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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bset.h"
   5 #include "btree_journal_iter.h"
   6 #include "journal_io.h"
   7
   8 #include <linux/sort.h>
   9
  10 /*
  11  * For managing keys we read from the journal: until journal replay works normal
  12  * btree lookups need to be able to find and return keys from the journal where
  13  * they overwrite what's in the btree, so we have a special iterator and
  14  * operations for the regular btree iter code to use:
  15  */
  16
  17 static int __journal_key_cmp(enum btree_id      l_btree_id,
  18                              unsigned           l_level,
  19                              struct bpos        l_pos,
  20                              const struct journal_key *r)
  21 {
  22         return (cmp_int(l_btree_id,     r->btree_id) ?:
  23                 cmp_int(l_level,        r->level) ?:
  24                 bpos_cmp(l_pos, r->k->k.p));
  25 }
  26
  27 static int journal_key_cmp(const struct journal_key *l, const struct journal_key *r)
  28 {
  29         return __journal_key_cmp(l->btree_id, l->level, l->k->k.p, r);
  30 }
  31
  32 static inline size_t idx_to_pos(struct journal_keys *keys, size_t idx)
  33 {
  34         size_t gap_size = keys->size - keys->nr;
  35
  36         if (idx >= keys->gap)
  37                 idx += gap_size;
  38         return idx;
  39 }
  40
  41 static inline struct journal_key *idx_to_key(struct journal_keys *keys, size_t idx)
  42 {
  43         return keys->d + idx_to_pos(keys, idx);
  44 }
  45
  46 static size_t __bch2_journal_key_search(struct journal_keys *keys,
  47                                         enum btree_id id, unsigned level,
  48                                         struct bpos pos)
  49 {
  50         size_t l = 0, r = keys->nr, m;
  51
  52         while (l < r) {
  53                 m = l + ((r - l) >> 1);
  54                 if (__journal_key_cmp(id, level, pos, idx_to_key(keys, m)) > 0)
  55                         l = m + 1;
  56                 else
  57                         r = m;
  58         }
  59
  60         BUG_ON(l < keys->nr &&
  61                __journal_key_cmp(id, level, pos, idx_to_key(keys, l)) > 0);
  62
  63         BUG_ON(l &&
  64                __journal_key_cmp(id, level, pos, idx_to_key(keys, l - 1)) <= 0);
  65
  66         return l;
  67 }
  68
  69 static size_t bch2_journal_key_search(struct journal_keys *keys,
  70                                       enum btree_id id, unsigned level,
  71                                       struct bpos pos)
  72 {
  73         return idx_to_pos(keys, __bch2_journal_key_search(keys, id, level, pos));
  74 }
  75
  76 /* Returns first non-overwritten key >= search key: */
  77 struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *c, enum btree_id btree_id,
  78                                            unsigned level, struct bpos pos,
  79                                            struct bpos end_pos, size_t *idx)
  80 {
  81         struct journal_keys *keys = &c->journal_keys;
  82         unsigned iters = 0;
  83         struct journal_key *k;
  84
  85         BUG_ON(*idx > keys->nr);
  86 search:
  87         if (!*idx)
  88                 *idx = __bch2_journal_key_search(keys, btree_id, level, pos);
  89
  90         while (*idx &&
  91                __journal_key_cmp(btree_id, level, end_pos, idx_to_key(keys, *idx - 1)) <= 0) {
  92                 --(*idx);
  93                 iters++;
  94                 if (iters == 10) {
  95                         *idx = 0;
  96                         goto search;
  97                 }
  98         }
  99
 100         while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
 101                 if (__journal_key_cmp(btree_id, level, end_pos, k) < 0)
 102                         return NULL;
 103
 104                 if (k->overwritten) {
 105                         (*idx)++;
 106                         continue;
 107                 }
 108
 109                 if (__journal_key_cmp(btree_id, level, pos, k) <= 0)
 110                         return k->k;
 111
 112                 (*idx)++;
 113                 iters++;
 114                 if (iters == 10) {
 115                         *idx = 0;
 116                         goto search;
 117                 }
 118         }
 119
 120         return NULL;
 121 }
 122
 123 struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree_id,
 124                                            unsigned level, struct bpos pos)
 125 {
 126         size_t idx = 0;
 127
 128         return bch2_journal_keys_peek_upto(c, btree_id, level, pos, pos, &idx);
 129 }
 130
 131 static void journal_iters_fix(struct bch_fs *c)
 132 {
 133         struct journal_keys *keys = &c->journal_keys;
 134         /* The key we just inserted is immediately before the gap: */
 135         size_t gap_end = keys->gap + (keys->size - keys->nr);
 136         struct btree_and_journal_iter *iter;
 137
 138         /*
 139          * If an iterator points one after the key we just inserted, decrement
 140          * the iterator so it points at the key we just inserted - if the
 141          * decrement was unnecessary, bch2_btree_and_journal_iter_peek() will
 142          * handle that:
 143          */
 144         list_for_each_entry(iter, &c->journal_iters, journal.list)
 145                 if (iter->journal.idx == gap_end)
 146                         iter->journal.idx = keys->gap - 1;
 147 }
 148
 149 static void journal_iters_move_gap(struct bch_fs *c, size_t old_gap, size_t new_gap)
 150 {
 151         struct journal_keys *keys = &c->journal_keys;
 152         struct journal_iter *iter;
 153         size_t gap_size = keys->size - keys->nr;
 154
 155         list_for_each_entry(iter, &c->journal_iters, list) {
 156                 if (iter->idx > old_gap)
 157                         iter->idx -= gap_size;
 158                 if (iter->idx >= new_gap)
 159                         iter->idx += gap_size;
 160         }
 161 }
 162
 163 int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
 164                                  unsigned level, struct bkey_i *k)
 165 {
 166         struct journal_key n = {
 167                 .btree_id       = id,
 168                 .level          = level,
 169                 .k              = k,
 170                 .allocated      = true,
 171                 /*
 172                  * Ensure these keys are done last by journal replay, to unblock
 173                  * journal reclaim:
 174                  */
 175                 .journal_seq    = U32_MAX,
 176         };
 177         struct journal_keys *keys = &c->journal_keys;
 178         size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
 179
 180         BUG_ON(test_bit(BCH_FS_rw, &c->flags));
 181
 182         if (idx < keys->size &&
 183             journal_key_cmp(&n, &keys->d[idx]) == 0) {
 184                 if (keys->d[idx].allocated)
 185                         kfree(keys->d[idx].k);
 186                 keys->d[idx] = n;
 187                 return 0;
 188         }
 189
 190         if (idx > keys->gap)
 191                 idx -= keys->size - keys->nr;
 192
 193         if (keys->nr == keys->size) {
 194                 struct journal_keys new_keys = {
 195                         .nr                     = keys->nr,
 196                         .size                   = max_t(size_t, keys->size, 8) * 2,
 197                 };
 198
 199                 new_keys.d = kvmalloc_array(new_keys.size, sizeof(new_keys.d[0]), GFP_KERNEL);
 200                 if (!new_keys.d) {
 201                         bch_err(c, "%s: error allocating new key array (size %zu)",
 202                                 __func__, new_keys.size);
 203                         return -BCH_ERR_ENOMEM_journal_key_insert;
 204                 }
 205
 206                 /* Since @keys was full, there was no gap: */
 207                 memcpy(new_keys.d, keys->d, sizeof(keys->d[0]) * keys->nr);
 208                 kvfree(keys->d);
 209                 keys->d         = new_keys.d;
 210                 keys->nr        = new_keys.nr;
 211                 keys->size      = new_keys.size;
 212
 213                 /* And now the gap is at the end: */
 214                 keys->gap       = keys->nr;
 215         }
 216
 217         journal_iters_move_gap(c, keys->gap, idx);
 218
 219         move_gap(keys->d, keys->nr, keys->size, keys->gap, idx);
 220         keys->gap = idx;
 221
 222         keys->nr++;
 223         keys->d[keys->gap++] = n;
 224
 225         journal_iters_fix(c);
 226
 227         return 0;
 228 }
 229
 230 /*
 231  * Can only be used from the recovery thread while we're still RO - can't be
 232  * used once we've got RW, as journal_keys is at that point used by multiple
 233  * threads:
 234  */
 235 int bch2_journal_key_insert(struct bch_fs *c, enum btree_id id,
 236                             unsigned level, struct bkey_i *k)
 237 {
 238         struct bkey_i *n;
 239         int ret;
 240
 241         n = kmalloc(bkey_bytes(&k->k), GFP_KERNEL);
 242         if (!n)
 243                 return -BCH_ERR_ENOMEM_journal_key_insert;
 244
 245         bkey_copy(n, k);
 246         ret = bch2_journal_key_insert_take(c, id, level, n);
 247         if (ret)
 248                 kfree(n);
 249         return ret;
 250 }
 251
 252 int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
 253                             unsigned level, struct bpos pos)
 254 {
 255         struct bkey_i whiteout;
 256
 257         bkey_init(&whiteout.k);
 258         whiteout.k.p = pos;
 259
 260         return bch2_journal_key_insert(c, id, level, &whiteout);
 261 }
 262
 263 void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
 264                                   unsigned level, struct bpos pos)
 265 {
 266         struct journal_keys *keys = &c->journal_keys;
 267         size_t idx = bch2_journal_key_search(keys, btree, level, pos);
 268
 269         if (idx < keys->size &&
 270             keys->d[idx].btree_id       == btree &&
 271             keys->d[idx].level          == level &&
 272             bpos_eq(keys->d[idx].k->k.p, pos))
 273                 keys->d[idx].overwritten = true;
 274 }
 275
 276 static void bch2_journal_iter_advance(struct journal_iter *iter)
 277 {
 278         if (iter->idx < iter->keys->size) {
 279                 iter->idx++;
 280                 if (iter->idx == iter->keys->gap)
 281                         iter->idx += iter->keys->size - iter->keys->nr;
 282         }
 283 }
 284
 285 static struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
 286 {
 287         struct journal_key *k = iter->keys->d + iter->idx;
 288
 289         while (k < iter->keys->d + iter->keys->size &&
 290                k->btree_id      == iter->btree_id &&
 291                k->level         == iter->level) {
 292                 if (!k->overwritten)
 293                         return bkey_i_to_s_c(k->k);
 294
 295                 bch2_journal_iter_advance(iter);
 296                 k = iter->keys->d + iter->idx;
 297         }
 298
 299         return bkey_s_c_null;
 300 }
 301
 302 static void bch2_journal_iter_exit(struct journal_iter *iter)
 303 {
 304         list_del(&iter->list);
 305 }
 306
 307 static void bch2_journal_iter_init(struct bch_fs *c,
 308                                    struct journal_iter *iter,
 309                                    enum btree_id id, unsigned level,
 310                                    struct bpos pos)
 311 {
 312         iter->btree_id  = id;
 313         iter->level     = level;
 314         iter->keys      = &c->journal_keys;
 315         iter->idx       = bch2_journal_key_search(&c->journal_keys, id, level, pos);
 316 }
 317
 318 static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
 319 {
 320         return bch2_btree_node_iter_peek_unpack(&iter->node_iter,
 321                                                 iter->b, &iter->unpacked);
 322 }
 323
 324 static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
 325 {
 326         bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
 327 }
 328
 329 void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
 330 {
 331         if (bpos_eq(iter->pos, SPOS_MAX))
 332                 iter->at_end = true;
 333         else
 334                 iter->pos = bpos_successor(iter->pos);
 335 }
 336
 337 struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
 338 {
 339         struct bkey_s_c btree_k, journal_k, ret;
 340 again:
 341         if (iter->at_end)
 342                 return bkey_s_c_null;
 343
 344         while ((btree_k = bch2_journal_iter_peek_btree(iter)).k &&
 345                bpos_lt(btree_k.k->p, iter->pos))
 346                 bch2_journal_iter_advance_btree(iter);
 347
 348         while ((journal_k = bch2_journal_iter_peek(&iter->journal)).k &&
 349                bpos_lt(journal_k.k->p, iter->pos))
 350                 bch2_journal_iter_advance(&iter->journal);
 351
 352         ret = journal_k.k &&
 353                 (!btree_k.k || bpos_le(journal_k.k->p, btree_k.k->p))
 354                 ? journal_k
 355                 : btree_k;
 356
 357         if (ret.k && iter->b && bpos_gt(ret.k->p, iter->b->data->max_key))
 358                 ret = bkey_s_c_null;
 359
 360         if (ret.k) {
 361                 iter->pos = ret.k->p;
 362                 if (bkey_deleted(ret.k)) {
 363                         bch2_btree_and_journal_iter_advance(iter);
 364                         goto again;
 365                 }
 366         } else {
 367                 iter->pos = SPOS_MAX;
 368                 iter->at_end = true;
 369         }
 370
 371         return ret;
 372 }
 373
 374 void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *iter)
 375 {
 376         bch2_journal_iter_exit(&iter->journal);
 377 }
 378
 379 void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
 380                                                   struct bch_fs *c,
 381                                                   struct btree *b,
 382                                                   struct btree_node_iter node_iter,
 383                                                   struct bpos pos)
 384 {
 385         memset(iter, 0, sizeof(*iter));
 386
 387         iter->b = b;
 388         iter->node_iter = node_iter;
 389         bch2_journal_iter_init(c, &iter->journal, b->c.btree_id, b->c.level, pos);
 390         INIT_LIST_HEAD(&iter->journal.list);
 391         iter->pos = b->data->min_key;
 392         iter->at_end = false;
 393 }
 394
 395 /*
 396  * this version is used by btree_gc before filesystem has gone RW and
 397  * multithreaded, so uses the journal_iters list:
 398  */
 399 void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
 400                                                 struct bch_fs *c,
 401                                                 struct btree *b)
 402 {
 403         struct btree_node_iter node_iter;
 404
 405         bch2_btree_node_iter_init_from_start(&node_iter, b);
 406         __bch2_btree_and_journal_iter_init_node_iter(iter, c, b, node_iter, b->data->min_key);
 407         list_add(&iter->journal.list, &c->journal_iters);
 408 }
 409
 410 /* sort and dedup all keys in the journal: */
 411
 412 void bch2_journal_entries_free(struct bch_fs *c)
 413 {
 414         struct journal_replay **i;
 415         struct genradix_iter iter;
 416
 417         genradix_for_each(&c->journal_entries, iter, i)
 418                 if (*i)
 419                         kvpfree(*i, offsetof(struct journal_replay, j) +
 420                                 vstruct_bytes(&(*i)->j));
 421         genradix_free(&c->journal_entries);
 422 }
 423
 424 /*
 425  * When keys compare equal, oldest compares first:
 426  */
 427 static int journal_sort_key_cmp(const void *_l, const void *_r)
 428 {
 429         const struct journal_key *l = _l;
 430         const struct journal_key *r = _r;
 431
 432         return  journal_key_cmp(l, r) ?:
 433                 cmp_int(l->journal_seq, r->journal_seq) ?:
 434                 cmp_int(l->journal_offset, r->journal_offset);
 435 }
 436
 437 void bch2_journal_keys_put(struct bch_fs *c)
 438 {
 439         struct journal_keys *keys = &c->journal_keys;
 440         struct journal_key *i;
 441
 442         BUG_ON(atomic_read(&keys->ref) <= 0);
 443
 444         if (!atomic_dec_and_test(&keys->ref))
 445                 return;
 446
 447         move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
 448         keys->gap = keys->nr;
 449
 450         for (i = keys->d; i < keys->d + keys->nr; i++)
 451                 if (i->allocated)
 452                         kfree(i->k);
 453
 454         kvfree(keys->d);
 455         keys->d = NULL;
 456         keys->nr = keys->gap = keys->size = 0;
 457
 458         bch2_journal_entries_free(c);
 459 }
 460
 461 static void __journal_keys_sort(struct journal_keys *keys)
 462 {
 463         struct journal_key *src, *dst;
 464
 465         sort(keys->d, keys->nr, sizeof(keys->d[0]), journal_sort_key_cmp, NULL);
 466
 467         src = dst = keys->d;
 468         while (src < keys->d + keys->nr) {
 469                 while (src + 1 < keys->d + keys->nr &&
 470                        !journal_key_cmp(src, src + 1))
 471                         src++;
 472
 473                 *dst++ = *src++;
 474         }
 475
 476         keys->nr = dst - keys->d;
 477 }
 478
 479 int bch2_journal_keys_sort(struct bch_fs *c)
 480 {
 481         struct genradix_iter iter;
 482         struct journal_replay *i, **_i;
 483         struct jset_entry *entry;
 484         struct bkey_i *k;
 485         struct journal_keys *keys = &c->journal_keys;
 486         size_t nr_keys = 0, nr_read = 0;
 487
 488         genradix_for_each(&c->journal_entries, iter, _i) {
 489                 i = *_i;
 490
 491                 if (!i || i->ignore)
 492                         continue;
 493
 494                 for_each_jset_key(k, entry, &i->j)
 495                         nr_keys++;
 496         }
 497
 498         if (!nr_keys)
 499                 return 0;
 500
 501         keys->size = roundup_pow_of_two(nr_keys);
 502
 503         keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
 504         if (!keys->d) {
 505                 bch_err(c, "Failed to allocate buffer for sorted journal keys (%zu keys); trying slowpath",
 506                         nr_keys);
 507
 508                 do {
 509                         keys->size >>= 1;
 510                         keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
 511                 } while (!keys->d && keys->size > nr_keys / 8);
 512
 513                 if (!keys->d) {
 514                         bch_err(c, "Failed to allocate %zu size buffer for sorted journal keys; exiting",
 515                                 keys->size);
 516                         return -BCH_ERR_ENOMEM_journal_keys_sort;
 517                 }
 518         }
 519
 520         genradix_for_each(&c->journal_entries, iter, _i) {
 521                 i = *_i;
 522
 523                 if (!i || i->ignore)
 524                         continue;
 525
 526                 cond_resched();
 527
 528                 for_each_jset_key(k, entry, &i->j) {
 529                         if (keys->nr == keys->size) {
 530                                 __journal_keys_sort(keys);
 531
 532                                 if (keys->nr > keys->size * 7 / 8) {
 533                                         bch_err(c, "Too many journal keys for slowpath; have %zu compacted, buf size %zu, processed %zu/%zu",
 534                                                 keys->nr, keys->size, nr_read, nr_keys);
 535                                         return -BCH_ERR_ENOMEM_journal_keys_sort;
 536                                 }
 537                         }
 538
 539                         keys->d[keys->nr++] = (struct journal_key) {
 540                                 .btree_id       = entry->btree_id,
 541                                 .level          = entry->level,
 542                                 .k              = k,
 543                                 .journal_seq    = le64_to_cpu(i->j.seq),
 544                                 .journal_offset = k->_data - i->j._data,
 545                         };
 546
 547                         nr_read++;
 548                 }
 549         }
 550
 551         __journal_keys_sort(keys);
 552         keys->gap = keys->nr;
 553
 554         bch_verbose(c, "Journal keys: %zu read, %zu after sorting and compacting", nr_keys, keys->nr);
 555         return 0;
 556 }