fs/bcachefs/btree_key_cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_cache.h"
   5 #include "btree_iter.h"
   6 #include "btree_key_cache.h"
   7 #include "btree_locking.h"
   8 #include "btree_update.h"
   9 #include "errcode.h"
  10 #include "error.h"
  11 #include "journal.h"
  12 #include "journal_reclaim.h"
  13 #include "trace.h"
  14
  15 #include <linux/sched/mm.h>
  16
  17 static inline bool btree_uses_pcpu_readers(enum btree_id id)
  18 {
  19         return id == BTREE_ID_subvolumes;
  20 }
  21
  22 static struct kmem_cache *bch2_key_cache;
  23
  24 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  25                                        const void *obj)
  26 {
  27         const struct bkey_cached *ck = obj;
  28         const struct bkey_cached_key *key = arg->key;
  29
  30         return ck->key.btree_id != key->btree_id ||
  31                 !bpos_eq(ck->key.pos, key->pos);
  32 }
  33
  34 static const struct rhashtable_params bch2_btree_key_cache_params = {
  35         .head_offset            = offsetof(struct bkey_cached, hash),
  36         .key_offset             = offsetof(struct bkey_cached, key),
  37         .key_len                = sizeof(struct bkey_cached_key),
  38         .obj_cmpfn              = bch2_btree_key_cache_cmp_fn,
  39         .automatic_shrinking    = true,
  40 };
  41
  42 static inline void btree_path_cached_set(struct btree_trans *trans, struct btree_path *path,
  43                                          struct bkey_cached *ck,
  44                                          enum btree_node_locked_type lock_held)
  45 {
  46         path->l[0].lock_seq     = six_lock_seq(&ck->c.lock);
  47         path->l[0].b            = (void *) ck;
  48         mark_btree_node_locked(trans, path, 0, lock_held);
  49 }
  50
  51 __flatten
  52 inline struct bkey_cached *
  53 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
  54 {
  55         struct bkey_cached_key key = {
  56                 .btree_id       = btree_id,
  57                 .pos            = pos,
  58         };
  59
  60         return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
  61                                       bch2_btree_key_cache_params);
  62 }
  63
  64 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
  65 {
  66         if (!six_trylock_intent(&ck->c.lock))
  67                 return false;
  68
  69         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
  70                 six_unlock_intent(&ck->c.lock);
  71                 return false;
  72         }
  73
  74         if (!six_trylock_write(&ck->c.lock)) {
  75                 six_unlock_intent(&ck->c.lock);
  76                 return false;
  77         }
  78
  79         return true;
  80 }
  81
  82 static bool bkey_cached_evict(struct btree_key_cache *c,
  83                               struct bkey_cached *ck)
  84 {
  85         bool ret = !rhashtable_remove_fast(&c->table, &ck->hash,
  86                                       bch2_btree_key_cache_params);
  87         if (ret) {
  88                 memset(&ck->key, ~0, sizeof(ck->key));
  89                 atomic_long_dec(&c->nr_keys);
  90         }
  91
  92         return ret;
  93 }
  94
  95 static void __bkey_cached_free(struct rcu_pending *pending, struct rcu_head *rcu)
  96 {
  97         struct bch_fs *c = container_of(pending->srcu, struct bch_fs, btree_trans_barrier);
  98         struct bkey_cached *ck = container_of(rcu, struct bkey_cached, rcu);
  99
 100         this_cpu_dec(*c->btree_key_cache.nr_pending);
 101         kmem_cache_free(bch2_key_cache, ck);
 102 }
 103
 104 static void bkey_cached_free(struct btree_key_cache *bc,
 105                              struct bkey_cached *ck)
 106 {
 107         kfree(ck->k);
 108         ck->k           = NULL;
 109         ck->u64s        = 0;
 110
 111         six_unlock_write(&ck->c.lock);
 112         six_unlock_intent(&ck->c.lock);
 113
 114         bool pcpu_readers = ck->c.lock.readers != NULL;
 115         rcu_pending_enqueue(&bc->pending[pcpu_readers], &ck->rcu);
 116         this_cpu_inc(*bc->nr_pending);
 117 }
 118
 119 static struct bkey_cached *__bkey_cached_alloc(unsigned key_u64s, gfp_t gfp)
 120 {
 121         gfp |= __GFP_ACCOUNT|__GFP_RECLAIMABLE;
 122
 123         struct bkey_cached *ck = kmem_cache_zalloc(bch2_key_cache, gfp);
 124         if (unlikely(!ck))
 125                 return NULL;
 126         ck->k = kmalloc(key_u64s * sizeof(u64), gfp);
 127         if (unlikely(!ck->k)) {
 128                 kmem_cache_free(bch2_key_cache, ck);
 129                 return NULL;
 130         }
 131         ck->u64s = key_u64s;
 132         return ck;
 133 }
 134
 135 static struct bkey_cached *
 136 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path, unsigned key_u64s)
 137 {
 138         struct bch_fs *c = trans->c;
 139         struct btree_key_cache *bc = &c->btree_key_cache;
 140         bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
 141         int ret;
 142
 143         struct bkey_cached *ck = container_of_or_null(
 144                                 rcu_pending_dequeue(&bc->pending[pcpu_readers]),
 145                                 struct bkey_cached, rcu);
 146         if (ck)
 147                 goto lock;
 148
 149         ck = allocate_dropping_locks(trans, ret,
 150                                      __bkey_cached_alloc(key_u64s, _gfp));
 151         if (ret) {
 152                 if (ck)
 153                         kfree(ck->k);
 154                 kmem_cache_free(bch2_key_cache, ck);
 155                 return ERR_PTR(ret);
 156         }
 157
 158         if (ck) {
 159                 bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
 160                 ck->c.cached = true;
 161                 goto lock;
 162         }
 163
 164         ck = container_of_or_null(rcu_pending_dequeue_from_all(&bc->pending[pcpu_readers]),
 165                                   struct bkey_cached, rcu);
 166         if (ck)
 167                 goto lock;
 168 lock:
 169         six_lock_intent(&ck->c.lock, NULL, NULL);
 170         six_lock_write(&ck->c.lock, NULL, NULL);
 171         return ck;
 172 }
 173
 174 static struct bkey_cached *
 175 bkey_cached_reuse(struct btree_key_cache *c)
 176 {
 177         struct bucket_table *tbl;
 178         struct rhash_head *pos;
 179         struct bkey_cached *ck;
 180         unsigned i;
 181
 182         rcu_read_lock();
 183         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 184         for (i = 0; i < tbl->size; i++)
 185                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 186                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 187                             bkey_cached_lock_for_evict(ck)) {
 188                                 if (bkey_cached_evict(c, ck))
 189                                         goto out;
 190                                 six_unlock_write(&ck->c.lock);
 191                                 six_unlock_intent(&ck->c.lock);
 192                         }
 193                 }
 194         ck = NULL;
 195 out:
 196         rcu_read_unlock();
 197         return ck;
 198 }
 199
 200 static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *path,
 201                                   struct bkey_s_c k)
 202 {
 203         struct bch_fs *c = trans->c;
 204         struct btree_key_cache *bc = &c->btree_key_cache;
 205
 206         /*
 207          * bch2_varint_decode can read past the end of the buffer by at
 208          * most 7 bytes (it won't be used):
 209          */
 210         unsigned key_u64s = k.k->u64s + 1;
 211
 212         /*
 213          * Allocate some extra space so that the transaction commit path is less
 214          * likely to have to reallocate, since that requires a transaction
 215          * restart:
 216          */
 217         key_u64s = min(256U, (key_u64s * 3) / 2);
 218         key_u64s = roundup_pow_of_two(key_u64s);
 219
 220         struct bkey_cached *ck = bkey_cached_alloc(trans, path, key_u64s);
 221         int ret = PTR_ERR_OR_ZERO(ck);
 222         if (ret)
 223                 return ret;
 224
 225         if (unlikely(!ck)) {
 226                 ck = bkey_cached_reuse(bc);
 227                 if (unlikely(!ck)) {
 228                         bch_err(c, "error allocating memory for key cache item, btree %s",
 229                                 bch2_btree_id_str(path->btree_id));
 230                         return -BCH_ERR_ENOMEM_btree_key_cache_create;
 231                 }
 232         }
 233
 234         ck->c.level             = 0;
 235         ck->c.btree_id          = path->btree_id;
 236         ck->key.btree_id        = path->btree_id;
 237         ck->key.pos             = path->pos;
 238         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 239
 240         if (unlikely(key_u64s > ck->u64s)) {
 241                 mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
 242
 243                 struct bkey_i *new_k = allocate_dropping_locks(trans, ret,
 244                                 kmalloc(key_u64s * sizeof(u64), _gfp));
 245                 if (unlikely(!new_k)) {
 246                         bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 247                                 bch2_btree_id_str(ck->key.btree_id), key_u64s);
 248                         ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
 249                 } else if (ret) {
 250                         kfree(new_k);
 251                         goto err;
 252                 }
 253
 254                 kfree(ck->k);
 255                 ck->k = new_k;
 256                 ck->u64s = key_u64s;
 257         }
 258
 259         bkey_reassemble(ck->k, k);
 260
 261         ret = rhashtable_lookup_insert_fast(&bc->table, &ck->hash, bch2_btree_key_cache_params);
 262         if (unlikely(ret)) /* raced with another fill? */
 263                 goto err;
 264
 265         atomic_long_inc(&bc->nr_keys);
 266         six_unlock_write(&ck->c.lock);
 267
 268         enum six_lock_type lock_want = __btree_lock_want(path, 0);
 269         if (lock_want == SIX_LOCK_read)
 270                 six_lock_downgrade(&ck->c.lock);
 271         btree_path_cached_set(trans, path, ck, (enum btree_node_locked_type) lock_want);
 272         path->uptodate = BTREE_ITER_UPTODATE;
 273         return 0;
 274 err:
 275         bkey_cached_free(bc, ck);
 276         mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
 277
 278         return ret;
 279 }
 280
 281 static noinline int btree_key_cache_fill(struct btree_trans *trans,
 282                                          struct btree_path *ck_path,
 283                                          unsigned flags)
 284 {
 285         if (flags & BTREE_ITER_cached_nofill) {
 286                 ck_path->uptodate = BTREE_ITER_UPTODATE;
 287                 return 0;
 288         }
 289
 290         struct bch_fs *c = trans->c;
 291         struct btree_iter iter;
 292         struct bkey_s_c k;
 293         int ret;
 294
 295         bch2_trans_iter_init(trans, &iter, ck_path->btree_id, ck_path->pos,
 296                              BTREE_ITER_key_cache_fill|
 297                              BTREE_ITER_cached_nofill);
 298         iter.flags &= ~BTREE_ITER_with_journal;
 299         k = bch2_btree_iter_peek_slot(&iter);
 300         ret = bkey_err(k);
 301         if (ret)
 302                 goto err;
 303
 304         /* Recheck after btree lookup, before allocating: */
 305         ret = bch2_btree_key_cache_find(c, ck_path->btree_id, ck_path->pos) ? -EEXIST : 0;
 306         if (unlikely(ret))
 307                 goto out;
 308
 309         ret = btree_key_cache_create(trans, ck_path, k);
 310         if (ret)
 311                 goto err;
 312 out:
 313         /* We're not likely to need this iterator again: */
 314         bch2_set_btree_iter_dontneed(&iter);
 315 err:
 316         bch2_trans_iter_exit(trans, &iter);
 317         return ret;
 318 }
 319
 320 static inline int btree_path_traverse_cached_fast(struct btree_trans *trans,
 321                                                   struct btree_path *path)
 322 {
 323         struct bch_fs *c = trans->c;
 324         struct bkey_cached *ck;
 325 retry:
 326         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 327         if (!ck)
 328                 return -ENOENT;
 329
 330         enum six_lock_type lock_want = __btree_lock_want(path, 0);
 331
 332         int ret = btree_node_lock(trans, path, (void *) ck, 0, lock_want, _THIS_IP_);
 333         if (ret)
 334                 return ret;
 335
 336         if (ck->key.btree_id != path->btree_id ||
 337             !bpos_eq(ck->key.pos, path->pos)) {
 338                 six_unlock_type(&ck->c.lock, lock_want);
 339                 goto retry;
 340         }
 341
 342         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 343                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 344
 345         btree_path_cached_set(trans, path, ck, (enum btree_node_locked_type) lock_want);
 346         path->uptodate = BTREE_ITER_UPTODATE;
 347         return 0;
 348 }
 349
 350 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 351                                     unsigned flags)
 352 {
 353         EBUG_ON(path->level);
 354
 355         path->l[1].b = NULL;
 356
 357         int ret;
 358         do {
 359                 ret = btree_path_traverse_cached_fast(trans, path);
 360                 if (unlikely(ret == -ENOENT))
 361                         ret = btree_key_cache_fill(trans, path, flags);
 362         } while (ret == -EEXIST);
 363
 364         if (unlikely(ret)) {
 365                 path->uptodate = BTREE_ITER_NEED_TRAVERSE;
 366                 if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 367                         btree_node_unlock(trans, path, 0);
 368                         path->l[0].b = ERR_PTR(ret);
 369                 }
 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_intent);
 393         b_iter.flags &= ~BTREE_ITER_with_key_cache;
 394
 395         ret = bch2_btree_iter_traverse(&c_iter);
 396         if (ret)
 397                 goto out;
 398
 399         ck = (void *) btree_iter_path(trans, &c_iter)->l[0].b;
 400         if (!ck)
 401                 goto out;
 402
 403         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 404                 if (evict)
 405                         goto evict;
 406                 goto out;
 407         }
 408
 409         if (journal_seq && ck->journal.seq != journal_seq)
 410                 goto out;
 411
 412         trans->journal_res.seq = ck->journal.seq;
 413
 414         /*
 415          * If we're at the end of the journal, we really want to free up space
 416          * in the journal right away - we don't want to pin that old journal
 417          * sequence number with a new btree node write, we want to re-journal
 418          * the update
 419          */
 420         if (ck->journal.seq == journal_last_seq(j))
 421                 commit_flags |= BCH_WATERMARK_reclaim;
 422
 423         if (ck->journal.seq != journal_last_seq(j) ||
 424             !test_bit(JOURNAL_space_low, &c->journal.flags))
 425                 commit_flags |= BCH_TRANS_COMMIT_no_journal_res;
 426
 427         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 428                 bch2_trans_update(trans, &b_iter, ck->k,
 429                                   BTREE_UPDATE_key_cache_reclaim|
 430                                   BTREE_UPDATE_internal_snapshot_node|
 431                                   BTREE_TRIGGER_norun) ?:
 432                 bch2_trans_commit(trans, NULL, NULL,
 433                                   BCH_TRANS_COMMIT_no_check_rw|
 434                                   BCH_TRANS_COMMIT_no_enospc|
 435                                   commit_flags);
 436
 437         bch2_fs_fatal_err_on(ret &&
 438                              !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 439                              !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
 440                              !bch2_journal_error(j), c,
 441                              "flushing key cache: %s", bch2_err_str(ret));
 442         if (ret)
 443                 goto out;
 444
 445         bch2_journal_pin_drop(j, &ck->journal);
 446
 447         struct btree_path *path = btree_iter_path(trans, &c_iter);
 448         BUG_ON(!btree_node_locked(path, 0));
 449
 450         if (!evict) {
 451                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 452                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 453                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 454                 }
 455         } else {
 456                 struct btree_path *path2;
 457                 unsigned i;
 458 evict:
 459                 trans_for_each_path(trans, path2, i)
 460                         if (path2 != path)
 461                                 __bch2_btree_path_unlock(trans, path2);
 462
 463                 bch2_btree_node_lock_write_nofail(trans, path, &ck->c);
 464
 465                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 466                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 467                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 468                 }
 469
 470                 mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
 471                 if (bkey_cached_evict(&c->btree_key_cache, ck)) {
 472                         bkey_cached_free(&c->btree_key_cache, ck);
 473                 } else {
 474                         six_unlock_write(&ck->c.lock);
 475                         six_unlock_intent(&ck->c.lock);
 476                 }
 477         }
 478 out:
 479         bch2_trans_iter_exit(trans, &b_iter);
 480         bch2_trans_iter_exit(trans, &c_iter);
 481         return ret;
 482 }
 483
 484 int bch2_btree_key_cache_journal_flush(struct journal *j,
 485                                 struct journal_entry_pin *pin, u64 seq)
 486 {
 487         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 488         struct bkey_cached *ck =
 489                 container_of(pin, struct bkey_cached, journal);
 490         struct bkey_cached_key key;
 491         struct btree_trans *trans = bch2_trans_get(c);
 492         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 493         int ret = 0;
 494
 495         btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
 496         key = ck->key;
 497
 498         if (ck->journal.seq != seq ||
 499             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 500                 six_unlock_read(&ck->c.lock);
 501                 goto unlock;
 502         }
 503
 504         if (ck->seq != seq) {
 505                 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
 506                                         bch2_btree_key_cache_journal_flush);
 507                 six_unlock_read(&ck->c.lock);
 508                 goto unlock;
 509         }
 510         six_unlock_read(&ck->c.lock);
 511
 512         ret = lockrestart_do(trans,
 513                 btree_key_cache_flush_pos(trans, key, seq,
 514                                 BCH_TRANS_COMMIT_journal_reclaim, false));
 515 unlock:
 516         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 517
 518         bch2_trans_put(trans);
 519         return ret;
 520 }
 521
 522 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 523                                   unsigned flags,
 524                                   struct btree_insert_entry *insert_entry)
 525 {
 526         struct bch_fs *c = trans->c;
 527         struct bkey_cached *ck = (void *) (trans->paths + insert_entry->path)->l[0].b;
 528         struct bkey_i *insert = insert_entry->k;
 529         bool kick_reclaim = false;
 530
 531         BUG_ON(insert->k.u64s > ck->u64s);
 532
 533         bkey_copy(ck->k, insert);
 534
 535         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 536                 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
 537                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 538                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 539
 540                 if (bch2_nr_btree_keys_need_flush(c))
 541                         kick_reclaim = true;
 542         }
 543
 544         /*
 545          * To minimize lock contention, we only add the journal pin here and
 546          * defer pin updates to the flush callback via ->seq. Be careful not to
 547          * update ->seq on nojournal commits because we don't want to update the
 548          * pin to a seq that doesn't include journal updates on disk. Otherwise
 549          * we risk losing the update after a crash.
 550          *
 551          * The only exception is if the pin is not active in the first place. We
 552          * have to add the pin because journal reclaim drives key cache
 553          * flushing. The flush callback will not proceed unless ->seq matches
 554          * the latest pin, so make sure it starts with a consistent value.
 555          */
 556         if (!(insert_entry->flags & BTREE_UPDATE_nojournal) ||
 557             !journal_pin_active(&ck->journal)) {
 558                 ck->seq = trans->journal_res.seq;
 559         }
 560         bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 561                              &ck->journal, bch2_btree_key_cache_journal_flush);
 562
 563         if (kick_reclaim)
 564                 journal_reclaim_kick(&c->journal);
 565         return true;
 566 }
 567
 568 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 569                                struct btree_path *path)
 570 {
 571         struct bch_fs *c = trans->c;
 572         struct btree_key_cache *bc = &c->btree_key_cache;
 573         struct bkey_cached *ck = (void *) path->l[0].b;
 574
 575         /*
 576          * We just did an update to the btree, bypassing the key cache: the key
 577          * cache key is now stale and must be dropped, even if dirty:
 578          */
 579         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 580                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 581                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 582                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 583         }
 584
 585         bkey_cached_evict(bc, ck);
 586         bkey_cached_free(bc, ck);
 587
 588         mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
 589         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 590         path->should_be_locked = false;
 591 }
 592
 593 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 594                                            struct shrink_control *sc)
 595 {
 596         struct bch_fs *c = shrink->private_data;
 597         struct btree_key_cache *bc = &c->btree_key_cache;
 598         struct bucket_table *tbl;
 599         struct bkey_cached *ck;
 600         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 601         unsigned iter, start;
 602         int srcu_idx;
 603
 604         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 605         rcu_read_lock();
 606
 607         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 608
 609         /*
 610          * Scanning is expensive while a rehash is in progress - most elements
 611          * will be on the new hashtable, if it's in progress
 612          *
 613          * A rehash could still start while we're scanning - that's ok, we'll
 614          * still see most elements.
 615          */
 616         if (unlikely(tbl->nest)) {
 617                 rcu_read_unlock();
 618                 srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 619                 return SHRINK_STOP;
 620         }
 621
 622         iter = bc->shrink_iter;
 623         if (iter >= tbl->size)
 624                 iter = 0;
 625         start = iter;
 626
 627         do {
 628                 struct rhash_head *pos, *next;
 629
 630                 pos = rht_ptr_rcu(&tbl->buckets[iter]);
 631
 632                 while (!rht_is_a_nulls(pos)) {
 633                         next = rht_dereference_bucket_rcu(pos->next, tbl, iter);
 634                         ck = container_of(pos, struct bkey_cached, hash);
 635
 636                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 637                                 bc->skipped_dirty++;
 638                         } else if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags)) {
 639                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 640                                 bc->skipped_accessed++;
 641                         } else if (!bkey_cached_lock_for_evict(ck)) {
 642                                 bc->skipped_lock_fail++;
 643                         } else if (bkey_cached_evict(bc, ck)) {
 644                                 bkey_cached_free(bc, ck);
 645                                 bc->freed++;
 646                                 freed++;
 647                         } else {
 648                                 six_unlock_write(&ck->c.lock);
 649                                 six_unlock_intent(&ck->c.lock);
 650                         }
 651
 652                         scanned++;
 653                         if (scanned >= nr)
 654                                 goto out;
 655
 656                         pos = next;
 657                 }
 658
 659                 iter++;
 660                 if (iter >= tbl->size)
 661                         iter = 0;
 662         } while (scanned < nr && iter != start);
 663 out:
 664         bc->shrink_iter = iter;
 665
 666         rcu_read_unlock();
 667         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 668
 669         return freed;
 670 }
 671
 672 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 673                                             struct shrink_control *sc)
 674 {
 675         struct bch_fs *c = shrink->private_data;
 676         struct btree_key_cache *bc = &c->btree_key_cache;
 677         long nr = atomic_long_read(&bc->nr_keys) -
 678                 atomic_long_read(&bc->nr_dirty);
 679
 680         /*
 681          * Avoid hammering our shrinker too much if it's nearly empty - the
 682          * shrinker code doesn't take into account how big our cache is, if it's
 683          * mostly empty but the system is under memory pressure it causes nasty
 684          * lock contention:
 685          */
 686         nr -= 128;
 687
 688         return max(0L, nr);
 689 }
 690
 691 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 692 {
 693         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 694         struct bucket_table *tbl;
 695         struct bkey_cached *ck;
 696         struct rhash_head *pos;
 697         LIST_HEAD(items);
 698         unsigned i;
 699
 700         shrinker_free(bc->shrink);
 701
 702         /*
 703          * The loop is needed to guard against racing with rehash:
 704          */
 705         while (atomic_long_read(&bc->nr_keys)) {
 706                 rcu_read_lock();
 707                 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 708                 if (tbl) {
 709                         if (tbl->nest) {
 710                                 /* wait for in progress rehash */
 711                                 rcu_read_unlock();
 712                                 mutex_lock(&bc->table.mutex);
 713                                 mutex_unlock(&bc->table.mutex);
 714                                 rcu_read_lock();
 715                                 continue;
 716                         }
 717                         for (i = 0; i < tbl->size; i++)
 718                                 while (pos = rht_ptr_rcu(&tbl->buckets[i]), !rht_is_a_nulls(pos)) {
 719                                         ck = container_of(pos, struct bkey_cached, hash);
 720                                         BUG_ON(!bkey_cached_evict(bc, ck));
 721                                         kfree(ck->k);
 722                                         kmem_cache_free(bch2_key_cache, ck);
 723                                 }
 724                 }
 725                 rcu_read_unlock();
 726         }
 727
 728         if (atomic_long_read(&bc->nr_dirty) &&
 729             !bch2_journal_error(&c->journal) &&
 730             test_bit(BCH_FS_was_rw, &c->flags))
 731                 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
 732                       atomic_long_read(&bc->nr_dirty));
 733
 734         if (atomic_long_read(&bc->nr_keys))
 735                 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
 736                       atomic_long_read(&bc->nr_keys));
 737
 738         if (bc->table_init_done)
 739                 rhashtable_destroy(&bc->table);
 740
 741         rcu_pending_exit(&bc->pending[0]);
 742         rcu_pending_exit(&bc->pending[1]);
 743
 744         free_percpu(bc->nr_pending);
 745 }
 746
 747 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 748 {
 749 }
 750
 751 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
 752 {
 753         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 754         struct shrinker *shrink;
 755
 756         bc->nr_pending = alloc_percpu(size_t);
 757         if (!bc->nr_pending)
 758                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
 759
 760         if (rcu_pending_init(&bc->pending[0], &c->btree_trans_barrier, __bkey_cached_free) ||
 761             rcu_pending_init(&bc->pending[1], &c->btree_trans_barrier, __bkey_cached_free))
 762                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
 763
 764         if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
 765                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
 766
 767         bc->table_init_done = true;
 768
 769         shrink = shrinker_alloc(0, "%s-btree_key_cache", c->name);
 770         if (!shrink)
 771                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
 772         bc->shrink = shrink;
 773         shrink->count_objects   = bch2_btree_key_cache_count;
 774         shrink->scan_objects    = bch2_btree_key_cache_scan;
 775         shrink->batch           = 1 << 14;
 776         shrink->seeks           = 0;
 777         shrink->private_data    = c;
 778         shrinker_register(shrink);
 779         return 0;
 780 }
 781
 782 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *bc)
 783 {
 784         printbuf_tabstop_push(out, 24);
 785         printbuf_tabstop_push(out, 12);
 786
 787         prt_printf(out, "keys:\t%lu\r\n",               atomic_long_read(&bc->nr_keys));
 788         prt_printf(out, "dirty:\t%lu\r\n",              atomic_long_read(&bc->nr_dirty));
 789         prt_printf(out, "table size:\t%u\r\n",          bc->table.tbl->size);
 790         prt_newline(out);
 791         prt_printf(out, "shrinker:\n");
 792         prt_printf(out, "requested_to_free:\t%lu\r\n",  bc->requested_to_free);
 793         prt_printf(out, "freed:\t%lu\r\n",              bc->freed);
 794         prt_printf(out, "skipped_dirty:\t%lu\r\n",      bc->skipped_dirty);
 795         prt_printf(out, "skipped_accessed:\t%lu\r\n",   bc->skipped_accessed);
 796         prt_printf(out, "skipped_lock_fail:\t%lu\r\n",  bc->skipped_lock_fail);
 797         prt_newline(out);
 798         prt_printf(out, "pending:\t%zu\r\n",            per_cpu_sum(bc->nr_pending));
 799 }
 800
 801 void bch2_btree_key_cache_exit(void)
 802 {
 803         kmem_cache_destroy(bch2_key_cache);
 804 }
 805
 806 int __init bch2_btree_key_cache_init(void)
 807 {
 808         bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
 809         if (!bch2_key_cache)
 810                 return -ENOMEM;
 811
 812         return 0;
 813 }