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

   1
   2 #include "bcachefs.h"
   3 #include "btree_cache.h"
   4 #include "btree_iter.h"
   5 #include "btree_key_cache.h"
   6 #include "btree_locking.h"
   7 #include "btree_update.h"
   8 #include "error.h"
   9 #include "journal.h"
  10 #include "journal_reclaim.h"
  11
  12 #include <trace/events/bcachefs.h>
  13
  14 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  15                                        const void *obj)
  16 {
  17         const struct bkey_cached *ck = obj;
  18         const struct bkey_cached_key *key = arg->key;
  19
  20         return cmp_int(ck->key.btree_id, key->btree_id) ?:
  21                 bkey_cmp(ck->key.pos, key->pos);
  22 }
  23
  24 static const struct rhashtable_params bch2_btree_key_cache_params = {
  25         .head_offset    = offsetof(struct bkey_cached, hash),
  26         .key_offset     = offsetof(struct bkey_cached, key),
  27         .key_len        = sizeof(struct bkey_cached_key),
  28         .obj_cmpfn      = bch2_btree_key_cache_cmp_fn,
  29 };
  30
  31 __flatten
  32 inline struct bkey_cached *
  33 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
  34 {
  35         struct bkey_cached_key key = {
  36                 .btree_id       = btree_id,
  37                 .pos            = pos,
  38         };
  39
  40         return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
  41                                       bch2_btree_key_cache_params);
  42 }
  43
  44 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
  45 {
  46         if (!six_trylock_intent(&ck->c.lock))
  47                 return false;
  48
  49         if (!six_trylock_write(&ck->c.lock)) {
  50                 six_unlock_intent(&ck->c.lock);
  51                 return false;
  52         }
  53
  54         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
  55                 six_unlock_write(&ck->c.lock);
  56                 six_unlock_intent(&ck->c.lock);
  57                 return false;
  58         }
  59
  60         return true;
  61 }
  62
  63 static void bkey_cached_evict(struct btree_key_cache *c,
  64                               struct bkey_cached *ck)
  65 {
  66         BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
  67                                       bch2_btree_key_cache_params));
  68         memset(&ck->key, ~0, sizeof(ck->key));
  69 }
  70
  71 static void bkey_cached_free(struct btree_key_cache *c,
  72                              struct bkey_cached *ck)
  73 {
  74         list_move(&ck->list, &c->freed);
  75
  76         kfree(ck->k);
  77         ck->k           = NULL;
  78         ck->u64s        = 0;
  79
  80         six_unlock_write(&ck->c.lock);
  81         six_unlock_intent(&ck->c.lock);
  82 }
  83
  84 static struct bkey_cached *
  85 bkey_cached_alloc(struct btree_key_cache *c)
  86 {
  87         struct bkey_cached *ck;
  88
  89         list_for_each_entry(ck, &c->freed, list)
  90                 if (bkey_cached_lock_for_evict(ck))
  91                         return ck;
  92
  93         list_for_each_entry(ck, &c->clean, list)
  94                 if (bkey_cached_lock_for_evict(ck)) {
  95                         bkey_cached_evict(c, ck);
  96                         return ck;
  97                 }
  98
  99         ck = kzalloc(sizeof(*ck), GFP_NOFS);
 100         if (!ck)
 101                 return NULL;
 102
 103         INIT_LIST_HEAD(&ck->list);
 104         six_lock_init(&ck->c.lock);
 105         BUG_ON(!six_trylock_intent(&ck->c.lock));
 106         BUG_ON(!six_trylock_write(&ck->c.lock));
 107
 108         return ck;
 109 }
 110
 111 static struct bkey_cached *
 112 btree_key_cache_create(struct btree_key_cache *c,
 113                        enum btree_id btree_id,
 114                        struct bpos pos)
 115 {
 116         struct bkey_cached *ck;
 117
 118         ck = bkey_cached_alloc(c);
 119         if (!ck)
 120                 return ERR_PTR(-ENOMEM);
 121
 122         ck->c.level             = 0;
 123         ck->c.btree_id          = btree_id;
 124         ck->key.btree_id        = btree_id;
 125         ck->key.pos             = pos;
 126         ck->valid               = false;
 127
 128         BUG_ON(ck->flags);
 129
 130         if (rhashtable_lookup_insert_fast(&c->table,
 131                                           &ck->hash,
 132                                           bch2_btree_key_cache_params)) {
 133                 /* We raced with another fill: */
 134                 bkey_cached_free(c, ck);
 135                 return NULL;
 136         }
 137
 138         list_move(&ck->list, &c->clean);
 139         six_unlock_write(&ck->c.lock);
 140
 141         return ck;
 142 }
 143
 144 static int btree_key_cache_fill(struct btree_trans *trans,
 145                                 struct btree_iter *ck_iter,
 146                                 struct bkey_cached *ck)
 147 {
 148         struct btree_iter *iter;
 149         struct bkey_s_c k;
 150         unsigned new_u64s = 0;
 151         struct bkey_i *new_k = NULL;
 152         int ret;
 153
 154         iter = bch2_trans_get_iter(trans, ck->key.btree_id,
 155                                    ck->key.pos, BTREE_ITER_SLOTS);
 156         if (IS_ERR(iter))
 157                 return PTR_ERR(iter);
 158
 159         k = bch2_btree_iter_peek_slot(iter);
 160         ret = bkey_err(k);
 161         if (ret) {
 162                 bch2_trans_iter_put(trans, iter);
 163                 return ret;
 164         }
 165
 166         if (!bch2_btree_node_relock(ck_iter, 0)) {
 167                 bch2_trans_iter_put(trans, iter);
 168                 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 169                 return -EINTR;
 170         }
 171
 172         if (k.k->u64s > ck->u64s) {
 173                 new_u64s = roundup_pow_of_two(k.k->u64s);
 174                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 175                 if (!new_k) {
 176                         bch2_trans_iter_put(trans, iter);
 177                         return -ENOMEM;
 178                 }
 179         }
 180
 181         bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
 182         if (new_k) {
 183                 kfree(ck->k);
 184                 ck->u64s = new_u64s;
 185                 ck->k = new_k;
 186         }
 187
 188         bkey_reassemble(ck->k, k);
 189         ck->valid = true;
 190         bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
 191
 192         /* We're not likely to need this iterator again: */
 193         bch2_trans_iter_free(trans, iter);
 194
 195         return 0;
 196 }
 197
 198 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 199 {
 200         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 201         const struct btree_iter *iter = p;
 202
 203         return ck->key.btree_id == iter->btree_id &&
 204                 !bkey_cmp(ck->key.pos, iter->pos) ? 0 : -1;
 205 }
 206
 207 __flatten
 208 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
 209 {
 210         struct btree_trans *trans = iter->trans;
 211         struct bch_fs *c = trans->c;
 212         struct bkey_cached *ck;
 213         int ret = 0;
 214
 215         BUG_ON(iter->level);
 216
 217         if (btree_node_locked(iter, 0)) {
 218                 ck = (void *) iter->l[0].b;
 219                 goto fill;
 220         }
 221 retry:
 222         ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
 223         if (!ck) {
 224                 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
 225                         iter->l[0].b = NULL;
 226                         return 0;
 227                 }
 228
 229                 mutex_lock(&c->btree_key_cache.lock);
 230                 ck = btree_key_cache_create(&c->btree_key_cache,
 231                                             iter->btree_id, iter->pos);
 232                 mutex_unlock(&c->btree_key_cache.lock);
 233
 234                 ret = PTR_ERR_OR_ZERO(ck);
 235                 if (ret)
 236                         goto err;
 237                 if (!ck)
 238                         goto retry;
 239
 240                 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
 241                 iter->locks_want = 1;
 242         } else {
 243                 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
 244
 245                 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
 246                                      bkey_cached_check_fn, iter, _THIS_IP_)) {
 247                         if (ck->key.btree_id != iter->btree_id ||
 248                             bkey_cmp(ck->key.pos, iter->pos)) {
 249                                 goto retry;
 250                         }
 251
 252                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 253                         ret = -EINTR;
 254                         goto err;
 255                 }
 256
 257                 if (ck->key.btree_id != iter->btree_id ||
 258                     bkey_cmp(ck->key.pos, iter->pos)) {
 259                         six_unlock_type(&ck->c.lock, lock_want);
 260                         goto retry;
 261                 }
 262
 263                 mark_btree_node_locked(iter, 0, lock_want);
 264         }
 265
 266         iter->l[0].lock_seq     = ck->c.lock.state.seq;
 267         iter->l[0].b            = (void *) ck;
 268 fill:
 269         if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
 270                 if (!btree_node_intent_locked(iter, 0))
 271                         bch2_btree_iter_upgrade(iter, 1);
 272                 if (!btree_node_intent_locked(iter, 0)) {
 273                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 274                         ret = -EINTR;
 275                         goto err;
 276                 }
 277
 278                 ret = btree_key_cache_fill(trans, iter, ck);
 279                 if (ret)
 280                         goto err;
 281         }
 282
 283         iter->uptodate = BTREE_ITER_NEED_PEEK;
 284         bch2_btree_iter_downgrade(iter);
 285         return ret;
 286 err:
 287         if (ret != -EINTR) {
 288                 btree_node_unlock(iter, 0);
 289                 iter->flags |= BTREE_ITER_ERROR;
 290                 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
 291         }
 292         return ret;
 293 }
 294
 295 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 296                                      struct bkey_cached_key key,
 297                                      u64 journal_seq,
 298                                      bool evict)
 299 {
 300         struct bch_fs *c = trans->c;
 301         struct journal *j = &c->journal;
 302         struct btree_iter *c_iter = NULL, *b_iter = NULL;
 303         struct bkey_cached *ck;
 304         int ret;
 305
 306         b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 307                                      BTREE_ITER_SLOTS|
 308                                      BTREE_ITER_INTENT);
 309         ret = PTR_ERR_OR_ZERO(b_iter);
 310         if (ret)
 311                 goto out;
 312
 313         c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 314                                      BTREE_ITER_CACHED|
 315                                      BTREE_ITER_CACHED_NOFILL|
 316                                      BTREE_ITER_CACHED_NOCREATE|
 317                                      BTREE_ITER_INTENT);
 318         ret = PTR_ERR_OR_ZERO(c_iter);
 319         if (ret)
 320                 goto out;
 321 retry:
 322         ret = bch2_btree_iter_traverse(c_iter);
 323         if (ret)
 324                 goto err;
 325
 326         ck = (void *) c_iter->l[0].b;
 327         if (!ck ||
 328             (journal_seq && ck->journal.seq != journal_seq))
 329                 goto out;
 330
 331         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 332                 if (!evict)
 333                         goto out;
 334                 goto evict;
 335         }
 336
 337         ret   = bch2_btree_iter_traverse(b_iter) ?:
 338                 bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
 339                 bch2_trans_commit(trans, NULL, NULL,
 340                                   BTREE_INSERT_NOUNLOCK|
 341                                   BTREE_INSERT_NOCHECK_RW|
 342                                   BTREE_INSERT_NOFAIL|
 343                                   BTREE_INSERT_USE_RESERVE|
 344                                   BTREE_INSERT_USE_ALLOC_RESERVE|
 345                                   BTREE_INSERT_JOURNAL_RESERVED|
 346                                   BTREE_INSERT_JOURNAL_RECLAIM);
 347 err:
 348         if (ret == -EINTR)
 349                 goto retry;
 350
 351         BUG_ON(ret && !bch2_journal_error(j));
 352
 353         if (ret)
 354                 goto out;
 355
 356         bch2_journal_pin_drop(j, &ck->journal);
 357         bch2_journal_preres_put(j, &ck->res);
 358         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 359
 360         if (!evict) {
 361                 mutex_lock(&c->btree_key_cache.lock);
 362                 list_move_tail(&ck->list, &c->btree_key_cache.clean);
 363                 mutex_unlock(&c->btree_key_cache.lock);
 364         } else {
 365 evict:
 366                 BUG_ON(!btree_node_intent_locked(c_iter, 0));
 367
 368                 mark_btree_node_unlocked(c_iter, 0);
 369                 c_iter->l[0].b = NULL;
 370
 371                 six_lock_write(&ck->c.lock, NULL, NULL);
 372
 373                 mutex_lock(&c->btree_key_cache.lock);
 374                 bkey_cached_evict(&c->btree_key_cache, ck);
 375                 bkey_cached_free(&c->btree_key_cache, ck);
 376                 mutex_unlock(&c->btree_key_cache.lock);
 377         }
 378 out:
 379         bch2_trans_iter_put(trans, b_iter);
 380         bch2_trans_iter_put(trans, c_iter);
 381         return ret;
 382 }
 383
 384 static void btree_key_cache_journal_flush(struct journal *j,
 385                                           struct journal_entry_pin *pin,
 386                                           u64 seq)
 387 {
 388         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 389         struct bkey_cached *ck =
 390                 container_of(pin, struct bkey_cached, journal);
 391         struct bkey_cached_key key;
 392         struct btree_trans trans;
 393
 394         six_lock_read(&ck->c.lock, NULL, NULL);
 395         key = ck->key;
 396
 397         if (ck->journal.seq != seq ||
 398             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 399                 six_unlock_read(&ck->c.lock);
 400                 return;
 401         }
 402         six_unlock_read(&ck->c.lock);
 403
 404         bch2_trans_init(&trans, c, 0, 0);
 405         btree_key_cache_flush_pos(&trans, key, seq, false);
 406         bch2_trans_exit(&trans);
 407 }
 408
 409 /*
 410  * Flush and evict a key from the key cache:
 411  */
 412 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 413                                enum btree_id id, struct bpos pos)
 414 {
 415         struct bch_fs *c = trans->c;
 416         struct bkey_cached_key key = { id, pos };
 417
 418         /* Fastpath - assume it won't be found: */
 419         if (!bch2_btree_key_cache_find(c, id, pos))
 420                 return 0;
 421
 422         return btree_key_cache_flush_pos(trans, key, 0, true);
 423 }
 424
 425 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 426                                   struct btree_iter *iter,
 427                                   struct bkey_i *insert)
 428 {
 429         struct bch_fs *c = trans->c;
 430         struct bkey_cached *ck = (void *) iter->l[0].b;
 431
 432         BUG_ON(insert->u64s > ck->u64s);
 433
 434         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 435                 int difference;
 436
 437                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 438
 439                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 440                 if (difference > 0) {
 441                         trans->journal_preres.u64s      -= difference;
 442                         ck->res.u64s                    += difference;
 443                 }
 444         }
 445
 446         bkey_copy(ck->k, insert);
 447         ck->valid = true;
 448
 449         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 450                 mutex_lock(&c->btree_key_cache.lock);
 451                 list_del_init(&ck->list);
 452
 453                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 454                 mutex_unlock(&c->btree_key_cache.lock);
 455         }
 456
 457         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 458                                 &ck->journal, btree_key_cache_journal_flush);
 459         return true;
 460 }
 461
 462 #ifdef CONFIG_BCACHEFS_DEBUG
 463 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
 464                                enum btree_id id, struct bpos pos)
 465 {
 466         BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
 467 }
 468 #endif
 469
 470 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *c)
 471 {
 472         struct bkey_cached *ck, *n;
 473
 474         mutex_lock(&c->lock);
 475         list_for_each_entry_safe(ck, n, &c->clean, list) {
 476                 kfree(ck->k);
 477                 kfree(ck);
 478         }
 479         list_for_each_entry_safe(ck, n, &c->freed, list)
 480                 kfree(ck);
 481         mutex_unlock(&c->lock);
 482
 483         rhashtable_destroy(&c->table);
 484 }
 485
 486 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 487 {
 488         mutex_init(&c->lock);
 489         INIT_LIST_HEAD(&c->freed);
 490         INIT_LIST_HEAD(&c->clean);
 491 }
 492
 493 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
 494 {
 495         return rhashtable_init(&c->table, &bch2_btree_key_cache_params);
 496 }
 497
 498 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 499 {
 500         struct bucket_table *tbl;
 501         struct bkey_cached *ck;
 502         struct rhash_head *pos;
 503         size_t i;
 504
 505         mutex_lock(&c->lock);
 506         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 507
 508         for (i = 0; i < tbl->size; i++) {
 509                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 510                         pr_buf(out, "%s:",
 511                                bch2_btree_ids[ck->key.btree_id]);
 512                         bch2_bpos_to_text(out, ck->key.pos);
 513
 514                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 515                                 pr_buf(out, " journal seq %llu", ck->journal.seq);
 516                         pr_buf(out, "\n");
 517                 }
 518         }
 519         mutex_unlock(&c->lock);
 520 }