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 bch_fs *c,
 150                        enum btree_id btree_id,
 151                        struct bpos pos)
 152 {
 153         struct btree_key_cache *bc = &c->btree_key_cache;
 154         struct bkey_cached *ck;
 155         bool was_new = true;
 156
 157         ck = bkey_cached_alloc(bc);
 158
 159         if (unlikely(!ck)) {
 160                 ck = bkey_cached_reuse(bc);
 161                 if (unlikely(!ck)) {
 162                         bch_err(c, "error allocating memory for key cache item, btree %s",
 163                                 bch2_btree_ids[btree_id]);
 164                         return ERR_PTR(-ENOMEM);
 165                 }
 166
 167                 was_new = false;
 168         }
 169
 170         if (btree_id == BTREE_ID_subvolumes)
 171                 six_lock_pcpu_alloc(&ck->c.lock);
 172         else
 173                 six_lock_pcpu_free(&ck->c.lock);
 174
 175         ck->c.level             = 0;
 176         ck->c.btree_id          = btree_id;
 177         ck->key.btree_id        = btree_id;
 178         ck->key.pos             = pos;
 179         ck->valid               = false;
 180         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 181
 182         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 183                                           &ck->hash,
 184                                           bch2_btree_key_cache_params))) {
 185                 /* We raced with another fill: */
 186
 187                 if (likely(was_new)) {
 188                         six_unlock_write(&ck->c.lock);
 189                         six_unlock_intent(&ck->c.lock);
 190                         kfree(ck);
 191                 } else {
 192                         mutex_lock(&bc->lock);
 193                         bkey_cached_free(bc, ck);
 194                         mutex_unlock(&bc->lock);
 195                 }
 196
 197                 return NULL;
 198         }
 199
 200         atomic_long_inc(&bc->nr_keys);
 201
 202         six_unlock_write(&ck->c.lock);
 203
 204         return ck;
 205 }
 206
 207 static int btree_key_cache_fill(struct btree_trans *trans,
 208                                 struct btree_path *ck_path,
 209                                 struct bkey_cached *ck)
 210 {
 211         struct btree_path *path;
 212         struct bkey_s_c k;
 213         unsigned new_u64s = 0;
 214         struct bkey_i *new_k = NULL;
 215         struct bkey u;
 216         int ret;
 217
 218         path = bch2_path_get(trans, ck->key.btree_id,
 219                              ck->key.pos, 0, 0, 0, _THIS_IP_);
 220         ret = bch2_btree_path_traverse(trans, path, 0);
 221         if (ret)
 222                 goto err;
 223
 224         k = bch2_btree_path_peek_slot(path, &u);
 225
 226         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 227                 trace_trans_restart_relock_key_cache_fill(trans->fn,
 228                                 _THIS_IP_, ck_path->btree_id, &ck_path->pos);
 229                 ret = btree_trans_restart(trans);
 230                 goto err;
 231         }
 232
 233         /*
 234          * bch2_varint_decode can read past the end of the buffer by at
 235          * most 7 bytes (it won't be used):
 236          */
 237         new_u64s = k.k->u64s + 1;
 238
 239         if (new_u64s > ck->u64s) {
 240                 new_u64s = roundup_pow_of_two(new_u64s);
 241                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 242                 if (!new_k) {
 243                         bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 244                                 bch2_btree_ids[ck->key.btree_id], new_u64s);
 245                         ret = -ENOMEM;
 246                         goto err;
 247                 }
 248         }
 249
 250         /*
 251          * XXX: not allowed to be holding read locks when we take a write lock,
 252          * currently
 253          */
 254         bch2_btree_node_lock_write(trans, ck_path, ck_path->l[0].b);
 255         if (new_k) {
 256                 kfree(ck->k);
 257                 ck->u64s = new_u64s;
 258                 ck->k = new_k;
 259         }
 260
 261         bkey_reassemble(ck->k, k);
 262         ck->valid = true;
 263         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 264
 265         /* We're not likely to need this iterator again: */
 266         path->preserve = false;
 267 err:
 268         bch2_path_put(trans, path, 0);
 269         return ret;
 270 }
 271
 272 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 273 {
 274         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 275         const struct btree_path *path = p;
 276
 277         return ck->key.btree_id == path->btree_id &&
 278                 !bpos_cmp(ck->key.pos, path->pos) ? 0 : -1;
 279 }
 280
 281 __flatten
 282 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 283                                     unsigned flags)
 284 {
 285         struct bch_fs *c = trans->c;
 286         struct bkey_cached *ck;
 287         int ret = 0;
 288
 289         BUG_ON(path->level);
 290
 291         path->l[1].b = NULL;
 292
 293         if (bch2_btree_node_relock(trans, path, 0)) {
 294                 ck = (void *) path->l[0].b;
 295                 goto fill;
 296         }
 297 retry:
 298         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 299         if (!ck) {
 300                 if (flags & BTREE_ITER_CACHED_NOCREATE) {
 301                         path->l[0].b = NULL;
 302                         return 0;
 303                 }
 304
 305                 ck = btree_key_cache_create(c, path->btree_id, path->pos);
 306                 ret = PTR_ERR_OR_ZERO(ck);
 307                 if (ret)
 308                         goto err;
 309                 if (!ck)
 310                         goto retry;
 311
 312                 mark_btree_node_locked(path, 0, SIX_LOCK_intent);
 313                 path->locks_want = 1;
 314         } else {
 315                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 316
 317                 if (!btree_node_lock(trans, path, (void *) ck, path->pos, 0,
 318                                      lock_want,
 319                                      bkey_cached_check_fn, path, _THIS_IP_)) {
 320                         if (!trans->restarted)
 321                                 goto retry;
 322
 323                         ret = -EINTR;
 324                         goto err;
 325                 }
 326
 327                 if (ck->key.btree_id != path->btree_id ||
 328                     bpos_cmp(ck->key.pos, path->pos)) {
 329                         six_unlock_type(&ck->c.lock, lock_want);
 330                         goto retry;
 331                 }
 332
 333                 mark_btree_node_locked(path, 0, lock_want);
 334         }
 335
 336         path->l[0].lock_seq     = ck->c.lock.state.seq;
 337         path->l[0].b            = (void *) ck;
 338 fill:
 339         if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
 340                 if (!path->locks_want &&
 341                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 342                         trace_transaction_restart_ip(trans->fn, _THIS_IP_);
 343                         ret = btree_trans_restart(trans);
 344                         goto err;
 345                 }
 346
 347                 ret = btree_key_cache_fill(trans, path, ck);
 348                 if (ret)
 349                         goto err;
 350         }
 351
 352         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 353                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 354
 355         path->uptodate = BTREE_ITER_UPTODATE;
 356         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 357
 358         return ret;
 359 err:
 360         if (ret != -EINTR) {
 361                 btree_node_unlock(path, 0);
 362                 path->l[0].b = BTREE_ITER_NO_NODE_ERROR;
 363         }
 364         return ret;
 365 }
 366
 367 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 368                                      struct bkey_cached_key key,
 369                                      u64 journal_seq,
 370                                      unsigned commit_flags,
 371                                      bool evict)
 372 {
 373         struct bch_fs *c = trans->c;
 374         struct journal *j = &c->journal;
 375         struct btree_iter c_iter, b_iter;
 376         struct bkey_cached *ck = NULL;
 377         int ret;
 378
 379         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 380                              BTREE_ITER_SLOTS|
 381                              BTREE_ITER_INTENT|
 382                              BTREE_ITER_ALL_SNAPSHOTS);
 383         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 384                              BTREE_ITER_CACHED|
 385                              BTREE_ITER_CACHED_NOFILL|
 386                              BTREE_ITER_CACHED_NOCREATE|
 387                              BTREE_ITER_INTENT);
 388         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 389
 390         ret = bch2_btree_iter_traverse(&c_iter);
 391         if (ret)
 392                 goto out;
 393
 394         ck = (void *) c_iter.path->l[0].b;
 395         if (!ck)
 396                 goto out;
 397
 398         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 399                 if (evict)
 400                         goto evict;
 401                 goto out;
 402         }
 403
 404         BUG_ON(!ck->valid);
 405
 406         if (journal_seq && ck->journal.seq != journal_seq)
 407                 goto out;
 408
 409         /*
 410          * Since journal reclaim depends on us making progress here, and the
 411          * allocator/copygc depend on journal reclaim making progress, we need
 412          * to be using alloc reserves:
 413          * */
 414         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 415                 bch2_trans_update(trans, &b_iter, ck->k,
 416                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 417                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 418                                   BTREE_TRIGGER_NORUN) ?:
 419                 bch2_trans_commit(trans, NULL, NULL,
 420                                   BTREE_INSERT_NOCHECK_RW|
 421                                   BTREE_INSERT_NOFAIL|
 422                                   BTREE_INSERT_USE_RESERVE|
 423                                   (ck->journal.seq == journal_last_seq(j)
 424                                    ? BTREE_INSERT_JOURNAL_RESERVED
 425                                    : 0)|
 426                                   commit_flags);
 427         if (ret) {
 428                 bch2_fs_fatal_err_on(ret != -EINTR &&
 429                                      ret != -EAGAIN &&
 430                                      !bch2_journal_error(j), c,
 431                         "error flushing key cache: %i", ret);
 432                 goto out;
 433         }
 434
 435         bch2_journal_pin_drop(j, &ck->journal);
 436         bch2_journal_preres_put(j, &ck->res);
 437
 438         BUG_ON(!btree_node_locked(c_iter.path, 0));
 439
 440         if (!evict) {
 441                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 442                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 443                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 444                 }
 445         } else {
 446 evict:
 447                 BUG_ON(!btree_node_intent_locked(c_iter.path, 0));
 448
 449                 mark_btree_node_unlocked(c_iter.path, 0);
 450                 c_iter.path->l[0].b = NULL;
 451
 452                 six_lock_write(&ck->c.lock, NULL, NULL);
 453
 454                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 455                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 456                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 457                 }
 458
 459                 bkey_cached_evict(&c->btree_key_cache, ck);
 460
 461                 mutex_lock(&c->btree_key_cache.lock);
 462                 bkey_cached_free(&c->btree_key_cache, ck);
 463                 mutex_unlock(&c->btree_key_cache.lock);
 464         }
 465 out:
 466         bch2_trans_iter_exit(trans, &b_iter);
 467         bch2_trans_iter_exit(trans, &c_iter);
 468         return ret;
 469 }
 470
 471 int bch2_btree_key_cache_journal_flush(struct journal *j,
 472                                 struct journal_entry_pin *pin, u64 seq)
 473 {
 474         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 475         struct bkey_cached *ck =
 476                 container_of(pin, struct bkey_cached, journal);
 477         struct bkey_cached_key key;
 478         int ret = 0;
 479
 480         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 481
 482         six_lock_read(&ck->c.lock, NULL, NULL);
 483         key = ck->key;
 484
 485         if (ck->journal.seq != seq ||
 486             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 487                 six_unlock_read(&ck->c.lock);
 488                 goto unlock;
 489         }
 490         six_unlock_read(&ck->c.lock);
 491
 492         ret = bch2_trans_do(c, NULL, NULL, 0,
 493                 btree_key_cache_flush_pos(&trans, key, seq,
 494                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 495 unlock:
 496         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 497
 498         return ret;
 499 }
 500
 501 /*
 502  * Flush and evict a key from the key cache:
 503  */
 504 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 505                                enum btree_id id, struct bpos pos)
 506 {
 507         struct bch_fs *c = trans->c;
 508         struct bkey_cached_key key = { id, pos };
 509
 510         /* Fastpath - assume it won't be found: */
 511         if (!bch2_btree_key_cache_find(c, id, pos))
 512                 return 0;
 513
 514         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 515 }
 516
 517 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 518                                   struct btree_path *path,
 519                                   struct bkey_i *insert)
 520 {
 521         struct bch_fs *c = trans->c;
 522         struct bkey_cached *ck = (void *) path->l[0].b;
 523         bool kick_reclaim = false;
 524
 525         BUG_ON(insert->u64s > ck->u64s);
 526
 527         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 528                 int difference;
 529
 530                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 531
 532                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 533                 if (difference > 0) {
 534                         trans->journal_preres.u64s      -= difference;
 535                         ck->res.u64s                    += difference;
 536                 }
 537         }
 538
 539         bkey_copy(ck->k, insert);
 540         ck->valid = true;
 541
 542         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 543                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 544                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 545
 546                 if (bch2_nr_btree_keys_need_flush(c))
 547                         kick_reclaim = true;
 548         }
 549
 550         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 551                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 552
 553         if (kick_reclaim)
 554                 journal_reclaim_kick(&c->journal);
 555         return true;
 556 }
 557
 558 #ifdef CONFIG_BCACHEFS_DEBUG
 559 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
 560                                enum btree_id id, struct bpos pos)
 561 {
 562         BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
 563 }
 564 #endif
 565
 566 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 567                                            struct shrink_control *sc)
 568 {
 569         struct bch_fs *c = container_of(shrink, struct bch_fs,
 570                                         btree_key_cache.shrink);
 571         struct btree_key_cache *bc = &c->btree_key_cache;
 572         struct bucket_table *tbl;
 573         struct bkey_cached *ck, *t;
 574         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 575         unsigned start, flags;
 576         int srcu_idx;
 577
 578         /* Return -1 if we can't do anything right now */
 579         if (sc->gfp_mask & __GFP_FS)
 580                 mutex_lock(&bc->lock);
 581         else if (!mutex_trylock(&bc->lock))
 582                 return -1;
 583
 584         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 585         flags = memalloc_nofs_save();
 586
 587         /*
 588          * Newest freed entries are at the end of the list - once we hit one
 589          * that's too new to be freed, we can bail out:
 590          */
 591         list_for_each_entry_safe(ck, t, &bc->freed, list) {
 592                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 593                                                  ck->btree_trans_barrier_seq))
 594                         break;
 595
 596                 list_del(&ck->list);
 597                 kmem_cache_free(bch2_key_cache, ck);
 598                 bc->nr_freed--;
 599                 scanned++;
 600                 freed++;
 601         }
 602
 603         if (scanned >= nr)
 604                 goto out;
 605
 606         rcu_read_lock();
 607         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 608         if (bc->shrink_iter >= tbl->size)
 609                 bc->shrink_iter = 0;
 610         start = bc->shrink_iter;
 611
 612         do {
 613                 struct rhash_head *pos, *next;
 614
 615                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 616
 617                 while (!rht_is_a_nulls(pos)) {
 618                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 619                         ck = container_of(pos, struct bkey_cached, hash);
 620
 621                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 622                                 goto next;
 623
 624                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 625                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 626                         else if (bkey_cached_lock_for_evict(ck)) {
 627                                 bkey_cached_evict(bc, ck);
 628                                 bkey_cached_free(bc, ck);
 629                         }
 630
 631                         scanned++;
 632                         if (scanned >= nr)
 633                                 break;
 634 next:
 635                         pos = next;
 636                 }
 637
 638                 bc->shrink_iter++;
 639                 if (bc->shrink_iter >= tbl->size)
 640                         bc->shrink_iter = 0;
 641         } while (scanned < nr && bc->shrink_iter != start);
 642
 643         rcu_read_unlock();
 644 out:
 645         memalloc_nofs_restore(flags);
 646         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 647         mutex_unlock(&bc->lock);
 648
 649         return freed;
 650 }
 651
 652 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 653                                             struct shrink_control *sc)
 654 {
 655         struct bch_fs *c = container_of(shrink, struct bch_fs,
 656                                         btree_key_cache.shrink);
 657         struct btree_key_cache *bc = &c->btree_key_cache;
 658         long nr = atomic_long_read(&bc->nr_keys) -
 659                 atomic_long_read(&bc->nr_dirty);
 660
 661         return max(0L, nr);
 662 }
 663
 664 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 665 {
 666         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 667         struct bucket_table *tbl;
 668         struct bkey_cached *ck, *n;
 669         struct rhash_head *pos;
 670         unsigned i;
 671
 672         if (bc->shrink.list.next)
 673                 unregister_shrinker(&bc->shrink);
 674
 675         mutex_lock(&bc->lock);
 676
 677         rcu_read_lock();
 678         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 679         if (tbl)
 680                 for (i = 0; i < tbl->size; i++)
 681                         rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 682                                 bkey_cached_evict(bc, ck);
 683                                 list_add(&ck->list, &bc->freed);
 684                         }
 685         rcu_read_unlock();
 686
 687         list_for_each_entry_safe(ck, n, &bc->freed, list) {
 688                 cond_resched();
 689
 690                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 691                 bch2_journal_preres_put(&c->journal, &ck->res);
 692
 693                 list_del(&ck->list);
 694                 kfree(ck->k);
 695                 kmem_cache_free(bch2_key_cache, ck);
 696         }
 697
 698         BUG_ON(atomic_long_read(&bc->nr_dirty) &&
 699                !bch2_journal_error(&c->journal) &&
 700                test_bit(BCH_FS_WAS_RW, &c->flags));
 701         BUG_ON(atomic_long_read(&bc->nr_keys));
 702
 703         mutex_unlock(&bc->lock);
 704
 705         if (bc->table_init_done)
 706                 rhashtable_destroy(&bc->table);
 707 }
 708
 709 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 710 {
 711         mutex_init(&c->lock);
 712         INIT_LIST_HEAD(&c->freed);
 713 }
 714
 715 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
 716 {
 717         int ret;
 718
 719         ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
 720         if (ret)
 721                 return ret;
 722
 723         c->table_init_done = true;
 724
 725         c->shrink.seeks                 = 1;
 726         c->shrink.count_objects         = bch2_btree_key_cache_count;
 727         c->shrink.scan_objects          = bch2_btree_key_cache_scan;
 728         return register_shrinker(&c->shrink);
 729 }
 730
 731 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 732 {
 733         pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
 734         pr_buf(out, "nr_keys:\t%zu\n",  atomic_long_read(&c->nr_keys));
 735         pr_buf(out, "nr_dirty:\t%zu\n", atomic_long_read(&c->nr_dirty));
 736 }
 737
 738 void bch2_btree_key_cache_exit(void)
 739 {
 740         if (bch2_key_cache)
 741                 kmem_cache_destroy(bch2_key_cache);
 742 }
 743
 744 int __init bch2_btree_key_cache_init(void)
 745 {
 746         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
 747         if (!bch2_key_cache)
 748                 return -ENOMEM;
 749
 750         return 0;
 751 }