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                 bc->nr_freed_pcpu++;
  95         } else {
  96                 list_move_tail(&ck->list, &bc->freed_nonpcpu);
  97                 bc->nr_freed_nonpcpu++;
  98         }
  99         atomic_long_inc(&bc->nr_freed);
 100
 101         kfree(ck->k);
 102         ck->k           = NULL;
 103         ck->u64s        = 0;
 104
 105         six_unlock_write(&ck->c.lock);
 106         six_unlock_intent(&ck->c.lock);
 107 }
 108
 109 #ifdef __KERNEL__
 110 static void __bkey_cached_move_to_freelist_ordered(struct btree_key_cache *bc,
 111                                                    struct bkey_cached *ck)
 112 {
 113         struct bkey_cached *pos;
 114
 115         bc->nr_freed_nonpcpu++;
 116
 117         list_for_each_entry_reverse(pos, &bc->freed_nonpcpu, list) {
 118                 if (ULONG_CMP_GE(ck->btree_trans_barrier_seq,
 119                                  pos->btree_trans_barrier_seq)) {
 120                         list_move(&ck->list, &pos->list);
 121                         return;
 122                 }
 123         }
 124
 125         list_move(&ck->list, &bc->freed_nonpcpu);
 126 }
 127 #endif
 128
 129 static void bkey_cached_move_to_freelist(struct btree_key_cache *bc,
 130                                          struct bkey_cached *ck)
 131 {
 132         BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
 133
 134         if (!ck->c.lock.readers) {
 135 #ifdef __KERNEL__
 136                 struct btree_key_cache_freelist *f;
 137                 bool freed = false;
 138
 139                 preempt_disable();
 140                 f = this_cpu_ptr(bc->pcpu_freed);
 141
 142                 if (f->nr < ARRAY_SIZE(f->objs)) {
 143                         f->objs[f->nr++] = ck;
 144                         freed = true;
 145                 }
 146                 preempt_enable();
 147
 148                 if (!freed) {
 149                         mutex_lock(&bc->lock);
 150                         preempt_disable();
 151                         f = this_cpu_ptr(bc->pcpu_freed);
 152
 153                         while (f->nr > ARRAY_SIZE(f->objs) / 2) {
 154                                 struct bkey_cached *ck2 = f->objs[--f->nr];
 155
 156                                 __bkey_cached_move_to_freelist_ordered(bc, ck2);
 157                         }
 158                         preempt_enable();
 159
 160                         __bkey_cached_move_to_freelist_ordered(bc, ck);
 161                         mutex_unlock(&bc->lock);
 162                 }
 163 #else
 164                 mutex_lock(&bc->lock);
 165                 list_move_tail(&ck->list, &bc->freed_nonpcpu);
 166                 bc->nr_freed_nonpcpu++;
 167                 mutex_unlock(&bc->lock);
 168 #endif
 169         } else {
 170                 mutex_lock(&bc->lock);
 171                 list_move_tail(&ck->list, &bc->freed_pcpu);
 172                 bc->nr_freed_pcpu++;
 173                 mutex_unlock(&bc->lock);
 174         }
 175 }
 176
 177 static void bkey_cached_free_fast(struct btree_key_cache *bc,
 178                                   struct bkey_cached *ck)
 179 {
 180         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 181
 182         ck->btree_trans_barrier_seq =
 183                 start_poll_synchronize_srcu(&c->btree_trans_barrier);
 184
 185         list_del_init(&ck->list);
 186         atomic_long_inc(&bc->nr_freed);
 187
 188         kfree(ck->k);
 189         ck->k           = NULL;
 190         ck->u64s        = 0;
 191
 192         bkey_cached_move_to_freelist(bc, ck);
 193
 194         six_unlock_write(&ck->c.lock);
 195         six_unlock_intent(&ck->c.lock);
 196 }
 197
 198 static struct bkey_cached *
 199 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path,
 200                   bool *was_new)
 201 {
 202         struct bch_fs *c = trans->c;
 203         struct btree_key_cache *bc = &c->btree_key_cache;
 204         struct bkey_cached *ck = NULL;
 205         bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
 206         int ret;
 207
 208         if (!pcpu_readers) {
 209 #ifdef __KERNEL__
 210                 struct btree_key_cache_freelist *f;
 211
 212                 preempt_disable();
 213                 f = this_cpu_ptr(bc->pcpu_freed);
 214                 if (f->nr)
 215                         ck = f->objs[--f->nr];
 216                 preempt_enable();
 217
 218                 if (!ck) {
 219                         mutex_lock(&bc->lock);
 220                         preempt_disable();
 221                         f = this_cpu_ptr(bc->pcpu_freed);
 222
 223                         while (!list_empty(&bc->freed_nonpcpu) &&
 224                                f->nr < ARRAY_SIZE(f->objs) / 2) {
 225                                 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 226                                 list_del_init(&ck->list);
 227                                 bc->nr_freed_nonpcpu--;
 228                                 f->objs[f->nr++] = ck;
 229                         }
 230
 231                         ck = f->nr ? f->objs[--f->nr] : NULL;
 232                         preempt_enable();
 233                         mutex_unlock(&bc->lock);
 234                 }
 235 #else
 236                 mutex_lock(&bc->lock);
 237                 if (!list_empty(&bc->freed_nonpcpu)) {
 238                         ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 239                         list_del_init(&ck->list);
 240                         bc->nr_freed_nonpcpu--;
 241                 }
 242                 mutex_unlock(&bc->lock);
 243 #endif
 244         } else {
 245                 mutex_lock(&bc->lock);
 246                 if (!list_empty(&bc->freed_pcpu)) {
 247                         ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
 248                         list_del_init(&ck->list);
 249                         bc->nr_freed_pcpu--;
 250                 }
 251                 mutex_unlock(&bc->lock);
 252         }
 253
 254         if (ck) {
 255                 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent, _THIS_IP_);
 256                 if (unlikely(ret)) {
 257                         bkey_cached_move_to_freelist(bc, ck);
 258                         return ERR_PTR(ret);
 259                 }
 260
 261                 path->l[0].b = (void *) ck;
 262                 path->l[0].lock_seq = six_lock_seq(&ck->c.lock);
 263                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
 264
 265                 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
 266                 if (unlikely(ret)) {
 267                         btree_node_unlock(trans, path, 0);
 268                         bkey_cached_move_to_freelist(bc, ck);
 269                         return ERR_PTR(ret);
 270                 }
 271
 272                 return ck;
 273         }
 274
 275         ck = allocate_dropping_locks(trans, ret,
 276                         kmem_cache_zalloc(bch2_key_cache, _gfp));
 277         if (ret) {
 278                 kmem_cache_free(bch2_key_cache, ck);
 279                 return ERR_PTR(ret);
 280         }
 281
 282         if (!ck)
 283                 return NULL;
 284
 285         INIT_LIST_HEAD(&ck->list);
 286         bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
 287
 288         ck->c.cached = true;
 289         BUG_ON(!six_trylock_intent(&ck->c.lock));
 290         BUG_ON(!six_trylock_write(&ck->c.lock));
 291         *was_new = true;
 292         return ck;
 293 }
 294
 295 static struct bkey_cached *
 296 bkey_cached_reuse(struct btree_key_cache *c)
 297 {
 298         struct bucket_table *tbl;
 299         struct rhash_head *pos;
 300         struct bkey_cached *ck;
 301         unsigned i;
 302
 303         mutex_lock(&c->lock);
 304         rcu_read_lock();
 305         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 306         for (i = 0; i < tbl->size; i++)
 307                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 308                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 309                             bkey_cached_lock_for_evict(ck)) {
 310                                 bkey_cached_evict(c, ck);
 311                                 goto out;
 312                         }
 313                 }
 314         ck = NULL;
 315 out:
 316         rcu_read_unlock();
 317         mutex_unlock(&c->lock);
 318         return ck;
 319 }
 320
 321 static struct bkey_cached *
 322 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
 323 {
 324         struct bch_fs *c = trans->c;
 325         struct btree_key_cache *bc = &c->btree_key_cache;
 326         struct bkey_cached *ck;
 327         bool was_new = false;
 328
 329         ck = bkey_cached_alloc(trans, path, &was_new);
 330         if (IS_ERR(ck))
 331                 return ck;
 332
 333         if (unlikely(!ck)) {
 334                 ck = bkey_cached_reuse(bc);
 335                 if (unlikely(!ck)) {
 336                         bch_err(c, "error allocating memory for key cache item, btree %s",
 337                                 bch2_btree_id_str(path->btree_id));
 338                         return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
 339                 }
 340
 341                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
 342         }
 343
 344         ck->c.level             = 0;
 345         ck->c.btree_id          = path->btree_id;
 346         ck->key.btree_id        = path->btree_id;
 347         ck->key.pos             = path->pos;
 348         ck->valid               = false;
 349         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 350
 351         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 352                                           &ck->hash,
 353                                           bch2_btree_key_cache_params))) {
 354                 /* We raced with another fill: */
 355
 356                 if (likely(was_new)) {
 357                         six_unlock_write(&ck->c.lock);
 358                         six_unlock_intent(&ck->c.lock);
 359                         kfree(ck);
 360                 } else {
 361                         bkey_cached_free_fast(bc, ck);
 362                 }
 363
 364                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
 365                 return NULL;
 366         }
 367
 368         atomic_long_inc(&bc->nr_keys);
 369
 370         six_unlock_write(&ck->c.lock);
 371
 372         return ck;
 373 }
 374
 375 static int btree_key_cache_fill(struct btree_trans *trans,
 376                                 struct btree_path *ck_path,
 377                                 struct bkey_cached *ck)
 378 {
 379         struct btree_iter iter;
 380         struct bkey_s_c k;
 381         unsigned new_u64s = 0;
 382         struct bkey_i *new_k = NULL;
 383         int ret;
 384
 385         bch2_trans_iter_init(trans, &iter, ck->key.btree_id, ck->key.pos,
 386                              BTREE_ITER_key_cache_fill|
 387                              BTREE_ITER_cached_nofill);
 388         iter.flags &= ~BTREE_ITER_with_journal;
 389         k = bch2_btree_iter_peek_slot(&iter);
 390         ret = bkey_err(k);
 391         if (ret)
 392                 goto err;
 393
 394         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 395                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 396                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 397                 goto err;
 398         }
 399
 400         /*
 401          * bch2_varint_decode can read past the end of the buffer by at
 402          * most 7 bytes (it won't be used):
 403          */
 404         new_u64s = k.k->u64s + 1;
 405
 406         /*
 407          * Allocate some extra space so that the transaction commit path is less
 408          * likely to have to reallocate, since that requires a transaction
 409          * restart:
 410          */
 411         new_u64s = min(256U, (new_u64s * 3) / 2);
 412
 413         if (new_u64s > ck->u64s) {
 414                 new_u64s = roundup_pow_of_two(new_u64s);
 415                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 416                 if (!new_k) {
 417                         bch2_trans_unlock(trans);
 418
 419                         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 420                         if (!new_k) {
 421                                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 422                                         bch2_btree_id_str(ck->key.btree_id), new_u64s);
 423                                 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
 424                                 goto err;
 425                         }
 426
 427                         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 428                                 kfree(new_k);
 429                                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 430                                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 431                                 goto err;
 432                         }
 433
 434                         ret = bch2_trans_relock(trans);
 435                         if (ret) {
 436                                 kfree(new_k);
 437                                 goto err;
 438                         }
 439                 }
 440         }
 441
 442         ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
 443         if (ret) {
 444                 kfree(new_k);
 445                 goto err;
 446         }
 447
 448         if (new_k) {
 449                 kfree(ck->k);
 450                 ck->u64s = new_u64s;
 451                 ck->k = new_k;
 452         }
 453
 454         bkey_reassemble(ck->k, k);
 455         ck->valid = true;
 456         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 457
 458         /* We're not likely to need this iterator again: */
 459         bch2_set_btree_iter_dontneed(&iter);
 460 err:
 461         bch2_trans_iter_exit(trans, &iter);
 462         return ret;
 463 }
 464
 465 static noinline int
 466 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
 467                                          unsigned flags)
 468 {
 469         struct bch_fs *c = trans->c;
 470         struct bkey_cached *ck;
 471         int ret = 0;
 472
 473         BUG_ON(path->level);
 474
 475         path->l[1].b = NULL;
 476
 477         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 478                 ck = (void *) path->l[0].b;
 479                 goto fill;
 480         }
 481 retry:
 482         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 483         if (!ck) {
 484                 ck = btree_key_cache_create(trans, path);
 485                 ret = PTR_ERR_OR_ZERO(ck);
 486                 if (ret)
 487                         goto err;
 488                 if (!ck)
 489                         goto retry;
 490
 491                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
 492                 path->locks_want = 1;
 493         } else {
 494                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 495
 496                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 497                                       lock_want, _THIS_IP_);
 498                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 499                         goto err;
 500
 501                 BUG_ON(ret);
 502
 503                 if (ck->key.btree_id != path->btree_id ||
 504                     !bpos_eq(ck->key.pos, path->pos)) {
 505                         six_unlock_type(&ck->c.lock, lock_want);
 506                         goto retry;
 507                 }
 508
 509                 mark_btree_node_locked(trans, path, 0,
 510                                        (enum btree_node_locked_type) lock_want);
 511         }
 512
 513         path->l[0].lock_seq     = six_lock_seq(&ck->c.lock);
 514         path->l[0].b            = (void *) ck;
 515 fill:
 516         path->uptodate = BTREE_ITER_UPTODATE;
 517
 518         if (!ck->valid && !(flags & BTREE_ITER_cached_nofill)) {
 519                 ret =   bch2_btree_path_upgrade(trans, path, 1) ?:
 520                         btree_key_cache_fill(trans, path, ck) ?:
 521                         bch2_btree_path_relock(trans, path, _THIS_IP_);
 522                 if (ret)
 523                         goto err;
 524
 525                 path->uptodate = BTREE_ITER_UPTODATE;
 526         }
 527
 528         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 529                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 530
 531         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 532         BUG_ON(path->uptodate);
 533
 534         return ret;
 535 err:
 536         path->uptodate = BTREE_ITER_NEED_TRAVERSE;
 537         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 538                 btree_node_unlock(trans, path, 0);
 539                 path->l[0].b = ERR_PTR(ret);
 540         }
 541         return ret;
 542 }
 543
 544 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 545                                     unsigned flags)
 546 {
 547         struct bch_fs *c = trans->c;
 548         struct bkey_cached *ck;
 549         int ret = 0;
 550
 551         EBUG_ON(path->level);
 552
 553         path->l[1].b = NULL;
 554
 555         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 556                 ck = (void *) path->l[0].b;
 557                 goto fill;
 558         }
 559 retry:
 560         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 561         if (!ck) {
 562                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 563         } else {
 564                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 565
 566                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 567                                       lock_want, _THIS_IP_);
 568                 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
 569
 570                 if (ret)
 571                         return ret;
 572
 573                 if (ck->key.btree_id != path->btree_id ||
 574                     !bpos_eq(ck->key.pos, path->pos)) {
 575                         six_unlock_type(&ck->c.lock, lock_want);
 576                         goto retry;
 577                 }
 578
 579                 mark_btree_node_locked(trans, path, 0,
 580                                        (enum btree_node_locked_type) lock_want);
 581         }
 582
 583         path->l[0].lock_seq     = six_lock_seq(&ck->c.lock);
 584         path->l[0].b            = (void *) ck;
 585 fill:
 586         if (!ck->valid)
 587                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 588
 589         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 590                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 591
 592         path->uptodate = BTREE_ITER_UPTODATE;
 593         EBUG_ON(!ck->valid);
 594         EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 595
 596         return ret;
 597 }
 598
 599 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 600                                      struct bkey_cached_key key,
 601                                      u64 journal_seq,
 602                                      unsigned commit_flags,
 603                                      bool evict)
 604 {
 605         struct bch_fs *c = trans->c;
 606         struct journal *j = &c->journal;
 607         struct btree_iter c_iter, b_iter;
 608         struct bkey_cached *ck = NULL;
 609         int ret;
 610
 611         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 612                              BTREE_ITER_slots|
 613                              BTREE_ITER_intent|
 614                              BTREE_ITER_all_snapshots);
 615         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 616                              BTREE_ITER_cached|
 617                              BTREE_ITER_intent);
 618         b_iter.flags &= ~BTREE_ITER_with_key_cache;
 619
 620         ret = bch2_btree_iter_traverse(&c_iter);
 621         if (ret)
 622                 goto out;
 623
 624         ck = (void *) btree_iter_path(trans, &c_iter)->l[0].b;
 625         if (!ck)
 626                 goto out;
 627
 628         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 629                 if (evict)
 630                         goto evict;
 631                 goto out;
 632         }
 633
 634         BUG_ON(!ck->valid);
 635
 636         if (journal_seq && ck->journal.seq != journal_seq)
 637                 goto out;
 638
 639         trans->journal_res.seq = ck->journal.seq;
 640
 641         /*
 642          * If we're at the end of the journal, we really want to free up space
 643          * in the journal right away - we don't want to pin that old journal
 644          * sequence number with a new btree node write, we want to re-journal
 645          * the update
 646          */
 647         if (ck->journal.seq == journal_last_seq(j))
 648                 commit_flags |= BCH_WATERMARK_reclaim;
 649
 650         if (ck->journal.seq != journal_last_seq(j) ||
 651             !test_bit(JOURNAL_space_low, &c->journal.flags))
 652                 commit_flags |= BCH_TRANS_COMMIT_no_journal_res;
 653
 654         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 655                 bch2_trans_update(trans, &b_iter, ck->k,
 656                                   BTREE_UPDATE_key_cache_reclaim|
 657                                   BTREE_UPDATE_internal_snapshot_node|
 658                                   BTREE_TRIGGER_norun) ?:
 659                 bch2_trans_commit(trans, NULL, NULL,
 660                                   BCH_TRANS_COMMIT_no_check_rw|
 661                                   BCH_TRANS_COMMIT_no_enospc|
 662                                   commit_flags);
 663
 664         bch2_fs_fatal_err_on(ret &&
 665                              !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 666                              !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
 667                              !bch2_journal_error(j), c,
 668                              "flushing key cache: %s", bch2_err_str(ret));
 669         if (ret)
 670                 goto out;
 671
 672         bch2_journal_pin_drop(j, &ck->journal);
 673
 674         struct btree_path *path = btree_iter_path(trans, &c_iter);
 675         BUG_ON(!btree_node_locked(path, 0));
 676
 677         if (!evict) {
 678                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 679                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 680                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 681                 }
 682         } else {
 683                 struct btree_path *path2;
 684                 unsigned i;
 685 evict:
 686                 trans_for_each_path(trans, path2, i)
 687                         if (path2 != path)
 688                                 __bch2_btree_path_unlock(trans, path2);
 689
 690                 bch2_btree_node_lock_write_nofail(trans, path, &ck->c);
 691
 692                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 693                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 694                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 695                 }
 696
 697                 mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
 698                 bkey_cached_evict(&c->btree_key_cache, ck);
 699                 bkey_cached_free_fast(&c->btree_key_cache, ck);
 700         }
 701 out:
 702         bch2_trans_iter_exit(trans, &b_iter);
 703         bch2_trans_iter_exit(trans, &c_iter);
 704         return ret;
 705 }
 706
 707 int bch2_btree_key_cache_journal_flush(struct journal *j,
 708                                 struct journal_entry_pin *pin, u64 seq)
 709 {
 710         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 711         struct bkey_cached *ck =
 712                 container_of(pin, struct bkey_cached, journal);
 713         struct bkey_cached_key key;
 714         struct btree_trans *trans = bch2_trans_get(c);
 715         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 716         int ret = 0;
 717
 718         btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
 719         key = ck->key;
 720
 721         if (ck->journal.seq != seq ||
 722             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 723                 six_unlock_read(&ck->c.lock);
 724                 goto unlock;
 725         }
 726
 727         if (ck->seq != seq) {
 728                 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
 729                                         bch2_btree_key_cache_journal_flush);
 730                 six_unlock_read(&ck->c.lock);
 731                 goto unlock;
 732         }
 733         six_unlock_read(&ck->c.lock);
 734
 735         ret = lockrestart_do(trans,
 736                 btree_key_cache_flush_pos(trans, key, seq,
 737                                 BCH_TRANS_COMMIT_journal_reclaim, false));
 738 unlock:
 739         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 740
 741         bch2_trans_put(trans);
 742         return ret;
 743 }
 744
 745 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 746                                   unsigned flags,
 747                                   struct btree_insert_entry *insert_entry)
 748 {
 749         struct bch_fs *c = trans->c;
 750         struct bkey_cached *ck = (void *) (trans->paths + insert_entry->path)->l[0].b;
 751         struct bkey_i *insert = insert_entry->k;
 752         bool kick_reclaim = false;
 753
 754         BUG_ON(insert->k.u64s > ck->u64s);
 755
 756         bkey_copy(ck->k, insert);
 757         ck->valid = true;
 758
 759         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 760                 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
 761                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 762                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 763
 764                 if (bch2_nr_btree_keys_need_flush(c))
 765                         kick_reclaim = true;
 766         }
 767
 768         /*
 769          * To minimize lock contention, we only add the journal pin here and
 770          * defer pin updates to the flush callback via ->seq. Be careful not to
 771          * update ->seq on nojournal commits because we don't want to update the
 772          * pin to a seq that doesn't include journal updates on disk. Otherwise
 773          * we risk losing the update after a crash.
 774          *
 775          * The only exception is if the pin is not active in the first place. We
 776          * have to add the pin because journal reclaim drives key cache
 777          * flushing. The flush callback will not proceed unless ->seq matches
 778          * the latest pin, so make sure it starts with a consistent value.
 779          */
 780         if (!(insert_entry->flags & BTREE_UPDATE_nojournal) ||
 781             !journal_pin_active(&ck->journal)) {
 782                 ck->seq = trans->journal_res.seq;
 783         }
 784         bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 785                              &ck->journal, bch2_btree_key_cache_journal_flush);
 786
 787         if (kick_reclaim)
 788                 journal_reclaim_kick(&c->journal);
 789         return true;
 790 }
 791
 792 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 793                                struct btree_path *path)
 794 {
 795         struct bch_fs *c = trans->c;
 796         struct bkey_cached *ck = (void *) path->l[0].b;
 797
 798         BUG_ON(!ck->valid);
 799
 800         /*
 801          * We just did an update to the btree, bypassing the key cache: the key
 802          * cache key is now stale and must be dropped, even if dirty:
 803          */
 804         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 805                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 806                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 807                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 808         }
 809
 810         ck->valid = false;
 811 }
 812
 813 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 814                                            struct shrink_control *sc)
 815 {
 816         struct bch_fs *c = shrink->private_data;
 817         struct btree_key_cache *bc = &c->btree_key_cache;
 818         struct bucket_table *tbl;
 819         struct bkey_cached *ck, *t;
 820         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 821         unsigned start, flags;
 822         int srcu_idx;
 823
 824         mutex_lock(&bc->lock);
 825         bc->requested_to_free += sc->nr_to_scan;
 826
 827         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 828         flags = memalloc_nofs_save();
 829
 830         /*
 831          * Newest freed entries are at the end of the list - once we hit one
 832          * that's too new to be freed, we can bail out:
 833          */
 834         list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
 835                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 836                                                  ck->btree_trans_barrier_seq))
 837                         break;
 838
 839                 list_del(&ck->list);
 840                 six_lock_exit(&ck->c.lock);
 841                 kmem_cache_free(bch2_key_cache, ck);
 842                 atomic_long_dec(&bc->nr_freed);
 843                 freed++;
 844                 bc->nr_freed_nonpcpu--;
 845                 bc->freed++;
 846         }
 847
 848         list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
 849                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 850                                                  ck->btree_trans_barrier_seq))
 851                         break;
 852
 853                 list_del(&ck->list);
 854                 six_lock_exit(&ck->c.lock);
 855                 kmem_cache_free(bch2_key_cache, ck);
 856                 atomic_long_dec(&bc->nr_freed);
 857                 freed++;
 858                 bc->nr_freed_pcpu--;
 859                 bc->freed++;
 860         }
 861
 862         rcu_read_lock();
 863         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 864         if (bc->shrink_iter >= tbl->size)
 865                 bc->shrink_iter = 0;
 866         start = bc->shrink_iter;
 867
 868         do {
 869                 struct rhash_head *pos, *next;
 870
 871                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 872
 873                 while (!rht_is_a_nulls(pos)) {
 874                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 875                         ck = container_of(pos, struct bkey_cached, hash);
 876
 877                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 878                                 bc->skipped_dirty++;
 879                                 goto next;
 880                         } else if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags)) {
 881                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 882                                 bc->skipped_accessed++;
 883                                 goto next;
 884                         } else if (bkey_cached_lock_for_evict(ck)) {
 885                                 bkey_cached_evict(bc, ck);
 886                                 bkey_cached_free(bc, ck);
 887                                 bc->moved_to_freelist++;
 888                         } else {
 889                                 bc->skipped_lock_fail++;
 890                         }
 891
 892                         scanned++;
 893                         if (scanned >= nr)
 894                                 break;
 895 next:
 896                         pos = next;
 897                 }
 898
 899                 bc->shrink_iter++;
 900                 if (bc->shrink_iter >= tbl->size)
 901                         bc->shrink_iter = 0;
 902         } while (scanned < nr && bc->shrink_iter != start);
 903
 904         rcu_read_unlock();
 905         memalloc_nofs_restore(flags);
 906         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 907         mutex_unlock(&bc->lock);
 908
 909         return freed;
 910 }
 911
 912 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 913                                             struct shrink_control *sc)
 914 {
 915         struct bch_fs *c = shrink->private_data;
 916         struct btree_key_cache *bc = &c->btree_key_cache;
 917         long nr = atomic_long_read(&bc->nr_keys) -
 918                 atomic_long_read(&bc->nr_dirty);
 919
 920         return max(0L, nr);
 921 }
 922
 923 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 924 {
 925         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 926         struct bucket_table *tbl;
 927         struct bkey_cached *ck, *n;
 928         struct rhash_head *pos;
 929         LIST_HEAD(items);
 930         unsigned i;
 931 #ifdef __KERNEL__
 932         int cpu;
 933 #endif
 934
 935         shrinker_free(bc->shrink);
 936
 937         mutex_lock(&bc->lock);
 938
 939         /*
 940          * The loop is needed to guard against racing with rehash:
 941          */
 942         while (atomic_long_read(&bc->nr_keys)) {
 943                 rcu_read_lock();
 944                 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 945                 if (tbl)
 946                         for (i = 0; i < tbl->size; i++)
 947                                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 948                                         bkey_cached_evict(bc, ck);
 949                                         list_add(&ck->list, &items);
 950                                 }
 951                 rcu_read_unlock();
 952         }
 953
 954 #ifdef __KERNEL__
 955         if (bc->pcpu_freed) {
 956                 for_each_possible_cpu(cpu) {
 957                         struct btree_key_cache_freelist *f =
 958                                 per_cpu_ptr(bc->pcpu_freed, cpu);
 959
 960                         for (i = 0; i < f->nr; i++) {
 961                                 ck = f->objs[i];
 962                                 list_add(&ck->list, &items);
 963                         }
 964                 }
 965         }
 966 #endif
 967
 968         BUG_ON(list_count_nodes(&bc->freed_pcpu) != bc->nr_freed_pcpu);
 969         BUG_ON(list_count_nodes(&bc->freed_nonpcpu) != bc->nr_freed_nonpcpu);
 970
 971         list_splice(&bc->freed_pcpu,    &items);
 972         list_splice(&bc->freed_nonpcpu, &items);
 973
 974         mutex_unlock(&bc->lock);
 975
 976         list_for_each_entry_safe(ck, n, &items, list) {
 977                 cond_resched();
 978
 979                 list_del(&ck->list);
 980                 kfree(ck->k);
 981                 six_lock_exit(&ck->c.lock);
 982                 kmem_cache_free(bch2_key_cache, ck);
 983         }
 984
 985         if (atomic_long_read(&bc->nr_dirty) &&
 986             !bch2_journal_error(&c->journal) &&
 987             test_bit(BCH_FS_was_rw, &c->flags))
 988                 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
 989                       atomic_long_read(&bc->nr_dirty));
 990
 991         if (atomic_long_read(&bc->nr_keys))
 992                 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
 993                       atomic_long_read(&bc->nr_keys));
 994
 995         if (bc->table_init_done)
 996                 rhashtable_destroy(&bc->table);
 997
 998         free_percpu(bc->pcpu_freed);
 999 }
1000
1001 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1002 {
1003         mutex_init(&c->lock);
1004         INIT_LIST_HEAD(&c->freed_pcpu);
1005         INIT_LIST_HEAD(&c->freed_nonpcpu);
1006 }
1007
1008 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1009 {
1010         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1011         struct shrinker *shrink;
1012
1013 #ifdef __KERNEL__
1014         bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1015         if (!bc->pcpu_freed)
1016                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1017 #endif
1018
1019         if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1020                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1021
1022         bc->table_init_done = true;
1023
1024         shrink = shrinker_alloc(0, "%s-btree_key_cache", c->name);
1025         if (!shrink)
1026                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1027         bc->shrink = shrink;
1028         shrink->seeks           = 0;
1029         shrink->count_objects   = bch2_btree_key_cache_count;
1030         shrink->scan_objects    = bch2_btree_key_cache_scan;
1031         shrink->private_data    = c;
1032         shrinker_register(shrink);
1033         return 0;
1034 }
1035
1036 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *bc)
1037 {
1038         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1039
1040         printbuf_tabstop_push(out, 24);
1041         printbuf_tabstop_push(out, 12);
1042
1043         unsigned flags = memalloc_nofs_save();
1044         mutex_lock(&bc->lock);
1045         prt_printf(out, "keys:\t%lu\r\n",               atomic_long_read(&bc->nr_keys));
1046         prt_printf(out, "dirty:\t%lu\r\n",              atomic_long_read(&bc->nr_dirty));
1047         prt_printf(out, "freelist:\t%lu\r\n",           atomic_long_read(&bc->nr_freed));
1048         prt_printf(out, "nonpcpu freelist:\t%zu\r\n",   bc->nr_freed_nonpcpu);
1049         prt_printf(out, "pcpu freelist:\t%zu\r\n",      bc->nr_freed_pcpu);
1050
1051         prt_printf(out, "\nshrinker:\n");
1052         prt_printf(out, "requested_to_free:\t%lu\r\n",  bc->requested_to_free);
1053         prt_printf(out, "freed:\t%lu\r\n",              bc->freed);
1054         prt_printf(out, "moved_to_freelist:\t%lu\r\n",  bc->moved_to_freelist);
1055         prt_printf(out, "skipped_dirty:\t%lu\r\n",      bc->skipped_dirty);
1056         prt_printf(out, "skipped_accessed:\t%lu\r\n",   bc->skipped_accessed);
1057         prt_printf(out, "skipped_lock_fail:\t%lu\r\n",  bc->skipped_lock_fail);
1058
1059         prt_printf(out, "srcu seq:\t%lu\r\n",           get_state_synchronize_srcu(&c->btree_trans_barrier));
1060
1061         struct bkey_cached *ck;
1062         unsigned iter = 0;
1063         list_for_each_entry(ck, &bc->freed_nonpcpu, list) {
1064                 prt_printf(out, "freed_nonpcpu:\t%lu\r\n", ck->btree_trans_barrier_seq);
1065                 if (++iter > 10)
1066                         break;
1067         }
1068
1069         iter = 0;
1070         list_for_each_entry(ck, &bc->freed_pcpu, list) {
1071                 prt_printf(out, "freed_pcpu:\t%lu\r\n", ck->btree_trans_barrier_seq);
1072                 if (++iter > 10)
1073                         break;
1074         }
1075         mutex_unlock(&bc->lock);
1076         memalloc_flags_restore(flags);
1077 }
1078
1079 void bch2_btree_key_cache_exit(void)
1080 {
1081         kmem_cache_destroy(bch2_key_cache);
1082 }
1083
1084 int __init bch2_btree_key_cache_init(void)
1085 {
1086         bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
1087         if (!bch2_key_cache)
1088                 return -ENOMEM;
1089
1090         return 0;
1091 }