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 };
  40
  41 __flatten
  42 inline struct bkey_cached *
  43 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
  44 {
  45         struct bkey_cached_key key = {
  46                 .btree_id       = btree_id,
  47                 .pos            = pos,
  48         };
  49
  50         return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
  51                                       bch2_btree_key_cache_params);
  52 }
  53
  54 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
  55 {
  56         if (!six_trylock_intent(&ck->c.lock))
  57                 return false;
  58
  59         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
  60                 six_unlock_intent(&ck->c.lock);
  61                 return false;
  62         }
  63
  64         if (!six_trylock_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 #ifdef __KERNEL__
 107 static void __bkey_cached_move_to_freelist_ordered(struct btree_key_cache *bc,
 108                                                    struct bkey_cached *ck)
 109 {
 110         struct bkey_cached *pos;
 111
 112         list_for_each_entry_reverse(pos, &bc->freed_nonpcpu, list) {
 113                 if (ULONG_CMP_GE(ck->btree_trans_barrier_seq,
 114                                  pos->btree_trans_barrier_seq)) {
 115                         list_move(&ck->list, &pos->list);
 116                         return;
 117                 }
 118         }
 119
 120         list_move(&ck->list, &bc->freed_nonpcpu);
 121 }
 122 #endif
 123
 124 static void bkey_cached_move_to_freelist(struct btree_key_cache *bc,
 125                                          struct bkey_cached *ck)
 126 {
 127         BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
 128
 129         if (!ck->c.lock.readers) {
 130 #ifdef __KERNEL__
 131                 struct btree_key_cache_freelist *f;
 132                 bool freed = false;
 133
 134                 preempt_disable();
 135                 f = this_cpu_ptr(bc->pcpu_freed);
 136
 137                 if (f->nr < ARRAY_SIZE(f->objs)) {
 138                         f->objs[f->nr++] = ck;
 139                         freed = true;
 140                 }
 141                 preempt_enable();
 142
 143                 if (!freed) {
 144                         mutex_lock(&bc->lock);
 145                         preempt_disable();
 146                         f = this_cpu_ptr(bc->pcpu_freed);
 147
 148                         while (f->nr > ARRAY_SIZE(f->objs) / 2) {
 149                                 struct bkey_cached *ck2 = f->objs[--f->nr];
 150
 151                                 __bkey_cached_move_to_freelist_ordered(bc, ck2);
 152                         }
 153                         preempt_enable();
 154
 155                         __bkey_cached_move_to_freelist_ordered(bc, ck);
 156                         mutex_unlock(&bc->lock);
 157                 }
 158 #else
 159                 mutex_lock(&bc->lock);
 160                 list_move_tail(&ck->list, &bc->freed_nonpcpu);
 161                 mutex_unlock(&bc->lock);
 162 #endif
 163         } else {
 164                 mutex_lock(&bc->lock);
 165                 list_move_tail(&ck->list, &bc->freed_pcpu);
 166                 mutex_unlock(&bc->lock);
 167         }
 168 }
 169
 170 static void bkey_cached_free_fast(struct btree_key_cache *bc,
 171                                   struct bkey_cached *ck)
 172 {
 173         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 174
 175         ck->btree_trans_barrier_seq =
 176                 start_poll_synchronize_srcu(&c->btree_trans_barrier);
 177
 178         list_del_init(&ck->list);
 179         atomic_long_inc(&bc->nr_freed);
 180
 181         kfree(ck->k);
 182         ck->k           = NULL;
 183         ck->u64s        = 0;
 184
 185         bkey_cached_move_to_freelist(bc, ck);
 186
 187         six_unlock_write(&ck->c.lock);
 188         six_unlock_intent(&ck->c.lock);
 189 }
 190
 191 static struct bkey_cached *
 192 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path,
 193                   bool *was_new)
 194 {
 195         struct bch_fs *c = trans->c;
 196         struct btree_key_cache *bc = &c->btree_key_cache;
 197         struct bkey_cached *ck = NULL;
 198         bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
 199         int ret;
 200
 201         if (!pcpu_readers) {
 202 #ifdef __KERNEL__
 203                 struct btree_key_cache_freelist *f;
 204
 205                 preempt_disable();
 206                 f = this_cpu_ptr(bc->pcpu_freed);
 207                 if (f->nr)
 208                         ck = f->objs[--f->nr];
 209                 preempt_enable();
 210
 211                 if (!ck) {
 212                         mutex_lock(&bc->lock);
 213                         preempt_disable();
 214                         f = this_cpu_ptr(bc->pcpu_freed);
 215
 216                         while (!list_empty(&bc->freed_nonpcpu) &&
 217                                f->nr < ARRAY_SIZE(f->objs) / 2) {
 218                                 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 219                                 list_del_init(&ck->list);
 220                                 f->objs[f->nr++] = ck;
 221                         }
 222
 223                         ck = f->nr ? f->objs[--f->nr] : NULL;
 224                         preempt_enable();
 225                         mutex_unlock(&bc->lock);
 226                 }
 227 #else
 228                 mutex_lock(&bc->lock);
 229                 if (!list_empty(&bc->freed_nonpcpu)) {
 230                         ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 231                         list_del_init(&ck->list);
 232                 }
 233                 mutex_unlock(&bc->lock);
 234 #endif
 235         } else {
 236                 mutex_lock(&bc->lock);
 237                 if (!list_empty(&bc->freed_pcpu)) {
 238                         ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
 239                         list_del_init(&ck->list);
 240                 }
 241                 mutex_unlock(&bc->lock);
 242         }
 243
 244         if (ck) {
 245                 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent, _THIS_IP_);
 246                 if (unlikely(ret)) {
 247                         bkey_cached_move_to_freelist(bc, ck);
 248                         return ERR_PTR(ret);
 249                 }
 250
 251                 path->l[0].b = (void *) ck;
 252                 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
 253                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
 254
 255                 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
 256                 if (unlikely(ret)) {
 257                         btree_node_unlock(trans, path, 0);
 258                         bkey_cached_move_to_freelist(bc, ck);
 259                         return ERR_PTR(ret);
 260                 }
 261
 262                 return ck;
 263         }
 264
 265         ck = allocate_dropping_locks(trans, ret,
 266                         kmem_cache_zalloc(bch2_key_cache, _gfp));
 267         if (ret) {
 268                 kmem_cache_free(bch2_key_cache, ck);
 269                 return ERR_PTR(ret);
 270         }
 271
 272         if (!ck)
 273                 return NULL;
 274
 275         INIT_LIST_HEAD(&ck->list);
 276         bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
 277
 278         ck->c.cached = true;
 279         BUG_ON(!six_trylock_intent(&ck->c.lock));
 280         BUG_ON(!six_trylock_write(&ck->c.lock));
 281         *was_new = true;
 282         return ck;
 283 }
 284
 285 static struct bkey_cached *
 286 bkey_cached_reuse(struct btree_key_cache *c)
 287 {
 288         struct bucket_table *tbl;
 289         struct rhash_head *pos;
 290         struct bkey_cached *ck;
 291         unsigned i;
 292
 293         mutex_lock(&c->lock);
 294         rcu_read_lock();
 295         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 296         for (i = 0; i < tbl->size; i++)
 297                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 298                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 299                             bkey_cached_lock_for_evict(ck)) {
 300                                 bkey_cached_evict(c, ck);
 301                                 goto out;
 302                         }
 303                 }
 304         ck = NULL;
 305 out:
 306         rcu_read_unlock();
 307         mutex_unlock(&c->lock);
 308         return ck;
 309 }
 310
 311 static struct bkey_cached *
 312 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
 313 {
 314         struct bch_fs *c = trans->c;
 315         struct btree_key_cache *bc = &c->btree_key_cache;
 316         struct bkey_cached *ck;
 317         bool was_new = false;
 318
 319         ck = bkey_cached_alloc(trans, path, &was_new);
 320         if (IS_ERR(ck))
 321                 return ck;
 322
 323         if (unlikely(!ck)) {
 324                 ck = bkey_cached_reuse(bc);
 325                 if (unlikely(!ck)) {
 326                         bch_err(c, "error allocating memory for key cache item, btree %s",
 327                                 bch2_btree_id_str(path->btree_id));
 328                         return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
 329                 }
 330
 331                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
 332         }
 333
 334         ck->c.level             = 0;
 335         ck->c.btree_id          = path->btree_id;
 336         ck->key.btree_id        = path->btree_id;
 337         ck->key.pos             = path->pos;
 338         ck->valid               = false;
 339         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 340
 341         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 342                                           &ck->hash,
 343                                           bch2_btree_key_cache_params))) {
 344                 /* We raced with another fill: */
 345
 346                 if (likely(was_new)) {
 347                         six_unlock_write(&ck->c.lock);
 348                         six_unlock_intent(&ck->c.lock);
 349                         kfree(ck);
 350                 } else {
 351                         bkey_cached_free_fast(bc, ck);
 352                 }
 353
 354                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
 355                 return NULL;
 356         }
 357
 358         atomic_long_inc(&bc->nr_keys);
 359
 360         six_unlock_write(&ck->c.lock);
 361
 362         return ck;
 363 }
 364
 365 static int btree_key_cache_fill(struct btree_trans *trans,
 366                                 struct btree_path *ck_path,
 367                                 struct bkey_cached *ck)
 368 {
 369         struct btree_iter iter;
 370         struct bkey_s_c k;
 371         unsigned new_u64s = 0;
 372         struct bkey_i *new_k = NULL;
 373         int ret;
 374
 375         k = bch2_bkey_get_iter(trans, &iter, ck->key.btree_id, ck->key.pos,
 376                                BTREE_ITER_KEY_CACHE_FILL|
 377                                BTREE_ITER_CACHED_NOFILL);
 378         ret = bkey_err(k);
 379         if (ret)
 380                 goto err;
 381
 382         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 383                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 384                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 385                 goto err;
 386         }
 387
 388         /*
 389          * bch2_varint_decode can read past the end of the buffer by at
 390          * most 7 bytes (it won't be used):
 391          */
 392         new_u64s = k.k->u64s + 1;
 393
 394         /*
 395          * Allocate some extra space so that the transaction commit path is less
 396          * likely to have to reallocate, since that requires a transaction
 397          * restart:
 398          */
 399         new_u64s = min(256U, (new_u64s * 3) / 2);
 400
 401         if (new_u64s > ck->u64s) {
 402                 new_u64s = roundup_pow_of_two(new_u64s);
 403                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 404                 if (!new_k) {
 405                         bch2_trans_unlock(trans);
 406
 407                         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 408                         if (!new_k) {
 409                                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 410                                         bch2_btree_id_str(ck->key.btree_id), new_u64s);
 411                                 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
 412                                 goto err;
 413                         }
 414
 415                         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 416                                 kfree(new_k);
 417                                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 418                                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 419                                 goto err;
 420                         }
 421
 422                         ret = bch2_trans_relock(trans);
 423                         if (ret) {
 424                                 kfree(new_k);
 425                                 goto err;
 426                         }
 427                 }
 428         }
 429
 430         ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
 431         if (ret) {
 432                 kfree(new_k);
 433                 goto err;
 434         }
 435
 436         if (new_k) {
 437                 kfree(ck->k);
 438                 ck->u64s = new_u64s;
 439                 ck->k = new_k;
 440         }
 441
 442         bkey_reassemble(ck->k, k);
 443         ck->valid = true;
 444         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 445
 446         /* We're not likely to need this iterator again: */
 447         set_btree_iter_dontneed(&iter);
 448 err:
 449         bch2_trans_iter_exit(trans, &iter);
 450         return ret;
 451 }
 452
 453 static noinline int
 454 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
 455                                          unsigned flags)
 456 {
 457         struct bch_fs *c = trans->c;
 458         struct bkey_cached *ck;
 459         int ret = 0;
 460
 461         BUG_ON(path->level);
 462
 463         path->l[1].b = NULL;
 464
 465         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 466                 ck = (void *) path->l[0].b;
 467                 goto fill;
 468         }
 469 retry:
 470         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 471         if (!ck) {
 472                 ck = btree_key_cache_create(trans, path);
 473                 ret = PTR_ERR_OR_ZERO(ck);
 474                 if (ret)
 475                         goto err;
 476                 if (!ck)
 477                         goto retry;
 478
 479                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
 480                 path->locks_want = 1;
 481         } else {
 482                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 483
 484                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 485                                       lock_want, _THIS_IP_);
 486                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 487                         goto err;
 488
 489                 BUG_ON(ret);
 490
 491                 if (ck->key.btree_id != path->btree_id ||
 492                     !bpos_eq(ck->key.pos, path->pos)) {
 493                         six_unlock_type(&ck->c.lock, lock_want);
 494                         goto retry;
 495                 }
 496
 497                 mark_btree_node_locked(trans, path, 0,
 498                                        (enum btree_node_locked_type) lock_want);
 499         }
 500
 501         path->l[0].lock_seq     = six_lock_seq(&ck->c.lock);
 502         path->l[0].b            = (void *) ck;
 503 fill:
 504         path->uptodate = BTREE_ITER_UPTODATE;
 505
 506         if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
 507                 /*
 508                  * Using the underscore version because we haven't set
 509                  * path->uptodate yet:
 510                  */
 511                 if (!path->locks_want &&
 512                     !__bch2_btree_path_upgrade(trans, path, 1, NULL)) {
 513                         trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
 514                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
 515                         goto err;
 516                 }
 517
 518                 ret = btree_key_cache_fill(trans, path, ck);
 519                 if (ret)
 520                         goto err;
 521
 522                 ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
 523                 if (ret)
 524                         goto err;
 525
 526                 path->uptodate = BTREE_ITER_UPTODATE;
 527         }
 528
 529         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 530                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 531
 532         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 533         BUG_ON(path->uptodate);
 534
 535         return ret;
 536 err:
 537         path->uptodate = BTREE_ITER_NEED_TRAVERSE;
 538         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 539                 btree_node_unlock(trans, path, 0);
 540                 path->l[0].b = ERR_PTR(ret);
 541         }
 542         return ret;
 543 }
 544
 545 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 546                                     unsigned flags)
 547 {
 548         struct bch_fs *c = trans->c;
 549         struct bkey_cached *ck;
 550         int ret = 0;
 551
 552         EBUG_ON(path->level);
 553
 554         path->l[1].b = NULL;
 555
 556         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 557                 ck = (void *) path->l[0].b;
 558                 goto fill;
 559         }
 560 retry:
 561         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 562         if (!ck) {
 563                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 564         } else {
 565                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 566
 567                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 568                                       lock_want, _THIS_IP_);
 569                 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
 570
 571                 if (ret)
 572                         return ret;
 573
 574                 if (ck->key.btree_id != path->btree_id ||
 575                     !bpos_eq(ck->key.pos, path->pos)) {
 576                         six_unlock_type(&ck->c.lock, lock_want);
 577                         goto retry;
 578                 }
 579
 580                 mark_btree_node_locked(trans, path, 0,
 581                                        (enum btree_node_locked_type) lock_want);
 582         }
 583
 584         path->l[0].lock_seq     = six_lock_seq(&ck->c.lock);
 585         path->l[0].b            = (void *) ck;
 586 fill:
 587         if (!ck->valid)
 588                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 589
 590         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 591                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 592
 593         path->uptodate = BTREE_ITER_UPTODATE;
 594         EBUG_ON(!ck->valid);
 595         EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 596
 597         return ret;
 598 }
 599
 600 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 601                                      struct bkey_cached_key key,
 602                                      u64 journal_seq,
 603                                      unsigned commit_flags,
 604                                      bool evict)
 605 {
 606         struct bch_fs *c = trans->c;
 607         struct journal *j = &c->journal;
 608         struct btree_iter c_iter, b_iter;
 609         struct bkey_cached *ck = NULL;
 610         int ret;
 611
 612         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 613                              BTREE_ITER_SLOTS|
 614                              BTREE_ITER_INTENT|
 615                              BTREE_ITER_ALL_SNAPSHOTS);
 616         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 617                              BTREE_ITER_CACHED|
 618                              BTREE_ITER_INTENT);
 619         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 620
 621         ret = bch2_btree_iter_traverse(&c_iter);
 622         if (ret)
 623                 goto out;
 624
 625         ck = (void *) c_iter.path->l[0].b;
 626         if (!ck)
 627                 goto out;
 628
 629         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 630                 if (evict)
 631                         goto evict;
 632                 goto out;
 633         }
 634
 635         BUG_ON(!ck->valid);
 636
 637         if (journal_seq && ck->journal.seq != journal_seq)
 638                 goto out;
 639
 640         /*
 641          * Since journal reclaim depends on us making progress here, and the
 642          * allocator/copygc depend on journal reclaim making progress, we need
 643          * to be using alloc reserves:
 644          */
 645         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 646                 bch2_trans_update(trans, &b_iter, ck->k,
 647                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 648                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 649                                   BTREE_TRIGGER_NORUN) ?:
 650                 bch2_trans_commit(trans, NULL, NULL,
 651                                   BTREE_INSERT_NOCHECK_RW|
 652                                   BTREE_INSERT_NOFAIL|
 653                                   (ck->journal.seq == journal_last_seq(j)
 654                                    ? BCH_WATERMARK_reclaim
 655                                    : 0)|
 656                                   commit_flags);
 657
 658         bch2_fs_fatal_err_on(ret &&
 659                              !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 660                              !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
 661                              !bch2_journal_error(j), c,
 662                              "error flushing key cache: %s", bch2_err_str(ret));
 663         if (ret)
 664                 goto out;
 665
 666         bch2_journal_pin_drop(j, &ck->journal);
 667         bch2_journal_preres_put(j, &ck->res);
 668
 669         BUG_ON(!btree_node_locked(c_iter.path, 0));
 670
 671         if (!evict) {
 672                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 673                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 674                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 675                 }
 676         } else {
 677                 struct btree_path *path2;
 678 evict:
 679                 trans_for_each_path(trans, path2)
 680                         if (path2 != c_iter.path)
 681                                 __bch2_btree_path_unlock(trans, path2);
 682
 683                 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
 684
 685                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 686                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 687                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 688                 }
 689
 690                 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
 691                 bkey_cached_evict(&c->btree_key_cache, ck);
 692                 bkey_cached_free_fast(&c->btree_key_cache, ck);
 693         }
 694 out:
 695         bch2_trans_iter_exit(trans, &b_iter);
 696         bch2_trans_iter_exit(trans, &c_iter);
 697         return ret;
 698 }
 699
 700 int bch2_btree_key_cache_journal_flush(struct journal *j,
 701                                 struct journal_entry_pin *pin, u64 seq)
 702 {
 703         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 704         struct bkey_cached *ck =
 705                 container_of(pin, struct bkey_cached, journal);
 706         struct bkey_cached_key key;
 707         struct btree_trans *trans = bch2_trans_get(c);
 708         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 709         int ret = 0;
 710
 711         btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
 712         key = ck->key;
 713
 714         if (ck->journal.seq != seq ||
 715             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 716                 six_unlock_read(&ck->c.lock);
 717                 goto unlock;
 718         }
 719
 720         if (ck->seq != seq) {
 721                 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
 722                                         bch2_btree_key_cache_journal_flush);
 723                 six_unlock_read(&ck->c.lock);
 724                 goto unlock;
 725         }
 726         six_unlock_read(&ck->c.lock);
 727
 728         ret = commit_do(trans, NULL, NULL, 0,
 729                 btree_key_cache_flush_pos(trans, key, seq,
 730                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 731 unlock:
 732         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 733
 734         bch2_trans_put(trans);
 735         return ret;
 736 }
 737
 738 /*
 739  * Flush and evict a key from the key cache:
 740  */
 741 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 742                                enum btree_id id, struct bpos pos)
 743 {
 744         struct bch_fs *c = trans->c;
 745         struct bkey_cached_key key = { id, pos };
 746
 747         /* Fastpath - assume it won't be found: */
 748         if (!bch2_btree_key_cache_find(c, id, pos))
 749                 return 0;
 750
 751         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 752 }
 753
 754 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 755                                   unsigned flags,
 756                                   struct btree_insert_entry *insert_entry)
 757 {
 758         struct bch_fs *c = trans->c;
 759         struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
 760         struct bkey_i *insert = insert_entry->k;
 761         bool kick_reclaim = false;
 762
 763         BUG_ON(insert->k.u64s > ck->u64s);
 764
 765         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 766                 int difference;
 767
 768                 BUG_ON(jset_u64s(insert->k.u64s) > trans->journal_preres.u64s);
 769
 770                 difference = jset_u64s(insert->k.u64s) - ck->res.u64s;
 771                 if (difference > 0) {
 772                         trans->journal_preres.u64s      -= difference;
 773                         ck->res.u64s                    += difference;
 774                 }
 775         }
 776
 777         bkey_copy(ck->k, insert);
 778         ck->valid = true;
 779
 780         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 781                 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
 782                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 783                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 784
 785                 if (bch2_nr_btree_keys_need_flush(c))
 786                         kick_reclaim = true;
 787         }
 788
 789         /*
 790          * To minimize lock contention, we only add the journal pin here and
 791          * defer pin updates to the flush callback via ->seq. Be careful not to
 792          * update ->seq on nojournal commits because we don't want to update the
 793          * pin to a seq that doesn't include journal updates on disk. Otherwise
 794          * we risk losing the update after a crash.
 795          *
 796          * The only exception is if the pin is not active in the first place. We
 797          * have to add the pin because journal reclaim drives key cache
 798          * flushing. The flush callback will not proceed unless ->seq matches
 799          * the latest pin, so make sure it starts with a consistent value.
 800          */
 801         if (!(insert_entry->flags & BTREE_UPDATE_NOJOURNAL) ||
 802             !journal_pin_active(&ck->journal)) {
 803                 ck->seq = trans->journal_res.seq;
 804         }
 805         bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 806                              &ck->journal, bch2_btree_key_cache_journal_flush);
 807
 808         if (kick_reclaim)
 809                 journal_reclaim_kick(&c->journal);
 810         return true;
 811 }
 812
 813 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 814                                struct btree_path *path)
 815 {
 816         struct bch_fs *c = trans->c;
 817         struct bkey_cached *ck = (void *) path->l[0].b;
 818
 819         BUG_ON(!ck->valid);
 820
 821         /*
 822          * We just did an update to the btree, bypassing the key cache: the key
 823          * cache key is now stale and must be dropped, even if dirty:
 824          */
 825         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 826                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 827                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 828                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 829         }
 830
 831         ck->valid = false;
 832 }
 833
 834 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 835                                            struct shrink_control *sc)
 836 {
 837         struct bch_fs *c = shrink->private_data;
 838         struct btree_key_cache *bc = &c->btree_key_cache;
 839         struct bucket_table *tbl;
 840         struct bkey_cached *ck, *t;
 841         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 842         unsigned start, flags;
 843         int srcu_idx;
 844
 845         mutex_lock(&bc->lock);
 846         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 847         flags = memalloc_nofs_save();
 848
 849         /*
 850          * Newest freed entries are at the end of the list - once we hit one
 851          * that's too new to be freed, we can bail out:
 852          */
 853         list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
 854                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 855                                                  ck->btree_trans_barrier_seq))
 856                         break;
 857
 858                 list_del(&ck->list);
 859                 six_lock_exit(&ck->c.lock);
 860                 kmem_cache_free(bch2_key_cache, ck);
 861                 atomic_long_dec(&bc->nr_freed);
 862                 scanned++;
 863                 freed++;
 864         }
 865
 866         if (scanned >= nr)
 867                 goto out;
 868
 869         list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
 870                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 871                                                  ck->btree_trans_barrier_seq))
 872                         break;
 873
 874                 list_del(&ck->list);
 875                 six_lock_exit(&ck->c.lock);
 876                 kmem_cache_free(bch2_key_cache, ck);
 877                 atomic_long_dec(&bc->nr_freed);
 878                 scanned++;
 879                 freed++;
 880         }
 881
 882         if (scanned >= nr)
 883                 goto out;
 884
 885         rcu_read_lock();
 886         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 887         if (bc->shrink_iter >= tbl->size)
 888                 bc->shrink_iter = 0;
 889         start = bc->shrink_iter;
 890
 891         do {
 892                 struct rhash_head *pos, *next;
 893
 894                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 895
 896                 while (!rht_is_a_nulls(pos)) {
 897                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 898                         ck = container_of(pos, struct bkey_cached, hash);
 899
 900                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 901                                 goto next;
 902
 903                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 904                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 905                         else if (bkey_cached_lock_for_evict(ck)) {
 906                                 bkey_cached_evict(bc, ck);
 907                                 bkey_cached_free(bc, ck);
 908                         }
 909
 910                         scanned++;
 911                         if (scanned >= nr)
 912                                 break;
 913 next:
 914                         pos = next;
 915                 }
 916
 917                 bc->shrink_iter++;
 918                 if (bc->shrink_iter >= tbl->size)
 919                         bc->shrink_iter = 0;
 920         } while (scanned < nr && bc->shrink_iter != start);
 921
 922         rcu_read_unlock();
 923 out:
 924         memalloc_nofs_restore(flags);
 925         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 926         mutex_unlock(&bc->lock);
 927
 928         return freed;
 929 }
 930
 931 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 932                                             struct shrink_control *sc)
 933 {
 934         struct bch_fs *c = shrink->private_data;
 935         struct btree_key_cache *bc = &c->btree_key_cache;
 936         long nr = atomic_long_read(&bc->nr_keys) -
 937                 atomic_long_read(&bc->nr_dirty);
 938
 939         return max(0L, nr);
 940 }
 941
 942 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 943 {
 944         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 945         struct bucket_table *tbl;
 946         struct bkey_cached *ck, *n;
 947         struct rhash_head *pos;
 948         LIST_HEAD(items);
 949         unsigned i;
 950 #ifdef __KERNEL__
 951         int cpu;
 952 #endif
 953
 954         shrinker_free(bc->shrink);
 955
 956         mutex_lock(&bc->lock);
 957
 958         /*
 959          * The loop is needed to guard against racing with rehash:
 960          */
 961         while (atomic_long_read(&bc->nr_keys)) {
 962                 rcu_read_lock();
 963                 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 964                 if (tbl)
 965                         for (i = 0; i < tbl->size; i++)
 966                                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 967                                         bkey_cached_evict(bc, ck);
 968                                         list_add(&ck->list, &items);
 969                                 }
 970                 rcu_read_unlock();
 971         }
 972
 973 #ifdef __KERNEL__
 974         for_each_possible_cpu(cpu) {
 975                 struct btree_key_cache_freelist *f =
 976                         per_cpu_ptr(bc->pcpu_freed, cpu);
 977
 978                 for (i = 0; i < f->nr; i++) {
 979                         ck = f->objs[i];
 980                         list_add(&ck->list, &items);
 981                 }
 982         }
 983 #endif
 984
 985         list_splice(&bc->freed_pcpu,    &items);
 986         list_splice(&bc->freed_nonpcpu, &items);
 987
 988         mutex_unlock(&bc->lock);
 989
 990         list_for_each_entry_safe(ck, n, &items, list) {
 991                 cond_resched();
 992
 993                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 994                 bch2_journal_preres_put(&c->journal, &ck->res);
 995
 996                 list_del(&ck->list);
 997                 kfree(ck->k);
 998                 six_lock_exit(&ck->c.lock);
 999                 kmem_cache_free(bch2_key_cache, ck);
1000         }
1001
1002         if (atomic_long_read(&bc->nr_dirty) &&
1003             !bch2_journal_error(&c->journal) &&
1004             test_bit(BCH_FS_WAS_RW, &c->flags))
1005                 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
1006                       atomic_long_read(&bc->nr_dirty));
1007
1008         if (atomic_long_read(&bc->nr_keys))
1009                 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
1010                       atomic_long_read(&bc->nr_keys));
1011
1012         if (bc->table_init_done)
1013                 rhashtable_destroy(&bc->table);
1014
1015         free_percpu(bc->pcpu_freed);
1016 }
1017
1018 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1019 {
1020         mutex_init(&c->lock);
1021         INIT_LIST_HEAD(&c->freed_pcpu);
1022         INIT_LIST_HEAD(&c->freed_nonpcpu);
1023 }
1024
1025 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1026 {
1027         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1028         struct shrinker *shrink;
1029
1030 #ifdef __KERNEL__
1031         bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1032         if (!bc->pcpu_freed)
1033                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1034 #endif
1035
1036         if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1037                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1038
1039         bc->table_init_done = true;
1040
1041         shrink = shrinker_alloc(0, "%s-btree_key_cache", c->name);
1042         if (!shrink)
1043                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1044         bc->shrink = shrink;
1045         shrink->seeks           = 0;
1046         shrink->count_objects   = bch2_btree_key_cache_count;
1047         shrink->scan_objects    = bch2_btree_key_cache_scan;
1048         shrink->private_data    = c;
1049         shrinker_register(shrink);
1050         return 0;
1051 }
1052
1053 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1054 {
1055         prt_printf(out, "nr_freed:\t%lu",       atomic_long_read(&c->nr_freed));
1056         prt_newline(out);
1057         prt_printf(out, "nr_keys:\t%lu",        atomic_long_read(&c->nr_keys));
1058         prt_newline(out);
1059         prt_printf(out, "nr_dirty:\t%lu",       atomic_long_read(&c->nr_dirty));
1060         prt_newline(out);
1061 }
1062
1063 void bch2_btree_key_cache_exit(void)
1064 {
1065         kmem_cache_destroy(bch2_key_cache);
1066 }
1067
1068 int __init bch2_btree_key_cache_init(void)
1069 {
1070         bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
1071         if (!bch2_key_cache)
1072                 return -ENOMEM;
1073
1074         return 0;
1075 }