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         if (btree_id == BTREE_ID_subvolumes)
 167                 six_lock_pcpu_alloc(&ck->c.lock);
 168         else
 169                 six_lock_pcpu_free(&ck->c.lock);
 170
 171         ck->c.level             = 0;
 172         ck->c.btree_id          = btree_id;
 173         ck->key.btree_id        = btree_id;
 174         ck->key.pos             = pos;
 175         ck->valid               = false;
 176         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 177
 178         if (unlikely(rhashtable_lookup_insert_fast(&c->table,
 179                                           &ck->hash,
 180                                           bch2_btree_key_cache_params))) {
 181                 /* We raced with another fill: */
 182
 183                 if (likely(was_new)) {
 184                         six_unlock_write(&ck->c.lock);
 185                         six_unlock_intent(&ck->c.lock);
 186                         kfree(ck);
 187                 } else {
 188                         mutex_lock(&c->lock);
 189                         bkey_cached_free(c, ck);
 190                         mutex_unlock(&c->lock);
 191                 }
 192
 193                 return NULL;
 194         }
 195
 196         atomic_long_inc(&c->nr_keys);
 197
 198         six_unlock_write(&ck->c.lock);
 199
 200         return ck;
 201 }
 202
 203 static int btree_key_cache_fill(struct btree_trans *trans,
 204                                 struct btree_path *ck_path,
 205                                 struct bkey_cached *ck)
 206 {
 207         struct btree_iter iter;
 208         struct bkey_s_c k;
 209         unsigned new_u64s = 0;
 210         struct bkey_i *new_k = NULL;
 211         int ret;
 212
 213         bch2_trans_iter_init(trans, &iter, ck->key.btree_id,
 214                              ck->key.pos, BTREE_ITER_SLOTS);
 215         k = bch2_btree_iter_peek_slot(&iter);
 216         ret = bkey_err(k);
 217         if (ret)
 218                 goto err;
 219
 220         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 221                 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 222                 ret = btree_trans_restart(trans);
 223                 goto err;
 224         }
 225
 226         /*
 227          * bch2_varint_decode can read past the end of the buffer by at
 228          * most 7 bytes (it won't be used):
 229          */
 230         new_u64s = k.k->u64s + 1;
 231
 232         if (new_u64s > ck->u64s) {
 233                 new_u64s = roundup_pow_of_two(new_u64s);
 234                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 235                 if (!new_k) {
 236                         ret = -ENOMEM;
 237                         goto err;
 238                 }
 239         }
 240
 241         /*
 242          * XXX: not allowed to be holding read locks when we take a write lock,
 243          * currently
 244          */
 245         bch2_btree_node_lock_write(trans, ck_path, ck_path->l[0].b);
 246         if (new_k) {
 247                 kfree(ck->k);
 248                 ck->u64s = new_u64s;
 249                 ck->k = new_k;
 250         }
 251
 252         bkey_reassemble(ck->k, k);
 253         ck->valid = true;
 254         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 255
 256         /* We're not likely to need this iterator again: */
 257         set_btree_iter_dontneed(&iter);
 258 err:
 259         bch2_trans_iter_exit(trans, &iter);
 260         return ret;
 261 }
 262
 263 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 264 {
 265         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 266         const struct btree_path *path = p;
 267
 268         return ck->key.btree_id == path->btree_id &&
 269                 !bpos_cmp(ck->key.pos, path->pos) ? 0 : -1;
 270 }
 271
 272 __flatten
 273 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 274                                     unsigned flags)
 275 {
 276         struct bch_fs *c = trans->c;
 277         struct bkey_cached *ck;
 278         int ret = 0;
 279
 280         BUG_ON(path->level);
 281
 282         path->l[1].b = NULL;
 283
 284         if (bch2_btree_node_relock(trans, path, 0)) {
 285                 ck = (void *) path->l[0].b;
 286                 goto fill;
 287         }
 288 retry:
 289         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 290         if (!ck) {
 291                 if (flags & BTREE_ITER_CACHED_NOCREATE) {
 292                         path->l[0].b = NULL;
 293                         return 0;
 294                 }
 295
 296                 ck = btree_key_cache_create(&c->btree_key_cache,
 297                                             path->btree_id, path->pos);
 298                 ret = PTR_ERR_OR_ZERO(ck);
 299                 if (ret)
 300                         goto err;
 301                 if (!ck)
 302                         goto retry;
 303
 304                 mark_btree_node_locked(path, 0, SIX_LOCK_intent);
 305                 path->locks_want = 1;
 306         } else {
 307                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 308
 309                 if (!btree_node_lock(trans, path, (void *) ck, path->pos, 0,
 310                                      lock_want,
 311                                      bkey_cached_check_fn, path, _THIS_IP_)) {
 312                         if (!trans->restarted)
 313                                 goto retry;
 314
 315                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 316                         ret = -EINTR;
 317                         goto err;
 318                 }
 319
 320                 if (ck->key.btree_id != path->btree_id ||
 321                     bpos_cmp(ck->key.pos, path->pos)) {
 322                         six_unlock_type(&ck->c.lock, lock_want);
 323                         goto retry;
 324                 }
 325
 326                 mark_btree_node_locked(path, 0, lock_want);
 327         }
 328
 329         path->l[0].lock_seq     = ck->c.lock.state.seq;
 330         path->l[0].b            = (void *) ck;
 331 fill:
 332         if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
 333                 if (!path->locks_want &&
 334                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 335                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 336                         ret = btree_trans_restart(trans);
 337                         goto err;
 338                 }
 339
 340                 ret = btree_key_cache_fill(trans, path, ck);
 341                 if (ret)
 342                         goto err;
 343         }
 344
 345         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 346                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 347
 348         path->uptodate = BTREE_ITER_UPTODATE;
 349         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 350
 351         return ret;
 352 err:
 353         if (ret != -EINTR) {
 354                 btree_node_unlock(path, 0);
 355                 path->l[0].b = BTREE_ITER_NO_NODE_ERROR;
 356         }
 357         return ret;
 358 }
 359
 360 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 361                                      struct bkey_cached_key key,
 362                                      u64 journal_seq,
 363                                      unsigned commit_flags,
 364                                      bool evict)
 365 {
 366         struct bch_fs *c = trans->c;
 367         struct journal *j = &c->journal;
 368         struct btree_iter c_iter, b_iter;
 369         struct bkey_cached *ck = NULL;
 370         int ret;
 371
 372         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 373                              BTREE_ITER_SLOTS|
 374                              BTREE_ITER_INTENT|
 375                              BTREE_ITER_ALL_SNAPSHOTS);
 376         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 377                              BTREE_ITER_CACHED|
 378                              BTREE_ITER_CACHED_NOFILL|
 379                              BTREE_ITER_CACHED_NOCREATE|
 380                              BTREE_ITER_INTENT);
 381         ret = bch2_btree_iter_traverse(&c_iter);
 382         if (ret)
 383                 goto out;
 384
 385         ck = (void *) c_iter.path->l[0].b;
 386         if (!ck ||
 387             (journal_seq && ck->journal.seq != journal_seq))
 388                 goto out;
 389
 390         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 391                 if (!evict)
 392                         goto out;
 393                 goto evict;
 394         }
 395
 396         /*
 397          * Since journal reclaim depends on us making progress here, and the
 398          * allocator/copygc depend on journal reclaim making progress, we need
 399          * to be using alloc reserves:
 400          * */
 401         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 402                 bch2_trans_update(trans, &b_iter, ck->k,
 403                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 404                                   BTREE_TRIGGER_NORUN) ?:
 405                 bch2_trans_commit(trans, NULL, NULL,
 406                                   BTREE_INSERT_NOCHECK_RW|
 407                                   BTREE_INSERT_NOFAIL|
 408                                   BTREE_INSERT_USE_RESERVE|
 409                                   (ck->journal.seq == journal_last_seq(j)
 410                                    ? BTREE_INSERT_JOURNAL_RESERVED
 411                                    : 0)|
 412                                   commit_flags);
 413         if (ret) {
 414                 bch2_fs_fatal_err_on(ret != -EINTR &&
 415                                      ret != -EAGAIN &&
 416                                      !bch2_journal_error(j), c,
 417                         "error flushing key cache: %i", ret);
 418                 goto out;
 419         }
 420
 421         bch2_journal_pin_drop(j, &ck->journal);
 422         bch2_journal_preres_put(j, &ck->res);
 423
 424         BUG_ON(!btree_node_locked(c_iter.path, 0));
 425
 426         if (!evict) {
 427                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 428                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 429                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 430                 }
 431         } else {
 432 evict:
 433                 BUG_ON(!btree_node_intent_locked(c_iter.path, 0));
 434
 435                 mark_btree_node_unlocked(c_iter.path, 0);
 436                 c_iter.path->l[0].b = NULL;
 437
 438                 six_lock_write(&ck->c.lock, NULL, NULL);
 439
 440                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 441                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 442                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 443                 }
 444
 445                 bkey_cached_evict(&c->btree_key_cache, ck);
 446
 447                 mutex_lock(&c->btree_key_cache.lock);
 448                 bkey_cached_free(&c->btree_key_cache, ck);
 449                 mutex_unlock(&c->btree_key_cache.lock);
 450         }
 451 out:
 452         bch2_trans_iter_exit(trans, &b_iter);
 453         bch2_trans_iter_exit(trans, &c_iter);
 454         return ret;
 455 }
 456
 457 int bch2_btree_key_cache_journal_flush(struct journal *j,
 458                                 struct journal_entry_pin *pin, u64 seq)
 459 {
 460         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 461         struct bkey_cached *ck =
 462                 container_of(pin, struct bkey_cached, journal);
 463         struct bkey_cached_key key;
 464         int ret = 0;
 465
 466         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 467
 468         six_lock_read(&ck->c.lock, NULL, NULL);
 469         key = ck->key;
 470
 471         if (ck->journal.seq != seq ||
 472             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 473                 six_unlock_read(&ck->c.lock);
 474                 goto unlock;
 475         }
 476         six_unlock_read(&ck->c.lock);
 477
 478         ret = bch2_trans_do(c, NULL, NULL, 0,
 479                 btree_key_cache_flush_pos(&trans, key, seq,
 480                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 481 unlock:
 482         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 483
 484         return ret;
 485 }
 486
 487 /*
 488  * Flush and evict a key from the key cache:
 489  */
 490 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 491                                enum btree_id id, struct bpos pos)
 492 {
 493         struct bch_fs *c = trans->c;
 494         struct bkey_cached_key key = { id, pos };
 495
 496         /* Fastpath - assume it won't be found: */
 497         if (!bch2_btree_key_cache_find(c, id, pos))
 498                 return 0;
 499
 500         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 501 }
 502
 503 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 504                                   struct btree_path *path,
 505                                   struct bkey_i *insert)
 506 {
 507         struct bch_fs *c = trans->c;
 508         struct bkey_cached *ck = (void *) path->l[0].b;
 509         bool kick_reclaim = false;
 510
 511         BUG_ON(insert->u64s > ck->u64s);
 512
 513         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 514                 int difference;
 515
 516                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 517
 518                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 519                 if (difference > 0) {
 520                         trans->journal_preres.u64s      -= difference;
 521                         ck->res.u64s                    += difference;
 522                 }
 523         }
 524
 525         bkey_copy(ck->k, insert);
 526         ck->valid = true;
 527
 528         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 529                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 530                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 531
 532                 if (bch2_nr_btree_keys_need_flush(c))
 533                         kick_reclaim = true;
 534         }
 535
 536         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 537                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 538
 539         if (kick_reclaim)
 540                 journal_reclaim_kick(&c->journal);
 541         return true;
 542 }
 543
 544 #ifdef CONFIG_BCACHEFS_DEBUG
 545 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
 546                                enum btree_id id, struct bpos pos)
 547 {
 548         BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
 549 }
 550 #endif
 551
 552 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 553                                            struct shrink_control *sc)
 554 {
 555         struct bch_fs *c = container_of(shrink, struct bch_fs,
 556                                         btree_key_cache.shrink);
 557         struct btree_key_cache *bc = &c->btree_key_cache;
 558         struct bucket_table *tbl;
 559         struct bkey_cached *ck, *t;
 560         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 561         unsigned start, flags;
 562         int srcu_idx;
 563
 564         /* Return -1 if we can't do anything right now */
 565         if (sc->gfp_mask & __GFP_FS)
 566                 mutex_lock(&bc->lock);
 567         else if (!mutex_trylock(&bc->lock))
 568                 return -1;
 569
 570         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 571         flags = memalloc_nofs_save();
 572
 573         /*
 574          * Newest freed entries are at the end of the list - once we hit one
 575          * that's too new to be freed, we can bail out:
 576          */
 577         list_for_each_entry_safe(ck, t, &bc->freed, list) {
 578                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 579                                                  ck->btree_trans_barrier_seq))
 580                         break;
 581
 582                 list_del(&ck->list);
 583                 kmem_cache_free(bch2_key_cache, ck);
 584                 bc->nr_freed--;
 585                 scanned++;
 586                 freed++;
 587         }
 588
 589         if (scanned >= nr)
 590                 goto out;
 591
 592         rcu_read_lock();
 593         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 594         if (bc->shrink_iter >= tbl->size)
 595                 bc->shrink_iter = 0;
 596         start = bc->shrink_iter;
 597
 598         do {
 599                 struct rhash_head *pos, *next;
 600
 601                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 602
 603                 while (!rht_is_a_nulls(pos)) {
 604                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 605                         ck = container_of(pos, struct bkey_cached, hash);
 606
 607                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 608                                 goto next;
 609
 610                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 611                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 612                         else if (bkey_cached_lock_for_evict(ck)) {
 613                                 bkey_cached_evict(bc, ck);
 614                                 bkey_cached_free(bc, ck);
 615                         }
 616
 617                         scanned++;
 618                         if (scanned >= nr)
 619                                 break;
 620 next:
 621                         pos = next;
 622                 }
 623
 624                 bc->shrink_iter++;
 625                 if (bc->shrink_iter >= tbl->size)
 626                         bc->shrink_iter = 0;
 627         } while (scanned < nr && bc->shrink_iter != start);
 628
 629         rcu_read_unlock();
 630 out:
 631         memalloc_nofs_restore(flags);
 632         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 633         mutex_unlock(&bc->lock);
 634
 635         return freed;
 636 }
 637
 638 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 639                                             struct shrink_control *sc)
 640 {
 641         struct bch_fs *c = container_of(shrink, struct bch_fs,
 642                                         btree_key_cache.shrink);
 643         struct btree_key_cache *bc = &c->btree_key_cache;
 644         long nr = atomic_long_read(&bc->nr_keys) -
 645                 atomic_long_read(&bc->nr_dirty);
 646
 647         return max(0L, nr);
 648 }
 649
 650 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 651 {
 652         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 653         struct bucket_table *tbl;
 654         struct bkey_cached *ck, *n;
 655         struct rhash_head *pos;
 656         unsigned i;
 657
 658         if (bc->shrink.list.next)
 659                 unregister_shrinker(&bc->shrink);
 660
 661         mutex_lock(&bc->lock);
 662
 663         rcu_read_lock();
 664         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 665         if (tbl)
 666                 for (i = 0; i < tbl->size; i++)
 667                         rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 668                                 bkey_cached_evict(bc, ck);
 669                                 list_add(&ck->list, &bc->freed);
 670                         }
 671         rcu_read_unlock();
 672
 673         list_for_each_entry_safe(ck, n, &bc->freed, list) {
 674                 cond_resched();
 675
 676                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 677                 bch2_journal_preres_put(&c->journal, &ck->res);
 678
 679                 list_del(&ck->list);
 680                 kfree(ck->k);
 681                 kmem_cache_free(bch2_key_cache, ck);
 682         }
 683
 684         BUG_ON(atomic_long_read(&bc->nr_dirty) &&
 685                !bch2_journal_error(&c->journal) &&
 686                test_bit(BCH_FS_WAS_RW, &c->flags));
 687         BUG_ON(atomic_long_read(&bc->nr_keys));
 688
 689         mutex_unlock(&bc->lock);
 690
 691         if (bc->table_init_done)
 692                 rhashtable_destroy(&bc->table);
 693 }
 694
 695 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 696 {
 697         mutex_init(&c->lock);
 698         INIT_LIST_HEAD(&c->freed);
 699 }
 700
 701 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
 702 {
 703         int ret;
 704
 705         ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
 706         if (ret)
 707                 return ret;
 708
 709         c->table_init_done = true;
 710
 711         c->shrink.seeks                 = 1;
 712         c->shrink.count_objects         = bch2_btree_key_cache_count;
 713         c->shrink.scan_objects          = bch2_btree_key_cache_scan;
 714         return register_shrinker(&c->shrink);
 715 }
 716
 717 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 718 {
 719         pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
 720         pr_buf(out, "nr_keys:\t%zu\n",  atomic_long_read(&c->nr_keys));
 721         pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
 722 }
 723
 724 void bch2_btree_key_cache_exit(void)
 725 {
 726         if (bch2_key_cache)
 727                 kmem_cache_destroy(bch2_key_cache);
 728 }
 729
 730 int __init bch2_btree_key_cache_init(void)
 731 {
 732         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
 733         if (!bch2_key_cache)
 734                 return -ENOMEM;
 735
 736         return 0;
 737 }