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                 preempt_disable();
 116                 f = this_cpu_ptr(bc->pcpu_freed);
 117
 118                 if (f->nr < ARRAY_SIZE(f->objs)) {
 119                         f->objs[f->nr++] = ck;
 120                         freed = true;
 121                 }
 122                 preempt_enable();
 123
 124                 if (!freed) {
 125                         mutex_lock(&bc->lock);
 126                         preempt_disable();
 127                         f = this_cpu_ptr(bc->pcpu_freed);
 128
 129                         while (f->nr > ARRAY_SIZE(f->objs) / 2) {
 130                                 struct bkey_cached *ck2 = f->objs[--f->nr];
 131
 132                                 list_move_tail(&ck2->list, &bc->freed_nonpcpu);
 133                         }
 134                         preempt_enable();
 135
 136                         list_move_tail(&ck->list, &bc->freed_nonpcpu);
 137                         mutex_unlock(&bc->lock);
 138                 }
 139         } else {
 140                 mutex_lock(&bc->lock);
 141                 list_move_tail(&ck->list, &bc->freed_pcpu);
 142                 mutex_unlock(&bc->lock);
 143         }
 144 }
 145
 146 static void bkey_cached_free_fast(struct btree_key_cache *bc,
 147                                   struct bkey_cached *ck)
 148 {
 149         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 150
 151         ck->btree_trans_barrier_seq =
 152                 start_poll_synchronize_srcu(&c->btree_trans_barrier);
 153
 154         list_del_init(&ck->list);
 155         atomic_long_inc(&bc->nr_freed);
 156
 157         kfree(ck->k);
 158         ck->k           = NULL;
 159         ck->u64s        = 0;
 160
 161         bkey_cached_move_to_freelist(bc, ck);
 162
 163         six_unlock_write(&ck->c.lock);
 164         six_unlock_intent(&ck->c.lock);
 165 }
 166
 167 static struct bkey_cached *
 168 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path)
 169 {
 170         struct bch_fs *c = trans->c;
 171         struct btree_key_cache *bc = &c->btree_key_cache;
 172         struct bkey_cached *ck = NULL;
 173         struct btree_key_cache_freelist *f;
 174         bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
 175
 176         if (!pcpu_readers) {
 177                 preempt_disable();
 178                 f = this_cpu_ptr(bc->pcpu_freed);
 179                 if (f->nr)
 180                         ck = f->objs[--f->nr];
 181                 preempt_enable();
 182
 183                 if (!ck) {
 184                         mutex_lock(&bc->lock);
 185                         preempt_disable();
 186                         f = this_cpu_ptr(bc->pcpu_freed);
 187
 188                         while (!list_empty(&bc->freed_nonpcpu) &&
 189                                f->nr < ARRAY_SIZE(f->objs) / 2) {
 190                                 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 191                                 list_del_init(&ck->list);
 192                                 f->objs[f->nr++] = ck;
 193                         }
 194
 195                         ck = f->nr ? f->objs[--f->nr] : NULL;
 196                         preempt_enable();
 197                         mutex_unlock(&bc->lock);
 198                 }
 199         } else {
 200                 mutex_lock(&bc->lock);
 201                 if (!list_empty(&bc->freed_pcpu)) {
 202                         ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
 203                         list_del_init(&ck->list);
 204                 }
 205                 mutex_unlock(&bc->lock);
 206         }
 207
 208         if (ck) {
 209                 int ret;
 210
 211                 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent);
 212                 if (unlikely(ret)) {
 213                         bkey_cached_move_to_freelist(bc, ck);
 214                         return ERR_PTR(ret);
 215                 }
 216
 217                 path->l[0].b = (void *) ck;
 218                 path->l[0].lock_seq = ck->c.lock.state.seq;
 219                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 220
 221                 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
 222                 if (unlikely(ret)) {
 223                         btree_node_unlock(trans, path, 0);
 224                         bkey_cached_move_to_freelist(bc, ck);
 225                         return ERR_PTR(ret);
 226                 }
 227
 228                 return ck;
 229         }
 230
 231         ck = kmem_cache_alloc(bch2_key_cache, GFP_NOFS|__GFP_ZERO);
 232         if (likely(ck)) {
 233                 INIT_LIST_HEAD(&ck->list);
 234                 __six_lock_init(&ck->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 235                 if (pcpu_readers)
 236                         six_lock_pcpu_alloc(&ck->c.lock);
 237
 238                 ck->c.cached = true;
 239                 BUG_ON(!six_trylock_intent(&ck->c.lock));
 240                 BUG_ON(!six_trylock_write(&ck->c.lock));
 241                 return ck;
 242         }
 243
 244         return NULL;
 245 }
 246
 247 static struct bkey_cached *
 248 bkey_cached_reuse(struct btree_key_cache *c)
 249 {
 250         struct bucket_table *tbl;
 251         struct rhash_head *pos;
 252         struct bkey_cached *ck;
 253         unsigned i;
 254
 255         rcu_read_lock();
 256         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 257         for (i = 0; i < tbl->size; i++)
 258                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 259                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 260                             bkey_cached_lock_for_evict(ck)) {
 261                                 bkey_cached_evict(c, ck);
 262                                 rcu_read_unlock();
 263                                 return ck;
 264                         }
 265                 }
 266         rcu_read_unlock();
 267
 268         return NULL;
 269 }
 270
 271 static struct bkey_cached *
 272 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
 273 {
 274         struct bch_fs *c = trans->c;
 275         struct btree_key_cache *bc = &c->btree_key_cache;
 276         struct bkey_cached *ck;
 277         bool was_new = true;
 278
 279         ck = bkey_cached_alloc(trans, path);
 280         if (unlikely(IS_ERR(ck)))
 281                 return ck;
 282
 283         if (unlikely(!ck)) {
 284                 ck = bkey_cached_reuse(bc);
 285                 if (unlikely(!ck)) {
 286                         bch_err(c, "error allocating memory for key cache item, btree %s",
 287                                 bch2_btree_ids[path->btree_id]);
 288                         return ERR_PTR(-ENOMEM);
 289                 }
 290
 291                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 292                 was_new = false;
 293         } else {
 294                 if (path->btree_id == BTREE_ID_subvolumes)
 295                         six_lock_pcpu_alloc(&ck->c.lock);
 296         }
 297
 298         ck->c.level             = 0;
 299         ck->c.btree_id          = path->btree_id;
 300         ck->key.btree_id        = path->btree_id;
 301         ck->key.pos             = path->pos;
 302         ck->valid               = false;
 303         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 304
 305         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 306                                           &ck->hash,
 307                                           bch2_btree_key_cache_params))) {
 308                 /* We raced with another fill: */
 309
 310                 if (likely(was_new)) {
 311                         six_unlock_write(&ck->c.lock);
 312                         six_unlock_intent(&ck->c.lock);
 313                         mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
 314                         kfree(ck);
 315                 } else {
 316                         bkey_cached_free_fast(bc, ck);
 317                 }
 318
 319                 return NULL;
 320         }
 321
 322         atomic_long_inc(&bc->nr_keys);
 323
 324         six_unlock_write(&ck->c.lock);
 325
 326         return ck;
 327 }
 328
 329 static int btree_key_cache_fill(struct btree_trans *trans,
 330                                 struct btree_path *ck_path,
 331                                 struct bkey_cached *ck)
 332 {
 333         struct btree_path *path;
 334         struct bkey_s_c k;
 335         unsigned new_u64s = 0;
 336         struct bkey_i *new_k = NULL;
 337         struct bkey u;
 338         int ret;
 339
 340         path = bch2_path_get(trans, ck->key.btree_id,
 341                              ck->key.pos, 0, 0, 0, _THIS_IP_);
 342         ret = bch2_btree_path_traverse(trans, path, 0);
 343         if (ret)
 344                 goto err;
 345
 346         k = bch2_btree_path_peek_slot(path, &u);
 347
 348         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 349                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 350                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
 351                 goto err;
 352         }
 353
 354         /*
 355          * bch2_varint_decode can read past the end of the buffer by at
 356          * most 7 bytes (it won't be used):
 357          */
 358         new_u64s = k.k->u64s + 1;
 359
 360         /*
 361          * Allocate some extra space so that the transaction commit path is less
 362          * likely to have to reallocate, since that requires a transaction
 363          * restart:
 364          */
 365         new_u64s = min(256U, (new_u64s * 3) / 2);
 366
 367         if (new_u64s > ck->u64s) {
 368                 new_u64s = roundup_pow_of_two(new_u64s);
 369                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 370                 if (!new_k) {
 371                         bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 372                                 bch2_btree_ids[ck->key.btree_id], new_u64s);
 373                         ret = -ENOMEM;
 374                         goto err;
 375                 }
 376         }
 377
 378         ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
 379         if (ret) {
 380                 kfree(new_k);
 381                 goto err;
 382         }
 383
 384         if (new_k) {
 385                 kfree(ck->k);
 386                 ck->u64s = new_u64s;
 387                 ck->k = new_k;
 388         }
 389
 390         bkey_reassemble(ck->k, k);
 391         ck->valid = true;
 392         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 393
 394         /* We're not likely to need this iterator again: */
 395         path->preserve = false;
 396 err:
 397         bch2_path_put(trans, path, 0);
 398         return ret;
 399 }
 400
 401 noinline static int
 402 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
 403                                          unsigned flags)
 404 {
 405         struct bch_fs *c = trans->c;
 406         struct bkey_cached *ck;
 407         int ret = 0;
 408
 409         BUG_ON(path->level);
 410
 411         path->l[1].b = NULL;
 412
 413         if (bch2_btree_node_relock(trans, path, 0)) {
 414                 ck = (void *) path->l[0].b;
 415                 goto fill;
 416         }
 417 retry:
 418         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 419         if (!ck) {
 420                 ck = btree_key_cache_create(trans, path);
 421                 ret = PTR_ERR_OR_ZERO(ck);
 422                 if (ret)
 423                         goto err;
 424                 if (!ck)
 425                         goto retry;
 426
 427                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 428                 path->locks_want = 1;
 429         } else {
 430                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 431
 432                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 433                                       lock_want, _THIS_IP_);
 434                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 435                         goto err;
 436
 437                 BUG_ON(ret);
 438
 439                 if (ck->key.btree_id != path->btree_id ||
 440                     bpos_cmp(ck->key.pos, path->pos)) {
 441                         six_unlock_type(&ck->c.lock, lock_want);
 442                         goto retry;
 443                 }
 444
 445                 mark_btree_node_locked(trans, path, 0, lock_want);
 446         }
 447
 448         path->l[0].lock_seq     = ck->c.lock.state.seq;
 449         path->l[0].b            = (void *) ck;
 450 fill:
 451         if (!ck->valid) {
 452                 /*
 453                  * Using the underscore version because we haven't set
 454                  * path->uptodate yet:
 455                  */
 456                 if (!path->locks_want &&
 457                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 458                         trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
 459                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
 460                         goto err;
 461                 }
 462
 463                 ret = btree_key_cache_fill(trans, path, ck);
 464                 if (ret)
 465                         goto err;
 466         }
 467
 468         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 469                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 470
 471         path->uptodate = BTREE_ITER_UPTODATE;
 472         BUG_ON(!ck->valid);
 473         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 474
 475         return ret;
 476 err:
 477         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 478                 btree_node_unlock(trans, path, 0);
 479                 path->l[0].b = ERR_PTR(ret);
 480         }
 481         return ret;
 482 }
 483
 484 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 485                                     unsigned flags)
 486 {
 487         struct bch_fs *c = trans->c;
 488         struct bkey_cached *ck;
 489         int ret = 0;
 490
 491         EBUG_ON(path->level);
 492
 493         path->l[1].b = NULL;
 494
 495         if (bch2_btree_node_relock(trans, path, 0)) {
 496                 ck = (void *) path->l[0].b;
 497                 goto fill;
 498         }
 499 retry:
 500         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 501         if (!ck) {
 502                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 503         } else {
 504                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 505
 506                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 507                                       lock_want, _THIS_IP_);
 508                 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
 509
 510                 if (ret)
 511                         return ret;
 512
 513                 if (ck->key.btree_id != path->btree_id ||
 514                     bpos_cmp(ck->key.pos, path->pos)) {
 515                         six_unlock_type(&ck->c.lock, lock_want);
 516                         goto retry;
 517                 }
 518
 519                 mark_btree_node_locked(trans, path, 0, lock_want);
 520         }
 521
 522         path->l[0].lock_seq     = ck->c.lock.state.seq;
 523         path->l[0].b            = (void *) ck;
 524 fill:
 525         if (!ck->valid)
 526                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 527
 528         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 529                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 530
 531         path->uptodate = BTREE_ITER_UPTODATE;
 532         EBUG_ON(!ck->valid);
 533         EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 534
 535         return ret;
 536 }
 537
 538 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 539                                      struct bkey_cached_key key,
 540                                      u64 journal_seq,
 541                                      unsigned commit_flags,
 542                                      bool evict)
 543 {
 544         struct bch_fs *c = trans->c;
 545         struct journal *j = &c->journal;
 546         struct btree_iter c_iter, b_iter;
 547         struct bkey_cached *ck = NULL;
 548         int ret;
 549
 550         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 551                              BTREE_ITER_SLOTS|
 552                              BTREE_ITER_INTENT|
 553                              BTREE_ITER_ALL_SNAPSHOTS);
 554         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 555                              BTREE_ITER_CACHED|
 556                              BTREE_ITER_INTENT);
 557         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 558
 559         ret = bch2_btree_iter_traverse(&c_iter);
 560         if (ret)
 561                 goto out;
 562
 563         ck = (void *) c_iter.path->l[0].b;
 564         if (!ck)
 565                 goto out;
 566
 567         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 568                 if (evict)
 569                         goto evict;
 570                 goto out;
 571         }
 572
 573         BUG_ON(!ck->valid);
 574
 575         if (journal_seq && ck->journal.seq != journal_seq)
 576                 goto out;
 577
 578         /*
 579          * Since journal reclaim depends on us making progress here, and the
 580          * allocator/copygc depend on journal reclaim making progress, we need
 581          * to be using alloc reserves:
 582          * */
 583         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 584                 bch2_trans_update(trans, &b_iter, ck->k,
 585                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 586                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 587                                   BTREE_TRIGGER_NORUN) ?:
 588                 bch2_trans_commit(trans, NULL, NULL,
 589                                   BTREE_INSERT_NOCHECK_RW|
 590                                   BTREE_INSERT_NOFAIL|
 591                                   BTREE_INSERT_USE_RESERVE|
 592                                   (ck->journal.seq == journal_last_seq(j)
 593                                    ? JOURNAL_WATERMARK_reserved
 594                                    : 0)|
 595                                   commit_flags);
 596
 597         bch2_fs_fatal_err_on(ret &&
 598                              !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 599                              !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
 600                              !bch2_journal_error(j), c,
 601                              "error flushing key cache: %s", bch2_err_str(ret));
 602         if (ret)
 603                 goto out;
 604
 605         bch2_journal_pin_drop(j, &ck->journal);
 606         bch2_journal_preres_put(j, &ck->res);
 607
 608         BUG_ON(!btree_node_locked(c_iter.path, 0));
 609
 610         if (!evict) {
 611                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 612                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 613                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 614                 }
 615         } else {
 616                 struct btree_path *path2;
 617 evict:
 618                 trans_for_each_path(trans, path2)
 619                         if (path2 != c_iter.path)
 620                                 __bch2_btree_path_unlock(trans, path2);
 621
 622                 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
 623
 624                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 625                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 626                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 627                 }
 628
 629                 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
 630                 bkey_cached_evict(&c->btree_key_cache, ck);
 631                 bkey_cached_free_fast(&c->btree_key_cache, ck);
 632         }
 633 out:
 634         bch2_trans_iter_exit(trans, &b_iter);
 635         bch2_trans_iter_exit(trans, &c_iter);
 636         return ret;
 637 }
 638
 639 int bch2_btree_key_cache_journal_flush(struct journal *j,
 640                                 struct journal_entry_pin *pin, u64 seq)
 641 {
 642         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 643         struct bkey_cached *ck =
 644                 container_of(pin, struct bkey_cached, journal);
 645         struct bkey_cached_key key;
 646         struct btree_trans trans;
 647         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 648         int ret = 0;
 649
 650         bch2_trans_init(&trans, c, 0, 0);
 651
 652         btree_node_lock_nopath_nofail(&trans, &ck->c, SIX_LOCK_read);
 653         key = ck->key;
 654
 655         if (ck->journal.seq != seq ||
 656             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 657                 six_unlock_read(&ck->c.lock);
 658                 goto unlock;
 659         }
 660         six_unlock_read(&ck->c.lock);
 661
 662         ret = commit_do(&trans, NULL, NULL, 0,
 663                 btree_key_cache_flush_pos(&trans, key, seq,
 664                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 665 unlock:
 666         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 667
 668         bch2_trans_exit(&trans);
 669         return ret;
 670 }
 671
 672 /*
 673  * Flush and evict a key from the key cache:
 674  */
 675 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 676                                enum btree_id id, struct bpos pos)
 677 {
 678         struct bch_fs *c = trans->c;
 679         struct bkey_cached_key key = { id, pos };
 680
 681         /* Fastpath - assume it won't be found: */
 682         if (!bch2_btree_key_cache_find(c, id, pos))
 683                 return 0;
 684
 685         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 686 }
 687
 688 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 689                                   struct btree_path *path,
 690                                   struct bkey_i *insert)
 691 {
 692         struct bch_fs *c = trans->c;
 693         struct bkey_cached *ck = (void *) path->l[0].b;
 694         bool kick_reclaim = false;
 695
 696         BUG_ON(insert->u64s > ck->u64s);
 697
 698         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 699                 int difference;
 700
 701                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 702
 703                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 704                 if (difference > 0) {
 705                         trans->journal_preres.u64s      -= difference;
 706                         ck->res.u64s                    += difference;
 707                 }
 708         }
 709
 710         bkey_copy(ck->k, insert);
 711         ck->valid = true;
 712
 713         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 714                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 715                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 716
 717                 if (bch2_nr_btree_keys_need_flush(c))
 718                         kick_reclaim = true;
 719         }
 720
 721         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 722                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 723
 724         if (kick_reclaim)
 725                 journal_reclaim_kick(&c->journal);
 726         return true;
 727 }
 728
 729 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 730                                struct btree_path *path)
 731 {
 732         struct bch_fs *c = trans->c;
 733         struct bkey_cached *ck = (void *) path->l[0].b;
 734
 735         BUG_ON(!ck->valid);
 736
 737         /*
 738          * We just did an update to the btree, bypassing the key cache: the key
 739          * cache key is now stale and must be dropped, even if dirty:
 740          */
 741         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 742                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 743                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 744                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 745         }
 746
 747         ck->valid = false;
 748 }
 749
 750 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 751                                            struct shrink_control *sc)
 752 {
 753         struct bch_fs *c = container_of(shrink, struct bch_fs,
 754                                         btree_key_cache.shrink);
 755         struct btree_key_cache *bc = &c->btree_key_cache;
 756         struct bucket_table *tbl;
 757         struct bkey_cached *ck, *t;
 758         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 759         unsigned start, flags;
 760         int srcu_idx;
 761
 762         /* Return -1 if we can't do anything right now */
 763         if (sc->gfp_mask & __GFP_FS)
 764                 mutex_lock(&bc->lock);
 765         else if (!mutex_trylock(&bc->lock))
 766                 return -1;
 767
 768         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 769         flags = memalloc_nofs_save();
 770
 771         /*
 772          * Newest freed entries are at the end of the list - once we hit one
 773          * that's too new to be freed, we can bail out:
 774          */
 775         list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
 776                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 777                                                  ck->btree_trans_barrier_seq))
 778                         break;
 779
 780                 list_del(&ck->list);
 781                 six_lock_pcpu_free(&ck->c.lock);
 782                 kmem_cache_free(bch2_key_cache, ck);
 783                 atomic_long_dec(&bc->nr_freed);
 784                 scanned++;
 785                 freed++;
 786         }
 787
 788         if (scanned >= nr)
 789                 goto out;
 790
 791         list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
 792                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 793                                                  ck->btree_trans_barrier_seq))
 794                         break;
 795
 796                 list_del(&ck->list);
 797                 six_lock_pcpu_free(&ck->c.lock);
 798                 kmem_cache_free(bch2_key_cache, ck);
 799                 atomic_long_dec(&bc->nr_freed);
 800                 scanned++;
 801                 freed++;
 802         }
 803
 804         if (scanned >= nr)
 805                 goto out;
 806
 807         rcu_read_lock();
 808         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 809         if (bc->shrink_iter >= tbl->size)
 810                 bc->shrink_iter = 0;
 811         start = bc->shrink_iter;
 812
 813         do {
 814                 struct rhash_head *pos, *next;
 815
 816                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 817
 818                 while (!rht_is_a_nulls(pos)) {
 819                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 820                         ck = container_of(pos, struct bkey_cached, hash);
 821
 822                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 823                                 goto next;
 824
 825                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 826                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 827                         else if (bkey_cached_lock_for_evict(ck)) {
 828                                 bkey_cached_evict(bc, ck);
 829                                 bkey_cached_free(bc, ck);
 830                         }
 831
 832                         scanned++;
 833                         if (scanned >= nr)
 834                                 break;
 835 next:
 836                         pos = next;
 837                 }
 838
 839                 bc->shrink_iter++;
 840                 if (bc->shrink_iter >= tbl->size)
 841                         bc->shrink_iter = 0;
 842         } while (scanned < nr && bc->shrink_iter != start);
 843
 844         rcu_read_unlock();
 845 out:
 846         memalloc_nofs_restore(flags);
 847         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 848         mutex_unlock(&bc->lock);
 849
 850         return freed;
 851 }
 852
 853 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 854                                             struct shrink_control *sc)
 855 {
 856         struct bch_fs *c = container_of(shrink, struct bch_fs,
 857                                         btree_key_cache.shrink);
 858         struct btree_key_cache *bc = &c->btree_key_cache;
 859         long nr = atomic_long_read(&bc->nr_keys) -
 860                 atomic_long_read(&bc->nr_dirty);
 861
 862         return max(0L, nr);
 863 }
 864
 865 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 866 {
 867         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 868         struct bucket_table *tbl;
 869         struct bkey_cached *ck, *n;
 870         struct rhash_head *pos;
 871         unsigned i;
 872         int cpu;
 873
 874         if (bc->shrink.list.next)
 875                 unregister_shrinker(&bc->shrink);
 876
 877         mutex_lock(&bc->lock);
 878
 879         rcu_read_lock();
 880         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 881         if (tbl)
 882                 for (i = 0; i < tbl->size; i++)
 883                         rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 884                                 bkey_cached_evict(bc, ck);
 885                                 list_add(&ck->list, &bc->freed_nonpcpu);
 886                         }
 887         rcu_read_unlock();
 888
 889         for_each_possible_cpu(cpu) {
 890                 struct btree_key_cache_freelist *f =
 891                         per_cpu_ptr(bc->pcpu_freed, cpu);
 892
 893                 for (i = 0; i < f->nr; i++) {
 894                         ck = f->objs[i];
 895                         list_add(&ck->list, &bc->freed_nonpcpu);
 896                 }
 897         }
 898
 899         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 900
 901         list_for_each_entry_safe(ck, n, &bc->freed_nonpcpu, list) {
 902                 cond_resched();
 903
 904                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 905                 bch2_journal_preres_put(&c->journal, &ck->res);
 906
 907                 list_del(&ck->list);
 908                 kfree(ck->k);
 909                 six_lock_pcpu_free(&ck->c.lock);
 910                 kmem_cache_free(bch2_key_cache, ck);
 911         }
 912
 913         BUG_ON(atomic_long_read(&bc->nr_dirty) &&
 914                !bch2_journal_error(&c->journal) &&
 915                test_bit(BCH_FS_WAS_RW, &c->flags));
 916         BUG_ON(atomic_long_read(&bc->nr_keys));
 917
 918         mutex_unlock(&bc->lock);
 919
 920         if (bc->table_init_done)
 921                 rhashtable_destroy(&bc->table);
 922
 923         free_percpu(bc->pcpu_freed);
 924 }
 925
 926 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 927 {
 928         mutex_init(&c->lock);
 929         INIT_LIST_HEAD(&c->freed_pcpu);
 930         INIT_LIST_HEAD(&c->freed_nonpcpu);
 931 }
 932
 933 static void bch2_btree_key_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 934 {
 935         struct btree_key_cache *bc =
 936                 container_of(shrink, struct btree_key_cache, shrink);
 937
 938         bch2_btree_key_cache_to_text(out, bc);
 939 }
 940
 941 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
 942 {
 943         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 944         int ret;
 945
 946         bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
 947         if (!bc->pcpu_freed)
 948                 return -ENOMEM;
 949
 950         ret = rhashtable_init(&bc->table, &bch2_btree_key_cache_params);
 951         if (ret)
 952                 return ret;
 953
 954         bc->table_init_done = true;
 955
 956         bc->shrink.seeks                = 1;
 957         bc->shrink.count_objects        = bch2_btree_key_cache_count;
 958         bc->shrink.scan_objects         = bch2_btree_key_cache_scan;
 959         bc->shrink.to_text              = bch2_btree_key_cache_shrinker_to_text;
 960         return register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name);
 961 }
 962
 963 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 964 {
 965         prt_printf(out, "nr_freed:\t%zu\n",     atomic_long_read(&c->nr_freed));
 966         prt_printf(out, "nr_keys:\t%lu\n",      atomic_long_read(&c->nr_keys));
 967         prt_printf(out, "nr_dirty:\t%lu\n",     atomic_long_read(&c->nr_dirty));
 968 }
 969
 970 void bch2_btree_key_cache_exit(void)
 971 {
 972         if (bch2_key_cache)
 973                 kmem_cache_destroy(bch2_key_cache);
 974 }
 975
 976 int __init bch2_btree_key_cache_init(void)
 977 {
 978         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
 979         if (!bch2_key_cache)
 980                 return -ENOMEM;
 981
 982         return 0;
 983 }