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 struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *c, enum btree_id btree_id,
  77                                            unsigned level, struct bpos pos,
  78                                            struct bpos end_pos, size_t *idx)
  79 {
  80         struct journal_keys *keys = &c->journal_keys;
  81         unsigned iters = 0;
  82         struct journal_key *k;
  83 search:
  84         if (!*idx)
  85                 *idx = __bch2_journal_key_search(keys, btree_id, level, pos);
  86
  87         while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
  88                 if (__journal_key_cmp(btree_id, level, end_pos, k) < 0)
  89                         return NULL;
  90
  91                 if (__journal_key_cmp(btree_id, level, pos, k) <= 0 &&
  92                     !k->overwritten)
  93                         return k->k;
  94
  95                 (*idx)++;
  96                 iters++;
  97                 if (iters == 10) {
  98                         *idx = 0;
  99                         goto search;
 100                 }
 101         }
 102
 103         return NULL;
 104 }
 105
 106 struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree_id,
 107                                            unsigned level, struct bpos pos)
 108 {
 109         size_t idx = 0;
 110
 111         return bch2_journal_keys_peek_upto(c, btree_id, level, pos, pos, &idx);
 112 }
 113
 114 static void journal_iters_fix(struct bch_fs *c)
 115 {
 116         struct journal_keys *keys = &c->journal_keys;
 117         /* The key we just inserted is immediately before the gap: */
 118         size_t gap_end = keys->gap + (keys->size - keys->nr);
 119         struct btree_and_journal_iter *iter;
 120
 121         /*
 122          * If an iterator points one after the key we just inserted, decrement
 123          * the iterator so it points at the key we just inserted - if the
 124          * decrement was unnecessary, bch2_btree_and_journal_iter_peek() will
 125          * handle that:
 126          */
 127         list_for_each_entry(iter, &c->journal_iters, journal.list)
 128                 if (iter->journal.idx == gap_end)
 129                         iter->journal.idx = keys->gap - 1;
 130 }
 131
 132 static void journal_iters_move_gap(struct bch_fs *c, size_t old_gap, size_t new_gap)
 133 {
 134         struct journal_keys *keys = &c->journal_keys;
 135         struct journal_iter *iter;
 136         size_t gap_size = keys->size - keys->nr;
 137
 138         list_for_each_entry(iter, &c->journal_iters, list) {
 139                 if (iter->idx > old_gap)
 140                         iter->idx -= gap_size;
 141                 if (iter->idx >= new_gap)
 142                         iter->idx += gap_size;
 143         }
 144 }
 145
 146 int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
 147                                  unsigned level, struct bkey_i *k)
 148 {
 149         struct journal_key n = {
 150                 .btree_id       = id,
 151                 .level          = level,
 152                 .k              = k,
 153                 .allocated      = true,
 154                 /*
 155                  * Ensure these keys are done last by journal replay, to unblock
 156                  * journal reclaim:
 157                  */
 158                 .journal_seq    = U32_MAX,
 159         };
 160         struct journal_keys *keys = &c->journal_keys;
 161         size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
 162
 163         BUG_ON(test_bit(BCH_FS_RW, &c->flags));
 164
 165         if (idx < keys->size &&
 166             journal_key_cmp(&n, &keys->d[idx]) == 0) {
 167                 if (keys->d[idx].allocated)
 168                         kfree(keys->d[idx].k);
 169                 keys->d[idx] = n;
 170                 return 0;
 171         }
 172
 173         if (idx > keys->gap)
 174                 idx -= keys->size - keys->nr;
 175
 176         if (keys->nr == keys->size) {
 177                 struct journal_keys new_keys = {
 178                         .nr                     = keys->nr,
 179                         .size                   = max_t(size_t, keys->size, 8) * 2,
 180                 };
 181
 182                 new_keys.d = kvmalloc_array(new_keys.size, sizeof(new_keys.d[0]), GFP_KERNEL);
 183                 if (!new_keys.d) {
 184                         bch_err(c, "%s: error allocating new key array (size %zu)",
 185                                 __func__, new_keys.size);
 186                         return -BCH_ERR_ENOMEM_journal_key_insert;
 187                 }
 188
 189                 /* Since @keys was full, there was no gap: */
 190                 memcpy(new_keys.d, keys->d, sizeof(keys->d[0]) * keys->nr);
 191                 kvfree(keys->d);
 192                 *keys = new_keys;
 193
 194                 /* And now the gap is at the end: */
 195                 keys->gap = keys->nr;
 196         }
 197
 198         journal_iters_move_gap(c, keys->gap, idx);
 199
 200         move_gap(keys->d, keys->nr, keys->size, keys->gap, idx);
 201         keys->gap = idx;
 202
 203         keys->nr++;
 204         keys->d[keys->gap++] = n;
 205
 206         journal_iters_fix(c);
 207
 208         return 0;
 209 }
 210
 211 /*
 212  * Can only be used from the recovery thread while we're still RO - can't be
 213  * used once we've got RW, as journal_keys is at that point used by multiple
 214  * threads:
 215  */
 216 int bch2_journal_key_insert(struct bch_fs *c, enum btree_id id,
 217                             unsigned level, struct bkey_i *k)
 218 {
 219         struct bkey_i *n;
 220         int ret;
 221
 222         n = kmalloc(bkey_bytes(&k->k), GFP_KERNEL);
 223         if (!n)
 224                 return -BCH_ERR_ENOMEM_journal_key_insert;
 225
 226         bkey_copy(n, k);
 227         ret = bch2_journal_key_insert_take(c, id, level, n);
 228         if (ret)
 229                 kfree(n);
 230         return ret;
 231 }
 232
 233 int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
 234                             unsigned level, struct bpos pos)
 235 {
 236         struct bkey_i whiteout;
 237
 238         bkey_init(&whiteout.k);
 239         whiteout.k.p = pos;
 240
 241         return bch2_journal_key_insert(c, id, level, &whiteout);
 242 }
 243
 244 void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
 245                                   unsigned level, struct bpos pos)
 246 {
 247         struct journal_keys *keys = &c->journal_keys;
 248         size_t idx = bch2_journal_key_search(keys, btree, level, pos);
 249
 250         if (idx < keys->size &&
 251             keys->d[idx].btree_id       == btree &&
 252             keys->d[idx].level          == level &&
 253             bpos_eq(keys->d[idx].k->k.p, pos))
 254                 keys->d[idx].overwritten = true;
 255 }
 256
 257 static void bch2_journal_iter_advance(struct journal_iter *iter)
 258 {
 259         if (iter->idx < iter->keys->size) {
 260                 iter->idx++;
 261                 if (iter->idx == iter->keys->gap)
 262                         iter->idx += iter->keys->size - iter->keys->nr;
 263         }
 264 }
 265
 266 static struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
 267 {
 268         struct journal_key *k = iter->keys->d + iter->idx;
 269
 270         while (k < iter->keys->d + iter->keys->size &&
 271                k->btree_id      == iter->btree_id &&
 272                k->level         == iter->level) {
 273                 if (!k->overwritten)
 274                         return bkey_i_to_s_c(k->k);
 275
 276                 bch2_journal_iter_advance(iter);
 277                 k = iter->keys->d + iter->idx;
 278         }
 279
 280         return bkey_s_c_null;
 281 }
 282
 283 static void bch2_journal_iter_exit(struct journal_iter *iter)
 284 {
 285         list_del(&iter->list);
 286 }
 287
 288 static void bch2_journal_iter_init(struct bch_fs *c,
 289                                    struct journal_iter *iter,
 290                                    enum btree_id id, unsigned level,
 291                                    struct bpos pos)
 292 {
 293         iter->btree_id  = id;
 294         iter->level     = level;
 295         iter->keys      = &c->journal_keys;
 296         iter->idx       = bch2_journal_key_search(&c->journal_keys, id, level, pos);
 297 }
 298
 299 static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
 300 {
 301         return bch2_btree_node_iter_peek_unpack(&iter->node_iter,
 302                                                 iter->b, &iter->unpacked);
 303 }
 304
 305 static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
 306 {
 307         bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
 308 }
 309
 310 void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
 311 {
 312         if (bpos_eq(iter->pos, SPOS_MAX))
 313                 iter->at_end = true;
 314         else
 315                 iter->pos = bpos_successor(iter->pos);
 316 }
 317
 318 struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
 319 {
 320         struct bkey_s_c btree_k, journal_k, ret;
 321 again:
 322         if (iter->at_end)
 323                 return bkey_s_c_null;
 324
 325         while ((btree_k = bch2_journal_iter_peek_btree(iter)).k &&
 326                bpos_lt(btree_k.k->p, iter->pos))
 327                 bch2_journal_iter_advance_btree(iter);
 328
 329         while ((journal_k = bch2_journal_iter_peek(&iter->journal)).k &&
 330                bpos_lt(journal_k.k->p, iter->pos))
 331                 bch2_journal_iter_advance(&iter->journal);
 332
 333         ret = journal_k.k &&
 334                 (!btree_k.k || bpos_le(journal_k.k->p, btree_k.k->p))
 335                 ? journal_k
 336                 : btree_k;
 337
 338         if (ret.k && iter->b && bpos_gt(ret.k->p, iter->b->data->max_key))
 339                 ret = bkey_s_c_null;
 340
 341         if (ret.k) {
 342                 iter->pos = ret.k->p;
 343                 if (bkey_deleted(ret.k)) {
 344                         bch2_btree_and_journal_iter_advance(iter);
 345                         goto again;
 346                 }
 347         } else {
 348                 iter->pos = SPOS_MAX;
 349                 iter->at_end = true;
 350         }
 351
 352         return ret;
 353 }
 354
 355 void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *iter)
 356 {
 357         bch2_journal_iter_exit(&iter->journal);
 358 }
 359
 360 void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
 361                                                   struct bch_fs *c,
 362                                                   struct btree *b,
 363                                                   struct btree_node_iter node_iter,
 364                                                   struct bpos pos)
 365 {
 366         memset(iter, 0, sizeof(*iter));
 367
 368         iter->b = b;
 369         iter->node_iter = node_iter;
 370         bch2_journal_iter_init(c, &iter->journal, b->c.btree_id, b->c.level, pos);
 371         INIT_LIST_HEAD(&iter->journal.list);
 372         iter->pos = b->data->min_key;
 373         iter->at_end = false;
 374 }
 375
 376 /*
 377  * this version is used by btree_gc before filesystem has gone RW and
 378  * multithreaded, so uses the journal_iters list:
 379  */
 380 void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
 381                                                 struct bch_fs *c,
 382                                                 struct btree *b)
 383 {
 384         struct btree_node_iter node_iter;
 385
 386         bch2_btree_node_iter_init_from_start(&node_iter, b);
 387         __bch2_btree_and_journal_iter_init_node_iter(iter, c, b, node_iter, b->data->min_key);
 388         list_add(&iter->journal.list, &c->journal_iters);
 389 }
 390
 391 /* sort and dedup all keys in the journal: */
 392
 393 void bch2_journal_entries_free(struct bch_fs *c)
 394 {
 395         struct journal_replay **i;
 396         struct genradix_iter iter;
 397
 398         genradix_for_each(&c->journal_entries, iter, i)
 399                 if (*i)
 400                         kvpfree(*i, offsetof(struct journal_replay, j) +
 401                                 vstruct_bytes(&(*i)->j));
 402         genradix_free(&c->journal_entries);
 403 }
 404
 405 /*
 406  * When keys compare equal, oldest compares first:
 407  */
 408 static int journal_sort_key_cmp(const void *_l, const void *_r)
 409 {
 410         const struct journal_key *l = _l;
 411         const struct journal_key *r = _r;
 412
 413         return  journal_key_cmp(l, r) ?:
 414                 cmp_int(l->journal_seq, r->journal_seq) ?:
 415                 cmp_int(l->journal_offset, r->journal_offset);
 416 }
 417
 418 void bch2_journal_keys_free(struct journal_keys *keys)
 419 {
 420         struct journal_key *i;
 421
 422         move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
 423         keys->gap = keys->nr;
 424
 425         for (i = keys->d; i < keys->d + keys->nr; i++)
 426                 if (i->allocated)
 427                         kfree(i->k);
 428
 429         kvfree(keys->d);
 430         keys->d = NULL;
 431         keys->nr = keys->gap = keys->size = 0;
 432 }
 433
 434 static void __journal_keys_sort(struct journal_keys *keys)
 435 {
 436         struct journal_key *src, *dst;
 437
 438         sort(keys->d, keys->nr, sizeof(keys->d[0]), journal_sort_key_cmp, NULL);
 439
 440         src = dst = keys->d;
 441         while (src < keys->d + keys->nr) {
 442                 while (src + 1 < keys->d + keys->nr &&
 443                        src[0].btree_id  == src[1].btree_id &&
 444                        src[0].level     == src[1].level &&
 445                        bpos_eq(src[0].k->k.p, src[1].k->k.p))
 446                         src++;
 447
 448                 *dst++ = *src++;
 449         }
 450
 451         keys->nr = dst - keys->d;
 452 }
 453
 454 int bch2_journal_keys_sort(struct bch_fs *c)
 455 {
 456         struct genradix_iter iter;
 457         struct journal_replay *i, **_i;
 458         struct jset_entry *entry;
 459         struct bkey_i *k;
 460         struct journal_keys *keys = &c->journal_keys;
 461         size_t nr_keys = 0, nr_read = 0;
 462
 463         genradix_for_each(&c->journal_entries, iter, _i) {
 464                 i = *_i;
 465
 466                 if (!i || i->ignore)
 467                         continue;
 468
 469                 for_each_jset_key(k, entry, &i->j)
 470                         nr_keys++;
 471         }
 472
 473         if (!nr_keys)
 474                 return 0;
 475
 476         keys->size = roundup_pow_of_two(nr_keys);
 477
 478         keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
 479         if (!keys->d) {
 480                 bch_err(c, "Failed to allocate buffer for sorted journal keys (%zu keys); trying slowpath",
 481                         nr_keys);
 482
 483                 do {
 484                         keys->size >>= 1;
 485                         keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
 486                 } while (!keys->d && keys->size > nr_keys / 8);
 487
 488                 if (!keys->d) {
 489                         bch_err(c, "Failed to allocate %zu size buffer for sorted journal keys; exiting",
 490                                 keys->size);
 491                         return -BCH_ERR_ENOMEM_journal_keys_sort;
 492                 }
 493         }
 494
 495         genradix_for_each(&c->journal_entries, iter, _i) {
 496                 i = *_i;
 497
 498                 if (!i || i->ignore)
 499                         continue;
 500
 501                 cond_resched();
 502
 503                 for_each_jset_key(k, entry, &i->j) {
 504                         if (keys->nr == keys->size) {
 505                                 __journal_keys_sort(keys);
 506
 507                                 if (keys->nr > keys->size * 7 / 8) {
 508                                         bch_err(c, "Too many journal keys for slowpath; have %zu compacted, buf size %zu, processed %zu/%zu",
 509                                                 keys->nr, keys->size, nr_read, nr_keys);
 510                                         return -BCH_ERR_ENOMEM_journal_keys_sort;
 511                                 }
 512                         }
 513
 514                         keys->d[keys->nr++] = (struct journal_key) {
 515                                 .btree_id       = entry->btree_id,
 516                                 .level          = entry->level,
 517                                 .k              = k,
 518                                 .journal_seq    = le64_to_cpu(i->j.seq),
 519                                 .journal_offset = k->_data - i->j._data,
 520                         };
 521
 522                         nr_read++;
 523                 }
 524         }
 525
 526         __journal_keys_sort(keys);
 527         keys->gap = keys->nr;
 528
 529         bch_verbose(c, "Journal keys: %zu read, %zu after sorting and compacting", nr_keys, keys->nr);
 530         return 0;
 531 }