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 <linux/sched/mm.h>
  13 #include <trace/events/bcachefs.h>
  14
  15 static struct kmem_cache *bch2_key_cache;
  16
  17 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  18                                        const void *obj)
  19 {
  20         const struct bkey_cached *ck = obj;
  21         const struct bkey_cached_key *key = arg->key;
  22
  23         return cmp_int(ck->key.btree_id, key->btree_id) ?:
  24                 bpos_cmp(ck->key.pos, key->pos);
  25 }
  26
  27 static const struct rhashtable_params bch2_btree_key_cache_params = {
  28         .head_offset    = offsetof(struct bkey_cached, hash),
  29         .key_offset     = offsetof(struct bkey_cached, key),
  30         .key_len        = sizeof(struct bkey_cached_key),
  31         .obj_cmpfn      = bch2_btree_key_cache_cmp_fn,
  32 };
  33
  34 __flatten
  35 inline struct bkey_cached *
  36 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
  37 {
  38         struct bkey_cached_key key = {
  39                 .btree_id       = btree_id,
  40                 .pos            = pos,
  41         };
  42
  43         return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
  44                                       bch2_btree_key_cache_params);
  45 }
  46
  47 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
  48 {
  49         if (!six_trylock_intent(&ck->c.lock))
  50                 return false;
  51
  52         if (!six_trylock_write(&ck->c.lock)) {
  53                 six_unlock_intent(&ck->c.lock);
  54                 return false;
  55         }
  56
  57         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
  58                 six_unlock_write(&ck->c.lock);
  59                 six_unlock_intent(&ck->c.lock);
  60                 return false;
  61         }
  62
  63         return true;
  64 }
  65
  66 static void bkey_cached_evict(struct btree_key_cache *c,
  67                               struct bkey_cached *ck)
  68 {
  69         BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
  70                                       bch2_btree_key_cache_params));
  71         memset(&ck->key, ~0, sizeof(ck->key));
  72
  73         atomic_long_dec(&c->nr_keys);
  74 }
  75
  76 static void bkey_cached_free(struct btree_key_cache *bc,
  77                              struct bkey_cached *ck)
  78 {
  79         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
  80
  81         BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
  82
  83         ck->btree_trans_barrier_seq =
  84                 start_poll_synchronize_srcu(&c->btree_trans_barrier);
  85
  86         list_move_tail(&ck->list, &bc->freed);
  87         bc->nr_freed++;
  88
  89         kfree(ck->k);
  90         ck->k           = NULL;
  91         ck->u64s        = 0;
  92
  93         six_unlock_write(&ck->c.lock);
  94         six_unlock_intent(&ck->c.lock);
  95 }
  96
  97 static struct bkey_cached *
  98 bkey_cached_alloc(struct btree_key_cache *c)
  99 {
 100         struct bkey_cached *ck;
 101
 102         ck = kmem_cache_alloc(bch2_key_cache, GFP_NOFS|__GFP_ZERO);
 103         if (likely(ck)) {
 104                 INIT_LIST_HEAD(&ck->list);
 105                 six_lock_init(&ck->c.lock);
 106                 BUG_ON(!six_trylock_intent(&ck->c.lock));
 107                 BUG_ON(!six_trylock_write(&ck->c.lock));
 108                 return ck;
 109         }
 110
 111         return NULL;
 112 }
 113
 114 static struct bkey_cached *
 115 bkey_cached_reuse(struct btree_key_cache *c)
 116 {
 117         struct bucket_table *tbl;
 118         struct rhash_head *pos;
 119         struct bkey_cached *ck;
 120         unsigned i;
 121
 122         mutex_lock(&c->lock);
 123         list_for_each_entry_reverse(ck, &c->freed, list)
 124                 if (bkey_cached_lock_for_evict(ck)) {
 125                         c->nr_freed--;
 126                         list_del(&ck->list);
 127                         mutex_unlock(&c->lock);
 128                         return ck;
 129                 }
 130         mutex_unlock(&c->lock);
 131
 132         rcu_read_lock();
 133         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 134         for (i = 0; i < tbl->size; i++)
 135                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 136                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 137                             bkey_cached_lock_for_evict(ck)) {
 138                                 bkey_cached_evict(c, ck);
 139                                 rcu_read_unlock();
 140                                 return ck;
 141                         }
 142                 }
 143         rcu_read_unlock();
 144
 145         return NULL;
 146 }
 147
 148 static struct bkey_cached *
 149 btree_key_cache_create(struct btree_key_cache *c,
 150                        enum btree_id btree_id,
 151                        struct bpos pos)
 152 {
 153         struct bkey_cached *ck;
 154         bool was_new = true;
 155
 156         ck = bkey_cached_alloc(c);
 157
 158         if (unlikely(!ck)) {
 159                 ck = bkey_cached_reuse(c);
 160                 if (unlikely(!ck))
 161                         return ERR_PTR(-ENOMEM);
 162
 163                 was_new = false;
 164         }
 165
 166         ck->c.level             = 0;
 167         ck->c.btree_id          = btree_id;
 168         ck->key.btree_id        = btree_id;
 169         ck->key.pos             = pos;
 170         ck->valid               = false;
 171         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 172
 173         if (unlikely(rhashtable_lookup_insert_fast(&c->table,
 174                                           &ck->hash,
 175                                           bch2_btree_key_cache_params))) {
 176                 /* We raced with another fill: */
 177
 178                 if (likely(was_new)) {
 179                         six_unlock_write(&ck->c.lock);
 180                         six_unlock_intent(&ck->c.lock);
 181                         kfree(ck);
 182                 } else {
 183                         mutex_lock(&c->lock);
 184                         bkey_cached_free(c, ck);
 185                         mutex_unlock(&c->lock);
 186                 }
 187
 188                 return NULL;
 189         }
 190
 191         atomic_long_inc(&c->nr_keys);
 192
 193         six_unlock_write(&ck->c.lock);
 194
 195         return ck;
 196 }
 197
 198 static int btree_key_cache_fill(struct btree_trans *trans,
 199                                 struct btree_iter *ck_iter,
 200                                 struct bkey_cached *ck)
 201 {
 202         struct btree_iter *iter;
 203         struct bkey_s_c k;
 204         unsigned new_u64s = 0;
 205         struct bkey_i *new_k = NULL;
 206         int ret;
 207
 208         iter = bch2_trans_get_iter(trans, ck->key.btree_id,
 209                                    ck->key.pos, BTREE_ITER_SLOTS);
 210         k = bch2_btree_iter_peek_slot(iter);
 211         ret = bkey_err(k);
 212         if (ret)
 213                 goto err;
 214
 215         if (!bch2_btree_node_relock(ck_iter, 0)) {
 216                 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 217                 ret = -EINTR;
 218                 goto err;
 219         }
 220
 221         if (k.k->u64s > ck->u64s) {
 222                 new_u64s = roundup_pow_of_two(k.k->u64s);
 223                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 224                 if (!new_k) {
 225                         ret = -ENOMEM;
 226                         goto err;
 227                 }
 228         }
 229
 230         bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
 231         if (new_k) {
 232                 kfree(ck->k);
 233                 ck->u64s = new_u64s;
 234                 ck->k = new_k;
 235         }
 236
 237         bkey_reassemble(ck->k, k);
 238         ck->valid = true;
 239         bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
 240
 241         /* We're not likely to need this iterator again: */
 242         set_btree_iter_dontneed(trans, iter);
 243 err:
 244         bch2_trans_iter_put(trans, iter);
 245         return ret;
 246 }
 247
 248 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 249 {
 250         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 251         const struct btree_iter *iter = p;
 252
 253         return ck->key.btree_id == iter->btree_id &&
 254                 !bpos_cmp(ck->key.pos, iter->pos) ? 0 : -1;
 255 }
 256
 257 __flatten
 258 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
 259 {
 260         struct btree_trans *trans = iter->trans;
 261         struct bch_fs *c = trans->c;
 262         struct bkey_cached *ck;
 263         int ret = 0;
 264
 265         BUG_ON(iter->level);
 266
 267         if (btree_node_locked(iter, 0)) {
 268                 ck = (void *) iter->l[0].b;
 269                 goto fill;
 270         }
 271 retry:
 272         ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
 273         if (!ck) {
 274                 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
 275                         iter->l[0].b = NULL;
 276                         return 0;
 277                 }
 278
 279                 ck = btree_key_cache_create(&c->btree_key_cache,
 280                                             iter->btree_id, iter->pos);
 281                 ret = PTR_ERR_OR_ZERO(ck);
 282                 if (ret)
 283                         goto err;
 284                 if (!ck)
 285                         goto retry;
 286
 287                 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
 288                 iter->locks_want = 1;
 289         } else {
 290                 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
 291
 292                 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
 293                                      bkey_cached_check_fn, iter, _THIS_IP_)) {
 294                         if (ck->key.btree_id != iter->btree_id ||
 295                             bpos_cmp(ck->key.pos, iter->pos)) {
 296                                 goto retry;
 297                         }
 298
 299                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 300                         ret = -EINTR;
 301                         goto err;
 302                 }
 303
 304                 if (ck->key.btree_id != iter->btree_id ||
 305                     bpos_cmp(ck->key.pos, iter->pos)) {
 306                         six_unlock_type(&ck->c.lock, lock_want);
 307                         goto retry;
 308                 }
 309
 310                 mark_btree_node_locked(iter, 0, lock_want);
 311         }
 312
 313         iter->l[0].lock_seq     = ck->c.lock.state.seq;
 314         iter->l[0].b            = (void *) ck;
 315 fill:
 316         if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
 317                 if (!btree_node_intent_locked(iter, 0))
 318                         bch2_btree_iter_upgrade(iter, 1);
 319                 if (!btree_node_intent_locked(iter, 0)) {
 320                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 321                         ret = -EINTR;
 322                         goto err;
 323                 }
 324
 325                 ret = btree_key_cache_fill(trans, iter, ck);
 326                 if (ret)
 327                         goto err;
 328         }
 329
 330         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 331                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 332
 333         iter->uptodate = BTREE_ITER_NEED_PEEK;
 334
 335         if (!(iter->flags & BTREE_ITER_INTENT))
 336                 bch2_btree_iter_downgrade(iter);
 337         else if (!iter->locks_want) {
 338                 if (!__bch2_btree_iter_upgrade(iter, 1))
 339                         ret = -EINTR;
 340         }
 341
 342         return ret;
 343 err:
 344         if (ret != -EINTR) {
 345                 btree_node_unlock(iter, 0);
 346                 iter->flags |= BTREE_ITER_ERROR;
 347                 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
 348         }
 349         return ret;
 350 }
 351
 352 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 353                                      struct bkey_cached_key key,
 354                                      u64 journal_seq,
 355                                      unsigned commit_flags,
 356                                      bool evict)
 357 {
 358         struct bch_fs *c = trans->c;
 359         struct journal *j = &c->journal;
 360         struct btree_iter *c_iter = NULL, *b_iter = NULL;
 361         struct bkey_cached *ck = NULL;
 362         int ret;
 363
 364         b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 365                                      BTREE_ITER_SLOTS|
 366                                      BTREE_ITER_INTENT);
 367         c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 368                                      BTREE_ITER_CACHED|
 369                                      BTREE_ITER_CACHED_NOFILL|
 370                                      BTREE_ITER_CACHED_NOCREATE|
 371                                      BTREE_ITER_INTENT);
 372 retry:
 373         ret = bch2_btree_iter_traverse(c_iter);
 374         if (ret)
 375                 goto err;
 376
 377         ck = (void *) c_iter->l[0].b;
 378         if (!ck ||
 379             (journal_seq && ck->journal.seq != journal_seq))
 380                 goto out;
 381
 382         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 383                 if (!evict)
 384                         goto out;
 385                 goto evict;
 386         }
 387
 388         ret   = bch2_btree_iter_traverse(b_iter) ?:
 389                 bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
 390                 bch2_trans_commit(trans, NULL, NULL,
 391                                   BTREE_INSERT_NOUNLOCK|
 392                                   BTREE_INSERT_NOCHECK_RW|
 393                                   BTREE_INSERT_NOFAIL|
 394                                   (ck->journal.seq == journal_last_seq(j)
 395                                    ? BTREE_INSERT_JOURNAL_RESERVED
 396                                    : 0)|
 397                                   commit_flags);
 398 err:
 399         if (ret == -EINTR)
 400                 goto retry;
 401
 402         if (ret == -EAGAIN)
 403                 goto out;
 404
 405         if (ret) {
 406                 bch2_fs_fatal_err_on(!bch2_journal_error(j), c,
 407                         "error flushing key cache: %i", ret);
 408                 goto out;
 409         }
 410
 411         bch2_journal_pin_drop(j, &ck->journal);
 412         bch2_journal_preres_put(j, &ck->res);
 413
 414         BUG_ON(!btree_node_locked(c_iter, 0));
 415
 416         if (!evict) {
 417                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 418                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 419                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 420                 }
 421         } else {
 422 evict:
 423                 BUG_ON(!btree_node_intent_locked(c_iter, 0));
 424
 425                 mark_btree_node_unlocked(c_iter, 0);
 426                 c_iter->l[0].b = NULL;
 427
 428                 six_lock_write(&ck->c.lock, NULL, NULL);
 429
 430                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 431                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 432                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 433                 }
 434
 435                 bkey_cached_evict(&c->btree_key_cache, ck);
 436
 437                 mutex_lock(&c->btree_key_cache.lock);
 438                 bkey_cached_free(&c->btree_key_cache, ck);
 439                 mutex_unlock(&c->btree_key_cache.lock);
 440         }
 441 out:
 442         bch2_trans_iter_put(trans, b_iter);
 443         bch2_trans_iter_put(trans, c_iter);
 444         return ret;
 445 }
 446
 447 int bch2_btree_key_cache_journal_flush(struct journal *j,
 448                                 struct journal_entry_pin *pin, u64 seq)
 449 {
 450         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 451         struct bkey_cached *ck =
 452                 container_of(pin, struct bkey_cached, journal);
 453         struct bkey_cached_key key;
 454         struct btree_trans trans;
 455         int ret = 0;
 456
 457         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 458
 459         six_lock_read(&ck->c.lock, NULL, NULL);
 460         key = ck->key;
 461
 462         if (ck->journal.seq != seq ||
 463             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 464                 six_unlock_read(&ck->c.lock);
 465                 goto unlock;
 466         }
 467         six_unlock_read(&ck->c.lock);
 468
 469         bch2_trans_init(&trans, c, 0, 0);
 470         ret = btree_key_cache_flush_pos(&trans, key, seq,
 471                                   BTREE_INSERT_JOURNAL_RECLAIM, false);
 472         bch2_trans_exit(&trans);
 473 unlock:
 474         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 475
 476         return ret;
 477 }
 478
 479 /*
 480  * Flush and evict a key from the key cache:
 481  */
 482 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 483                                enum btree_id id, struct bpos pos)
 484 {
 485         struct bch_fs *c = trans->c;
 486         struct bkey_cached_key key = { id, pos };
 487
 488         /* Fastpath - assume it won't be found: */
 489         if (!bch2_btree_key_cache_find(c, id, pos))
 490                 return 0;
 491
 492         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 493 }
 494
 495 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 496                                   struct btree_iter *iter,
 497                                   struct bkey_i *insert)
 498 {
 499         struct bch_fs *c = trans->c;
 500         struct bkey_cached *ck = (void *) iter->l[0].b;
 501         bool kick_reclaim = false;
 502
 503         BUG_ON(insert->u64s > ck->u64s);
 504
 505         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 506                 int difference;
 507
 508                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 509
 510                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 511                 if (difference > 0) {
 512                         trans->journal_preres.u64s      -= difference;
 513                         ck->res.u64s                    += difference;
 514                 }
 515         }
 516
 517         bkey_copy(ck->k, insert);
 518         ck->valid = true;
 519
 520         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 521                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 522                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 523
 524                 if (bch2_nr_btree_keys_need_flush(c))
 525                         kick_reclaim = true;
 526         }
 527
 528         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 529                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 530
 531         if (kick_reclaim)
 532                 journal_reclaim_kick(&c->journal);
 533         return true;
 534 }
 535
 536 #ifdef CONFIG_BCACHEFS_DEBUG
 537 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
 538                                enum btree_id id, struct bpos pos)
 539 {
 540         BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
 541 }
 542 #endif
 543
 544 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 545                                            struct shrink_control *sc)
 546 {
 547         struct bch_fs *c = container_of(shrink, struct bch_fs,
 548                                         btree_key_cache.shrink);
 549         struct btree_key_cache *bc = &c->btree_key_cache;
 550         struct bucket_table *tbl;
 551         struct bkey_cached *ck, *t;
 552         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 553         unsigned start, flags;
 554         int srcu_idx;
 555
 556         /* Return -1 if we can't do anything right now */
 557         if (sc->gfp_mask & __GFP_FS)
 558                 mutex_lock(&bc->lock);
 559         else if (!mutex_trylock(&bc->lock))
 560                 return -1;
 561
 562         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 563         flags = memalloc_nofs_save();
 564
 565         /*
 566          * Newest freed entries are at the end of the list - once we hit one
 567          * that's too new to be freed, we can bail out:
 568          */
 569         list_for_each_entry_safe(ck, t, &bc->freed, list) {
 570                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 571                                                  ck->btree_trans_barrier_seq))
 572                         break;
 573
 574                 list_del(&ck->list);
 575                 kmem_cache_free(bch2_key_cache, ck);
 576                 bc->nr_freed--;
 577                 scanned++;
 578                 freed++;
 579         }
 580
 581         if (scanned >= nr)
 582                 goto out;
 583
 584         rcu_read_lock();
 585         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 586         if (bc->shrink_iter >= tbl->size)
 587                 bc->shrink_iter = 0;
 588         start = bc->shrink_iter;
 589
 590         do {
 591                 struct rhash_head *pos, *next;
 592
 593                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 594
 595                 while (!rht_is_a_nulls(pos)) {
 596                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 597                         ck = container_of(pos, struct bkey_cached, hash);
 598
 599                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 600                                 goto next;
 601
 602                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 603                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 604                         else if (bkey_cached_lock_for_evict(ck)) {
 605                                 bkey_cached_evict(bc, ck);
 606                                 bkey_cached_free(bc, ck);
 607                         }
 608
 609                         scanned++;
 610                         if (scanned >= nr)
 611                                 break;
 612 next:
 613                         pos = next;
 614                 }
 615
 616                 bc->shrink_iter++;
 617                 if (bc->shrink_iter >= tbl->size)
 618                         bc->shrink_iter = 0;
 619         } while (scanned < nr && bc->shrink_iter != start);
 620
 621         rcu_read_unlock();
 622 out:
 623         memalloc_nofs_restore(flags);
 624         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 625         mutex_unlock(&bc->lock);
 626
 627         return freed;
 628 }
 629
 630 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 631                                             struct shrink_control *sc)
 632 {
 633         struct bch_fs *c = container_of(shrink, struct bch_fs,
 634                                         btree_key_cache.shrink);
 635         struct btree_key_cache *bc = &c->btree_key_cache;
 636
 637         return atomic_long_read(&bc->nr_keys);
 638 }
 639
 640 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 641 {
 642         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 643         struct bucket_table *tbl;
 644         struct bkey_cached *ck, *n;
 645         struct rhash_head *pos;
 646         unsigned i;
 647
 648         if (bc->shrink.list.next)
 649                 unregister_shrinker(&bc->shrink);
 650
 651         mutex_lock(&bc->lock);
 652
 653         rcu_read_lock();
 654         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 655         for (i = 0; i < tbl->size; i++)
 656                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 657                         bkey_cached_evict(bc, ck);
 658                         list_add(&ck->list, &bc->freed);
 659                 }
 660         rcu_read_unlock();
 661
 662         list_for_each_entry_safe(ck, n, &bc->freed, list) {
 663                 cond_resched();
 664
 665                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 666                 bch2_journal_preres_put(&c->journal, &ck->res);
 667
 668                 list_del(&ck->list);
 669                 kfree(ck->k);
 670                 kmem_cache_free(bch2_key_cache, ck);
 671         }
 672
 673         BUG_ON(atomic_long_read(&bc->nr_dirty) && !bch2_journal_error(&c->journal));
 674         BUG_ON(atomic_long_read(&bc->nr_keys));
 675
 676         mutex_unlock(&bc->lock);
 677
 678         if (bc->table_init_done)
 679                 rhashtable_destroy(&bc->table);
 680 }
 681
 682 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 683 {
 684         mutex_init(&c->lock);
 685         INIT_LIST_HEAD(&c->freed);
 686 }
 687
 688 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
 689 {
 690         int ret;
 691
 692         ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
 693         if (ret)
 694                 return ret;
 695
 696         c->table_init_done = true;
 697
 698         c->shrink.seeks                 = 1;
 699         c->shrink.count_objects         = bch2_btree_key_cache_count;
 700         c->shrink.scan_objects          = bch2_btree_key_cache_scan;
 701         return register_shrinker(&c->shrink);
 702 }
 703
 704 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 705 {
 706         pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
 707         pr_buf(out, "nr_keys:\t%zu\n",  atomic_long_read(&c->nr_keys));
 708         pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
 709 }
 710
 711 void bch2_btree_key_cache_exit(void)
 712 {
 713         if (bch2_key_cache)
 714                 kmem_cache_destroy(bch2_key_cache);
 715 }
 716
 717 int __init bch2_btree_key_cache_init(void)
 718 {
 719         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
 720         if (!bch2_key_cache)
 721                 return -ENOMEM;
 722
 723         return 0;
 724 }