fs/bcachefs/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 #include "trace.h"
  12
  13 #include <linux/sched/mm.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                 lockdep_set_novalidate_class(&ck->c.lock);
 107                 BUG_ON(!six_trylock_intent(&ck->c.lock));
 108                 BUG_ON(!six_trylock_write(&ck->c.lock));
 109                 return ck;
 110         }
 111
 112         return NULL;
 113 }
 114
 115 static struct bkey_cached *
 116 bkey_cached_reuse(struct btree_key_cache *c)
 117 {
 118         struct bucket_table *tbl;
 119         struct rhash_head *pos;
 120         struct bkey_cached *ck;
 121         unsigned i;
 122
 123         mutex_lock(&c->lock);
 124         list_for_each_entry_reverse(ck, &c->freed, list)
 125                 if (bkey_cached_lock_for_evict(ck)) {
 126                         c->nr_freed--;
 127                         list_del(&ck->list);
 128                         mutex_unlock(&c->lock);
 129                         return ck;
 130                 }
 131         mutex_unlock(&c->lock);
 132
 133         rcu_read_lock();
 134         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 135         for (i = 0; i < tbl->size; i++)
 136                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 137                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 138                             bkey_cached_lock_for_evict(ck)) {
 139                                 bkey_cached_evict(c, ck);
 140                                 rcu_read_unlock();
 141                                 return ck;
 142                         }
 143                 }
 144         rcu_read_unlock();
 145
 146         return NULL;
 147 }
 148
 149 static struct bkey_cached *
 150 btree_key_cache_create(struct btree_key_cache *c,
 151                        enum btree_id btree_id,
 152                        struct bpos pos)
 153 {
 154         struct bkey_cached *ck;
 155         bool was_new = true;
 156
 157         ck = bkey_cached_alloc(c);
 158
 159         if (unlikely(!ck)) {
 160                 ck = bkey_cached_reuse(c);
 161                 if (unlikely(!ck))
 162                         return ERR_PTR(-ENOMEM);
 163
 164                 was_new = false;
 165         }
 166
 167         ck->c.level             = 0;
 168         ck->c.btree_id          = btree_id;
 169         ck->key.btree_id        = btree_id;
 170         ck->key.pos             = pos;
 171         ck->valid               = false;
 172         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 173
 174         if (unlikely(rhashtable_lookup_insert_fast(&c->table,
 175                                           &ck->hash,
 176                                           bch2_btree_key_cache_params))) {
 177                 /* We raced with another fill: */
 178
 179                 if (likely(was_new)) {
 180                         six_unlock_write(&ck->c.lock);
 181                         six_unlock_intent(&ck->c.lock);
 182                         kfree(ck);
 183                 } else {
 184                         mutex_lock(&c->lock);
 185                         bkey_cached_free(c, ck);
 186                         mutex_unlock(&c->lock);
 187                 }
 188
 189                 return NULL;
 190         }
 191
 192         atomic_long_inc(&c->nr_keys);
 193
 194         six_unlock_write(&ck->c.lock);
 195
 196         return ck;
 197 }
 198
 199 static int btree_key_cache_fill(struct btree_trans *trans,
 200                                 struct btree_iter *ck_iter,
 201                                 struct bkey_cached *ck)
 202 {
 203         struct btree_iter *iter;
 204         struct bkey_s_c k;
 205         unsigned new_u64s = 0;
 206         struct bkey_i *new_k = NULL;
 207         int ret;
 208
 209         iter = bch2_trans_get_iter(trans, ck->key.btree_id,
 210                                    ck->key.pos, BTREE_ITER_SLOTS);
 211         k = bch2_btree_iter_peek_slot(iter);
 212         ret = bkey_err(k);
 213         if (ret)
 214                 goto err;
 215
 216         if (!bch2_btree_node_relock(ck_iter, 0)) {
 217                 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 218                 ret = -EINTR;
 219                 goto err;
 220         }
 221
 222         if (k.k->u64s > ck->u64s) {
 223                 new_u64s = roundup_pow_of_two(k.k->u64s);
 224                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 225                 if (!new_k) {
 226                         ret = -ENOMEM;
 227                         goto err;
 228                 }
 229         }
 230
 231         bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
 232         if (new_k) {
 233                 kfree(ck->k);
 234                 ck->u64s = new_u64s;
 235                 ck->k = new_k;
 236         }
 237
 238         bkey_reassemble(ck->k, k);
 239         ck->valid = true;
 240         bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
 241
 242         /* We're not likely to need this iterator again: */
 243         set_btree_iter_dontneed(trans, iter);
 244 err:
 245         bch2_trans_iter_put(trans, iter);
 246         return ret;
 247 }
 248
 249 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 250 {
 251         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 252         const struct btree_iter *iter = p;
 253
 254         return ck->key.btree_id == iter->btree_id &&
 255                 !bpos_cmp(ck->key.pos, iter->pos) ? 0 : -1;
 256 }
 257
 258 __flatten
 259 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
 260 {
 261         struct btree_trans *trans = iter->trans;
 262         struct bch_fs *c = trans->c;
 263         struct bkey_cached *ck;
 264         int ret = 0;
 265
 266         BUG_ON(iter->level);
 267
 268         if (btree_node_locked(iter, 0)) {
 269                 ck = (void *) iter->l[0].b;
 270                 goto fill;
 271         }
 272 retry:
 273         ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
 274         if (!ck) {
 275                 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
 276                         iter->l[0].b = NULL;
 277                         return 0;
 278                 }
 279
 280                 ck = btree_key_cache_create(&c->btree_key_cache,
 281                                             iter->btree_id, iter->pos);
 282                 ret = PTR_ERR_OR_ZERO(ck);
 283                 if (ret)
 284                         goto err;
 285                 if (!ck)
 286                         goto retry;
 287
 288                 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
 289                 iter->locks_want = 1;
 290         } else {
 291                 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
 292
 293                 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
 294                                      bkey_cached_check_fn, iter, _THIS_IP_)) {
 295                         if (ck->key.btree_id != iter->btree_id ||
 296                             bpos_cmp(ck->key.pos, iter->pos)) {
 297                                 goto retry;
 298                         }
 299
 300                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 301                         ret = -EINTR;
 302                         goto err;
 303                 }
 304
 305                 if (ck->key.btree_id != iter->btree_id ||
 306                     bpos_cmp(ck->key.pos, iter->pos)) {
 307                         six_unlock_type(&ck->c.lock, lock_want);
 308                         goto retry;
 309                 }
 310
 311                 mark_btree_node_locked(iter, 0, lock_want);
 312         }
 313
 314         iter->l[0].lock_seq     = ck->c.lock.state.seq;
 315         iter->l[0].b            = (void *) ck;
 316 fill:
 317         if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
 318                 if (!btree_node_intent_locked(iter, 0))
 319                         bch2_btree_iter_upgrade(iter, 1);
 320                 if (!btree_node_intent_locked(iter, 0)) {
 321                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 322                         ret = -EINTR;
 323                         goto err;
 324                 }
 325
 326                 ret = btree_key_cache_fill(trans, iter, ck);
 327                 if (ret)
 328                         goto err;
 329         }
 330
 331         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 332                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 333
 334         iter->uptodate = BTREE_ITER_NEED_PEEK;
 335
 336         if (!(iter->flags & BTREE_ITER_INTENT))
 337                 bch2_btree_iter_downgrade(iter);
 338         else if (!iter->locks_want) {
 339                 if (!__bch2_btree_iter_upgrade(iter, 1))
 340                         ret = -EINTR;
 341         }
 342
 343         return ret;
 344 err:
 345         if (ret != -EINTR) {
 346                 btree_node_unlock(iter, 0);
 347                 iter->flags |= BTREE_ITER_ERROR;
 348                 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
 349         }
 350         return ret;
 351 }
 352
 353 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 354                                      struct bkey_cached_key key,
 355                                      u64 journal_seq,
 356                                      unsigned commit_flags,
 357                                      bool evict)
 358 {
 359         struct bch_fs *c = trans->c;
 360         struct journal *j = &c->journal;
 361         struct btree_iter *c_iter = NULL, *b_iter = NULL;
 362         struct bkey_cached *ck = NULL;
 363         int ret;
 364
 365         b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 366                                      BTREE_ITER_SLOTS|
 367                                      BTREE_ITER_INTENT);
 368         c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 369                                      BTREE_ITER_CACHED|
 370                                      BTREE_ITER_CACHED_NOFILL|
 371                                      BTREE_ITER_CACHED_NOCREATE|
 372                                      BTREE_ITER_INTENT);
 373 retry:
 374         ret = bch2_btree_iter_traverse(c_iter);
 375         if (ret)
 376                 goto err;
 377
 378         ck = (void *) c_iter->l[0].b;
 379         if (!ck ||
 380             (journal_seq && ck->journal.seq != journal_seq))
 381                 goto out;
 382
 383         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 384                 if (!evict)
 385                         goto out;
 386                 goto evict;
 387         }
 388
 389         ret   = bch2_btree_iter_traverse(b_iter) ?:
 390                 bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
 391                 bch2_trans_commit(trans, NULL, NULL,
 392                                   BTREE_INSERT_NOUNLOCK|
 393                                   BTREE_INSERT_NOCHECK_RW|
 394                                   BTREE_INSERT_NOFAIL|
 395                                   (ck->journal.seq == journal_last_seq(j)
 396                                    ? BTREE_INSERT_JOURNAL_RESERVED
 397                                    : 0)|
 398                                   commit_flags);
 399 err:
 400         if (ret == -EINTR)
 401                 goto retry;
 402
 403         if (ret == -EAGAIN)
 404                 goto out;
 405
 406         if (ret) {
 407                 bch2_fs_fatal_err_on(!bch2_journal_error(j), c,
 408                         "error flushing key cache: %i", ret);
 409                 goto out;
 410         }
 411
 412         bch2_journal_pin_drop(j, &ck->journal);
 413         bch2_journal_preres_put(j, &ck->res);
 414
 415         BUG_ON(!btree_node_locked(c_iter, 0));
 416
 417         if (!evict) {
 418                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 419                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 420                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 421                 }
 422         } else {
 423 evict:
 424                 BUG_ON(!btree_node_intent_locked(c_iter, 0));
 425
 426                 mark_btree_node_unlocked(c_iter, 0);
 427                 c_iter->l[0].b = NULL;
 428
 429                 six_lock_write(&ck->c.lock, NULL, NULL);
 430
 431                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 432                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 433                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 434                 }
 435
 436                 bkey_cached_evict(&c->btree_key_cache, ck);
 437
 438                 mutex_lock(&c->btree_key_cache.lock);
 439                 bkey_cached_free(&c->btree_key_cache, ck);
 440                 mutex_unlock(&c->btree_key_cache.lock);
 441         }
 442 out:
 443         bch2_trans_iter_put(trans, b_iter);
 444         bch2_trans_iter_put(trans, c_iter);
 445         return ret;
 446 }
 447
 448 int bch2_btree_key_cache_journal_flush(struct journal *j,
 449                                 struct journal_entry_pin *pin, u64 seq)
 450 {
 451         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 452         struct bkey_cached *ck =
 453                 container_of(pin, struct bkey_cached, journal);
 454         struct bkey_cached_key key;
 455         struct btree_trans trans;
 456         int ret = 0;
 457
 458         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 459
 460         six_lock_read(&ck->c.lock, NULL, NULL);
 461         key = ck->key;
 462
 463         if (ck->journal.seq != seq ||
 464             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 465                 six_unlock_read(&ck->c.lock);
 466                 goto unlock;
 467         }
 468         six_unlock_read(&ck->c.lock);
 469
 470         bch2_trans_init(&trans, c, 0, 0);
 471         ret = btree_key_cache_flush_pos(&trans, key, seq,
 472                                   BTREE_INSERT_JOURNAL_RECLAIM, false);
 473         bch2_trans_exit(&trans);
 474 unlock:
 475         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 476
 477         return ret;
 478 }
 479
 480 /*
 481  * Flush and evict a key from the key cache:
 482  */
 483 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 484                                enum btree_id id, struct bpos pos)
 485 {
 486         struct bch_fs *c = trans->c;
 487         struct bkey_cached_key key = { id, pos };
 488
 489         /* Fastpath - assume it won't be found: */
 490         if (!bch2_btree_key_cache_find(c, id, pos))
 491                 return 0;
 492
 493         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 494 }
 495
 496 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 497                                   struct btree_iter *iter,
 498                                   struct bkey_i *insert)
 499 {
 500         struct bch_fs *c = trans->c;
 501         struct bkey_cached *ck = (void *) iter->l[0].b;
 502         bool kick_reclaim = false;
 503
 504         BUG_ON(insert->u64s > ck->u64s);
 505
 506         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 507                 int difference;
 508
 509                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 510
 511                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 512                 if (difference > 0) {
 513                         trans->journal_preres.u64s      -= difference;
 514                         ck->res.u64s                    += difference;
 515                 }
 516         }
 517
 518         bkey_copy(ck->k, insert);
 519         ck->valid = true;
 520
 521         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 522                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 523                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 524
 525                 if (bch2_nr_btree_keys_need_flush(c))
 526                         kick_reclaim = true;
 527         }
 528
 529         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 530                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 531
 532         if (kick_reclaim)
 533                 journal_reclaim_kick(&c->journal);
 534         return true;
 535 }
 536
 537 #ifdef CONFIG_BCACHEFS_DEBUG
 538 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
 539                                enum btree_id id, struct bpos pos)
 540 {
 541         BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
 542 }
 543 #endif
 544
 545 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 546                                            struct shrink_control *sc)
 547 {
 548         struct bch_fs *c = container_of(shrink, struct bch_fs,
 549                                         btree_key_cache.shrink);
 550         struct btree_key_cache *bc = &c->btree_key_cache;
 551         struct bucket_table *tbl;
 552         struct bkey_cached *ck, *t;
 553         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 554         unsigned start, flags;
 555         int srcu_idx;
 556
 557         /* Return -1 if we can't do anything right now */
 558         if (sc->gfp_mask & __GFP_FS)
 559                 mutex_lock(&bc->lock);
 560         else if (!mutex_trylock(&bc->lock))
 561                 return -1;
 562
 563         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 564         flags = memalloc_nofs_save();
 565
 566         /*
 567          * Newest freed entries are at the end of the list - once we hit one
 568          * that's too new to be freed, we can bail out:
 569          */
 570         list_for_each_entry_safe(ck, t, &bc->freed, list) {
 571                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 572                                                  ck->btree_trans_barrier_seq))
 573                         break;
 574
 575                 list_del(&ck->list);
 576                 kmem_cache_free(bch2_key_cache, ck);
 577                 bc->nr_freed--;
 578                 scanned++;
 579                 freed++;
 580         }
 581
 582         if (scanned >= nr)
 583                 goto out;
 584
 585         rcu_read_lock();
 586         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 587         if (bc->shrink_iter >= tbl->size)
 588                 bc->shrink_iter = 0;
 589         start = bc->shrink_iter;
 590
 591         do {
 592                 struct rhash_head *pos, *next;
 593
 594                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 595
 596                 while (!rht_is_a_nulls(pos)) {
 597                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 598                         ck = container_of(pos, struct bkey_cached, hash);
 599
 600                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 601                                 goto next;
 602
 603                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 604                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 605                         else if (bkey_cached_lock_for_evict(ck)) {
 606                                 bkey_cached_evict(bc, ck);
 607                                 bkey_cached_free(bc, ck);
 608                         }
 609
 610                         scanned++;
 611                         if (scanned >= nr)
 612                                 break;
 613 next:
 614                         pos = next;
 615                 }
 616
 617                 bc->shrink_iter++;
 618                 if (bc->shrink_iter >= tbl->size)
 619                         bc->shrink_iter = 0;
 620         } while (scanned < nr && bc->shrink_iter != start);
 621
 622         rcu_read_unlock();
 623 out:
 624         memalloc_nofs_restore(flags);
 625         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 626         mutex_unlock(&bc->lock);
 627
 628         return freed;
 629 }
 630
 631 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 632                                             struct shrink_control *sc)
 633 {
 634         struct bch_fs *c = container_of(shrink, struct bch_fs,
 635                                         btree_key_cache.shrink);
 636         struct btree_key_cache *bc = &c->btree_key_cache;
 637
 638         return atomic_long_read(&bc->nr_keys);
 639 }
 640
 641 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 642 {
 643         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 644         struct bucket_table *tbl;
 645         struct bkey_cached *ck, *n;
 646         struct rhash_head *pos;
 647         unsigned i;
 648
 649         if (bc->shrink.list.next)
 650                 unregister_shrinker(&bc->shrink);
 651
 652         mutex_lock(&bc->lock);
 653
 654         rcu_read_lock();
 655         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 656         for (i = 0; i < tbl->size; i++)
 657                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 658                         bkey_cached_evict(bc, ck);
 659                         list_add(&ck->list, &bc->freed);
 660                 }
 661         rcu_read_unlock();
 662
 663         list_for_each_entry_safe(ck, n, &bc->freed, list) {
 664                 cond_resched();
 665
 666                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 667                 bch2_journal_preres_put(&c->journal, &ck->res);
 668
 669                 list_del(&ck->list);
 670                 kfree(ck->k);
 671                 kmem_cache_free(bch2_key_cache, ck);
 672         }
 673
 674         BUG_ON(atomic_long_read(&bc->nr_dirty) && !bch2_journal_error(&c->journal));
 675         BUG_ON(atomic_long_read(&bc->nr_keys));
 676
 677         mutex_unlock(&bc->lock);
 678
 679         if (bc->table_init_done)
 680                 rhashtable_destroy(&bc->table);
 681 }
 682
 683 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 684 {
 685         mutex_init(&c->lock);
 686         INIT_LIST_HEAD(&c->freed);
 687 }
 688
 689 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
 690 {
 691         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 692         int ret;
 693
 694         ret = rhashtable_init(&bc->table, &bch2_btree_key_cache_params);
 695         if (ret)
 696                 return ret;
 697
 698         bc->table_init_done = true;
 699
 700         bc->shrink.seeks                = 1;
 701         bc->shrink.count_objects        = bch2_btree_key_cache_count;
 702         bc->shrink.scan_objects         = bch2_btree_key_cache_scan;
 703         return register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name);
 704 }
 705
 706 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 707 {
 708         pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
 709         pr_buf(out, "nr_keys:\t%zu\n",  atomic_long_read(&c->nr_keys));
 710         pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
 711 }
 712
 713 void bch2_btree_key_cache_exit(void)
 714 {
 715         if (bch2_key_cache)
 716                 kmem_cache_destroy(bch2_key_cache);
 717 }
 718
 719 int __init bch2_btree_key_cache_init(void)
 720 {
 721         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
 722         if (!bch2_key_cache)
 723                 return -ENOMEM;
 724
 725         return 0;
 726 }