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         /*
 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         bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
 237         if (new_k) {
 238                 kfree(ck->k);
 239                 ck->u64s = new_u64s;
 240                 ck->k = new_k;
 241         }
 242
 243         bkey_reassemble(ck->k, k);
 244         ck->valid = true;
 245         bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
 246
 247         /* We're not likely to need this iterator again: */
 248         set_btree_iter_dontneed(trans, iter);
 249 err:
 250         bch2_trans_iter_put(trans, iter);
 251         return ret;
 252 }
 253
 254 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 255 {
 256         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 257         const struct btree_iter *iter = p;
 258
 259         return ck->key.btree_id == iter->btree_id &&
 260                 !bpos_cmp(ck->key.pos, iter->pos) ? 0 : -1;
 261 }
 262
 263 __flatten
 264 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
 265 {
 266         struct btree_trans *trans = iter->trans;
 267         struct bch_fs *c = trans->c;
 268         struct bkey_cached *ck;
 269         int ret = 0;
 270
 271         BUG_ON(iter->level);
 272
 273         iter->l[1].b = NULL;
 274
 275         if (bch2_btree_node_relock(iter, 0)) {
 276                 ck = (void *) iter->l[0].b;
 277                 goto fill;
 278         }
 279 retry:
 280         ck = bch2_btree_key_cache_find(c, iter->btree_id, iter->pos);
 281         if (!ck) {
 282                 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
 283                         iter->l[0].b = NULL;
 284                         return 0;
 285                 }
 286
 287                 ck = btree_key_cache_create(&c->btree_key_cache,
 288                                             iter->btree_id, iter->pos);
 289                 ret = PTR_ERR_OR_ZERO(ck);
 290                 if (ret)
 291                         goto err;
 292                 if (!ck)
 293                         goto retry;
 294
 295                 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
 296                 iter->locks_want = 1;
 297         } else {
 298                 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
 299
 300                 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
 301                                      bkey_cached_check_fn, iter, _THIS_IP_)) {
 302                         if (ck->key.btree_id != iter->btree_id ||
 303                             bpos_cmp(ck->key.pos, iter->pos)) {
 304                                 goto retry;
 305                         }
 306
 307                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 308                         ret = -EINTR;
 309                         goto err;
 310                 }
 311
 312                 if (ck->key.btree_id != iter->btree_id ||
 313                     bpos_cmp(ck->key.pos, iter->pos)) {
 314                         six_unlock_type(&ck->c.lock, lock_want);
 315                         goto retry;
 316                 }
 317
 318                 mark_btree_node_locked(iter, 0, lock_want);
 319         }
 320
 321         iter->l[0].lock_seq     = ck->c.lock.state.seq;
 322         iter->l[0].b            = (void *) ck;
 323 fill:
 324         if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
 325                 if (!btree_node_intent_locked(iter, 0))
 326                         bch2_btree_iter_upgrade(iter, 1);
 327                 if (!btree_node_intent_locked(iter, 0)) {
 328                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 329                         ret = -EINTR;
 330                         goto err;
 331                 }
 332
 333                 ret = btree_key_cache_fill(trans, iter, ck);
 334                 if (ret)
 335                         goto err;
 336         }
 337
 338         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 339                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 340
 341         iter->uptodate = BTREE_ITER_NEED_PEEK;
 342
 343         if (!(iter->flags & BTREE_ITER_INTENT))
 344                 bch2_btree_iter_downgrade(iter);
 345         else if (!iter->locks_want) {
 346                 if (!__bch2_btree_iter_upgrade(iter, 1))
 347                         ret = -EINTR;
 348         }
 349
 350         return ret;
 351 err:
 352         if (ret != -EINTR) {
 353                 btree_node_unlock(iter, 0);
 354                 iter->flags |= BTREE_ITER_ERROR;
 355                 iter->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 = NULL, *b_iter = NULL;
 369         struct bkey_cached *ck = NULL;
 370         int ret;
 371
 372         b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 373                                      BTREE_ITER_SLOTS|
 374                                      BTREE_ITER_INTENT);
 375         c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 376                                      BTREE_ITER_CACHED|
 377                                      BTREE_ITER_CACHED_NOFILL|
 378                                      BTREE_ITER_CACHED_NOCREATE|
 379                                      BTREE_ITER_INTENT);
 380 retry:
 381         ret = bch2_btree_iter_traverse(c_iter);
 382         if (ret)
 383                 goto err;
 384
 385         ck = (void *) c_iter->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, BTREE_TRIGGER_NORUN) ?:
 403                 bch2_trans_commit(trans, NULL, NULL,
 404                                   BTREE_INSERT_NOUNLOCK|
 405                                   BTREE_INSERT_NOCHECK_RW|
 406                                   BTREE_INSERT_NOFAIL|
 407                                   BTREE_INSERT_USE_RESERVE|
 408                                   (ck->journal.seq == journal_last_seq(j)
 409                                    ? BTREE_INSERT_JOURNAL_RESERVED
 410                                    : 0)|
 411                                   commit_flags);
 412 err:
 413         if (ret == -EINTR)
 414                 goto retry;
 415
 416         if (ret == -EAGAIN)
 417                 goto out;
 418
 419         if (ret) {
 420                 bch2_fs_fatal_err_on(!bch2_journal_error(j), c,
 421                         "error flushing key cache: %i", ret);
 422                 goto out;
 423         }
 424
 425         bch2_journal_pin_drop(j, &ck->journal);
 426         bch2_journal_preres_put(j, &ck->res);
 427
 428         BUG_ON(!btree_node_locked(c_iter, 0));
 429
 430         if (!evict) {
 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         } else {
 436 evict:
 437                 BUG_ON(!btree_node_intent_locked(c_iter, 0));
 438
 439                 mark_btree_node_unlocked(c_iter, 0);
 440                 c_iter->l[0].b = NULL;
 441
 442                 six_lock_write(&ck->c.lock, NULL, NULL);
 443
 444                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 445                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 446                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 447                 }
 448
 449                 bkey_cached_evict(&c->btree_key_cache, ck);
 450
 451                 mutex_lock(&c->btree_key_cache.lock);
 452                 bkey_cached_free(&c->btree_key_cache, ck);
 453                 mutex_unlock(&c->btree_key_cache.lock);
 454         }
 455 out:
 456         bch2_trans_iter_put(trans, b_iter);
 457         bch2_trans_iter_put(trans, c_iter);
 458         return ret;
 459 }
 460
 461 int bch2_btree_key_cache_journal_flush(struct journal *j,
 462                                 struct journal_entry_pin *pin, u64 seq)
 463 {
 464         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 465         struct bkey_cached *ck =
 466                 container_of(pin, struct bkey_cached, journal);
 467         struct bkey_cached_key key;
 468         struct btree_trans trans;
 469         int ret = 0;
 470
 471         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 472
 473         six_lock_read(&ck->c.lock, NULL, NULL);
 474         key = ck->key;
 475
 476         if (ck->journal.seq != seq ||
 477             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 478                 six_unlock_read(&ck->c.lock);
 479                 goto unlock;
 480         }
 481         six_unlock_read(&ck->c.lock);
 482
 483         bch2_trans_init(&trans, c, 0, 0);
 484         ret = btree_key_cache_flush_pos(&trans, key, seq,
 485                                   BTREE_INSERT_JOURNAL_RECLAIM, false);
 486         bch2_trans_exit(&trans);
 487 unlock:
 488         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 489
 490         return ret;
 491 }
 492
 493 /*
 494  * Flush and evict a key from the key cache:
 495  */
 496 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 497                                enum btree_id id, struct bpos pos)
 498 {
 499         struct bch_fs *c = trans->c;
 500         struct bkey_cached_key key = { id, pos };
 501
 502         /* Fastpath - assume it won't be found: */
 503         if (!bch2_btree_key_cache_find(c, id, pos))
 504                 return 0;
 505
 506         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 507 }
 508
 509 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 510                                   struct btree_iter *iter,
 511                                   struct bkey_i *insert)
 512 {
 513         struct bch_fs *c = trans->c;
 514         struct bkey_cached *ck = (void *) iter->l[0].b;
 515         bool kick_reclaim = false;
 516
 517         BUG_ON(insert->u64s > ck->u64s);
 518
 519         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 520                 int difference;
 521
 522                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 523
 524                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 525                 if (difference > 0) {
 526                         trans->journal_preres.u64s      -= difference;
 527                         ck->res.u64s                    += difference;
 528                 }
 529         }
 530
 531         bkey_copy(ck->k, insert);
 532         ck->valid = true;
 533
 534         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 535                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 536                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 537
 538                 if (bch2_nr_btree_keys_need_flush(c))
 539                         kick_reclaim = true;
 540         }
 541
 542         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 543                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 544
 545         if (kick_reclaim)
 546                 journal_reclaim_kick(&c->journal);
 547         return true;
 548 }
 549
 550 #ifdef CONFIG_BCACHEFS_DEBUG
 551 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
 552                                enum btree_id id, struct bpos pos)
 553 {
 554         BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
 555 }
 556 #endif
 557
 558 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 559                                            struct shrink_control *sc)
 560 {
 561         struct bch_fs *c = container_of(shrink, struct bch_fs,
 562                                         btree_key_cache.shrink);
 563         struct btree_key_cache *bc = &c->btree_key_cache;
 564         struct bucket_table *tbl;
 565         struct bkey_cached *ck, *t;
 566         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 567         unsigned start, flags;
 568         int srcu_idx;
 569
 570         /* Return -1 if we can't do anything right now */
 571         if (sc->gfp_mask & __GFP_FS)
 572                 mutex_lock(&bc->lock);
 573         else if (!mutex_trylock(&bc->lock))
 574                 return -1;
 575
 576         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 577         flags = memalloc_nofs_save();
 578
 579         /*
 580          * Newest freed entries are at the end of the list - once we hit one
 581          * that's too new to be freed, we can bail out:
 582          */
 583         list_for_each_entry_safe(ck, t, &bc->freed, list) {
 584                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 585                                                  ck->btree_trans_barrier_seq))
 586                         break;
 587
 588                 list_del(&ck->list);
 589                 kmem_cache_free(bch2_key_cache, ck);
 590                 bc->nr_freed--;
 591                 scanned++;
 592                 freed++;
 593         }
 594
 595         if (scanned >= nr)
 596                 goto out;
 597
 598         rcu_read_lock();
 599         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 600         if (bc->shrink_iter >= tbl->size)
 601                 bc->shrink_iter = 0;
 602         start = bc->shrink_iter;
 603
 604         do {
 605                 struct rhash_head *pos, *next;
 606
 607                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 608
 609                 while (!rht_is_a_nulls(pos)) {
 610                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 611                         ck = container_of(pos, struct bkey_cached, hash);
 612
 613                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 614                                 goto next;
 615
 616                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 617                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 618                         else if (bkey_cached_lock_for_evict(ck)) {
 619                                 bkey_cached_evict(bc, ck);
 620                                 bkey_cached_free(bc, ck);
 621                         }
 622
 623                         scanned++;
 624                         if (scanned >= nr)
 625                                 break;
 626 next:
 627                         pos = next;
 628                 }
 629
 630                 bc->shrink_iter++;
 631                 if (bc->shrink_iter >= tbl->size)
 632                         bc->shrink_iter = 0;
 633         } while (scanned < nr && bc->shrink_iter != start);
 634
 635         rcu_read_unlock();
 636 out:
 637         memalloc_nofs_restore(flags);
 638         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 639         mutex_unlock(&bc->lock);
 640
 641         return freed;
 642 }
 643
 644 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 645                                             struct shrink_control *sc)
 646 {
 647         struct bch_fs *c = container_of(shrink, struct bch_fs,
 648                                         btree_key_cache.shrink);
 649         struct btree_key_cache *bc = &c->btree_key_cache;
 650         long nr = atomic_long_read(&bc->nr_keys) -
 651                 atomic_long_read(&bc->nr_dirty);
 652
 653         return max(0L, nr);
 654 }
 655
 656 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 657 {
 658         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 659         struct bucket_table *tbl;
 660         struct bkey_cached *ck, *n;
 661         struct rhash_head *pos;
 662         unsigned i;
 663
 664         if (bc->shrink.list.next)
 665                 unregister_shrinker(&bc->shrink);
 666
 667         mutex_lock(&bc->lock);
 668
 669         rcu_read_lock();
 670         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 671         for (i = 0; i < tbl->size; i++)
 672                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 673                         bkey_cached_evict(bc, ck);
 674                         list_add(&ck->list, &bc->freed);
 675                 }
 676         rcu_read_unlock();
 677
 678         list_for_each_entry_safe(ck, n, &bc->freed, list) {
 679                 cond_resched();
 680
 681                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 682                 bch2_journal_preres_put(&c->journal, &ck->res);
 683
 684                 list_del(&ck->list);
 685                 kfree(ck->k);
 686                 kmem_cache_free(bch2_key_cache, ck);
 687         }
 688
 689         BUG_ON(atomic_long_read(&bc->nr_dirty) &&
 690                !bch2_journal_error(&c->journal) &&
 691                test_bit(BCH_FS_WAS_RW, &c->flags));
 692         BUG_ON(atomic_long_read(&bc->nr_keys));
 693
 694         mutex_unlock(&bc->lock);
 695
 696         if (bc->table_init_done)
 697                 rhashtable_destroy(&bc->table);
 698 }
 699
 700 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 701 {
 702         mutex_init(&c->lock);
 703         INIT_LIST_HEAD(&c->freed);
 704 }
 705
 706 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
 707 {
 708         int ret;
 709
 710         ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
 711         if (ret)
 712                 return ret;
 713
 714         c->table_init_done = true;
 715
 716         c->shrink.seeks                 = 1;
 717         c->shrink.count_objects         = bch2_btree_key_cache_count;
 718         c->shrink.scan_objects          = bch2_btree_key_cache_scan;
 719         return register_shrinker(&c->shrink);
 720 }
 721
 722 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 723 {
 724         pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
 725         pr_buf(out, "nr_keys:\t%zu\n",  atomic_long_read(&c->nr_keys));
 726         pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
 727 }
 728
 729 void bch2_btree_key_cache_exit(void)
 730 {
 731         if (bch2_key_cache)
 732                 kmem_cache_destroy(bch2_key_cache);
 733 }
 734
 735 int __init bch2_btree_key_cache_init(void)
 736 {
 737         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
 738         if (!bch2_key_cache)
 739                 return -ENOMEM;
 740
 741         return 0;
 742 }