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         atomic_long_inc(&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 void bkey_cached_free_fast(struct btree_key_cache *bc,
  98                                   struct bkey_cached *ck)
  99 {
 100         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 101         struct btree_key_cache_freelist *f;
 102         bool freed = false;
 103
 104         BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
 105
 106         ck->btree_trans_barrier_seq =
 107                 start_poll_synchronize_srcu(&c->btree_trans_barrier);
 108
 109         list_del_init(&ck->list);
 110         atomic_long_inc(&bc->nr_freed);
 111
 112         kfree(ck->k);
 113         ck->k           = NULL;
 114         ck->u64s        = 0;
 115
 116         preempt_disable();
 117         f = this_cpu_ptr(bc->pcpu_freed);
 118
 119         if (f->nr < ARRAY_SIZE(f->objs)) {
 120                 f->objs[f->nr++] = ck;
 121                 freed = true;
 122         }
 123         preempt_enable();
 124
 125         if (!freed) {
 126                 mutex_lock(&bc->lock);
 127                 preempt_disable();
 128                 f = this_cpu_ptr(bc->pcpu_freed);
 129
 130                 while (f->nr > ARRAY_SIZE(f->objs) / 2) {
 131                         struct bkey_cached *ck2 = f->objs[--f->nr];
 132
 133                         list_move_tail(&ck2->list, &bc->freed);
 134                 }
 135                 preempt_enable();
 136
 137                 list_move_tail(&ck->list, &bc->freed);
 138                 mutex_unlock(&bc->lock);
 139         }
 140
 141         six_unlock_write(&ck->c.lock);
 142         six_unlock_intent(&ck->c.lock);
 143 }
 144
 145 static struct bkey_cached *
 146 bkey_cached_alloc(struct btree_key_cache *c)
 147 {
 148         struct bkey_cached *ck = NULL;
 149         struct btree_key_cache_freelist *f;
 150
 151         preempt_disable();
 152         f = this_cpu_ptr(c->pcpu_freed);
 153         if (f->nr)
 154                 ck = f->objs[--f->nr];
 155         preempt_enable();
 156
 157         if (!ck) {
 158                 mutex_lock(&c->lock);
 159                 preempt_disable();
 160                 f = this_cpu_ptr(c->pcpu_freed);
 161
 162                 while (!list_empty(&c->freed) &&
 163                        f->nr < ARRAY_SIZE(f->objs) / 2) {
 164                         ck = list_last_entry(&c->freed, struct bkey_cached, list);
 165                         list_del_init(&ck->list);
 166                         f->objs[f->nr++] = ck;
 167                 }
 168
 169                 ck = f->nr ? f->objs[--f->nr] : NULL;
 170                 preempt_enable();
 171                 mutex_unlock(&c->lock);
 172         }
 173
 174         if (ck) {
 175                 six_lock_intent(&ck->c.lock, NULL, NULL);
 176                 six_lock_write(&ck->c.lock, NULL, NULL);
 177                 return ck;
 178         }
 179
 180         ck = kmem_cache_alloc(bch2_key_cache, GFP_NOFS|__GFP_ZERO);
 181         if (likely(ck)) {
 182                 INIT_LIST_HEAD(&ck->list);
 183                 six_lock_init(&ck->c.lock);
 184                 BUG_ON(!six_trylock_intent(&ck->c.lock));
 185                 BUG_ON(!six_trylock_write(&ck->c.lock));
 186                 return ck;
 187         }
 188
 189         return NULL;
 190 }
 191
 192 static struct bkey_cached *
 193 bkey_cached_reuse(struct btree_key_cache *c)
 194 {
 195         struct bucket_table *tbl;
 196         struct rhash_head *pos;
 197         struct bkey_cached *ck;
 198         unsigned i;
 199
 200         rcu_read_lock();
 201         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 202         for (i = 0; i < tbl->size; i++)
 203                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 204                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 205                             bkey_cached_lock_for_evict(ck)) {
 206                                 bkey_cached_evict(c, ck);
 207                                 rcu_read_unlock();
 208                                 return ck;
 209                         }
 210                 }
 211         rcu_read_unlock();
 212
 213         return NULL;
 214 }
 215
 216 static struct bkey_cached *
 217 btree_key_cache_create(struct bch_fs *c,
 218                        enum btree_id btree_id,
 219                        struct bpos pos)
 220 {
 221         struct btree_key_cache *bc = &c->btree_key_cache;
 222         struct bkey_cached *ck;
 223         bool was_new = true;
 224
 225         ck = bkey_cached_alloc(bc);
 226
 227         if (unlikely(!ck)) {
 228                 ck = bkey_cached_reuse(bc);
 229                 if (unlikely(!ck)) {
 230                         bch_err(c, "error allocating memory for key cache item, btree %s",
 231                                 bch2_btree_ids[btree_id]);
 232                         return ERR_PTR(-ENOMEM);
 233                 }
 234
 235                 was_new = false;
 236         } else {
 237                 if (btree_id == BTREE_ID_subvolumes)
 238                         six_lock_pcpu_alloc(&ck->c.lock);
 239                 else
 240                         six_lock_pcpu_free(&ck->c.lock);
 241         }
 242
 243         ck->c.level             = 0;
 244         ck->c.btree_id          = btree_id;
 245         ck->key.btree_id        = btree_id;
 246         ck->key.pos             = pos;
 247         ck->valid               = false;
 248         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 249
 250         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 251                                           &ck->hash,
 252                                           bch2_btree_key_cache_params))) {
 253                 /* We raced with another fill: */
 254
 255                 if (likely(was_new)) {
 256                         six_unlock_write(&ck->c.lock);
 257                         six_unlock_intent(&ck->c.lock);
 258                         kfree(ck);
 259                 } else {
 260                         bkey_cached_free_fast(bc, ck);
 261                 }
 262
 263                 return NULL;
 264         }
 265
 266         atomic_long_inc(&bc->nr_keys);
 267
 268         six_unlock_write(&ck->c.lock);
 269
 270         return ck;
 271 }
 272
 273 static int btree_key_cache_fill(struct btree_trans *trans,
 274                                 struct btree_path *ck_path,
 275                                 struct bkey_cached *ck)
 276 {
 277         struct btree_path *path;
 278         struct bkey_s_c k;
 279         unsigned new_u64s = 0;
 280         struct bkey_i *new_k = NULL;
 281         struct bkey u;
 282         int ret;
 283
 284         path = bch2_path_get(trans, ck->key.btree_id,
 285                              ck->key.pos, 0, 0, 0, _THIS_IP_);
 286         ret = bch2_btree_path_traverse(trans, path, 0);
 287         if (ret)
 288                 goto err;
 289
 290         k = bch2_btree_path_peek_slot(path, &u);
 291
 292         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 293                 trace_trans_restart_relock_key_cache_fill(trans->fn,
 294                                 _THIS_IP_, ck_path->btree_id, &ck_path->pos);
 295                 ret = btree_trans_restart(trans);
 296                 goto err;
 297         }
 298
 299         /*
 300          * bch2_varint_decode can read past the end of the buffer by at
 301          * most 7 bytes (it won't be used):
 302          */
 303         new_u64s = k.k->u64s + 1;
 304
 305         /*
 306          * Allocate some extra space so that the transaction commit path is less
 307          * likely to have to reallocate, since that requires a transaction
 308          * restart:
 309          */
 310         new_u64s = min(256U, (new_u64s * 3) / 2);
 311
 312         if (new_u64s > ck->u64s) {
 313                 new_u64s = roundup_pow_of_two(new_u64s);
 314                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 315                 if (!new_k) {
 316                         bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 317                                 bch2_btree_ids[ck->key.btree_id], new_u64s);
 318                         ret = -ENOMEM;
 319                         goto err;
 320                 }
 321         }
 322
 323         /*
 324          * XXX: not allowed to be holding read locks when we take a write lock,
 325          * currently
 326          */
 327         bch2_btree_node_lock_write(trans, ck_path, ck_path->l[0].b);
 328         if (new_k) {
 329                 kfree(ck->k);
 330                 ck->u64s = new_u64s;
 331                 ck->k = new_k;
 332         }
 333
 334         bkey_reassemble(ck->k, k);
 335         ck->valid = true;
 336         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 337
 338         /* We're not likely to need this iterator again: */
 339         path->preserve = false;
 340 err:
 341         bch2_path_put(trans, path, 0);
 342         return ret;
 343 }
 344
 345 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 346 {
 347         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 348         const struct btree_path *path = p;
 349
 350         return ck->key.btree_id == path->btree_id &&
 351                 !bpos_cmp(ck->key.pos, path->pos) ? 0 : -1;
 352 }
 353
 354 __flatten
 355 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 356                                     unsigned flags)
 357 {
 358         struct bch_fs *c = trans->c;
 359         struct bkey_cached *ck;
 360         int ret = 0;
 361
 362         BUG_ON(path->level);
 363
 364         path->l[1].b = NULL;
 365
 366         if (bch2_btree_node_relock(trans, path, 0)) {
 367                 ck = (void *) path->l[0].b;
 368                 goto fill;
 369         }
 370 retry:
 371         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 372         if (!ck) {
 373                 if (flags & BTREE_ITER_CACHED_NOCREATE) {
 374                         path->l[0].b = NULL;
 375                         return 0;
 376                 }
 377
 378                 ck = btree_key_cache_create(c, path->btree_id, path->pos);
 379                 ret = PTR_ERR_OR_ZERO(ck);
 380                 if (ret)
 381                         goto err;
 382                 if (!ck)
 383                         goto retry;
 384
 385                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 386                 path->locks_want = 1;
 387         } else {
 388                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 389
 390                 if (!btree_node_lock(trans, path, (void *) ck, path->pos, 0,
 391                                      lock_want,
 392                                      bkey_cached_check_fn, path, _THIS_IP_)) {
 393                         if (!trans->restarted)
 394                                 goto retry;
 395
 396                         ret = -EINTR;
 397                         goto err;
 398                 }
 399
 400                 if (ck->key.btree_id != path->btree_id ||
 401                     bpos_cmp(ck->key.pos, path->pos)) {
 402                         six_unlock_type(&ck->c.lock, lock_want);
 403                         goto retry;
 404                 }
 405
 406                 mark_btree_node_locked(trans, path, 0, lock_want);
 407         }
 408
 409         path->l[0].lock_seq     = ck->c.lock.state.seq;
 410         path->l[0].b            = (void *) ck;
 411 fill:
 412         if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
 413                 if (!path->locks_want &&
 414                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 415                         trace_transaction_restart_ip(trans->fn, _THIS_IP_);
 416                         ret = btree_trans_restart(trans);
 417                         goto err;
 418                 }
 419
 420                 ret = btree_key_cache_fill(trans, path, ck);
 421                 if (ret)
 422                         goto err;
 423         }
 424
 425         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 426                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 427
 428         path->uptodate = BTREE_ITER_UPTODATE;
 429         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 430
 431         return ret;
 432 err:
 433         if (ret != -EINTR) {
 434                 btree_node_unlock(path, 0);
 435                 path->l[0].b = BTREE_ITER_NO_NODE_ERROR;
 436         }
 437         return ret;
 438 }
 439
 440 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 441                                      struct bkey_cached_key key,
 442                                      u64 journal_seq,
 443                                      unsigned commit_flags,
 444                                      bool evict)
 445 {
 446         struct bch_fs *c = trans->c;
 447         struct journal *j = &c->journal;
 448         struct btree_iter c_iter, b_iter;
 449         struct bkey_cached *ck = NULL;
 450         int ret;
 451
 452         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 453                              BTREE_ITER_SLOTS|
 454                              BTREE_ITER_INTENT|
 455                              BTREE_ITER_ALL_SNAPSHOTS);
 456         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 457                              BTREE_ITER_CACHED|
 458                              BTREE_ITER_CACHED_NOFILL|
 459                              BTREE_ITER_CACHED_NOCREATE|
 460                              BTREE_ITER_INTENT);
 461         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 462
 463         ret = bch2_btree_iter_traverse(&c_iter);
 464         if (ret)
 465                 goto out;
 466
 467         ck = (void *) c_iter.path->l[0].b;
 468         if (!ck)
 469                 goto out;
 470
 471         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 472                 if (evict)
 473                         goto evict;
 474                 goto out;
 475         }
 476
 477         BUG_ON(!ck->valid);
 478
 479         if (journal_seq && ck->journal.seq != journal_seq)
 480                 goto out;
 481
 482         /*
 483          * Since journal reclaim depends on us making progress here, and the
 484          * allocator/copygc depend on journal reclaim making progress, we need
 485          * to be using alloc reserves:
 486          * */
 487         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 488                 bch2_trans_update(trans, &b_iter, ck->k,
 489                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 490                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 491                                   BTREE_TRIGGER_NORUN) ?:
 492                 bch2_trans_commit(trans, NULL, NULL,
 493                                   BTREE_INSERT_NOCHECK_RW|
 494                                   BTREE_INSERT_NOFAIL|
 495                                   BTREE_INSERT_USE_RESERVE|
 496                                   (ck->journal.seq == journal_last_seq(j)
 497                                    ? JOURNAL_WATERMARK_reserved
 498                                    : 0)|
 499                                   commit_flags);
 500         if (ret) {
 501                 bch2_fs_fatal_err_on(ret != -EINTR &&
 502                                      ret != -EAGAIN &&
 503                                      !bch2_journal_error(j), c,
 504                         "error flushing key cache: %i", ret);
 505                 goto out;
 506         }
 507
 508         bch2_journal_pin_drop(j, &ck->journal);
 509         bch2_journal_preres_put(j, &ck->res);
 510
 511         BUG_ON(!btree_node_locked(c_iter.path, 0));
 512
 513         if (!evict) {
 514                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 515                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 516                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 517                 }
 518         } else {
 519 evict:
 520                 BUG_ON(!btree_node_intent_locked(c_iter.path, 0));
 521
 522                 mark_btree_node_unlocked(c_iter.path, 0);
 523                 c_iter.path->l[0].b = NULL;
 524
 525                 six_lock_write(&ck->c.lock, NULL, NULL);
 526
 527                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 528                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 529                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 530                 }
 531
 532                 bkey_cached_evict(&c->btree_key_cache, ck);
 533
 534                 bkey_cached_free_fast(&c->btree_key_cache, ck);
 535         }
 536 out:
 537         bch2_trans_iter_exit(trans, &b_iter);
 538         bch2_trans_iter_exit(trans, &c_iter);
 539         return ret;
 540 }
 541
 542 int bch2_btree_key_cache_journal_flush(struct journal *j,
 543                                 struct journal_entry_pin *pin, u64 seq)
 544 {
 545         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 546         struct bkey_cached *ck =
 547                 container_of(pin, struct bkey_cached, journal);
 548         struct bkey_cached_key key;
 549         int ret = 0;
 550
 551         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 552
 553         six_lock_read(&ck->c.lock, NULL, NULL);
 554         key = ck->key;
 555
 556         if (ck->journal.seq != seq ||
 557             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 558                 six_unlock_read(&ck->c.lock);
 559                 goto unlock;
 560         }
 561         six_unlock_read(&ck->c.lock);
 562
 563         ret = bch2_trans_do(c, NULL, NULL, 0,
 564                 btree_key_cache_flush_pos(&trans, key, seq,
 565                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 566 unlock:
 567         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 568
 569         return ret;
 570 }
 571
 572 /*
 573  * Flush and evict a key from the key cache:
 574  */
 575 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 576                                enum btree_id id, struct bpos pos)
 577 {
 578         struct bch_fs *c = trans->c;
 579         struct bkey_cached_key key = { id, pos };
 580
 581         /* Fastpath - assume it won't be found: */
 582         if (!bch2_btree_key_cache_find(c, id, pos))
 583                 return 0;
 584
 585         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 586 }
 587
 588 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 589                                   struct btree_path *path,
 590                                   struct bkey_i *insert)
 591 {
 592         struct bch_fs *c = trans->c;
 593         struct bkey_cached *ck = (void *) path->l[0].b;
 594         bool kick_reclaim = false;
 595
 596         BUG_ON(insert->u64s > ck->u64s);
 597
 598         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 599                 int difference;
 600
 601                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 602
 603                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 604                 if (difference > 0) {
 605                         trans->journal_preres.u64s      -= difference;
 606                         ck->res.u64s                    += difference;
 607                 }
 608         }
 609
 610         bkey_copy(ck->k, insert);
 611         ck->valid = true;
 612
 613         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 614                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 615                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 616
 617                 if (bch2_nr_btree_keys_need_flush(c))
 618                         kick_reclaim = true;
 619         }
 620
 621         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 622                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 623
 624         if (kick_reclaim)
 625                 journal_reclaim_kick(&c->journal);
 626         return true;
 627 }
 628
 629 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 630                                struct btree_path *path)
 631 {
 632         struct bkey_cached *ck = (void *) path->l[0].b;
 633
 634         ck->valid = false;
 635
 636         BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
 637 }
 638
 639 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 640                                            struct shrink_control *sc)
 641 {
 642         struct bch_fs *c = container_of(shrink, struct bch_fs,
 643                                         btree_key_cache.shrink);
 644         struct btree_key_cache *bc = &c->btree_key_cache;
 645         struct bucket_table *tbl;
 646         struct bkey_cached *ck, *t;
 647         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 648         unsigned start, flags;
 649         int srcu_idx;
 650
 651         /* Return -1 if we can't do anything right now */
 652         if (sc->gfp_mask & __GFP_FS)
 653                 mutex_lock(&bc->lock);
 654         else if (!mutex_trylock(&bc->lock))
 655                 return -1;
 656
 657         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 658         flags = memalloc_nofs_save();
 659
 660         /*
 661          * Newest freed entries are at the end of the list - once we hit one
 662          * that's too new to be freed, we can bail out:
 663          */
 664         list_for_each_entry_safe(ck, t, &bc->freed, list) {
 665                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 666                                                  ck->btree_trans_barrier_seq))
 667                         break;
 668
 669                 list_del(&ck->list);
 670                 kmem_cache_free(bch2_key_cache, ck);
 671                 atomic_long_dec(&bc->nr_freed);
 672                 scanned++;
 673                 freed++;
 674         }
 675
 676         if (scanned >= nr)
 677                 goto out;
 678
 679         rcu_read_lock();
 680         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 681         if (bc->shrink_iter >= tbl->size)
 682                 bc->shrink_iter = 0;
 683         start = bc->shrink_iter;
 684
 685         do {
 686                 struct rhash_head *pos, *next;
 687
 688                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 689
 690                 while (!rht_is_a_nulls(pos)) {
 691                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 692                         ck = container_of(pos, struct bkey_cached, hash);
 693
 694                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 695                                 goto next;
 696
 697                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 698                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 699                         else if (bkey_cached_lock_for_evict(ck)) {
 700                                 bkey_cached_evict(bc, ck);
 701                                 bkey_cached_free(bc, ck);
 702                         }
 703
 704                         scanned++;
 705                         if (scanned >= nr)
 706                                 break;
 707 next:
 708                         pos = next;
 709                 }
 710
 711                 bc->shrink_iter++;
 712                 if (bc->shrink_iter >= tbl->size)
 713                         bc->shrink_iter = 0;
 714         } while (scanned < nr && bc->shrink_iter != start);
 715
 716         rcu_read_unlock();
 717 out:
 718         memalloc_nofs_restore(flags);
 719         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 720         mutex_unlock(&bc->lock);
 721
 722         return freed;
 723 }
 724
 725 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 726                                             struct shrink_control *sc)
 727 {
 728         struct bch_fs *c = container_of(shrink, struct bch_fs,
 729                                         btree_key_cache.shrink);
 730         struct btree_key_cache *bc = &c->btree_key_cache;
 731         long nr = atomic_long_read(&bc->nr_keys) -
 732                 atomic_long_read(&bc->nr_dirty);
 733
 734         return max(0L, nr);
 735 }
 736
 737 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 738 {
 739         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 740         struct bucket_table *tbl;
 741         struct bkey_cached *ck, *n;
 742         struct rhash_head *pos;
 743         unsigned i;
 744         int cpu;
 745
 746         if (bc->shrink.list.next)
 747                 unregister_shrinker(&bc->shrink);
 748
 749         mutex_lock(&bc->lock);
 750
 751         rcu_read_lock();
 752         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 753         if (tbl)
 754                 for (i = 0; i < tbl->size; i++)
 755                         rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 756                                 bkey_cached_evict(bc, ck);
 757                                 list_add(&ck->list, &bc->freed);
 758                         }
 759         rcu_read_unlock();
 760
 761         for_each_possible_cpu(cpu) {
 762                 struct btree_key_cache_freelist *f =
 763                         per_cpu_ptr(bc->pcpu_freed, cpu);
 764
 765                 for (i = 0; i < f->nr; i++) {
 766                         ck = f->objs[i];
 767                         list_add(&ck->list, &bc->freed);
 768                 }
 769         }
 770
 771         list_for_each_entry_safe(ck, n, &bc->freed, list) {
 772                 cond_resched();
 773
 774                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 775                 bch2_journal_preres_put(&c->journal, &ck->res);
 776
 777                 list_del(&ck->list);
 778                 kfree(ck->k);
 779                 kmem_cache_free(bch2_key_cache, ck);
 780         }
 781
 782         BUG_ON(atomic_long_read(&bc->nr_dirty) &&
 783                !bch2_journal_error(&c->journal) &&
 784                test_bit(BCH_FS_WAS_RW, &c->flags));
 785         BUG_ON(atomic_long_read(&bc->nr_keys));
 786
 787         mutex_unlock(&bc->lock);
 788
 789         if (bc->table_init_done)
 790                 rhashtable_destroy(&bc->table);
 791
 792         free_percpu(bc->pcpu_freed);
 793 }
 794
 795 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 796 {
 797         mutex_init(&c->lock);
 798         INIT_LIST_HEAD(&c->freed);
 799 }
 800
 801 static void bch2_btree_key_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 802 {
 803         struct btree_key_cache *bc =
 804                 container_of(shrink, struct btree_key_cache, shrink);
 805
 806         bch2_btree_key_cache_to_text(out, bc);
 807 }
 808
 809 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
 810 {
 811         int ret;
 812
 813         c->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
 814         if (!c->pcpu_freed)
 815                 return -ENOMEM;
 816
 817         ret = rhashtable_init(&c->table, &bch2_btree_key_cache_params);
 818         if (ret)
 819                 return ret;
 820
 821         c->table_init_done = true;
 822
 823         c->shrink.seeks                 = 1;
 824         c->shrink.count_objects         = bch2_btree_key_cache_count;
 825         c->shrink.scan_objects          = bch2_btree_key_cache_scan;
 826         c->shrink.to_text               = bch2_btree_key_cache_shrinker_to_text;
 827         return register_shrinker(&c->shrink);
 828 }
 829
 830 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 831 {
 832         prt_printf(out, "nr_freed:\t%zu\n",     atomic_long_read(&c->nr_freed));
 833         prt_printf(out, "nr_keys:\t%lu\n",      atomic_long_read(&c->nr_keys));
 834         prt_printf(out, "nr_dirty:\t%lu\n",     atomic_long_read(&c->nr_dirty));
 835 }
 836
 837 void bch2_btree_key_cache_exit(void)
 838 {
 839         if (bch2_key_cache)
 840                 kmem_cache_destroy(bch2_key_cache);
 841 }
 842
 843 int __init bch2_btree_key_cache_init(void)
 844 {
 845         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
 846         if (!bch2_key_cache)
 847                 return -ENOMEM;
 848
 849         return 0;
 850 }