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         } else {
 169                 if (btree_id == BTREE_ID_subvolumes)
 170                         six_lock_pcpu_alloc(&ck->c.lock);
 171                 else
 172                         six_lock_pcpu_free(&ck->c.lock);
 173         }
 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         /*
 240          * Allocate some extra space so that the transaction commit path is less
 241          * likely to have to reallocate, since that requires a transaction
 242          * restart:
 243          */
 244         new_u64s = min(256U, (new_u64s * 3) / 2);
 245
 246         if (new_u64s > ck->u64s) {
 247                 new_u64s = roundup_pow_of_two(new_u64s);
 248                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 249                 if (!new_k) {
 250                         bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 251                                 bch2_btree_ids[ck->key.btree_id], new_u64s);
 252                         ret = -ENOMEM;
 253                         goto err;
 254                 }
 255         }
 256
 257         /*
 258          * XXX: not allowed to be holding read locks when we take a write lock,
 259          * currently
 260          */
 261         bch2_btree_node_lock_write(trans, ck_path, ck_path->l[0].b);
 262         if (new_k) {
 263                 kfree(ck->k);
 264                 ck->u64s = new_u64s;
 265                 ck->k = new_k;
 266         }
 267
 268         bkey_reassemble(ck->k, k);
 269         ck->valid = true;
 270         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 271
 272         /* We're not likely to need this iterator again: */
 273         path->preserve = false;
 274 err:
 275         bch2_path_put(trans, path, 0);
 276         return ret;
 277 }
 278
 279 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 280 {
 281         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 282         const struct btree_path *path = p;
 283
 284         return ck->key.btree_id == path->btree_id &&
 285                 !bpos_cmp(ck->key.pos, path->pos) ? 0 : -1;
 286 }
 287
 288 __flatten
 289 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 290                                     unsigned flags)
 291 {
 292         struct bch_fs *c = trans->c;
 293         struct bkey_cached *ck;
 294         int ret = 0;
 295
 296         BUG_ON(path->level);
 297
 298         path->l[1].b = NULL;
 299
 300         if (bch2_btree_node_relock(trans, path, 0)) {
 301                 ck = (void *) path->l[0].b;
 302                 goto fill;
 303         }
 304 retry:
 305         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 306         if (!ck) {
 307                 if (flags & BTREE_ITER_CACHED_NOCREATE) {
 308                         path->l[0].b = NULL;
 309                         return 0;
 310                 }
 311
 312                 ck = btree_key_cache_create(c, path->btree_id, path->pos);
 313                 ret = PTR_ERR_OR_ZERO(ck);
 314                 if (ret)
 315                         goto err;
 316                 if (!ck)
 317                         goto retry;
 318
 319                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 320                 path->locks_want = 1;
 321         } else {
 322                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 323
 324                 if (!btree_node_lock(trans, path, (void *) ck, path->pos, 0,
 325                                      lock_want,
 326                                      bkey_cached_check_fn, path, _THIS_IP_)) {
 327                         if (!trans->restarted)
 328                                 goto retry;
 329
 330                         ret = -EINTR;
 331                         goto err;
 332                 }
 333
 334                 if (ck->key.btree_id != path->btree_id ||
 335                     bpos_cmp(ck->key.pos, path->pos)) {
 336                         six_unlock_type(&ck->c.lock, lock_want);
 337                         goto retry;
 338                 }
 339
 340                 mark_btree_node_locked(trans, path, 0, lock_want);
 341         }
 342
 343         path->l[0].lock_seq     = ck->c.lock.state.seq;
 344         path->l[0].b            = (void *) ck;
 345 fill:
 346         if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
 347                 if (!path->locks_want &&
 348                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 349                         trace_transaction_restart_ip(trans->fn, _THIS_IP_);
 350                         ret = btree_trans_restart(trans);
 351                         goto err;
 352                 }
 353
 354                 ret = btree_key_cache_fill(trans, path, ck);
 355                 if (ret)
 356                         goto err;
 357         }
 358
 359         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 360                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 361
 362         path->uptodate = BTREE_ITER_UPTODATE;
 363         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 364
 365         return ret;
 366 err:
 367         if (ret != -EINTR) {
 368                 btree_node_unlock(path, 0);
 369                 path->l[0].b = BTREE_ITER_NO_NODE_ERROR;
 370         }
 371         return ret;
 372 }
 373
 374 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 375                                      struct bkey_cached_key key,
 376                                      u64 journal_seq,
 377                                      unsigned commit_flags,
 378                                      bool evict)
 379 {
 380         struct bch_fs *c = trans->c;
 381         struct journal *j = &c->journal;
 382         struct btree_iter c_iter, b_iter;
 383         struct bkey_cached *ck = NULL;
 384         int ret;
 385
 386         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 387                              BTREE_ITER_SLOTS|
 388                              BTREE_ITER_INTENT|
 389                              BTREE_ITER_ALL_SNAPSHOTS);
 390         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 391                              BTREE_ITER_CACHED|
 392                              BTREE_ITER_CACHED_NOFILL|
 393                              BTREE_ITER_CACHED_NOCREATE|
 394                              BTREE_ITER_INTENT);
 395         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 396
 397         ret = bch2_btree_iter_traverse(&c_iter);
 398         if (ret)
 399                 goto out;
 400
 401         ck = (void *) c_iter.path->l[0].b;
 402         if (!ck)
 403                 goto out;
 404
 405         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 406                 if (evict)
 407                         goto evict;
 408                 goto out;
 409         }
 410
 411         BUG_ON(!ck->valid);
 412
 413         if (journal_seq && ck->journal.seq != journal_seq)
 414                 goto out;
 415
 416         /*
 417          * Since journal reclaim depends on us making progress here, and the
 418          * allocator/copygc depend on journal reclaim making progress, we need
 419          * to be using alloc reserves:
 420          * */
 421         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 422                 bch2_trans_update(trans, &b_iter, ck->k,
 423                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 424                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 425                                   BTREE_TRIGGER_NORUN) ?:
 426                 bch2_trans_commit(trans, NULL, NULL,
 427                                   BTREE_INSERT_NOCHECK_RW|
 428                                   BTREE_INSERT_NOFAIL|
 429                                   BTREE_INSERT_USE_RESERVE|
 430                                   (ck->journal.seq == journal_last_seq(j)
 431                                    ? JOURNAL_WATERMARK_reserved
 432                                    : 0)|
 433                                   commit_flags);
 434         if (ret) {
 435                 bch2_fs_fatal_err_on(ret != -EINTR &&
 436                                      ret != -EAGAIN &&
 437                                      !bch2_journal_error(j), c,
 438                         "error flushing key cache: %i", ret);
 439                 goto out;
 440         }
 441
 442         bch2_journal_pin_drop(j, &ck->journal);
 443         bch2_journal_preres_put(j, &ck->res);
 444
 445         BUG_ON(!btree_node_locked(c_iter.path, 0));
 446
 447         if (!evict) {
 448                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 449                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 450                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 451                 }
 452         } else {
 453 evict:
 454                 BUG_ON(!btree_node_intent_locked(c_iter.path, 0));
 455
 456                 mark_btree_node_unlocked(c_iter.path, 0);
 457                 c_iter.path->l[0].b = NULL;
 458
 459                 six_lock_write(&ck->c.lock, NULL, NULL);
 460
 461                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 462                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 463                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 464                 }
 465
 466                 bkey_cached_evict(&c->btree_key_cache, ck);
 467
 468                 mutex_lock(&c->btree_key_cache.lock);
 469                 bkey_cached_free(&c->btree_key_cache, ck);
 470                 mutex_unlock(&c->btree_key_cache.lock);
 471         }
 472 out:
 473         bch2_trans_iter_exit(trans, &b_iter);
 474         bch2_trans_iter_exit(trans, &c_iter);
 475         return ret;
 476 }
 477
 478 int bch2_btree_key_cache_journal_flush(struct journal *j,
 479                                 struct journal_entry_pin *pin, u64 seq)
 480 {
 481         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 482         struct bkey_cached *ck =
 483                 container_of(pin, struct bkey_cached, journal);
 484         struct bkey_cached_key key;
 485         int ret = 0;
 486
 487         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 488
 489         six_lock_read(&ck->c.lock, NULL, NULL);
 490         key = ck->key;
 491
 492         if (ck->journal.seq != seq ||
 493             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 494                 six_unlock_read(&ck->c.lock);
 495                 goto unlock;
 496         }
 497         six_unlock_read(&ck->c.lock);
 498
 499         ret = bch2_trans_do(c, NULL, NULL, 0,
 500                 btree_key_cache_flush_pos(&trans, key, seq,
 501                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 502 unlock:
 503         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 504
 505         return ret;
 506 }
 507
 508 /*
 509  * Flush and evict a key from the key cache:
 510  */
 511 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 512                                enum btree_id id, struct bpos pos)
 513 {
 514         struct bch_fs *c = trans->c;
 515         struct bkey_cached_key key = { id, pos };
 516
 517         /* Fastpath - assume it won't be found: */
 518         if (!bch2_btree_key_cache_find(c, id, pos))
 519                 return 0;
 520
 521         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 522 }
 523
 524 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 525                                   struct btree_path *path,
 526                                   struct bkey_i *insert)
 527 {
 528         struct bch_fs *c = trans->c;
 529         struct bkey_cached *ck = (void *) path->l[0].b;
 530         bool kick_reclaim = false;
 531
 532         BUG_ON(insert->u64s > ck->u64s);
 533
 534         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 535                 int difference;
 536
 537                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 538
 539                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 540                 if (difference > 0) {
 541                         trans->journal_preres.u64s      -= difference;
 542                         ck->res.u64s                    += difference;
 543                 }
 544         }
 545
 546         bkey_copy(ck->k, insert);
 547         ck->valid = true;
 548
 549         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 550                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 551                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 552
 553                 if (bch2_nr_btree_keys_need_flush(c))
 554                         kick_reclaim = true;
 555         }
 556
 557         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 558                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 559
 560         if (kick_reclaim)
 561                 journal_reclaim_kick(&c->journal);
 562         return true;
 563 }
 564
 565 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 566                                            struct shrink_control *sc)
 567 {
 568         struct bch_fs *c = container_of(shrink, struct bch_fs,
 569                                         btree_key_cache.shrink);
 570         struct btree_key_cache *bc = &c->btree_key_cache;
 571         struct bucket_table *tbl;
 572         struct bkey_cached *ck, *t;
 573         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 574         unsigned start, flags;
 575         int srcu_idx;
 576
 577         /* Return -1 if we can't do anything right now */
 578         if (sc->gfp_mask & __GFP_FS)
 579                 mutex_lock(&bc->lock);
 580         else if (!mutex_trylock(&bc->lock))
 581                 return -1;
 582
 583         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 584         flags = memalloc_nofs_save();
 585
 586         /*
 587          * Newest freed entries are at the end of the list - once we hit one
 588          * that's too new to be freed, we can bail out:
 589          */
 590         list_for_each_entry_safe(ck, t, &bc->freed, list) {
 591                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 592                                                  ck->btree_trans_barrier_seq))
 593                         break;
 594
 595                 list_del(&ck->list);
 596                 kmem_cache_free(bch2_key_cache, ck);
 597                 bc->nr_freed--;
 598                 scanned++;
 599                 freed++;
 600         }
 601
 602         if (scanned >= nr)
 603                 goto out;
 604
 605         rcu_read_lock();
 606         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 607         if (bc->shrink_iter >= tbl->size)
 608                 bc->shrink_iter = 0;
 609         start = bc->shrink_iter;
 610
 611         do {
 612                 struct rhash_head *pos, *next;
 613
 614                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 615
 616                 while (!rht_is_a_nulls(pos)) {
 617                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 618                         ck = container_of(pos, struct bkey_cached, hash);
 619
 620                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 621                                 goto next;
 622
 623                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 624                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 625                         else if (bkey_cached_lock_for_evict(ck)) {
 626                                 bkey_cached_evict(bc, ck);
 627                                 bkey_cached_free(bc, ck);
 628                         }
 629
 630                         scanned++;
 631                         if (scanned >= nr)
 632                                 break;
 633 next:
 634                         pos = next;
 635                 }
 636
 637                 bc->shrink_iter++;
 638                 if (bc->shrink_iter >= tbl->size)
 639                         bc->shrink_iter = 0;
 640         } while (scanned < nr && bc->shrink_iter != start);
 641
 642         rcu_read_unlock();
 643 out:
 644         memalloc_nofs_restore(flags);
 645         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 646         mutex_unlock(&bc->lock);
 647
 648         return freed;
 649 }
 650
 651 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 652                                             struct shrink_control *sc)
 653 {
 654         struct bch_fs *c = container_of(shrink, struct bch_fs,
 655                                         btree_key_cache.shrink);
 656         struct btree_key_cache *bc = &c->btree_key_cache;
 657         long nr = atomic_long_read(&bc->nr_keys) -
 658                 atomic_long_read(&bc->nr_dirty);
 659
 660         return max(0L, nr);
 661 }
 662
 663 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 664 {
 665         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 666         struct bucket_table *tbl;
 667         struct bkey_cached *ck, *n;
 668         struct rhash_head *pos;
 669         unsigned i;
 670
 671         if (bc->shrink.list.next)
 672                 unregister_shrinker(&bc->shrink);
 673
 674         mutex_lock(&bc->lock);
 675
 676         rcu_read_lock();
 677         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 678         if (tbl)
 679                 for (i = 0; i < tbl->size; i++)
 680                         rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 681                                 bkey_cached_evict(bc, ck);
 682                                 list_add(&ck->list, &bc->freed);
 683                         }
 684         rcu_read_unlock();
 685
 686         list_for_each_entry_safe(ck, n, &bc->freed, list) {
 687                 cond_resched();
 688
 689                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 690                 bch2_journal_preres_put(&c->journal, &ck->res);
 691
 692                 list_del(&ck->list);
 693                 kfree(ck->k);
 694                 kmem_cache_free(bch2_key_cache, ck);
 695         }
 696
 697         BUG_ON(atomic_long_read(&bc->nr_dirty) &&
 698                !bch2_journal_error(&c->journal) &&
 699                test_bit(BCH_FS_WAS_RW, &c->flags));
 700         BUG_ON(atomic_long_read(&bc->nr_keys));
 701
 702         mutex_unlock(&bc->lock);
 703
 704         if (bc->table_init_done)
 705                 rhashtable_destroy(&bc->table);
 706 }
 707
 708 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 709 {
 710         mutex_init(&c->lock);
 711         INIT_LIST_HEAD(&c->freed);
 712 }
 713
 714 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
 715 {
 716         int ret;
 717
 718         ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
 719         if (ret)
 720                 return ret;
 721
 722         c->table_init_done = true;
 723
 724         c->shrink.seeks                 = 1;
 725         c->shrink.count_objects         = bch2_btree_key_cache_count;
 726         c->shrink.scan_objects          = bch2_btree_key_cache_scan;
 727         return register_shrinker(&c->shrink);
 728 }
 729
 730 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 731 {
 732         pr_buf(out, "nr_freed:\t%zu\n", c->nr_freed);
 733         pr_buf(out, "nr_keys:\t%lu\n",  atomic_long_read(&c->nr_keys));
 734         pr_buf(out, "nr_dirty:\t%lu\n", atomic_long_read(&c->nr_dirty));
 735 }
 736
 737 void bch2_btree_key_cache_exit(void)
 738 {
 739         if (bch2_key_cache)
 740                 kmem_cache_destroy(bch2_key_cache);
 741 }
 742
 743 int __init bch2_btree_key_cache_init(void)
 744 {
 745         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
 746         if (!bch2_key_cache)
 747                 return -ENOMEM;
 748
 749         return 0;
 750 }