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