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 "errcode.h"
   9 #include "error.h"
  10 #include "journal.h"
  11 #include "journal_reclaim.h"
  12
  13 #include <linux/sched/mm.h>
  14 #include <trace/events/bcachefs.h>
  15
  16 static inline bool btree_uses_pcpu_readers(enum btree_id id)
  17 {
  18         return id == BTREE_ID_subvolumes;
  19 }
  20
  21 static struct kmem_cache *bch2_key_cache;
  22
  23 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  24                                        const void *obj)
  25 {
  26         const struct bkey_cached *ck = obj;
  27         const struct bkey_cached_key *key = arg->key;
  28
  29         return cmp_int(ck->key.btree_id, key->btree_id) ?:
  30                 bpos_cmp(ck->key.pos, key->pos);
  31 }
  32
  33 static const struct rhashtable_params bch2_btree_key_cache_params = {
  34         .head_offset    = offsetof(struct bkey_cached, hash),
  35         .key_offset     = offsetof(struct bkey_cached, key),
  36         .key_len        = sizeof(struct bkey_cached_key),
  37         .obj_cmpfn      = bch2_btree_key_cache_cmp_fn,
  38 };
  39
  40 __flatten
  41 inline struct bkey_cached *
  42 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
  43 {
  44         struct bkey_cached_key key = {
  45                 .btree_id       = btree_id,
  46                 .pos            = pos,
  47         };
  48
  49         return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
  50                                       bch2_btree_key_cache_params);
  51 }
  52
  53 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
  54 {
  55         if (!six_trylock_intent(&ck->c.lock))
  56                 return false;
  57
  58         if (!six_trylock_write(&ck->c.lock)) {
  59                 six_unlock_intent(&ck->c.lock);
  60                 return false;
  61         }
  62
  63         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
  64                 six_unlock_write(&ck->c.lock);
  65                 six_unlock_intent(&ck->c.lock);
  66                 return false;
  67         }
  68
  69         return true;
  70 }
  71
  72 static void bkey_cached_evict(struct btree_key_cache *c,
  73                               struct bkey_cached *ck)
  74 {
  75         BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
  76                                       bch2_btree_key_cache_params));
  77         memset(&ck->key, ~0, sizeof(ck->key));
  78
  79         atomic_long_dec(&c->nr_keys);
  80 }
  81
  82 static void bkey_cached_free(struct btree_key_cache *bc,
  83                              struct bkey_cached *ck)
  84 {
  85         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
  86
  87         BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
  88
  89         ck->btree_trans_barrier_seq =
  90                 start_poll_synchronize_srcu(&c->btree_trans_barrier);
  91
  92         if (ck->c.lock.readers)
  93                 list_move_tail(&ck->list, &bc->freed_pcpu);
  94         else
  95                 list_move_tail(&ck->list, &bc->freed_nonpcpu);
  96         atomic_long_inc(&bc->nr_freed);
  97
  98         kfree(ck->k);
  99         ck->k           = NULL;
 100         ck->u64s        = 0;
 101
 102         six_unlock_write(&ck->c.lock);
 103         six_unlock_intent(&ck->c.lock);
 104 }
 105
 106 static void bkey_cached_move_to_freelist(struct btree_key_cache *bc,
 107                                          struct bkey_cached *ck)
 108 {
 109         struct btree_key_cache_freelist *f;
 110         bool freed = false;
 111
 112         BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
 113
 114         if (!ck->c.lock.readers) {
 115 #ifdef __KERNEL__
 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_nonpcpu);
 134                         }
 135                         preempt_enable();
 136
 137                         list_move_tail(&ck->list, &bc->freed_nonpcpu);
 138                         mutex_unlock(&bc->lock);
 139                 }
 140 #else
 141                 mutex_lock(&bc->lock);
 142                 list_move_tail(&ck->list, &bc->freed_nonpcpu);
 143                 mutex_unlock(&bc->lock);
 144 #endif
 145         } else {
 146                 mutex_lock(&bc->lock);
 147                 list_move_tail(&ck->list, &bc->freed_pcpu);
 148                 mutex_unlock(&bc->lock);
 149         }
 150 }
 151
 152 static void bkey_cached_free_fast(struct btree_key_cache *bc,
 153                                   struct bkey_cached *ck)
 154 {
 155         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 156
 157         ck->btree_trans_barrier_seq =
 158                 start_poll_synchronize_srcu(&c->btree_trans_barrier);
 159
 160         list_del_init(&ck->list);
 161         atomic_long_inc(&bc->nr_freed);
 162
 163         kfree(ck->k);
 164         ck->k           = NULL;
 165         ck->u64s        = 0;
 166
 167         bkey_cached_move_to_freelist(bc, ck);
 168
 169         six_unlock_write(&ck->c.lock);
 170         six_unlock_intent(&ck->c.lock);
 171 }
 172
 173 static struct bkey_cached *
 174 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path)
 175 {
 176         struct bch_fs *c = trans->c;
 177         struct btree_key_cache *bc = &c->btree_key_cache;
 178         struct bkey_cached *ck = NULL;
 179         struct btree_key_cache_freelist *f;
 180         bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
 181
 182         if (!pcpu_readers) {
 183 #ifdef __KERNEL__
 184                 preempt_disable();
 185                 f = this_cpu_ptr(bc->pcpu_freed);
 186                 if (f->nr)
 187                         ck = f->objs[--f->nr];
 188                 preempt_enable();
 189
 190                 if (!ck) {
 191                         mutex_lock(&bc->lock);
 192                         preempt_disable();
 193                         f = this_cpu_ptr(bc->pcpu_freed);
 194
 195                         while (!list_empty(&bc->freed_nonpcpu) &&
 196                                f->nr < ARRAY_SIZE(f->objs) / 2) {
 197                                 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 198                                 list_del_init(&ck->list);
 199                                 f->objs[f->nr++] = ck;
 200                         }
 201
 202                         ck = f->nr ? f->objs[--f->nr] : NULL;
 203                         preempt_enable();
 204                         mutex_unlock(&bc->lock);
 205                 }
 206 #else
 207                 mutex_lock(&bc->lock);
 208                 if (!list_empty(&bc->freed_nonpcpu)) {
 209                         ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 210                         list_del_init(&ck->list);
 211                 }
 212                 mutex_unlock(&bc->lock);
 213 #endif
 214         } else {
 215                 mutex_lock(&bc->lock);
 216                 if (!list_empty(&bc->freed_pcpu)) {
 217                         ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
 218                         list_del_init(&ck->list);
 219                 }
 220                 mutex_unlock(&bc->lock);
 221         }
 222
 223         if (ck) {
 224                 int ret;
 225
 226                 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent);
 227                 if (unlikely(ret)) {
 228                         bkey_cached_move_to_freelist(bc, ck);
 229                         return ERR_PTR(ret);
 230                 }
 231
 232                 path->l[0].b = (void *) ck;
 233                 path->l[0].lock_seq = ck->c.lock.state.seq;
 234                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 235
 236                 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
 237                 if (unlikely(ret)) {
 238                         btree_node_unlock(trans, path, 0);
 239                         bkey_cached_move_to_freelist(bc, ck);
 240                         return ERR_PTR(ret);
 241                 }
 242
 243                 return ck;
 244         }
 245
 246         /* GFP_NOFS because we're holding btree locks: */
 247         ck = kmem_cache_alloc(bch2_key_cache, GFP_NOFS|__GFP_ZERO);
 248         if (likely(ck)) {
 249                 INIT_LIST_HEAD(&ck->list);
 250                 __six_lock_init(&ck->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 251                 if (pcpu_readers)
 252                         six_lock_pcpu_alloc(&ck->c.lock);
 253
 254                 ck->c.cached = true;
 255                 BUG_ON(!six_trylock_intent(&ck->c.lock));
 256                 BUG_ON(!six_trylock_write(&ck->c.lock));
 257                 return ck;
 258         }
 259
 260         return NULL;
 261 }
 262
 263 static struct bkey_cached *
 264 bkey_cached_reuse(struct btree_key_cache *c)
 265 {
 266         struct bucket_table *tbl;
 267         struct rhash_head *pos;
 268         struct bkey_cached *ck;
 269         unsigned i;
 270
 271         mutex_lock(&c->lock);
 272         rcu_read_lock();
 273         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 274         for (i = 0; i < tbl->size; i++)
 275                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 276                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 277                             bkey_cached_lock_for_evict(ck)) {
 278                                 bkey_cached_evict(c, ck);
 279                                 goto out;
 280                         }
 281                 }
 282         ck = NULL;
 283 out:
 284         rcu_read_unlock();
 285         mutex_unlock(&c->lock);
 286         return ck;
 287 }
 288
 289 static struct bkey_cached *
 290 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
 291 {
 292         struct bch_fs *c = trans->c;
 293         struct btree_key_cache *bc = &c->btree_key_cache;
 294         struct bkey_cached *ck;
 295         bool was_new = true;
 296
 297         ck = bkey_cached_alloc(trans, path);
 298         if (unlikely(IS_ERR(ck)))
 299                 return ck;
 300
 301         if (unlikely(!ck)) {
 302                 ck = bkey_cached_reuse(bc);
 303                 if (unlikely(!ck)) {
 304                         bch_err(c, "error allocating memory for key cache item, btree %s",
 305                                 bch2_btree_ids[path->btree_id]);
 306                         return ERR_PTR(-ENOMEM);
 307                 }
 308
 309                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 310                 was_new = false;
 311         } else {
 312                 if (path->btree_id == BTREE_ID_subvolumes)
 313                         six_lock_pcpu_alloc(&ck->c.lock);
 314         }
 315
 316         ck->c.level             = 0;
 317         ck->c.btree_id          = path->btree_id;
 318         ck->key.btree_id        = path->btree_id;
 319         ck->key.pos             = path->pos;
 320         ck->valid               = false;
 321         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 322
 323         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 324                                           &ck->hash,
 325                                           bch2_btree_key_cache_params))) {
 326                 /* We raced with another fill: */
 327
 328                 if (likely(was_new)) {
 329                         six_unlock_write(&ck->c.lock);
 330                         six_unlock_intent(&ck->c.lock);
 331                         mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
 332                         kfree(ck);
 333                 } else {
 334                         bkey_cached_free_fast(bc, ck);
 335                 }
 336
 337                 return NULL;
 338         }
 339
 340         atomic_long_inc(&bc->nr_keys);
 341
 342         six_unlock_write(&ck->c.lock);
 343
 344         return ck;
 345 }
 346
 347 static int btree_key_cache_fill(struct btree_trans *trans,
 348                                 struct btree_path *ck_path,
 349                                 struct bkey_cached *ck)
 350 {
 351         struct btree_path *path;
 352         struct bkey_s_c k;
 353         unsigned new_u64s = 0;
 354         struct bkey_i *new_k = NULL;
 355         struct bkey u;
 356         int ret;
 357
 358         path = bch2_path_get(trans, ck->key.btree_id,
 359                              ck->key.pos, 0, 0, 0, _THIS_IP_);
 360         ret = bch2_btree_path_traverse(trans, path, 0);
 361         if (ret)
 362                 goto err;
 363
 364         k = bch2_btree_path_peek_slot(path, &u);
 365
 366         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 367                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 368                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
 369                 goto err;
 370         }
 371
 372         /*
 373          * bch2_varint_decode can read past the end of the buffer by at
 374          * most 7 bytes (it won't be used):
 375          */
 376         new_u64s = k.k->u64s + 1;
 377
 378         /*
 379          * Allocate some extra space so that the transaction commit path is less
 380          * likely to have to reallocate, since that requires a transaction
 381          * restart:
 382          */
 383         new_u64s = min(256U, (new_u64s * 3) / 2);
 384
 385         if (new_u64s > ck->u64s) {
 386                 new_u64s = roundup_pow_of_two(new_u64s);
 387                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 388                 if (!new_k) {
 389                         bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 390                                 bch2_btree_ids[ck->key.btree_id], new_u64s);
 391                         ret = -ENOMEM;
 392                         goto err;
 393                 }
 394         }
 395
 396         ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
 397         if (ret) {
 398                 kfree(new_k);
 399                 goto err;
 400         }
 401
 402         if (new_k) {
 403                 kfree(ck->k);
 404                 ck->u64s = new_u64s;
 405                 ck->k = new_k;
 406         }
 407
 408         bkey_reassemble(ck->k, k);
 409         ck->valid = true;
 410         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 411
 412         /* We're not likely to need this iterator again: */
 413         path->preserve = false;
 414 err:
 415         bch2_path_put(trans, path, 0);
 416         return ret;
 417 }
 418
 419 noinline static int
 420 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
 421                                          unsigned flags)
 422 {
 423         struct bch_fs *c = trans->c;
 424         struct bkey_cached *ck;
 425         int ret = 0;
 426
 427         BUG_ON(path->level);
 428
 429         path->l[1].b = NULL;
 430
 431         if (bch2_btree_node_relock(trans, path, 0)) {
 432                 ck = (void *) path->l[0].b;
 433                 goto fill;
 434         }
 435 retry:
 436         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 437         if (!ck) {
 438                 ck = btree_key_cache_create(trans, path);
 439                 ret = PTR_ERR_OR_ZERO(ck);
 440                 if (ret)
 441                         goto err;
 442                 if (!ck)
 443                         goto retry;
 444
 445                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 446                 path->locks_want = 1;
 447         } else {
 448                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 449
 450                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 451                                       lock_want, _THIS_IP_);
 452                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 453                         goto err;
 454
 455                 BUG_ON(ret);
 456
 457                 if (ck->key.btree_id != path->btree_id ||
 458                     bpos_cmp(ck->key.pos, path->pos)) {
 459                         six_unlock_type(&ck->c.lock, lock_want);
 460                         goto retry;
 461                 }
 462
 463                 mark_btree_node_locked(trans, path, 0, lock_want);
 464         }
 465
 466         path->l[0].lock_seq     = ck->c.lock.state.seq;
 467         path->l[0].b            = (void *) ck;
 468 fill:
 469         if (!ck->valid) {
 470                 /*
 471                  * Using the underscore version because we haven't set
 472                  * path->uptodate yet:
 473                  */
 474                 if (!path->locks_want &&
 475                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 476                         trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
 477                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
 478                         goto err;
 479                 }
 480
 481                 ret = btree_key_cache_fill(trans, path, ck);
 482                 if (ret)
 483                         goto err;
 484         }
 485
 486         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 487                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 488
 489         path->uptodate = BTREE_ITER_UPTODATE;
 490         BUG_ON(!ck->valid);
 491         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 492
 493         return ret;
 494 err:
 495         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 496                 btree_node_unlock(trans, path, 0);
 497                 path->l[0].b = ERR_PTR(ret);
 498         }
 499         return ret;
 500 }
 501
 502 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 503                                     unsigned flags)
 504 {
 505         struct bch_fs *c = trans->c;
 506         struct bkey_cached *ck;
 507         int ret = 0;
 508
 509         EBUG_ON(path->level);
 510
 511         path->l[1].b = NULL;
 512
 513         if (bch2_btree_node_relock(trans, path, 0)) {
 514                 ck = (void *) path->l[0].b;
 515                 goto fill;
 516         }
 517 retry:
 518         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 519         if (!ck) {
 520                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 521         } else {
 522                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 523
 524                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 525                                       lock_want, _THIS_IP_);
 526                 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
 527
 528                 if (ret)
 529                         return ret;
 530
 531                 if (ck->key.btree_id != path->btree_id ||
 532                     bpos_cmp(ck->key.pos, path->pos)) {
 533                         six_unlock_type(&ck->c.lock, lock_want);
 534                         goto retry;
 535                 }
 536
 537                 mark_btree_node_locked(trans, path, 0, lock_want);
 538         }
 539
 540         path->l[0].lock_seq     = ck->c.lock.state.seq;
 541         path->l[0].b            = (void *) ck;
 542 fill:
 543         if (!ck->valid)
 544                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 545
 546         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 547                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 548
 549         path->uptodate = BTREE_ITER_UPTODATE;
 550         EBUG_ON(!ck->valid);
 551         EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 552
 553         return ret;
 554 }
 555
 556 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 557                                      struct bkey_cached_key key,
 558                                      u64 journal_seq,
 559                                      unsigned commit_flags,
 560                                      bool evict)
 561 {
 562         struct bch_fs *c = trans->c;
 563         struct journal *j = &c->journal;
 564         struct btree_iter c_iter, b_iter;
 565         struct bkey_cached *ck = NULL;
 566         int ret;
 567
 568         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 569                              BTREE_ITER_SLOTS|
 570                              BTREE_ITER_INTENT|
 571                              BTREE_ITER_ALL_SNAPSHOTS);
 572         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 573                              BTREE_ITER_CACHED|
 574                              BTREE_ITER_INTENT);
 575         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 576
 577         ret = bch2_btree_iter_traverse(&c_iter);
 578         if (ret)
 579                 goto out;
 580
 581         ck = (void *) c_iter.path->l[0].b;
 582         if (!ck)
 583                 goto out;
 584
 585         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 586                 if (evict)
 587                         goto evict;
 588                 goto out;
 589         }
 590
 591         BUG_ON(!ck->valid);
 592
 593         if (journal_seq && ck->journal.seq != journal_seq)
 594                 goto out;
 595
 596         /*
 597          * Since journal reclaim depends on us making progress here, and the
 598          * allocator/copygc depend on journal reclaim making progress, we need
 599          * to be using alloc reserves:
 600          * */
 601         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 602                 bch2_trans_update(trans, &b_iter, ck->k,
 603                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 604                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 605                                   BTREE_TRIGGER_NORUN) ?:
 606                 bch2_trans_commit(trans, NULL, NULL,
 607                                   BTREE_INSERT_NOCHECK_RW|
 608                                   BTREE_INSERT_NOFAIL|
 609                                   BTREE_INSERT_USE_RESERVE|
 610                                   (ck->journal.seq == journal_last_seq(j)
 611                                    ? JOURNAL_WATERMARK_reserved
 612                                    : 0)|
 613                                   commit_flags);
 614
 615         bch2_fs_fatal_err_on(ret &&
 616                              !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 617                              !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
 618                              !bch2_journal_error(j), c,
 619                              "error flushing key cache: %s", bch2_err_str(ret));
 620         if (ret)
 621                 goto out;
 622
 623         bch2_journal_pin_drop(j, &ck->journal);
 624         bch2_journal_preres_put(j, &ck->res);
 625
 626         BUG_ON(!btree_node_locked(c_iter.path, 0));
 627
 628         if (!evict) {
 629                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 630                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 631                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 632                 }
 633         } else {
 634                 struct btree_path *path2;
 635 evict:
 636                 trans_for_each_path(trans, path2)
 637                         if (path2 != c_iter.path)
 638                                 __bch2_btree_path_unlock(trans, path2);
 639
 640                 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
 641
 642                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 643                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 644                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 645                 }
 646
 647                 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
 648                 bkey_cached_evict(&c->btree_key_cache, ck);
 649                 bkey_cached_free_fast(&c->btree_key_cache, ck);
 650         }
 651 out:
 652         bch2_trans_iter_exit(trans, &b_iter);
 653         bch2_trans_iter_exit(trans, &c_iter);
 654         return ret;
 655 }
 656
 657 int bch2_btree_key_cache_journal_flush(struct journal *j,
 658                                 struct journal_entry_pin *pin, u64 seq)
 659 {
 660         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 661         struct bkey_cached *ck =
 662                 container_of(pin, struct bkey_cached, journal);
 663         struct bkey_cached_key key;
 664         struct btree_trans trans;
 665         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 666         int ret = 0;
 667
 668         bch2_trans_init(&trans, c, 0, 0);
 669
 670         btree_node_lock_nopath_nofail(&trans, &ck->c, SIX_LOCK_read);
 671         key = ck->key;
 672
 673         if (ck->journal.seq != seq ||
 674             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 675                 six_unlock_read(&ck->c.lock);
 676                 goto unlock;
 677         }
 678         six_unlock_read(&ck->c.lock);
 679
 680         ret = commit_do(&trans, NULL, NULL, 0,
 681                 btree_key_cache_flush_pos(&trans, key, seq,
 682                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 683 unlock:
 684         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 685
 686         bch2_trans_exit(&trans);
 687         return ret;
 688 }
 689
 690 /*
 691  * Flush and evict a key from the key cache:
 692  */
 693 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 694                                enum btree_id id, struct bpos pos)
 695 {
 696         struct bch_fs *c = trans->c;
 697         struct bkey_cached_key key = { id, pos };
 698
 699         /* Fastpath - assume it won't be found: */
 700         if (!bch2_btree_key_cache_find(c, id, pos))
 701                 return 0;
 702
 703         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 704 }
 705
 706 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 707                                   struct btree_path *path,
 708                                   struct bkey_i *insert)
 709 {
 710         struct bch_fs *c = trans->c;
 711         struct bkey_cached *ck = (void *) path->l[0].b;
 712         bool kick_reclaim = false;
 713
 714         BUG_ON(insert->u64s > ck->u64s);
 715
 716         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 717                 int difference;
 718
 719                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 720
 721                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 722                 if (difference > 0) {
 723                         trans->journal_preres.u64s      -= difference;
 724                         ck->res.u64s                    += difference;
 725                 }
 726         }
 727
 728         bkey_copy(ck->k, insert);
 729         ck->valid = true;
 730
 731         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 732                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 733                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 734
 735                 if (bch2_nr_btree_keys_need_flush(c))
 736                         kick_reclaim = true;
 737         }
 738
 739         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 740                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 741
 742         if (kick_reclaim)
 743                 journal_reclaim_kick(&c->journal);
 744         return true;
 745 }
 746
 747 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 748                                struct btree_path *path)
 749 {
 750         struct bch_fs *c = trans->c;
 751         struct bkey_cached *ck = (void *) path->l[0].b;
 752
 753         BUG_ON(!ck->valid);
 754
 755         /*
 756          * We just did an update to the btree, bypassing the key cache: the key
 757          * cache key is now stale and must be dropped, even if dirty:
 758          */
 759         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 760                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 761                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 762                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 763         }
 764
 765         ck->valid = false;
 766 }
 767
 768 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 769                                            struct shrink_control *sc)
 770 {
 771         struct bch_fs *c = container_of(shrink, struct bch_fs,
 772                                         btree_key_cache.shrink);
 773         struct btree_key_cache *bc = &c->btree_key_cache;
 774         struct bucket_table *tbl;
 775         struct bkey_cached *ck, *t;
 776         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 777         unsigned start, flags;
 778         int srcu_idx;
 779
 780         mutex_lock(&bc->lock);
 781         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 782         flags = memalloc_nofs_save();
 783
 784         /*
 785          * Newest freed entries are at the end of the list - once we hit one
 786          * that's too new to be freed, we can bail out:
 787          */
 788         list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
 789                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 790                                                  ck->btree_trans_barrier_seq))
 791                         break;
 792
 793                 list_del(&ck->list);
 794                 six_lock_pcpu_free(&ck->c.lock);
 795                 kmem_cache_free(bch2_key_cache, ck);
 796                 atomic_long_dec(&bc->nr_freed);
 797                 scanned++;
 798                 freed++;
 799         }
 800
 801         if (scanned >= nr)
 802                 goto out;
 803
 804         list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
 805                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 806                                                  ck->btree_trans_barrier_seq))
 807                         break;
 808
 809                 list_del(&ck->list);
 810                 six_lock_pcpu_free(&ck->c.lock);
 811                 kmem_cache_free(bch2_key_cache, ck);
 812                 atomic_long_dec(&bc->nr_freed);
 813                 scanned++;
 814                 freed++;
 815         }
 816
 817         if (scanned >= nr)
 818                 goto out;
 819
 820         rcu_read_lock();
 821         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 822         if (bc->shrink_iter >= tbl->size)
 823                 bc->shrink_iter = 0;
 824         start = bc->shrink_iter;
 825
 826         do {
 827                 struct rhash_head *pos, *next;
 828
 829                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 830
 831                 while (!rht_is_a_nulls(pos)) {
 832                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 833                         ck = container_of(pos, struct bkey_cached, hash);
 834
 835                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 836                                 goto next;
 837
 838                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 839                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 840                         else if (bkey_cached_lock_for_evict(ck)) {
 841                                 bkey_cached_evict(bc, ck);
 842                                 bkey_cached_free(bc, ck);
 843                         }
 844
 845                         scanned++;
 846                         if (scanned >= nr)
 847                                 break;
 848 next:
 849                         pos = next;
 850                 }
 851
 852                 bc->shrink_iter++;
 853                 if (bc->shrink_iter >= tbl->size)
 854                         bc->shrink_iter = 0;
 855         } while (scanned < nr && bc->shrink_iter != start);
 856
 857         rcu_read_unlock();
 858 out:
 859         memalloc_nofs_restore(flags);
 860         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 861         mutex_unlock(&bc->lock);
 862
 863         return freed;
 864 }
 865
 866 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 867                                             struct shrink_control *sc)
 868 {
 869         struct bch_fs *c = container_of(shrink, struct bch_fs,
 870                                         btree_key_cache.shrink);
 871         struct btree_key_cache *bc = &c->btree_key_cache;
 872         long nr = atomic_long_read(&bc->nr_keys) -
 873                 atomic_long_read(&bc->nr_dirty);
 874
 875         return max(0L, nr);
 876 }
 877
 878 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 879 {
 880         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 881         struct bucket_table *tbl;
 882         struct bkey_cached *ck, *n;
 883         struct rhash_head *pos;
 884         unsigned i;
 885 #ifdef __KERNEL__
 886         int cpu;
 887 #endif
 888
 889         if (bc->shrink.list.next)
 890                 unregister_shrinker(&bc->shrink);
 891
 892         mutex_lock(&bc->lock);
 893
 894         rcu_read_lock();
 895         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 896         if (tbl)
 897                 for (i = 0; i < tbl->size; i++)
 898                         rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 899                                 bkey_cached_evict(bc, ck);
 900                                 list_add(&ck->list, &bc->freed_nonpcpu);
 901                         }
 902         rcu_read_unlock();
 903
 904 #ifdef __KERNEL__
 905         for_each_possible_cpu(cpu) {
 906                 struct btree_key_cache_freelist *f =
 907                         per_cpu_ptr(bc->pcpu_freed, cpu);
 908
 909                 for (i = 0; i < f->nr; i++) {
 910                         ck = f->objs[i];
 911                         list_add(&ck->list, &bc->freed_nonpcpu);
 912                 }
 913         }
 914 #endif
 915
 916         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 917
 918         list_for_each_entry_safe(ck, n, &bc->freed_nonpcpu, list) {
 919                 cond_resched();
 920
 921                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 922                 bch2_journal_preres_put(&c->journal, &ck->res);
 923
 924                 list_del(&ck->list);
 925                 kfree(ck->k);
 926                 six_lock_pcpu_free(&ck->c.lock);
 927                 kmem_cache_free(bch2_key_cache, ck);
 928         }
 929
 930         if (atomic_long_read(&bc->nr_dirty) &&
 931             !bch2_journal_error(&c->journal) &&
 932             test_bit(BCH_FS_WAS_RW, &c->flags))
 933                 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
 934                       atomic_long_read(&bc->nr_dirty));
 935
 936         if (atomic_long_read(&bc->nr_keys))
 937                 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
 938                       atomic_long_read(&bc->nr_keys));
 939
 940         mutex_unlock(&bc->lock);
 941
 942         if (bc->table_init_done)
 943                 rhashtable_destroy(&bc->table);
 944
 945         free_percpu(bc->pcpu_freed);
 946 }
 947
 948 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 949 {
 950         mutex_init(&c->lock);
 951         INIT_LIST_HEAD(&c->freed_pcpu);
 952         INIT_LIST_HEAD(&c->freed_nonpcpu);
 953 }
 954
 955 static void bch2_btree_key_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 956 {
 957         struct btree_key_cache *bc =
 958                 container_of(shrink, struct btree_key_cache, shrink);
 959
 960         bch2_btree_key_cache_to_text(out, bc);
 961 }
 962
 963 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
 964 {
 965         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 966         int ret;
 967
 968 #ifdef __KERNEL__
 969         bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
 970         if (!bc->pcpu_freed)
 971                 return -ENOMEM;
 972 #endif
 973
 974         ret = rhashtable_init(&bc->table, &bch2_btree_key_cache_params);
 975         if (ret)
 976                 return ret;
 977
 978         bc->table_init_done = true;
 979
 980         bc->shrink.seeks                = 1;
 981         bc->shrink.count_objects        = bch2_btree_key_cache_count;
 982         bc->shrink.scan_objects         = bch2_btree_key_cache_scan;
 983         bc->shrink.to_text              = bch2_btree_key_cache_shrinker_to_text;
 984         return register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name);
 985 }
 986
 987 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 988 {
 989         prt_printf(out, "nr_freed:\t%zu\n",     atomic_long_read(&c->nr_freed));
 990         prt_printf(out, "nr_keys:\t%lu\n",      atomic_long_read(&c->nr_keys));
 991         prt_printf(out, "nr_dirty:\t%lu\n",     atomic_long_read(&c->nr_dirty));
 992 }
 993
 994 void bch2_btree_key_cache_exit(void)
 995 {
 996         if (bch2_key_cache)
 997                 kmem_cache_destroy(bch2_key_cache);
 998 }
 999
1000 int __init bch2_btree_key_cache_init(void)
1001 {
1002         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
1003         if (!bch2_key_cache)
1004                 return -ENOMEM;
1005
1006         return 0;
1007 }