.flags = ln->header.flags;
 762         mn->header.max_entries = ln->header.max_entries;
 763         mn->header.value_size = ln->header.value_size;
 764
 765         redistribute3(ln, mn, rn);
 766
 767         /* patch up the parent */
 768         pn->keys[middle_index] = rn->keys[0];
 769         location = cpu_to_le64(dm_block_location(middle));
 770         __dm_bless_for_disk(&location);
 771         r = insert_at(sizeof(__le64), pn, middle_index,
 772                       le64_to_cpu(mn->keys[0]), &location);
 773         if (r) {
 774                 if (shadow_current(s) != left)
 775                         unlock_block(s->info, left);
 776
 777                 unlock_block(s->info, middle);
 778
 779                 if (shadow_current(s) != right)
 780                         unlock_block(s->info, right);
 781
 782                 return r;
 783         }
 784
 785
 786         /* patch up the spine */
 787         if (key < le64_to_cpu(mn->keys[0])) {
 788                 unlock_block(s->info, middle);
 789                 unlock_block(s->info, right);
 790                 s->nodes[1] = left;
 791         } else if (key < le64_to_cpu(rn->keys[0])) {
 792                 unlock_block(s->info, left);
 793                 unlock_block(s->info, right);
 794                 s->nodes[1] = middle;
 795         } else {
 796                 unlock_block(s->info, left);
 797                 unlock_block(s->info, middle);
 798                 s->nodes[1] = right;
 799         }
 800
 801         return 0;
 802 }
 803
 804 /*----------------------------------------------------------------*/
 805
 806 /*
 807  * Splits a node by creating two new children beneath the given node.
 808  *
 809  * Before:
 810  *        +----------+
 811  *        | A ++++++ |
 812  *        +----------+
 813  *
 814  *
 815  * After:
 816  *      +------------+
 817  *      | A (shadow) |
 818  *      +------------+
 819  *          |   |
 820  *   +------+   +----+
 821  *   |               |
 822  *   v               v
 823  * +-------+     +-------+
 824  * | B +++ |     | C +++ |
 825  * +-------+     +-------+
 826  */
 827 static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
 828 {
 829         int r;
 830         size_t size;
 831         unsigned nr_left, nr_right;
 832         struct dm_block *left, *right, *new_parent;
 833         struct btree_node *pn, *ln, *rn;
 834         __le64 val;
 835
 836         new_parent = shadow_current(s);
 837
 838         pn = dm_block_data(new_parent);
 839         size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
 840                 sizeof(__le64) : s->info->value_type.size;
 841
 842         /* create & init the left block */
 843         r = new_block(s->info, &left);
 844         if (r < 0)
 845                 return r;
 846
 847         ln = dm_block_data(left);
 848         nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
 849
 850         ln->header.flags = pn->header.flags;
 851         ln->header.nr_entries = cpu_to_le32(nr_left);
 852         ln->header.max_entries = pn->header.max_entries;
 853         ln->header.value_size = pn->header.value_size;
 854         memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
 855         memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
 856
 857         /* create & init the right block */
 858         r = new_block(s->info, &right);
 859         if (r < 0) {
 860                 unlock_block(s->info, left);
 861                 return r;
 862         }
 863
 864         rn = dm_block_data(right);
 865         nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
 866
 867         rn->header.flags = pn->header.flags;
 868         rn->header.nr_entries = cpu_to_le32(nr_right);
 869         rn->header.max_entries = pn->header.max_entries;
 870         rn->header.value_size = pn->header.value_size;
 871         memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
 872         memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
 873                nr_right * size);
 874
 875         /* new_parent should just point to l and r now */
 876         pn->header.flags = cpu_to_le32(INTERNAL_NODE);
 877         pn->header.nr_entries = cpu_to_le32(2);
 878         pn->header.max_entries = cpu_to_le32(
 879                 calc_max_entries(sizeof(__le64),
 880                                  dm_bm_block_size(
 881                                          dm_tm_get_bm(s->info->tm))));
 882         pn->header.value_size = cpu_to_le32(sizeof(__le64));
 883
 884         val = cpu_to_le64(dm_block_location(left));
 885         __dm_bless_for_disk(&val);
 886         pn->keys[0] = ln->keys[0];
 887         memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
 888
 889         val = cpu_to_le64(dm_block_location(right));
 890         __dm_bless_for_disk(&val);
 891         pn->keys[1] = rn->keys[0];
 892         memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
 893
 894         unlock_block(s->info, left);
 895         unlock_block(s->info, right);
 896         return 0;
 897 }
 898
 899 /*----------------------------------------------------------------*/
 900
 901 /*
 902  * Redistributes a node's entries with its left sibling.
 903  */
 904 static int rebalance_left(struct shadow_spine *s, struct dm_btree_value_type *vt,
 905                           unsigned parent_index, uint64_t key)
 906 {
 907         int r;
 908         struct dm_block *sib;
 909         struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
 910
 911         r = shadow_child(s->info, vt, parent, parent_index - 1, &sib);
 912         if (r)
 913                 return r;
 914
 915         left = dm_block_data(sib);
 916         right = dm_block_data(shadow_current(s));
 917         redistribute2(left, right);
 918         *key_ptr(parent, parent_index) = right->keys[0];
 919
 920         if (key < le64_to_cpu(right->keys[0])) {
 921                 unlock_block(s->info, s->nodes[1]);
 922                 s->nodes[1] = sib;
 923         } else {
 924                 unlock_block(s->info, sib);
 925         }
 926
 927         return 0;
 928 }
 929
 930 /*
 931  * Redistributes a nodes entries with its right sibling.
 932  */
 933 static int rebalance_right(struct shadow_spine *s, struct dm_btree_value_type *vt,
 934                            unsigned parent_index, uint64_t key)
 935 {
 936         int r;
 937         struct dm_block *sib;
 938         struct btree_node *left, *right, *parent = dm_block_data(shadow_parent(s));
 939
 940         r = shadow_child(s->info, vt, parent, parent_index + 1, &sib);
 941         if (r)
 942                 return r;
 943
 944         left = dm_block_data(shadow_current(s));
 945         right = dm_block_data(sib);
 946         redistribute2(left, right);
 947         *key_ptr(parent, parent_index + 1) = right->keys[0];
 948
 949         if (key < le64_to_cpu(right->keys[0])) {
 950                 unlock_block(s->info, sib);
 951         } else {
 952                 unlock_block(s->info, s->nodes[1]);
 953                 s->nodes[1] = sib;
 954         }
 955
 956         return 0;
 957 }
 958
 959 /*
 960  * Returns the number of spare entries in a node.
 961  */
 962 static int get_node_free_space(struct dm_btree_info *info, dm_block_t b, unsigned *space)
 963 {
 964         int r;
 965         unsigned nr_entries;
 966         struct dm_block *block;
 967         struct btree_node *node;
 968
 969         r = bn_read_lock(info, b, &block);
 970         if (r)
 971                 return r;
 972
 973         node = dm_block_data(block);
 974         nr_entries = le32_to_cpu(node->header.nr_entries);
 975         *space = le32_to_cpu(node->header.max_entries) - nr_entries;
 976
 977         unlock_block(info, block);
 978         return 0;
 979 }
 980
 981 /*
 982  * Make space in a node, either by moving some entries to a sibling,
 983  * or creating a new sibling node.  SPACE_THRESHOLD defines the minimum
 984  * number of free entries that must be in the sibling to make the move
 985  * worth while.  If the siblings are shared (eg, part of a snapshot),
 986  * then they are not touched, since this break sharing and so consume
 987  * more space than we save.
 988  */
 989 #define SPACE_THRESHOLD 8
 990 static int rebalance_or_split(struct shadow_spine *s, struct dm_btree_value_type *vt,
 991                               unsigned parent_index, uint64_t key)
 992 {
 993         int r;
 994         struct btree_node *parent = dm_block_data(shadow_parent(s));
 995         unsigned nr_parent = le32_to_cpu(parent->header.nr_entries);
 996         unsigned free_space;
 997         int left_shared = 0, right_shared = 0;
 998
 999         /* Should we move entries to the left sibling? */
1000         if (parent_index > 0) {
1001                 dm_block_t left_b = value64(parent, parent_index - 1);
1002                 r = dm_tm_block_is_shared(s->info->tm, left_b, &left_shared);
1003                 if (r)
1004                         return r;
1005
1006                 if (!left_shared) {
1007                         r = get_node_free_space(s->info, left_b, &free_space);
1008                         if (r)
1009                                 return r;
1010
1011                         if (free_space >= SPACE_THRESHOLD)
1012                                 return rebalance_left(s, vt, parent_index, key);
1013                 }
1014         }
1015
1016         /* Should we move entries to the right sibling? */
1017         if (parent_index < (nr_parent - 1)) {
1018                 dm_block_t right_b = value64(parent, parent_index + 1);
1019                 r = dm_tm_block_is_shared(s->info->tm, right_b, &right_shared);
1020                 if (r)
1021                         return r;
1022
1023                 if (!right_shared) {
1024                         r = get_node_free_space(s->info, right_b, &free_space);
1025                         if (r)
1026                                 return r;
1027
1028                         if (free_space >= SPACE_THRESHOLD)
1029                                 return rebalance_right(s, vt, parent_index, key);
1030                 }
1031         }
1032
1033         /*
1034          * We need to split the node, normally we split two nodes
1035          * into three.  But when inserting a sequence that is either
1036          * monotonically increasing or decreasing it's better to split
1037          * a single node into two.
1038          */
1039         if (left_shared || right_shared || (nr_parent <= 2) ||
1040             (parent_index == 0) || (parent_index + 1 == nr_parent)) {
1041                 return split_one_into_two(s, parent_index, vt, key);
1042         } else {
1043                 return split_two_into_three(s, parent_index, vt, key);
1044         }
1045 }
1046
1047 /*
1048  * Does the node contain a particular key?
1049  */
1050 static bool contains_key(struct btree_node *node, uint64_t key)
1051 {
1052         int i = lower_bound(node, key);
1053
1054         if (i >= 0 && le64_to_cpu(node->keys[i]) == key)
1055                 return true;
1056
1057         return false;
1058 }
1059
1060 /*
1061  * In general we preemptively make sure there's a free entry in every
1062  * node on the spine when doing an insert.  But we can avoid that with
1063  * leaf nodes if we know it's an overwrite.
1064  */
1065 static bool has_space_for_insert(struct btree_node *node, uint64_t key)
1066 {
1067         if (node->header.nr_entries == node->header.max_entries) {
1068                 if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1069                         /* we don't need space if it's an overwrite */
1070                         return contains_key(node, key);
1071                 }
1072
1073                 return false;
1074         }
1075
1076         return true;
1077 }
1078
1079 static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
1080                             struct dm_btree_value_type *vt,
1081                             uint64_t key, unsigned *index)
1082 {
1083         int r, i = *index, top = 1;
1084         struct btree_node *node;
1085
1086         for (;;) {
1087                 r = shadow_step(s, root, vt);
1088                 if (r < 0)
1089                         return r;
1090
1091                 node = dm_block_data(shadow_current(s));
1092
1093                 /*
1094                  * We have to patch up the parent node, ugly, but I don't
1095                  * see a way to do this automatically as part of the spine
1096                  * op.
1097                  */
1098                 if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
1099                         __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1100
1101                         __dm_bless_for_disk(&location);
1102                         memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1103                                     &location, sizeof(__le64));
1104                 }
1105
1106                 node = dm_block_data(shadow_current(s));
1107
1108                 if (!has_space_for_insert(node, key)) {
1109                         if (top)
1110                                 r = btree_split_beneath(s, key);
1111                         else
1112                                 r = rebalance_or_split(s, vt, i, key);
1113
1114                         if (r < 0)
1115                                 return r;
1116
1117                         /* making space can cause the current node to change */
1118                         node = dm_block_data(shadow_current(s));
1119                 }
1120
1121                 i = lower_bound(node, key);
1122
1123                 if (le32_to_cpu(node->header.flags) & LEAF_NODE)
1124                         break;
1125
1126                 if (i < 0) {
1127                         /* change the bounds on the lowest key */
1128                         node->keys[0] = cpu_to_le64(key);
1129                         i = 0;
1130                 }
1131
1132                 root = value64(node, i);
1133                 top = 0;
1134         }
1135
1136         if (i < 0 || le64_to_cpu(node->keys[i]) != key)
1137                 i++;
1138
1139         *index = i;
1140         return 0;
1141 }
1142
1143 static int __btree_get_overwrite_leaf(struct shadow_spine *s, dm_block_t root,
1144                                       uint64_t key, int *index)
1145 {
1146         int r, i = -1;
1147         struct btree_node *node;
1148
1149         *index = 0;
1150         for (;;) {
1151                 r = shadow_step(s, root, &s->info->value_type);
1152                 if (r < 0)
1153                         return r;
1154
1155                 node = dm_block_data(shadow_current(s));
1156
1157                 /*
1158                  * We have to patch up the parent node, ugly, but I don't
1159                  * see a way to do this automatically as part of the spine
1160                  * op.
1161                  */
1162                 if (shadow_has_parent(s) && i >= 0) {
1163                         __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
1164
1165                         __dm_bless_for_disk(&location);
1166                         memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
1167                                     &location, sizeof(__le64));
1168                 }
1169
1170                 node = dm_block_data(shadow_current(s));
1171                 i = lower_bound(node, key);
1172
1173                 BUG_ON(i < 0);
1174                 BUG_ON(i >= le32_to_cpu(node->header.nr_entries));
1175
1176                 if (le32_to_cpu(node->header.flags) & LEAF_NODE) {
1177                         if (key != le64_to_cpu(node->keys[i]))
1178                                 return -EINVAL;
1179                         break;
1180                 }
1181
1182                 root = value64(node, i);
1183         }
1184
1185         *index = i;
1186         return 0;
1187 }
1188
1189 int btree_get_overwrite_leaf(struct dm_btree_info *info, dm_block_t root,
1190                              uint64_t key, int *index,
1191                              dm_block_t *new_root, struct dm_block **leaf)
1192 {
1193         int r;
1194         struct shadow_spine spine;
1195
1196         BUG_ON(info->levels > 1);
1197         init_shadow_spine(&spine, info);
1198         r = __btree_get_overwrite_leaf(&spine, root, key, index);
1199         if (!r) {
1200                 *new_root = shadow_root(&spine);
1201                 *leaf = shadow_current(&spine);
1202
1203                 /*
1204                  * Decrement the count so exit_shadow_spine() doesn't
1205                  * unlock the leaf.
1206                  */
1207                 spine.count--;
1208         }
1209         exit_shadow_spine(&spine);
1210
1211         return r;
1212 }
1213
1214 static bool need_insert(struct btree_node *node, uint64_t *keys,
1215                         unsigned level, unsigned index)
1216 {
1217         return ((index >= le32_to_cpu(node->header.nr_entries)) ||
1218                 (le64_to_cpu(node->keys[index]) != keys[level]));
1219 }
1220
1221 static int insert(struct dm_btree_info *info, dm_block_t root,
1222                   uint64_t *keys, void *value, dm_block_t *new_root,
1223                   int *inserted)
1224                   __dm_written_to_disk(value)
1225 {
1226         int r;
1227         unsigned level, index = -1, last_level = info->levels - 1;
1228         dm_block_t block = root;
1229         struct shadow_spine spine;
1230         struct btree_node *n;
1231         struct dm_btree_value_type le64_type;
1232
1233         init_le64_type(info->tm, &le64_type);
1234         init_shadow_spine(&spine, info);
1235
1236         for (level = 0; level < (info->levels - 1); level++) {
1237                 r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
1238                 if (r < 0)
1239                         goto bad;
1240
1241                 n = dm_block_data(shadow_current(&spine));
1242
1243                 if (need_insert(n, keys, level, index)) {
1244                         dm_block_t new_tree;
1245                         __le64 new_le;
1246
1247                         r = dm_btree_empty(info, &new_tree);
1248                         if (r < 0)
1249                                 goto bad;
1250
1251                         new_le = cpu_to_le64(new_tree);
1252                         __dm_bless_for_disk(&new_le);
1253
1254                         r = insert_at(sizeof(uint64_t), n, index,
1255                                       keys[level], &new_le);
1256                         if (r)
1257                                 goto bad;
1258                 }
1259
1260                 if (level < last_level)
1261                         block = value64(n, index);
1262         }
1263
1264         r = btree_insert_raw(&spine, block, &info->value_type,
1265                              keys[level], &index);
1266         if (r < 0)
1267                 goto bad;
1268
1269         n = dm_block_data(shadow_current(&spine));
1270
1271         if (need_insert(n, keys, level, index)) {
1272                 if (inserted)
1273                         *inserted = 1;
1274
1275                 r = insert_at(info->value_type.size, n, index,
1276                               keys[level], value);
1277                 if (r)
1278                         goto bad_unblessed;
1279         } else {
1280                 if (inserted)
1281                         *inserted = 0;
1282
1283                 if (info->value_type.dec &&
1284                     (!info->value_type.equal ||
1285                      !info->value_type.equal(
1286                              info->value_type.context,
1287                              value_ptr(n, index),
1288                              value))) {
1289                         info->value_type.dec(info->value_type.context,
1290                                              value_ptr(n, index), 1);
1291                 }
1292                 memcpy_disk(value_ptr(n, index),
1293                             value, info->value_type.size);
1294         }
1295
1296         *new_root = shadow_root(&spine);
1297         exit_shadow_spine(&spine);
1298
1299         return 0;
1300
1301 bad:
1302         __dm_unbless_for_disk(value);
1303 bad_unblessed:
1304         exit_shadow_spine(&spine);
1305         return r;
1306 }
1307
1308 int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
1309                     uint64_t *keys, void *value, dm_block_t *new_root)
1310                     __dm_written_to_disk(value)
1311 {
1312         return insert(info, root, keys, value, new_root, NULL);
1313 }
1314 EXPORT_SYMBOL_GPL(dm_btree_insert);
1315
1316 int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
1317                            uint64_t *keys, void *value, dm_block_t *new_root,
1318                            int *inserted)
1319                            __dm_written_to_disk(value)
1320 {
1321         return insert(info, root, keys, value, new_root, inserted);
1322 }
1323 EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
1324
1325 /*----------------------------------------------------------------*/
1326
1327 static int find_key(struct ro_spine *s, dm_block_t block, bool find_highest,
1328                     uint64_t *result_key, dm_block_t *next_block)
1329 {
1330         int i, r;
1331         uint32_t flags;
1332
1333         do {
1334                 r = ro_step(s, block);
1335                 if (r < 0)
1336                         return r;
1337
1338                 flags = le32_to_cpu(ro_node(s)->header.flags);
1339                 i = le32_to_cpu(ro_node(s)->header.nr_entries);
1340                 if (!i)
1341                         return -ENODATA;
1342                 else
1343                         i--;
1344
1345                 if (find_highest)
1346                         *result_key = le64_to_cpu(ro_node(s)->keys[i]);
1347                 else
1348                         *result_key = le64_to_cpu(ro_node(s)->keys[0]);
1349
1350                 if (next_block || flags & INTERNAL_NODE) {
1351                         if (find_highest)
1352                                 block = value64(ro_node(s), i);
1353                         else
1354                                 block = value64(ro_node(s), 0);
1355                 }
1356
1357         } while (flags & INTERNAL_NODE);
1358
1359         if (next_block)
1360                 *next_block = block;
1361         return 0;
1362 }
1363
1364 static int dm_btree_find_key(struct dm_btree_info *info, dm_block_t root,
1365                              bool find_highest, uint64_t *result_keys)
1366 {
1367         int r = 0, count = 0, level;
1368         struct ro_spine spine;
1369
1370         init_ro_spine(&spine, info);
1371         for (level = 0; level < info->levels; level++) {
1372                 r = find_key(&spine, root, find_highest, result_keys + level,
1373                              level == info->levels - 1 ? NULL : &root);
1374                 if (r == -ENODATA) {
1375                         r = 0;
1376                         break;
1377
1378                 } else if (r)
1379                         break;
1380
1381                 count++;
1382         }
1383         exit_ro_spine(&spine);
1384
1385         return r ? r : count;
1386 }
1387
1388 int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
1389                               uint64_t *result_keys)
1390 {
1391         return dm_btree_find_key(info, root, true, result_keys);
1392 }
1393 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
1394
1395 int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
1396                              uint64_t *result_keys)
1397 {
1398         return dm_btree_find_key(info, root, false, result_keys);
1399 }
1400 EXPORT_SYMBOL_GPL(dm_btree_find_lowest_key);
1401
1402 /*----------------------------------------------------------------*/
1403
1404 /*
1405  * FIXME: We shouldn't use a recursive algorithm when we have limited stack
1406  * space.  Also this only works for single level trees.
1407  */
1408 static int walk_node(struct dm_btree_info *info, dm_block_t block,
1409                      int (*fn)(void *context, uint64_t *keys, void *leaf),
1410                      void *context)
1411 {
1412         int r;
1413         unsigned i, nr;
1414         struct dm_block *node;
1415         struct btree_node *n;
1416         uint64_t keys;
1417
1418         r = bn_read_lock(info, block, &node);
1419         if (r)
1420                 return r;
1421
1422         n = dm_block_data(node);
1423
1424         nr = le32_to_cpu(n->header.nr_entries);
1425         for (i = 0; i < nr; i++) {
1426                 if (le32_to_cpu(n->header.flags) & INTERNAL_NODE) {
1427                         r = walk_node(info, value64(n, i), fn, context);
1428                         if (r)
1429                                 goto out;
1430                 } else {
1431                         keys = le64_to_cpu(*key_ptr(n, i));
1432                         r = fn(context, &keys, value_ptr(n, i));
1433                         if (r)
1434                                 goto out;
1435                 }
1436         }
1437
1438 out:
1439         dm_tm_unlock(info->tm, node);
1440         return r;
1441 }
1442
1443 int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
1444                   int (*fn)(void *context, uint64_t *keys, void *leaf),
1445                   void *context)
1446 {
1447         BUG_ON(info->levels > 1);
1448         return walk_node(info, root, fn, context);
1449 }
1450 EXPORT_SYMBOL_GPL(dm_btree_walk);
1451
1452 /*----------------------------------------------------------------*/
1453
1454 static void prefetch_values(struct dm_btree_cursor *c)
1455 {
1456         unsigned i, nr;
1457         __le64 value_le;
1458         struct cursor_node *n = c->nodes + c->depth - 1;
1459         struct btree_node *bn = dm_block_data(n->b);
1460         struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
1461
1462         BUG_ON(c->info->value_type.size != sizeof(value_le));
1463
1464         nr = le32_to_cpu(bn->header.nr_entries);
1465         for (i = 0; i < nr; i++) {
1466                 memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
1467                 dm_bm_prefetch(bm, le64_to_cpu(value_le));
1468         }
1469 }
1470
1471 static bool leaf_node(struct dm_btree_cursor *c)
1472 {
1473         struct cursor_node *n = c->nodes + c->depth - 1;
1474         struct btree_node *bn = dm_block_data(n->b);
1475
1476         return le32_to_cpu(bn->header.flags) & LEAF_NODE;
1477 }
1478
1479 static int push_node(struct dm_btree_cursor *c, dm_block_t b)
1480 {
1481         int r;
1482         struct cursor_node *n = c->nodes + c->depth;
1483
1484         if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
1485                 DMERR("couldn't push cursor node, stack depth too high");
1486                 return -EINVAL;
1487         }
1488
1489         r = bn_read_lock(c->info, b, &n->b);
1490         if (r)
1491                 return r;
1492
1493         n->index = 0;
1494         c->depth++;
1495
1496         if (c->prefetch_leaves || !leaf_node(c))
1497                 prefetch_values(c);
1498
1499         return 0;
1500 }
1501
1502 static void pop_node(struct dm_btree_cursor *c)
1503 {
1504         c->depth--;
1505         unlock_block(c->info, c->nodes[c->depth].b);
1506 }
1507
1508 static int inc_or_backtrack(struct dm_btree_cursor *c)
1509 {
1510         struct cursor_node *n;
1511         struct btree_node *bn;
1512
1513         for (;;) {
1514                 if (!c->depth)
1515                         return -ENODATA;
1516
1517                 n = c->nodes + c->depth - 1;
1518                 bn = dm_block_data(n->b);
1519
1520                 n->index++;
1521                 if (n->index < le32_to_cpu(bn->header.nr_entries))
1522                         break;
1523
1524                 pop_node(c);
1525         }
1526
1527         return 0;
1528 }
1529
1530 static int find_leaf(struct dm_btree_cursor *c)
1531 {
1532         int r = 0;
1533         struct cursor_node *n;
1534         struct btree_node *bn;
1535         __le64 value_le;
1536
1537         for (;;) {
1538                 n = c->nodes + c->depth - 1;
1539                 bn = dm_block_data(n->b);
1540
1541                 if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
1542                         break;
1543
1544                 memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
1545                 r = push_node(c, le64_to_cpu(value_le));
1546                 if (r) {
1547                         DMERR("push_node failed");
1548                         break;
1549                 }
1550         }
1551
1552         if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
1553                 return -ENODATA;
1554
1555         return r;
1556 }
1557
1558 int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
1559                           bool prefetch_leaves, struct dm_btree_cursor *c)
1560 {
1561         int r;
1562
1563         c->info = info;
1564         c->root = root;
1565         c->depth = 0;
1566         c->prefetch_leaves = prefetch_leaves;
1567
1568         r = push_node(c, root);
1569         if (r)
1570                 return r;
1571
1572         return find_leaf(c);
1573 }
1574 EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
1575
1576 void dm_btree_cursor_end(struct dm_btree_cursor *c)
1577 {
1578         while (c->depth)
1579                 pop_node(c);
1580 }
1581 EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
1582
1583 int dm_btree_cursor_next(struct dm_btree_cursor *c)
1584 {
1585         int r = inc_or_backtrack(c);
1586         if (!r) {
1587                 r = find_leaf(c);
1588                 if (r)
1589                         DMERR("find_leaf failed");
1590         }
1591
1592         return r;
1593 }
1594 EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
1595
1596 int dm_btree_cursor_skip(struct dm_btree_cursor *c, uint32_t count)
1597 {
1598         int r = 0;
1599
1600         while (count-- && !r)
1601                 r = dm_btree_cursor_next(c);
1602
1603         return r;
1604 }
1605 EXPORT_SYMBOL_GPL(dm_btree_cursor_skip);
1606
1607 int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
1608 {
1609         if (c->depth) {
1610                 struct cursor_node *n = c->nodes + c->depth - 1;
1611                 struct btree_node *bn = dm_block_data(n->b);
1612
1613                 if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
1614                         return -EINVAL;
1615
1616                 *key = le64_to_cpu(*key_ptr(bn, n->index));
1617                 memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
1618                 return 0;
1619
1620         } else
1621                 return -ENODATA;
1622 }
1623 EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);