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

   1 /*
   2  * Copyright (C) 2011 Red Hat, Inc.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-btree.h"
   8 #include "dm-btree-internal.h"
   9 #include "dm-transaction-manager.h"
  10
  11 #include <linux/export.h>
  12
  13 /*
  14  * Removing an entry from a btree
  15  * ==============================
  16  *
  17  * A very important constraint for our btree is that no node, except the
  18  * root, may have fewer than a certain number of entries.
  19  * (MIN_ENTRIES <= nr_entries <= MAX_ENTRIES).
  20  *
  21  * Ensuring this is complicated by the way we want to only ever hold the
  22  * locks on 2 nodes concurrently, and only change nodes in a top to bottom
  23  * fashion.
  24  *
  25  * Each node may have a left or right sibling.  When decending the spine,
  26  * if a node contains only MIN_ENTRIES then we try and increase this to at
  27  * least MIN_ENTRIES + 1.  We do this in the following ways:
  28  *
  29  * [A] No siblings => this can only happen if the node is the root, in which
  30  *     case we copy the childs contents over the root.
  31  *
  32  * [B] No left sibling
  33  *     ==> rebalance(node, right sibling)
  34  *
  35  * [C] No right sibling
  36  *     ==> rebalance(left sibling, node)
  37  *
  38  * [D] Both siblings, total_entries(left, node, right) <= DEL_THRESHOLD
  39  *     ==> delete node adding it's contents to left and right
  40  *
  41  * [E] Both siblings, total_entries(left, node, right) > DEL_THRESHOLD
  42  *     ==> rebalance(left, node, right)
  43  *
  44  * After these operations it's possible that the our original node no
  45  * longer contains the desired sub tree.  For this reason this rebalancing
  46  * is performed on the children of the current node.  This also avoids
  47  * having a special case for the root.
  48  *
  49  * Once this rebalancing has occurred we can then step into the child node
  50  * for internal nodes.  Or delete the entry for leaf nodes.
  51  */
  52
  53 /*
  54  * Some little utilities for moving node data around.
  55  */
  56 static void node_shift(struct btree_node *n, int shift)
  57 {
  58         uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
  59         uint32_t value_size = le32_to_cpu(n->header.value_size);
  60
  61         if (shift < 0) {
  62                 shift = -shift;
  63                 BUG_ON(shift > nr_entries);
  64                 BUG_ON((void *) key_ptr(n, shift) >= value_ptr(n, shift));
  65                 memmove(key_ptr(n, 0),
  66                         key_ptr(n, shift),
  67                         (nr_entries - shift) * sizeof(__le64));
  68                 memmove(value_ptr(n, 0),
  69                         value_ptr(n, shift),
  70                         (nr_entries - shift) * value_size);
  71         } else {
  72                 BUG_ON(nr_entries + shift > le32_to_cpu(n->header.max_entries));
  73                 memmove(key_ptr(n, shift),
  74                         key_ptr(n, 0),
  75                         nr_entries * sizeof(__le64));
  76                 memmove(value_ptr(n, shift),
  77                         value_ptr(n, 0),
  78                         nr_entries * value_size);
  79         }
  80 }
  81
  82 static void node_copy(struct btree_node *left, struct btree_node *right, int shift)
  83 {
  84         uint32_t nr_left = le32_to_cpu(left->header.nr_entries);
  85         uint32_t value_size = le32_to_cpu(left->header.value_size);
  86         BUG_ON(value_size != le32_to_cpu(right->header.value_size));
  87
  88         if (shift < 0) {
  89                 shift = -shift;
  90                 BUG_ON(nr_left + shift > le32_to_cpu(left->header.max_entries));
  91                 memcpy(key_ptr(left, nr_left),
  92                        key_ptr(right, 0),
  93                        shift * sizeof(__le64));
  94                 memcpy(value_ptr(left, nr_left),
  95                        value_ptr(right, 0),
  96                        shift * value_size);
  97         } else {
  98                 BUG_ON(shift > le32_to_cpu(right->header.max_entries));
  99                 memcpy(key_ptr(right, 0),
 100                        key_ptr(left, nr_left - shift),
 101                        shift * sizeof(__le64));
 102                 memcpy(value_ptr(right, 0),
 103                        value_ptr(left, nr_left - shift),
 104                        shift * value_size);
 105         }
 106 }
 107
 108 /*
 109  * Delete a specific entry from a leaf node.
 110  */
 111 static void delete_at(struct btree_node *n, unsigned index)
 112 {
 113         unsigned nr_entries = le32_to_cpu(n->header.nr_entries);
 114         unsigned nr_to_copy = nr_entries - (index + 1);
 115         uint32_t value_size = le32_to_cpu(n->header.value_size);
 116         BUG_ON(index >= nr_entries);
 117
 118         if (nr_to_copy) {
 119                 memmove(key_ptr(n, index),
 120                         key_ptr(n, index + 1),
 121                         nr_to_copy * sizeof(__le64));
 122
 123                 memmove(value_ptr(n, index),
 124                         value_ptr(n, index + 1),
 125                         nr_to_copy * value_size);
 126         }
 127
 128         n->header.nr_entries = cpu_to_le32(nr_entries - 1);
 129 }
 130
 131 static unsigned merge_threshold(struct btree_node *n)
 132 {
 133         return le32_to_cpu(n->header.max_entries) / 3;
 134 }
 135
 136 struct child {
 137         unsigned index;
 138         struct dm_block *block;
 139         struct btree_node *n;
 140 };
 141
 142 static int init_child(struct dm_btree_info *info, struct dm_btree_value_type *vt,
 143                       struct btree_node *parent,
 144                       unsigned index, struct child *result)
 145 {
 146         int r, inc;
 147         dm_block_t root;
 148
 149         result->index = index;
 150         root = value64(parent, index);
 151
 152         r = dm_tm_shadow_block(info->tm, root, &btree_node_validator,
 153                                &result->block, &inc);
 154         if (r)
 155                 return r;
 156
 157         result->n = dm_block_data(result->block);
 158
 159         if (inc)
 160                 inc_children(info->tm, result->n, vt);
 161
 162         *((__le64 *) value_ptr(parent, index)) =
 163                 cpu_to_le64(dm_block_location(result->block));
 164
 165         return 0;
 166 }
 167
 168 static int exit_child(struct dm_btree_info *info, struct child *c)
 169 {
 170         return dm_tm_unlock(info->tm, c->block);
 171 }
 172
 173 static void shift(struct btree_node *left, struct btree_node *right, int count)
 174 {
 175         uint32_t nr_left = le32_to_cpu(left->header.nr_entries);
 176         uint32_t nr_right = le32_to_cpu(right->header.nr_entries);
 177         uint32_t max_entries = le32_to_cpu(left->header.max_entries);
 178         uint32_t r_max_entries = le32_to_cpu(right->header.max_entries);
 179
 180         BUG_ON(max_entries != r_max_entries);
 181         BUG_ON(nr_left - count > max_entries);
 182         BUG_ON(nr_right + count > max_entries);
 183
 184         if (!count)
 185                 return;
 186
 187         if (count > 0) {
 188                 node_shift(right, count);
 189                 node_copy(left, right, count);
 190         } else {
 191                 node_copy(left, right, count);
 192                 node_shift(right, count);
 193         }
 194
 195         left->header.nr_entries = cpu_to_le32(nr_left - count);
 196         right->header.nr_entries = cpu_to_le32(nr_right + count);
 197 }
 198
 199 static void __rebalance2(struct dm_btree_info *info, struct btree_node *parent,
 200                          struct child *l, struct child *r)
 201 {
 202         struct btree_node *left = l->n;
 203         struct btree_node *right = r->n;
 204         uint32_t nr_left = le32_to_cpu(left->header.nr_entries);
 205         uint32_t nr_right = le32_to_cpu(right->header.nr_entries);
 206         unsigned threshold = 2 * merge_threshold(left) + 1;
 207
 208         if (nr_left + nr_right < threshold) {
 209                 /*
 210                  * Merge
 211                  */
 212                 node_copy(left, right, -nr_right);
 213                 left->header.nr_entries = cpu_to_le32(nr_left + nr_right);
 214                 delete_at(parent, r->index);
 215
 216                 /*
 217                  * We need to decrement the right block, but not it's
 218                  * children, since they're still referenced by left.
 219                  */
 220                 dm_tm_dec(info->tm, dm_block_location(r->block));
 221         } else {
 222                 /*
 223                  * Rebalance.
 224                  */
 225                 unsigned target_left = (nr_left + nr_right) / 2;
 226                 shift(left, right, nr_left - target_left);
 227                 *key_ptr(parent, r->index) = right->keys[0];
 228         }
 229 }
 230
 231 static int rebalance2(struct shadow_spine *s, struct dm_btree_info *info,
 232                       struct dm_btree_value_type *vt, unsigned left_index)
 233 {
 234         int r;
 235         struct btree_node *parent;
 236         struct child left, right;
 237
 238         parent = dm_block_data(shadow_current(s));
 239
 240         r = init_child(info, vt, parent, left_index, &left);
 241         if (r)
 242                 return r;
 243
 244         r = init_child(info, vt, parent, left_index + 1, &right);
 245         if (r) {
 246                 exit_child(info, &left);
 247                 return r;
 248         }
 249
 250         __rebalance2(info, parent, &left, &right);
 251
 252         r = exit_child(info, &left);
 253         if (r) {
 254                 exit_child(info, &right);
 255                 return r;
 256         }
 257
 258         return exit_child(info, &right);
 259 }
 260
 261 /*
 262  * We dump as many entries from center as possible into left, then the rest
 263  * in right, then rebalance2.  This wastes some cpu, but I want something
 264  * simple atm.
 265  */
 266 static void delete_center_node(struct dm_btree_info *info, struct btree_node *parent,
 267                                struct child *l, struct child *c, struct child *r,
 268                                struct btree_node *left, struct btree_node *center, struct btree_node *right,
 269                                uint32_t nr_left, uint32_t nr_center, uint32_t nr_right)
 270 {
 271         uint32_t max_entries = le32_to_cpu(left->header.max_entries);
 272         unsigned shift = min(max_entries - nr_left, nr_center);
 273
 274         BUG_ON(nr_left + shift > max_entries);
 275         node_copy(left, center, -shift);
 276         left->header.nr_entries = cpu_to_le32(nr_left + shift);
 277
 278         if (shift != nr_center) {
 279                 shift = nr_center - shift;
 280                 BUG_ON((nr_right + shift) > max_entries);
 281                 node_shift(right, shift);
 282                 node_copy(center, right, shift);
 283                 right->header.nr_entries = cpu_to_le32(nr_right + shift);
 284         }
 285         *key_ptr(parent, r->index) = right->keys[0];
 286
 287         delete_at(parent, c->index);
 288         r->index--;
 289
 290         dm_tm_dec(info->tm, dm_block_location(c->block));
 291         __rebalance2(info, parent, l, r);
 292 }
 293
 294 /*
 295  * Redistributes entries among 3 sibling nodes.
 296  */
 297 static void redistribute3(struct dm_btree_info *info, struct btree_node *parent,
 298                           struct child *l, struct child *c, struct child *r,
 299                           struct btree_node *left, struct btree_node *center, struct btree_node *right,
 300                           uint32_t nr_left, uint32_t nr_center, uint32_t nr_right)
 301 {
 302         int s;
 303         uint32_t max_entries = le32_to_cpu(left->header.max_entries);
 304         unsigned target = (nr_left + nr_center + nr_right) / 3;
 305         BUG_ON(target > max_entries);
 306
 307         if (nr_left < nr_right) {
 308                 s = nr_left - target;
 309
 310                 if (s < 0 && nr_center < -s) {
 311                         /* not enough in central node */
 312                         shift(left, center, -nr_center);
 313                         s += nr_center;
 314                         shift(left, right, s);
 315                         nr_right += s;
 316                 } else
 317                         shift(left, center, s);
 318
 319                 shift(center, right, target - nr_right);
 320
 321         } else {
 322                 s = target - nr_right;
 323                 if (s > 0 && nr_center < s) {
 324                         /* not enough in central node */
 325                         shift(center, right, nr_center);
 326                         s -= nr_center;
 327                         shift(left, right, s);
 328                         nr_left -= s;
 329                 } else
 330                         shift(center, right, s);
 331
 332                 shift(left, center, nr_left - target);
 333         }
 334
 335         *key_ptr(parent, c->index) = center->keys[0];
 336         *key_ptr(parent, r->index) = right->keys[0];
 337 }
 338
 339 static void __rebalance3(struct dm_btree_info *info, struct btree_node *parent,
 340                          struct child *l, struct child *c, struct child *r)
 341 {
 342         struct btree_node *left = l->n;
 343         struct btree_node *center = c->n;
 344         struct btree_node *right = r->n;
 345
 346         uint32_t nr_left = le32_to_cpu(left->header.nr_entries);
 347         uint32_t nr_center = le32_to_cpu(center->header.nr_entries);
 348         uint32_t nr_right = le32_to_cpu(right->header.nr_entries);
 349
 350         unsigned threshold = merge_threshold(left) * 4 + 1;
 351
 352         BUG_ON(left->header.max_entries != center->header.max_entries);
 353         BUG_ON(center->header.max_entries != right->header.max_entries);
 354
 355         if ((nr_left + nr_center + nr_right) < threshold)
 356                 delete_center_node(info, parent, l, c, r, left, center, right,
 357                                    nr_left, nr_center, nr_right);
 358         else
 359                 redistribute3(info, parent, l, c, r, left, center, right,
 360                               nr_left, nr_center, nr_right);
 361 }
 362
 363 static int rebalance3(struct shadow_spine *s, struct dm_btree_info *info,
 364                       struct dm_btree_value_type *vt, unsigned left_index)
 365 {
 366         int r;
 367         struct btree_node *parent = dm_block_data(shadow_current(s));
 368         struct child left, center, right;
 369
 370         /*
 371          * FIXME: fill out an array?
 372          */
 373         r = init_child(info, vt, parent, left_index, &left);
 374         if (r)
 375                 return r;
 376
 377         r = init_child(info, vt, parent, left_index + 1, &center);
 378         if (r) {
 379                 exit_child(info, &left);
 380                 return r;
 381         }
 382
 383         r = init_child(info, vt, parent, left_index + 2, &right);
 384         if (r) {
 385                 exit_child(info, &left);
 386                 exit_child(info, &center);
 387                 return r;
 388         }
 389
 390         __rebalance3(info, parent, &left, &center, &right);
 391
 392         r = exit_child(info, &left);
 393         if (r) {
 394                 exit_child(info, &center);
 395                 exit_child(info, &right);
 396                 return r;
 397         }
 398
 399         r = exit_child(info, &center);
 400         if (r) {
 401                 exit_child(info, &right);
 402                 return r;
 403         }
 404
 405         r = exit_child(info, &right);
 406         if (r)
 407                 return r;
 408
 409         return 0;
 410 }
 411
 412 static int rebalance_children(struct shadow_spine *s,
 413                               struct dm_btree_info *info,
 414                               struct dm_btree_value_type *vt, uint64_t key)
 415 {
 416         int i, r, has_left_sibling, has_right_sibling;
 417         struct btree_node *n;
 418
 419         n = dm_block_data(shadow_current(s));
 420
 421         if (le32_to_cpu(n->header.nr_entries) == 1) {
 422                 struct dm_block *child;
 423                 dm_block_t b = value64(n, 0);
 424
 425                 r = dm_tm_read_lock(info->tm, b, &btree_node_validator, &child);
 426                 if (r)
 427                         return r;
 428
 429                 memcpy(n, dm_block_data(child),
 430                        dm_bm_block_size(dm_tm_get_bm(info->tm)));
 431                 r = dm_tm_unlock(info->tm, child);
 432                 if (r)
 433                         return r;
 434
 435                 dm_tm_dec(info->tm, dm_block_location(child));
 436                 return 0;
 437         }
 438
 439         i = lower_bound(n, key);
 440         if (i < 0)
 441                 return -ENODATA;
 442
 443         has_left_sibling = i > 0;
 444         has_right_sibling = i < (le32_to_cpu(n->header.nr_entries) - 1);
 445
 446         if (!has_left_sibling)
 447                 r = rebalance2(s, info, vt, i);
 448
 449         else if (!has_right_sibling)
 450                 r = rebalance2(s, info, vt, i - 1);
 451
 452         else
 453                 r = rebalance3(s, info, vt, i - 1);
 454
 455         return r;
 456 }
 457
 458 static int do_leaf(struct btree_node *n, uint64_t key, unsigned *index)
 459 {
 460         int i = lower_bound(n, key);
 461
 462         if ((i < 0) ||
 463             (i >= le32_to_cpu(n->header.nr_entries)) ||
 464             (le64_to_cpu(n->keys[i]) != key))
 465                 return -ENODATA;
 466
 467         *index = i;
 468
 469         return 0;
 470 }
 471
 472 /*
 473  * Prepares for removal from one level of the hierarchy.  The caller must
 474  * call delete_at() to remove the entry at index.
 475  */
 476 static int remove_raw(struct shadow_spine *s, struct dm_btree_info *info,
 477                       struct dm_btree_value_type *vt, dm_block_t root,
 478                       uint64_t key, unsigned *index)
 479 {
 480         int i = *index, r;
 481         struct btree_node *n;
 482
 483         for (;;) {
 484                 r = shadow_step(s, root, vt);
 485                 if (r < 0)
 486                         break;
 487
 488                 /*
 489                  * We have to patch up the parent node, ugly, but I don't
 490                  * see a way to do this automatically as part of the spine
 491                  * op.
 492                  */
 493                 if (shadow_has_parent(s)) {
 494                         __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
 495                         memcpy(value_ptr(dm_block_data(shadow_parent(s)), i),
 496                                &location, sizeof(__le64));
 497                 }
 498
 499                 n = dm_block_data(shadow_current(s));
 500
 501                 if (le32_to_cpu(n->header.flags) & LEAF_NODE)
 502                         return do_leaf(n, key, index);
 503
 504                 r = rebalance_children(s, info, vt, key);
 505                 if (r)
 506                         break;
 507
 508                 n = dm_block_data(shadow_current(s));
 509                 if (le32_to_cpu(n->header.flags) & LEAF_NODE)
 510                         return do_leaf(n, key, index);
 511
 512                 i = lower_bound(n, key);
 513
 514                 /*
 515                  * We know the key is present, or else
 516                  * rebalance_children would have returned
 517                  * -ENODATA
 518                  */
 519                 root = value64(n, i);
 520         }
 521
 522         return r;
 523 }
 524
 525 int dm_btree_remove(struct dm_btree_info *info, dm_block_t root,
 526                     uint64_t *keys, dm_block_t *new_root)
 527 {
 528         unsigned level, last_level = info->levels - 1;
 529         int index = 0, r = 0;
 530         struct shadow_spine spine;
 531         struct btree_node *n;
 532         struct dm_btree_value_type le64_vt;
 533
 534         init_le64_type(info->tm, &le64_vt);
 535         init_shadow_spine(&spine, info);
 536         for (level = 0; level < info->levels; level++) {
 537                 r = remove_raw(&spine, info,
 538                                (level == last_level ?
 539                                 &info->value_type : &le64_vt),
 540                                root, keys[level], (unsigned *)&index);
 541                 if (r < 0)
 542                         break;
 543
 544                 n = dm_block_data(shadow_current(&spine));
 545                 if (level != last_level) {
 546                         root = value64(n, index);
 547                         continue;
 548                 }
 549
 550                 BUG_ON(index < 0 || index >= le32_to_cpu(n->header.nr_entries));
 551
 552                 if (info->value_type.dec)
 553                         info->value_type.dec(info->value_type.context,
 554                                              value_ptr(n, index));
 555
 556                 delete_at(n, index);
 557         }
 558
 559         *new_root = shadow_root(&spine);
 560         exit_shadow_spine(&spine);
 561
 562         return r;
 563 }
 564 EXPORT_SYMBOL_GPL(dm_btree_remove);
 565
 566 /*----------------------------------------------------------------*/
 567
 568 static int remove_nearest(struct shadow_spine *s, struct dm_btree_info *info,
 569                           struct dm_btree_value_type *vt, dm_block_t root,
 570                           uint64_t key, int *index)
 571 {
 572         int i = *index, r;
 573         struct btree_node *n;
 574
 575         for (;;) {
 576                 r = shadow_step(s, root, vt);
 577                 if (r < 0)
 578                         break;
 579
 580                 /*
 581                  * We have to patch up the parent node, ugly, but I don't
 582                  * see a way to do this automatically as part of the spine
 583                  * op.
 584                  */
 585                 if (shadow_has_parent(s)) {
 586                         __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
 587                         memcpy(value_ptr(dm_block_data(shadow_parent(s)), i),
 588                                &location, sizeof(__le64));
 589                 }
 590
 591                 n = dm_block_data(shadow_current(s));
 592
 593                 if (le32_to_cpu(n->header.flags) & LEAF_NODE) {
 594                         *index = lower_bound(n, key);
 595                         return 0;
 596                 }
 597
 598                 r = rebalance_children(s, info, vt, key);
 599                 if (r)
 600                         break;
 601
 602                 n = dm_block_data(shadow_current(s));
 603                 if (le32_to_cpu(n->header.flags) & LEAF_NODE) {
 604                         *index = lower_bound(n, key);
 605                         return 0;
 606                 }
 607
 608                 i = lower_bound(n, key);
 609
 610                 /*
 611                  * We know the key is present, or else
 612                  * rebalance_children would have returned
 613                  * -ENODATA
 614                  */
 615                 root = value64(n, i);
 616         }
 617
 618         return r;
 619 }
 620
 621 static int remove_one(struct dm_btree_info *info, dm_block_t root,
 622                       uint64_t *keys, uint64_t end_key,
 623                       dm_block_t *new_root, unsigned *nr_removed)
 624 {
 625         unsigned level, last_level = info->levels - 1;
 626         int index = 0, r = 0;
 627         struct shadow_spine spine;
 628         struct btree_node *n;
 629         struct dm_btree_value_type le64_vt;
 630         uint64_t k;
 631
 632         init_le64_type(info->tm, &le64_vt);
 633         init_shadow_spine(&spine, info);
 634         for (level = 0; level < last_level; level++) {
 635                 r = remove_raw(&spine, info, &le64_vt,
 636                                root, keys[level], (unsigned *) &index);
 637                 if (r < 0)
 638                         goto out;
 639
 640                 n = dm_block_data(shadow_current(&spine));
 641                 root = value64(n, index);
 642         }
 643
 644         r = remove_nearest(&spine, info, &info->value_type,
 645                            root, keys[last_level], &index);
 646         if (r < 0)
 647                 goto out;
 648
 649         n = dm_block_data(shadow_current(&spine));
 650
 651         if (index < 0)
 652                 index = 0;
 653
 654         if (index >= le32_to_cpu(n->header.nr_entries)) {
 655                 r = -ENODATA;
 656                 goto out;
 657         }
 658
 659         k = le64_to_cpu(n->keys[index]);
 660         if (k >= keys[last_level] && k < end_key) {
 661                 if (info->value_type.dec)
 662                         info->value_type.dec(info->value_type.context,
 663                                              value_ptr(n, index));
 664
 665                 delete_at(n, index);
 666                 keys[last_level] = k + 1ull;
 667
 668         } else
 669                 r = -ENODATA;
 670
 671 out:
 672         *new_root = shadow_root(&spine);
 673         exit_shadow_spine(&spine);
 674
 675         return r;
 676 }
 677
 678 int dm_btree_remove_leaves(struct dm_btree_info *info, dm_block_t root,
 679                            uint64_t *first_key, uint64_t end_key,
 680                            dm_block_t *new_root, unsigned *nr_removed)
 681 {
 682         int r;
 683
 684         *nr_removed = 0;
 685         do {
 686                 r = remove_one(info, root, first_key, end_key, &root, nr_removed);
 687                 if (!r)
 688                         (*nr_removed)++;
 689         } while (!r);
 690
 691         *new_root = root;
 692         return r == -ENODATA ? 0 : r;
 693 }
 694 EXPORT_SYMBOL_GPL(dm_btree_remove_leaves);