git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/bkey_sort.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "bkey_on_stack.h"
   4 #include "bkey_sort.h"
   5 #include "bset.h"
   6 #include "extents.h"
   7
   8 /* too many iterators, need to clean this up */
   9
  10 /* btree_node_iter_large: */
  11
  12 #define btree_node_iter_cmp_heap(h, _l, _r) btree_node_iter_cmp(b, _l, _r)
  13
  14 static inline bool
  15 bch2_btree_node_iter_large_end(struct btree_node_iter_large *iter)
  16 {
  17         return !iter->used;
  18 }
  19
  20 static inline struct bkey_packed *
  21 bch2_btree_node_iter_large_peek_all(struct btree_node_iter_large *iter,
  22                                     struct btree *b)
  23 {
  24         return bch2_btree_node_iter_large_end(iter)
  25                 ? NULL
  26                 : __btree_node_offset_to_key(b, iter->data->k);
  27 }
  28
  29 static void
  30 bch2_btree_node_iter_large_advance(struct btree_node_iter_large *iter,
  31                                    struct btree *b)
  32 {
  33         iter->data->k += __btree_node_offset_to_key(b, iter->data->k)->u64s;
  34
  35         EBUG_ON(!iter->used);
  36         EBUG_ON(iter->data->k > iter->data->end);
  37
  38         if (iter->data->k == iter->data->end)
  39                 heap_del(iter, 0, btree_node_iter_cmp_heap, NULL);
  40         else
  41                 heap_sift_down(iter, 0, btree_node_iter_cmp_heap, NULL);
  42 }
  43
  44 static inline struct bkey_packed *
  45 bch2_btree_node_iter_large_next_all(struct btree_node_iter_large *iter,
  46                                     struct btree *b)
  47 {
  48         struct bkey_packed *ret = bch2_btree_node_iter_large_peek_all(iter, b);
  49
  50         if (ret)
  51                 bch2_btree_node_iter_large_advance(iter, b);
  52
  53         return ret;
  54 }
  55
  56 void bch2_btree_node_iter_large_push(struct btree_node_iter_large *iter,
  57                                      struct btree *b,
  58                                      const struct bkey_packed *k,
  59                                      const struct bkey_packed *end)
  60 {
  61         if (k != end) {
  62                 struct btree_node_iter_set n =
  63                         ((struct btree_node_iter_set) {
  64                                  __btree_node_key_to_offset(b, k),
  65                                  __btree_node_key_to_offset(b, end)
  66                          });
  67
  68                 __heap_add(iter, n, btree_node_iter_cmp_heap, NULL);
  69         }
  70 }
  71
  72 static void sort_key_next(struct btree_node_iter_large *iter,
  73                           struct btree *b,
  74                           struct btree_node_iter_set *i)
  75 {
  76         i->k += __btree_node_offset_to_key(b, i->k)->u64s;
  77
  78         while (i->k != i->end &&
  79                !__btree_node_offset_to_key(b, i->k)->u64s)
  80                 i->k++;
  81
  82         if (i->k == i->end)
  83                 *i = iter->data[--iter->used];
  84 }
  85
  86 /* regular sort_iters */
  87
  88 typedef int (*sort_cmp_fn)(struct btree *,
  89                            struct bkey_packed *,
  90                            struct bkey_packed *);
  91
  92 static inline void __sort_iter_sift(struct sort_iter *iter,
  93                                     unsigned from,
  94                                     sort_cmp_fn cmp)
  95 {
  96         unsigned i;
  97
  98         for (i = from;
  99              i + 1 < iter->used &&
 100              cmp(iter->b, iter->data[i].k, iter->data[i + 1].k) > 0;
 101              i++)
 102                 swap(iter->data[i], iter->data[i + 1]);
 103 }
 104
 105 static inline void sort_iter_sift(struct sort_iter *iter, sort_cmp_fn cmp)
 106 {
 107
 108         __sort_iter_sift(iter, 0, cmp);
 109 }
 110
 111 static inline void sort_iter_sort(struct sort_iter *iter, sort_cmp_fn cmp)
 112 {
 113         unsigned i = iter->used;
 114
 115         while (i--)
 116                 __sort_iter_sift(iter, i, cmp);
 117 }
 118
 119 static inline struct bkey_packed *sort_iter_peek(struct sort_iter *iter)
 120 {
 121         return iter->used ? iter->data->k : NULL;
 122 }
 123
 124 static inline void sort_iter_advance(struct sort_iter *iter, sort_cmp_fn cmp)
 125 {
 126         iter->data->k = bkey_next_skip_noops(iter->data->k, iter->data->end);
 127
 128         BUG_ON(iter->data->k > iter->data->end);
 129
 130         if (iter->data->k == iter->data->end)
 131                 array_remove_item(iter->data, iter->used, 0);
 132         else
 133                 sort_iter_sift(iter, cmp);
 134 }
 135
 136 static inline struct bkey_packed *sort_iter_next(struct sort_iter *iter,
 137                                                  sort_cmp_fn cmp)
 138 {
 139         struct bkey_packed *ret = sort_iter_peek(iter);
 140
 141         if (ret)
 142                 sort_iter_advance(iter, cmp);
 143
 144         return ret;
 145 }
 146
 147 /*
 148  * Returns true if l > r - unless l == r, in which case returns true if l is
 149  * older than r.
 150  *
 151  * Necessary for btree_sort_fixup() - if there are multiple keys that compare
 152  * equal in different sets, we have to process them newest to oldest.
 153  */
 154 #define key_sort_cmp(h, l, r)                                           \
 155 ({                                                                      \
 156         bkey_cmp_packed(b,                                              \
 157                         __btree_node_offset_to_key(b, (l).k),           \
 158                         __btree_node_offset_to_key(b, (r).k))           \
 159                                                                         \
 160         ?: (l).k - (r).k;                                               \
 161 })
 162
 163 static inline bool should_drop_next_key(struct btree_node_iter_large *iter,
 164                                         struct btree *b)
 165 {
 166         struct btree_node_iter_set *l = iter->data, *r = iter->data + 1;
 167         struct bkey_packed *k = __btree_node_offset_to_key(b, l->k);
 168
 169         if (bkey_whiteout(k))
 170                 return true;
 171
 172         if (iter->used < 2)
 173                 return false;
 174
 175         if (iter->used > 2 &&
 176             key_sort_cmp(iter, r[0], r[1]) >= 0)
 177                 r++;
 178
 179         /*
 180          * key_sort_cmp() ensures that when keys compare equal the older key
 181          * comes first; so if l->k compares equal to r->k then l->k is older and
 182          * should be dropped.
 183          */
 184         return !bkey_cmp_packed(b,
 185                                 __btree_node_offset_to_key(b, l->k),
 186                                 __btree_node_offset_to_key(b, r->k));
 187 }
 188
 189 struct btree_nr_keys bch2_key_sort_fix_overlapping(struct bset *dst,
 190                                         struct btree *b,
 191                                         struct btree_node_iter_large *iter)
 192 {
 193         struct bkey_packed *out = dst->start;
 194         struct btree_nr_keys nr;
 195
 196         memset(&nr, 0, sizeof(nr));
 197
 198         heap_resort(iter, key_sort_cmp, NULL);
 199
 200         while (!bch2_btree_node_iter_large_end(iter)) {
 201                 if (!should_drop_next_key(iter, b)) {
 202                         struct bkey_packed *k =
 203                                 __btree_node_offset_to_key(b, iter->data->k);
 204
 205                         bkey_copy(out, k);
 206                         btree_keys_account_key_add(&nr, 0, out);
 207                         out = bkey_next(out);
 208                 }
 209
 210                 sort_key_next(iter, b, iter->data);
 211                 heap_sift_down(iter, 0, key_sort_cmp, NULL);
 212         }
 213
 214         dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
 215         return nr;
 216 }
 217
 218 /*
 219  * If keys compare equal, compare by pointer order:
 220  *
 221  * Necessary for sort_fix_overlapping() - if there are multiple keys that
 222  * compare equal in different sets, we have to process them newest to oldest.
 223  */
 224 #define extent_sort_cmp(h, l, r)                                        \
 225 ({                                                                      \
 226         struct bkey _ul = bkey_unpack_key(b,                            \
 227                                 __btree_node_offset_to_key(b, (l).k));  \
 228         struct bkey _ur = bkey_unpack_key(b,                            \
 229                                 __btree_node_offset_to_key(b, (r).k));  \
 230                                                                         \
 231         bkey_cmp(bkey_start_pos(&_ul),                                  \
 232                  bkey_start_pos(&_ur)) ?: (r).k - (l).k;                \
 233 })
 234
 235 static inline void extent_sort_sift(struct btree_node_iter_large *iter,
 236                                     struct btree *b, size_t i)
 237 {
 238         heap_sift_down(iter, i, extent_sort_cmp, NULL);
 239 }
 240
 241 static inline void extent_sort_next(struct btree_node_iter_large *iter,
 242                                     struct btree *b,
 243                                     struct btree_node_iter_set *i)
 244 {
 245         sort_key_next(iter, b, i);
 246         heap_sift_down(iter, i - iter->data, extent_sort_cmp, NULL);
 247 }
 248
 249 static void extent_sort_advance_prev(struct bkey_format *f,
 250                                      struct btree_nr_keys *nr,
 251                                      struct bkey_packed *start,
 252                                      struct bkey_packed **prev)
 253 {
 254         if (*prev) {
 255                 bch2_bkey_pack(*prev, (void *) *prev, f);
 256
 257                 btree_keys_account_key_add(nr, 0, *prev);
 258                 *prev = bkey_next(*prev);
 259         } else {
 260                 *prev = start;
 261         }
 262 }
 263
 264 static void extent_sort_append(struct bch_fs *c,
 265                                struct bkey_format *f,
 266                                struct btree_nr_keys *nr,
 267                                struct bkey_packed *start,
 268                                struct bkey_packed **prev,
 269                                struct bkey_s k)
 270 {
 271         if (bkey_whiteout(k.k))
 272                 return;
 273
 274         /*
 275          * prev is always unpacked, for key merging - until right before we
 276          * advance it:
 277          */
 278
 279         if (*prev &&
 280             bch2_bkey_merge(c, bkey_i_to_s((void *) *prev), k) ==
 281             BCH_MERGE_MERGE)
 282                 return;
 283
 284         extent_sort_advance_prev(f, nr, start, prev);
 285
 286         bkey_reassemble((void *) *prev, k.s_c);
 287 }
 288
 289 struct btree_nr_keys bch2_extent_sort_fix_overlapping(struct bch_fs *c,
 290                                         struct bset *dst,
 291                                         struct btree *b,
 292                                         struct btree_node_iter_large *iter)
 293 {
 294         struct bkey_format *f = &b->format;
 295         struct btree_node_iter_set *_l = iter->data, *_r;
 296         struct bkey_packed *prev = NULL, *lk, *rk;
 297         struct bkey l_unpacked, r_unpacked;
 298         struct bkey_s l, r;
 299         struct btree_nr_keys nr;
 300         struct bkey_on_stack split;
 301
 302         memset(&nr, 0, sizeof(nr));
 303         bkey_on_stack_init(&split);
 304
 305         heap_resort(iter, extent_sort_cmp, NULL);
 306
 307         while (!bch2_btree_node_iter_large_end(iter)) {
 308                 lk = __btree_node_offset_to_key(b, _l->k);
 309                 l = __bkey_disassemble(b, lk, &l_unpacked);
 310
 311                 if (iter->used == 1) {
 312                         extent_sort_append(c, f, &nr, dst->start, &prev, l);
 313                         extent_sort_next(iter, b, _l);
 314                         continue;
 315                 }
 316
 317                 _r = iter->data + 1;
 318                 if (iter->used > 2 &&
 319                     extent_sort_cmp(iter, _r[0], _r[1]) >= 0)
 320                         _r++;
 321
 322                 rk = __btree_node_offset_to_key(b, _r->k);
 323                 r = __bkey_disassemble(b, rk, &r_unpacked);
 324
 325                 /* If current key and next key don't overlap, just append */
 326                 if (bkey_cmp(l.k->p, bkey_start_pos(r.k)) <= 0) {
 327                         extent_sort_append(c, f, &nr, dst->start, &prev, l);
 328                         extent_sort_next(iter, b, _l);
 329                         continue;
 330                 }
 331
 332                 /* Skip 0 size keys */
 333                 if (!r.k->size) {
 334                         extent_sort_next(iter, b, _r);
 335                         continue;
 336                 }
 337
 338                 /*
 339                  * overlap: keep the newer key and trim the older key so they
 340                  * don't overlap. comparing pointers tells us which one is
 341                  * newer, since the bsets are appended one after the other.
 342                  */
 343
 344                 /* can't happen because of comparison func */
 345                 BUG_ON(_l->k < _r->k &&
 346                        !bkey_cmp(bkey_start_pos(l.k), bkey_start_pos(r.k)));
 347
 348                 if (_l->k > _r->k) {
 349                         /* l wins, trim r */
 350                         if (bkey_cmp(l.k->p, r.k->p) >= 0) {
 351                                 sort_key_next(iter, b, _r);
 352                         } else {
 353                                 bch2_cut_front_s(l.k->p, r);
 354                                 extent_save(b, rk, r.k);
 355                         }
 356
 357                         extent_sort_sift(iter, b, _r - iter->data);
 358                 } else if (bkey_cmp(l.k->p, r.k->p) > 0) {
 359                         bkey_on_stack_realloc(&split, c, l.k->u64s);
 360
 361                         /*
 362                          * r wins, but it overlaps in the middle of l - split l:
 363                          */
 364                         bkey_reassemble(split.k, l.s_c);
 365                         bch2_cut_back(bkey_start_pos(r.k), split.k);
 366
 367                         bch2_cut_front_s(r.k->p, l);
 368                         extent_save(b, lk, l.k);
 369
 370                         extent_sort_sift(iter, b, 0);
 371
 372                         extent_sort_append(c, f, &nr, dst->start,
 373                                            &prev, bkey_i_to_s(split.k));
 374                 } else {
 375                         bch2_cut_back_s(bkey_start_pos(r.k), l);
 376                         extent_save(b, lk, l.k);
 377                 }
 378         }
 379
 380         extent_sort_advance_prev(f, &nr, dst->start, &prev);
 381
 382         dst->u64s = cpu_to_le16((u64 *) prev - dst->_data);
 383
 384         bkey_on_stack_exit(&split, c);
 385         return nr;
 386 }
 387
 388 /* Sort + repack in a new format: */
 389 struct btree_nr_keys
 390 bch2_sort_repack(struct bset *dst, struct btree *src,
 391                  struct btree_node_iter *src_iter,
 392                  struct bkey_format *out_f,
 393                  bool filter_whiteouts)
 394 {
 395         struct bkey_format *in_f = &src->format;
 396         struct bkey_packed *in, *out = vstruct_last(dst);
 397         struct btree_nr_keys nr;
 398
 399         memset(&nr, 0, sizeof(nr));
 400
 401         while ((in = bch2_btree_node_iter_next_all(src_iter, src))) {
 402                 if (filter_whiteouts && bkey_whiteout(in))
 403                         continue;
 404
 405                 if (bch2_bkey_transform(out_f, out, bkey_packed(in)
 406                                        ? in_f : &bch2_bkey_format_current, in))
 407                         out->format = KEY_FORMAT_LOCAL_BTREE;
 408                 else
 409                         bch2_bkey_unpack(src, (void *) out, in);
 410
 411                 btree_keys_account_key_add(&nr, 0, out);
 412                 out = bkey_next(out);
 413         }
 414
 415         dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
 416         return nr;
 417 }
 418
 419 /* Sort, repack, and merge: */
 420 struct btree_nr_keys
 421 bch2_sort_repack_merge(struct bch_fs *c,
 422                        struct bset *dst, struct btree *src,
 423                        struct btree_node_iter *iter,
 424                        struct bkey_format *out_f,
 425                        bool filter_whiteouts)
 426 {
 427         struct bkey_packed *prev = NULL, *k_packed;
 428         struct bkey_s k;
 429         struct btree_nr_keys nr;
 430         struct bkey unpacked;
 431
 432         memset(&nr, 0, sizeof(nr));
 433
 434         while ((k_packed = bch2_btree_node_iter_next_all(iter, src))) {
 435                 if (filter_whiteouts && bkey_whiteout(k_packed))
 436                         continue;
 437
 438                 k = __bkey_disassemble(src, k_packed, &unpacked);
 439
 440                 if (filter_whiteouts &&
 441                     bch2_bkey_normalize(c, k))
 442                         continue;
 443
 444                 extent_sort_append(c, out_f, &nr, vstruct_last(dst), &prev, k);
 445         }
 446
 447         extent_sort_advance_prev(out_f, &nr, vstruct_last(dst), &prev);
 448
 449         dst->u64s = cpu_to_le16((u64 *) prev - dst->_data);
 450         return nr;
 451 }
 452
 453 static inline int sort_keys_cmp(struct btree *b,
 454                                 struct bkey_packed *l,
 455                                 struct bkey_packed *r)
 456 {
 457         return bkey_cmp_packed(b, l, r) ?:
 458                 (int) bkey_whiteout(r) - (int) bkey_whiteout(l) ?:
 459                 (int) l->needs_whiteout - (int) r->needs_whiteout;
 460 }
 461
 462 unsigned bch2_sort_keys(struct bkey_packed *dst,
 463                         struct sort_iter *iter,
 464                         bool filter_whiteouts)
 465 {
 466         const struct bkey_format *f = &iter->b->format;
 467         struct bkey_packed *in, *next, *out = dst;
 468
 469         sort_iter_sort(iter, sort_keys_cmp);
 470
 471         while ((in = sort_iter_next(iter, sort_keys_cmp))) {
 472                 if (bkey_whiteout(in) &&
 473                     (filter_whiteouts || !in->needs_whiteout))
 474                         continue;
 475
 476                 if (bkey_whiteout(in) &&
 477                     (next = sort_iter_peek(iter)) &&
 478                     !bkey_cmp_packed(iter->b, in, next)) {
 479                         BUG_ON(in->needs_whiteout &&
 480                                next->needs_whiteout);
 481                         /*
 482                          * XXX racy, called with read lock from write path
 483                          *
 484                          * leads to spurious BUG_ON() in bkey_unpack_key() in
 485                          * debug mode
 486                          */
 487                         next->needs_whiteout |= in->needs_whiteout;
 488                         continue;
 489                 }
 490
 491                 if (bkey_whiteout(in)) {
 492                         memcpy_u64s(out, in, bkeyp_key_u64s(f, in));
 493                         set_bkeyp_val_u64s(f, out, 0);
 494                 } else {
 495                         bkey_copy(out, in);
 496                 }
 497                 out = bkey_next(out);
 498         }
 499
 500         return (u64 *) out - (u64 *) dst;
 501 }
 502
 503 static inline int sort_extents_cmp(struct btree *b,
 504                                    struct bkey_packed *l,
 505                                    struct bkey_packed *r)
 506 {
 507         return bkey_cmp_packed(b, l, r) ?:
 508                 (int) bkey_deleted(l) - (int) bkey_deleted(r);
 509 }
 510
 511 unsigned bch2_sort_extents(struct bkey_packed *dst,
 512                            struct sort_iter *iter,
 513                            bool filter_whiteouts)
 514 {
 515         struct bkey_packed *in, *out = dst;
 516
 517         sort_iter_sort(iter, sort_extents_cmp);
 518
 519         while ((in = sort_iter_next(iter, sort_extents_cmp))) {
 520                 if (bkey_deleted(in))
 521                         continue;
 522
 523                 if (bkey_whiteout(in) &&
 524                     (filter_whiteouts || !in->needs_whiteout))
 525                         continue;
 526
 527                 bkey_copy(out, in);
 528                 out = bkey_next(out);
 529         }
 530
 531         return (u64 *) out - (u64 *) dst;
 532 }
 533
 534 static inline int sort_key_whiteouts_cmp(struct btree *b,
 535                                          struct bkey_packed *l,
 536                                          struct bkey_packed *r)
 537 {
 538         return bkey_cmp_packed(b, l, r);
 539 }
 540
 541 unsigned bch2_sort_key_whiteouts(struct bkey_packed *dst,
 542                                  struct sort_iter *iter)
 543 {
 544         struct bkey_packed *in, *out = dst;
 545
 546         sort_iter_sort(iter, sort_key_whiteouts_cmp);
 547
 548         while ((in = sort_iter_next(iter, sort_key_whiteouts_cmp))) {
 549                 bkey_copy(out, in);
 550                 out = bkey_next(out);
 551         }
 552
 553         return (u64 *) out - (u64 *) dst;
 554 }
 555
 556 static inline int sort_extent_whiteouts_cmp(struct btree *b,
 557                                             struct bkey_packed *l,
 558                                             struct bkey_packed *r)
 559 {
 560         struct bkey ul = bkey_unpack_key(b, l);
 561         struct bkey ur = bkey_unpack_key(b, r);
 562
 563         return bkey_cmp(bkey_start_pos(&ul), bkey_start_pos(&ur));
 564 }
 565
 566 unsigned bch2_sort_extent_whiteouts(struct bkey_packed *dst,
 567                                     struct sort_iter *iter)
 568 {
 569         const struct bkey_format *f = &iter->b->format;
 570         struct bkey_packed *in, *out = dst;
 571         struct bkey_i l, r;
 572         bool prev = false, l_packed = false;
 573         u64 max_packed_size     = bkey_field_max(f, BKEY_FIELD_SIZE);
 574         u64 max_packed_offset   = bkey_field_max(f, BKEY_FIELD_OFFSET);
 575         u64 new_size;
 576
 577         max_packed_size = min_t(u64, max_packed_size, KEY_SIZE_MAX);
 578
 579         sort_iter_sort(iter, sort_extent_whiteouts_cmp);
 580
 581         while ((in = sort_iter_next(iter, sort_extent_whiteouts_cmp))) {
 582                 if (bkey_deleted(in))
 583                         continue;
 584
 585                 EBUG_ON(bkeyp_val_u64s(f, in));
 586                 EBUG_ON(in->type != KEY_TYPE_discard);
 587
 588                 r.k = bkey_unpack_key(iter->b, in);
 589
 590                 if (prev &&
 591                     bkey_cmp(l.k.p, bkey_start_pos(&r.k)) >= 0) {
 592                         if (bkey_cmp(l.k.p, r.k.p) >= 0)
 593                                 continue;
 594
 595                         new_size = l_packed
 596                                 ? min(max_packed_size, max_packed_offset -
 597                                       bkey_start_offset(&l.k))
 598                                 : KEY_SIZE_MAX;
 599
 600                         new_size = min(new_size, r.k.p.offset -
 601                                        bkey_start_offset(&l.k));
 602
 603                         BUG_ON(new_size < l.k.size);
 604
 605                         bch2_key_resize(&l.k, new_size);
 606
 607                         if (bkey_cmp(l.k.p, r.k.p) >= 0)
 608                                 continue;
 609
 610                         bch2_cut_front(l.k.p, &r);
 611                 }
 612
 613                 if (prev) {
 614                         if (!bch2_bkey_pack(out, &l, f)) {
 615                                 BUG_ON(l_packed);
 616                                 bkey_copy(out, &l);
 617                         }
 618                         out = bkey_next(out);
 619                 }
 620
 621                 l = r;
 622                 prev = true;
 623                 l_packed = bkey_packed(in);
 624         }
 625
 626         if (prev) {
 627                 if (!bch2_bkey_pack(out, &l, f)) {
 628                         BUG_ON(l_packed);
 629                         bkey_copy(out, &l);
 630                 }
 631                 out = bkey_next(out);
 632         }
 633
 634         return (u64 *) out - (u64 *) dst;
 635 }