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