git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/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 #include <linux/seq_buf.h>
  17
  18 static inline bool btree_uses_pcpu_readers(enum btree_id id)
  19 {
  20         return id == BTREE_ID_subvolumes;
  21 }
  22
  23 static struct kmem_cache *bch2_key_cache;
  24
  25 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  26                                        const void *obj)
  27 {
  28         const struct bkey_cached *ck = obj;
  29         const struct bkey_cached_key *key = arg->key;
  30
  31         return ck->key.btree_id != key->btree_id ||
  32                 !bpos_eq(ck->key.pos, key->pos);
  33 }
  34
  35 static const struct rhashtable_params bch2_btree_key_cache_params = {
  36         .head_offset    = offsetof(struct bkey_cached, hash),
  37         .key_offset     = offsetof(struct bkey_cached, key),
  38         .key_len        = sizeof(struct bkey_cached_key),
  39         .obj_cmpfn      = bch2_btree_key_cache_cmp_fn,
  40 };
  41
  42 __flatten
  43 inline struct bkey_cached *
  44 bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
  45 {
  46         struct bkey_cached_key key = {
  47                 .btree_id       = btree_id,
  48                 .pos            = pos,
  49         };
  50
  51         return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
  52                                       bch2_btree_key_cache_params);
  53 }
  54
  55 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
  56 {
  57         if (!six_trylock_intent(&ck->c.lock))
  58                 return false;
  59
  60         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
  61                 six_unlock_intent(&ck->c.lock);
  62                 return false;
  63         }
  64
  65         if (!six_trylock_write(&ck->c.lock)) {
  66                 six_unlock_intent(&ck->c.lock);
  67                 return false;
  68         }
  69
  70         return true;
  71 }
  72
  73 static void bkey_cached_evict(struct btree_key_cache *c,
  74                               struct bkey_cached *ck)
  75 {
  76         BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
  77                                       bch2_btree_key_cache_params));
  78         memset(&ck->key, ~0, sizeof(ck->key));
  79
  80         atomic_long_dec(&c->nr_keys);
  81 }
  82
  83 static void bkey_cached_free(struct btree_key_cache *bc,
  84                              struct bkey_cached *ck)
  85 {
  86         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
  87
  88         BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
  89
  90         ck->btree_trans_barrier_seq =
  91                 start_poll_synchronize_srcu(&c->btree_trans_barrier);
  92
  93         if (ck->c.lock.readers)
  94                 list_move_tail(&ck->list, &bc->freed_pcpu);
  95         else
  96                 list_move_tail(&ck->list, &bc->freed_nonpcpu);
  97         atomic_long_inc(&bc->nr_freed);
  98
  99         kfree(ck->k);
 100         ck->k           = NULL;
 101         ck->u64s        = 0;
 102
 103         six_unlock_write(&ck->c.lock);
 104         six_unlock_intent(&ck->c.lock);
 105 }
 106
 107 #ifdef __KERNEL__
 108 static void __bkey_cached_move_to_freelist_ordered(struct btree_key_cache *bc,
 109                                                    struct bkey_cached *ck)
 110 {
 111         struct bkey_cached *pos;
 112
 113         list_for_each_entry_reverse(pos, &bc->freed_nonpcpu, list) {
 114                 if (ULONG_CMP_GE(ck->btree_trans_barrier_seq,
 115                                  pos->btree_trans_barrier_seq)) {
 116                         list_move(&ck->list, &pos->list);
 117                         return;
 118                 }
 119         }
 120
 121         list_move(&ck->list, &bc->freed_nonpcpu);
 122 }
 123 #endif
 124
 125 static void bkey_cached_move_to_freelist(struct btree_key_cache *bc,
 126                                          struct bkey_cached *ck)
 127 {
 128         BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
 129
 130         if (!ck->c.lock.readers) {
 131 #ifdef __KERNEL__
 132                 struct btree_key_cache_freelist *f;
 133                 bool freed = false;
 134
 135                 preempt_disable();
 136                 f = this_cpu_ptr(bc->pcpu_freed);
 137
 138                 if (f->nr < ARRAY_SIZE(f->objs)) {
 139                         f->objs[f->nr++] = ck;
 140                         freed = true;
 141                 }
 142                 preempt_enable();
 143
 144                 if (!freed) {
 145                         mutex_lock(&bc->lock);
 146                         preempt_disable();
 147                         f = this_cpu_ptr(bc->pcpu_freed);
 148
 149                         while (f->nr > ARRAY_SIZE(f->objs) / 2) {
 150                                 struct bkey_cached *ck2 = f->objs[--f->nr];
 151
 152                                 __bkey_cached_move_to_freelist_ordered(bc, ck2);
 153                         }
 154                         preempt_enable();
 155
 156                         __bkey_cached_move_to_freelist_ordered(bc, ck);
 157                         mutex_unlock(&bc->lock);
 158                 }
 159 #else
 160                 mutex_lock(&bc->lock);
 161                 list_move_tail(&ck->list, &bc->freed_nonpcpu);
 162                 mutex_unlock(&bc->lock);
 163 #endif
 164         } else {
 165                 mutex_lock(&bc->lock);
 166                 list_move_tail(&ck->list, &bc->freed_pcpu);
 167                 mutex_unlock(&bc->lock);
 168         }
 169 }
 170
 171 static void bkey_cached_free_fast(struct btree_key_cache *bc,
 172                                   struct bkey_cached *ck)
 173 {
 174         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 175
 176         ck->btree_trans_barrier_seq =
 177                 start_poll_synchronize_srcu(&c->btree_trans_barrier);
 178
 179         list_del_init(&ck->list);
 180         atomic_long_inc(&bc->nr_freed);
 181
 182         kfree(ck->k);
 183         ck->k           = NULL;
 184         ck->u64s        = 0;
 185
 186         bkey_cached_move_to_freelist(bc, ck);
 187
 188         six_unlock_write(&ck->c.lock);
 189         six_unlock_intent(&ck->c.lock);
 190 }
 191
 192 static struct bkey_cached *
 193 bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path,
 194                   bool *was_new)
 195 {
 196         struct bch_fs *c = trans->c;
 197         struct btree_key_cache *bc = &c->btree_key_cache;
 198         struct bkey_cached *ck = NULL;
 199         bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
 200         int ret;
 201
 202         if (!pcpu_readers) {
 203 #ifdef __KERNEL__
 204                 struct btree_key_cache_freelist *f;
 205
 206                 preempt_disable();
 207                 f = this_cpu_ptr(bc->pcpu_freed);
 208                 if (f->nr)
 209                         ck = f->objs[--f->nr];
 210                 preempt_enable();
 211
 212                 if (!ck) {
 213                         mutex_lock(&bc->lock);
 214                         preempt_disable();
 215                         f = this_cpu_ptr(bc->pcpu_freed);
 216
 217                         while (!list_empty(&bc->freed_nonpcpu) &&
 218                                f->nr < ARRAY_SIZE(f->objs) / 2) {
 219                                 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 220                                 list_del_init(&ck->list);
 221                                 f->objs[f->nr++] = ck;
 222                         }
 223
 224                         ck = f->nr ? f->objs[--f->nr] : NULL;
 225                         preempt_enable();
 226                         mutex_unlock(&bc->lock);
 227                 }
 228 #else
 229                 mutex_lock(&bc->lock);
 230                 if (!list_empty(&bc->freed_nonpcpu)) {
 231                         ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 232                         list_del_init(&ck->list);
 233                 }
 234                 mutex_unlock(&bc->lock);
 235 #endif
 236         } else {
 237                 mutex_lock(&bc->lock);
 238                 if (!list_empty(&bc->freed_pcpu)) {
 239                         ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
 240                         list_del_init(&ck->list);
 241                 }
 242                 mutex_unlock(&bc->lock);
 243         }
 244
 245         if (ck) {
 246                 int ret;
 247
 248                 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent, _THIS_IP_);
 249                 if (unlikely(ret)) {
 250                         bkey_cached_move_to_freelist(bc, ck);
 251                         return ERR_PTR(ret);
 252                 }
 253
 254                 path->l[0].b = (void *) ck;
 255                 path->l[0].lock_seq = ck->c.lock.state.seq;
 256                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 257
 258                 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
 259                 if (unlikely(ret)) {
 260                         btree_node_unlock(trans, path, 0);
 261                         bkey_cached_move_to_freelist(bc, ck);
 262                         return ERR_PTR(ret);
 263                 }
 264
 265                 return ck;
 266         }
 267
 268         ck = kmem_cache_zalloc(bch2_key_cache, GFP_NOWAIT|__GFP_NOWARN);
 269         if (likely(ck))
 270                 goto init;
 271
 272         bch2_trans_unlock(trans);
 273
 274         ck = kmem_cache_zalloc(bch2_key_cache, GFP_KERNEL);
 275
 276         ret = bch2_trans_relock(trans);
 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 init:
 285         INIT_LIST_HEAD(&ck->list);
 286         bch2_btree_lock_init(&ck->c);
 287         if (pcpu_readers)
 288                 six_lock_pcpu_alloc(&ck->c.lock);
 289
 290         ck->c.cached = true;
 291         BUG_ON(!six_trylock_intent(&ck->c.lock));
 292         BUG_ON(!six_trylock_write(&ck->c.lock));
 293         *was_new = true;
 294         return ck;
 295 }
 296
 297 static struct bkey_cached *
 298 bkey_cached_reuse(struct btree_key_cache *c)
 299 {
 300         struct bucket_table *tbl;
 301         struct rhash_head *pos;
 302         struct bkey_cached *ck;
 303         unsigned i;
 304
 305         mutex_lock(&c->lock);
 306         rcu_read_lock();
 307         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 308         for (i = 0; i < tbl->size; i++)
 309                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 310                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 311                             bkey_cached_lock_for_evict(ck)) {
 312                                 bkey_cached_evict(c, ck);
 313                                 goto out;
 314                         }
 315                 }
 316         ck = NULL;
 317 out:
 318         rcu_read_unlock();
 319         mutex_unlock(&c->lock);
 320         return ck;
 321 }
 322
 323 static struct bkey_cached *
 324 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
 325 {
 326         struct bch_fs *c = trans->c;
 327         struct btree_key_cache *bc = &c->btree_key_cache;
 328         struct bkey_cached *ck;
 329         bool was_new = false;
 330
 331         ck = bkey_cached_alloc(trans, path, &was_new);
 332         if (IS_ERR(ck))
 333                 return ck;
 334
 335         if (unlikely(!ck)) {
 336                 ck = bkey_cached_reuse(bc);
 337                 if (unlikely(!ck)) {
 338                         bch_err(c, "error allocating memory for key cache item, btree %s",
 339                                 bch2_btree_ids[path->btree_id]);
 340                         return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
 341                 }
 342
 343                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 344         } else {
 345                 if (path->btree_id == BTREE_ID_subvolumes)
 346                         six_lock_pcpu_alloc(&ck->c.lock);
 347         }
 348
 349         ck->c.level             = 0;
 350         ck->c.btree_id          = path->btree_id;
 351         ck->key.btree_id        = path->btree_id;
 352         ck->key.pos             = path->pos;
 353         ck->valid               = false;
 354         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 355
 356         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 357                                           &ck->hash,
 358                                           bch2_btree_key_cache_params))) {
 359                 /* We raced with another fill: */
 360
 361                 if (likely(was_new)) {
 362                         six_unlock_write(&ck->c.lock);
 363                         six_unlock_intent(&ck->c.lock);
 364                         kfree(ck);
 365                 } else {
 366                         bkey_cached_free_fast(bc, ck);
 367                 }
 368
 369                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
 370                 return NULL;
 371         }
 372
 373         atomic_long_inc(&bc->nr_keys);
 374
 375         six_unlock_write(&ck->c.lock);
 376
 377         return ck;
 378 }
 379
 380 static int btree_key_cache_fill(struct btree_trans *trans,
 381                                 struct btree_path *ck_path,
 382                                 struct bkey_cached *ck)
 383 {
 384         struct btree_iter iter;
 385         struct bkey_s_c k;
 386         unsigned new_u64s = 0;
 387         struct bkey_i *new_k = NULL;
 388         int ret;
 389
 390         k = bch2_bkey_get_iter(trans, &iter, ck->key.btree_id, ck->key.pos,
 391                                BTREE_ITER_KEY_CACHE_FILL|
 392                                BTREE_ITER_CACHED_NOFILL);
 393         ret = bkey_err(k);
 394         if (ret)
 395                 goto err;
 396
 397         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 398                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 399                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 400                 goto err;
 401         }
 402
 403         /*
 404          * bch2_varint_decode can read past the end of the buffer by at
 405          * most 7 bytes (it won't be used):
 406          */
 407         new_u64s = k.k->u64s + 1;
 408
 409         /*
 410          * Allocate some extra space so that the transaction commit path is less
 411          * likely to have to reallocate, since that requires a transaction
 412          * restart:
 413          */
 414         new_u64s = min(256U, (new_u64s * 3) / 2);
 415
 416         if (new_u64s > ck->u64s) {
 417                 new_u64s = roundup_pow_of_two(new_u64s);
 418                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 419                 if (!new_k) {
 420                         bch2_trans_unlock(trans);
 421
 422                         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 423                         if (!new_k) {
 424                                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 425                                         bch2_btree_ids[ck->key.btree_id], new_u64s);
 426                                 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
 427                                 goto err;
 428                         }
 429
 430                         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 431                                 kfree(new_k);
 432                                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 433                                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 434                                 goto err;
 435                         }
 436
 437                         ret = bch2_trans_relock(trans);
 438                         if (ret) {
 439                                 kfree(new_k);
 440                                 goto err;
 441                         }
 442                 }
 443         }
 444
 445         ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
 446         if (ret) {
 447                 kfree(new_k);
 448                 goto err;
 449         }
 450
 451         if (new_k) {
 452                 kfree(ck->k);
 453                 ck->u64s = new_u64s;
 454                 ck->k = new_k;
 455         }
 456
 457         bkey_reassemble(ck->k, k);
 458         ck->valid = true;
 459         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 460
 461         /* We're not likely to need this iterator again: */
 462         set_btree_iter_dontneed(&iter);
 463 err:
 464         bch2_trans_iter_exit(trans, &iter);
 465         return ret;
 466 }
 467
 468 static noinline int
 469 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
 470                                          unsigned flags)
 471 {
 472         struct bch_fs *c = trans->c;
 473         struct bkey_cached *ck;
 474         int ret = 0;
 475
 476         BUG_ON(path->level);
 477
 478         path->l[1].b = NULL;
 479
 480         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 481                 ck = (void *) path->l[0].b;
 482                 goto fill;
 483         }
 484 retry:
 485         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 486         if (!ck) {
 487                 ck = btree_key_cache_create(trans, path);
 488                 ret = PTR_ERR_OR_ZERO(ck);
 489                 if (ret)
 490                         goto err;
 491                 if (!ck)
 492                         goto retry;
 493
 494                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 495                 path->locks_want = 1;
 496         } else {
 497                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 498
 499                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 500                                       lock_want, _THIS_IP_);
 501                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 502                         goto err;
 503
 504                 BUG_ON(ret);
 505
 506                 if (ck->key.btree_id != path->btree_id ||
 507                     !bpos_eq(ck->key.pos, path->pos)) {
 508                         six_unlock_type(&ck->c.lock, lock_want);
 509                         goto retry;
 510                 }
 511
 512                 mark_btree_node_locked(trans, path, 0, lock_want);
 513         }
 514
 515         path->l[0].lock_seq     = ck->c.lock.state.seq;
 516         path->l[0].b            = (void *) ck;
 517 fill:
 518         path->uptodate = BTREE_ITER_UPTODATE;
 519
 520         if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
 521                 /*
 522                  * Using the underscore version because we haven't set
 523                  * path->uptodate yet:
 524                  */
 525                 if (!path->locks_want &&
 526                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 527                         trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
 528                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
 529                         goto err;
 530                 }
 531
 532                 ret = btree_key_cache_fill(trans, path, ck);
 533                 if (ret)
 534                         goto err;
 535
 536                 ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
 537                 if (ret)
 538                         goto err;
 539
 540                 path->uptodate = BTREE_ITER_UPTODATE;
 541         }
 542
 543         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 544                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 545
 546         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 547         BUG_ON(path->uptodate);
 548
 549         return ret;
 550 err:
 551         path->uptodate = BTREE_ITER_NEED_TRAVERSE;
 552         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 553                 btree_node_unlock(trans, path, 0);
 554                 path->l[0].b = ERR_PTR(ret);
 555         }
 556         return ret;
 557 }
 558
 559 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 560                                     unsigned flags)
 561 {
 562         struct bch_fs *c = trans->c;
 563         struct bkey_cached *ck;
 564         int ret = 0;
 565
 566         EBUG_ON(path->level);
 567
 568         path->l[1].b = NULL;
 569
 570         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 571                 ck = (void *) path->l[0].b;
 572                 goto fill;
 573         }
 574 retry:
 575         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 576         if (!ck) {
 577                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 578         } else {
 579                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 580
 581                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 582                                       lock_want, _THIS_IP_);
 583                 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
 584
 585                 if (ret)
 586                         return ret;
 587
 588                 if (ck->key.btree_id != path->btree_id ||
 589                     !bpos_eq(ck->key.pos, path->pos)) {
 590                         six_unlock_type(&ck->c.lock, lock_want);
 591                         goto retry;
 592                 }
 593
 594                 mark_btree_node_locked(trans, path, 0, lock_want);
 595         }
 596
 597         path->l[0].lock_seq     = ck->c.lock.state.seq;
 598         path->l[0].b            = (void *) ck;
 599 fill:
 600         if (!ck->valid)
 601                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 602
 603         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 604                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 605
 606         path->uptodate = BTREE_ITER_UPTODATE;
 607         EBUG_ON(!ck->valid);
 608         EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 609
 610         return ret;
 611 }
 612
 613 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 614                                      struct bkey_cached_key key,
 615                                      u64 journal_seq,
 616                                      unsigned commit_flags,
 617                                      bool evict)
 618 {
 619         struct bch_fs *c = trans->c;
 620         struct journal *j = &c->journal;
 621         struct btree_iter c_iter, b_iter;
 622         struct bkey_cached *ck = NULL;
 623         int ret;
 624
 625         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 626                              BTREE_ITER_SLOTS|
 627                              BTREE_ITER_INTENT|
 628                              BTREE_ITER_ALL_SNAPSHOTS);
 629         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 630                              BTREE_ITER_CACHED|
 631                              BTREE_ITER_INTENT);
 632         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 633
 634         ret = bch2_btree_iter_traverse(&c_iter);
 635         if (ret)
 636                 goto out;
 637
 638         ck = (void *) c_iter.path->l[0].b;
 639         if (!ck)
 640                 goto out;
 641
 642         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 643                 if (evict)
 644                         goto evict;
 645                 goto out;
 646         }
 647
 648         BUG_ON(!ck->valid);
 649
 650         if (journal_seq && ck->journal.seq != journal_seq)
 651                 goto out;
 652
 653         /*
 654          * Since journal reclaim depends on us making progress here, and the
 655          * allocator/copygc depend on journal reclaim making progress, we need
 656          * to be using alloc reserves:
 657          */
 658         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 659                 bch2_trans_update(trans, &b_iter, ck->k,
 660                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 661                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 662                                   BTREE_TRIGGER_NORUN) ?:
 663                 bch2_trans_commit(trans, NULL, NULL,
 664                                   BTREE_INSERT_NOCHECK_RW|
 665                                   BTREE_INSERT_NOFAIL|
 666                                   BTREE_INSERT_USE_RESERVE|
 667                                   (ck->journal.seq == journal_last_seq(j)
 668                                    ? JOURNAL_WATERMARK_reserved
 669                                    : 0)|
 670                                   commit_flags);
 671
 672         bch2_fs_fatal_err_on(ret &&
 673                              !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 674                              !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
 675                              !bch2_journal_error(j), c,
 676                              "error flushing key cache: %s", bch2_err_str(ret));
 677         if (ret)
 678                 goto out;
 679
 680         bch2_journal_pin_drop(j, &ck->journal);
 681         bch2_journal_preres_put(j, &ck->res);
 682
 683         BUG_ON(!btree_node_locked(c_iter.path, 0));
 684
 685         if (!evict) {
 686                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 687                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 688                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 689                 }
 690         } else {
 691                 struct btree_path *path2;
 692 evict:
 693                 trans_for_each_path(trans, path2)
 694                         if (path2 != c_iter.path)
 695                                 __bch2_btree_path_unlock(trans, path2);
 696
 697                 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
 698
 699                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 700                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 701                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 702                 }
 703
 704                 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
 705                 bkey_cached_evict(&c->btree_key_cache, ck);
 706                 bkey_cached_free_fast(&c->btree_key_cache, ck);
 707         }
 708 out:
 709         bch2_trans_iter_exit(trans, &b_iter);
 710         bch2_trans_iter_exit(trans, &c_iter);
 711         return ret;
 712 }
 713
 714 int bch2_btree_key_cache_journal_flush(struct journal *j,
 715                                 struct journal_entry_pin *pin, u64 seq)
 716 {
 717         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 718         struct bkey_cached *ck =
 719                 container_of(pin, struct bkey_cached, journal);
 720         struct bkey_cached_key key;
 721         struct btree_trans trans;
 722         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 723         int ret = 0;
 724
 725         bch2_trans_init(&trans, c, 0, 0);
 726
 727         btree_node_lock_nopath_nofail(&trans, &ck->c, SIX_LOCK_read);
 728         key = ck->key;
 729
 730         if (ck->journal.seq != seq ||
 731             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 732                 six_unlock_read(&ck->c.lock);
 733                 goto unlock;
 734         }
 735
 736         if (ck->seq != seq) {
 737                 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
 738                                         bch2_btree_key_cache_journal_flush);
 739                 six_unlock_read(&ck->c.lock);
 740                 goto unlock;
 741         }
 742         six_unlock_read(&ck->c.lock);
 743
 744         ret = commit_do(&trans, NULL, NULL, 0,
 745                 btree_key_cache_flush_pos(&trans, key, seq,
 746                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 747 unlock:
 748         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 749
 750         bch2_trans_exit(&trans);
 751         return ret;
 752 }
 753
 754 /*
 755  * Flush and evict a key from the key cache:
 756  */
 757 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 758                                enum btree_id id, struct bpos pos)
 759 {
 760         struct bch_fs *c = trans->c;
 761         struct bkey_cached_key key = { id, pos };
 762
 763         /* Fastpath - assume it won't be found: */
 764         if (!bch2_btree_key_cache_find(c, id, pos))
 765                 return 0;
 766
 767         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 768 }
 769
 770 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 771                                   unsigned flags,
 772                                   struct btree_insert_entry *insert_entry)
 773 {
 774         struct bch_fs *c = trans->c;
 775         struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
 776         struct bkey_i *insert = insert_entry->k;
 777         bool kick_reclaim = false;
 778
 779         BUG_ON(insert->k.u64s > ck->u64s);
 780
 781         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 782                 int difference;
 783
 784                 BUG_ON(jset_u64s(insert->k.u64s) > trans->journal_preres.u64s);
 785
 786                 difference = jset_u64s(insert->k.u64s) - ck->res.u64s;
 787                 if (difference > 0) {
 788                         trans->journal_preres.u64s      -= difference;
 789                         ck->res.u64s                    += difference;
 790                 }
 791         }
 792
 793         bkey_copy(ck->k, insert);
 794         ck->valid = true;
 795
 796         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 797                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 798                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 799
 800                 if (bch2_nr_btree_keys_need_flush(c))
 801                         kick_reclaim = true;
 802         }
 803
 804         /*
 805          * To minimize lock contention, we only add the journal pin here and
 806          * defer pin updates to the flush callback via ->seq. Be careful not to
 807          * update ->seq on nojournal commits because we don't want to update the
 808          * pin to a seq that doesn't include journal updates on disk. Otherwise
 809          * we risk losing the update after a crash.
 810          *
 811          * The only exception is if the pin is not active in the first place. We
 812          * have to add the pin because journal reclaim drives key cache
 813          * flushing. The flush callback will not proceed unless ->seq matches
 814          * the latest pin, so make sure it starts with a consistent value.
 815          */
 816         if (!(insert_entry->flags & BTREE_UPDATE_NOJOURNAL) ||
 817             !journal_pin_active(&ck->journal)) {
 818                 ck->seq = trans->journal_res.seq;
 819         }
 820         bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 821                              &ck->journal, bch2_btree_key_cache_journal_flush);
 822
 823         if (kick_reclaim)
 824                 journal_reclaim_kick(&c->journal);
 825         return true;
 826 }
 827
 828 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 829                                struct btree_path *path)
 830 {
 831         struct bch_fs *c = trans->c;
 832         struct bkey_cached *ck = (void *) path->l[0].b;
 833
 834         BUG_ON(!ck->valid);
 835
 836         /*
 837          * We just did an update to the btree, bypassing the key cache: the key
 838          * cache key is now stale and must be dropped, even if dirty:
 839          */
 840         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 841                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 842                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 843                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 844         }
 845
 846         ck->valid = false;
 847 }
 848
 849 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 850                                            struct shrink_control *sc)
 851 {
 852         struct bch_fs *c = container_of(shrink, struct bch_fs,
 853                                         btree_key_cache.shrink);
 854         struct btree_key_cache *bc = &c->btree_key_cache;
 855         struct bucket_table *tbl;
 856         struct bkey_cached *ck, *t;
 857         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 858         unsigned start, flags;
 859         int srcu_idx;
 860
 861         mutex_lock(&bc->lock);
 862         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 863         flags = memalloc_nofs_save();
 864
 865         /*
 866          * Newest freed entries are at the end of the list - once we hit one
 867          * that's too new to be freed, we can bail out:
 868          */
 869         list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, 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_pcpu_free(&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         list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
 886                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 887                                                  ck->btree_trans_barrier_seq))
 888                         break;
 889
 890                 list_del(&ck->list);
 891                 six_lock_pcpu_free(&ck->c.lock);
 892                 kmem_cache_free(bch2_key_cache, ck);
 893                 atomic_long_dec(&bc->nr_freed);
 894                 scanned++;
 895                 freed++;
 896         }
 897
 898         if (scanned >= nr)
 899                 goto out;
 900
 901         rcu_read_lock();
 902         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 903         if (bc->shrink_iter >= tbl->size)
 904                 bc->shrink_iter = 0;
 905         start = bc->shrink_iter;
 906
 907         do {
 908                 struct rhash_head *pos, *next;
 909
 910                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 911
 912                 while (!rht_is_a_nulls(pos)) {
 913                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 914                         ck = container_of(pos, struct bkey_cached, hash);
 915
 916                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 917                                 goto next;
 918
 919                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 920                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 921                         else if (bkey_cached_lock_for_evict(ck)) {
 922                                 bkey_cached_evict(bc, ck);
 923                                 bkey_cached_free(bc, ck);
 924                         }
 925
 926                         scanned++;
 927                         if (scanned >= nr)
 928                                 break;
 929 next:
 930                         pos = next;
 931                 }
 932
 933                 bc->shrink_iter++;
 934                 if (bc->shrink_iter >= tbl->size)
 935                         bc->shrink_iter = 0;
 936         } while (scanned < nr && bc->shrink_iter != start);
 937
 938         rcu_read_unlock();
 939 out:
 940         memalloc_nofs_restore(flags);
 941         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 942         mutex_unlock(&bc->lock);
 943
 944         return freed;
 945 }
 946
 947 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 948                                             struct shrink_control *sc)
 949 {
 950         struct bch_fs *c = container_of(shrink, struct bch_fs,
 951                                         btree_key_cache.shrink);
 952         struct btree_key_cache *bc = &c->btree_key_cache;
 953         long nr = atomic_long_read(&bc->nr_keys) -
 954                 atomic_long_read(&bc->nr_dirty);
 955
 956         return max(0L, nr);
 957 }
 958
 959 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 960 {
 961         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 962         struct bucket_table *tbl;
 963         struct bkey_cached *ck, *n;
 964         struct rhash_head *pos;
 965         LIST_HEAD(items);
 966         unsigned i;
 967 #ifdef __KERNEL__
 968         int cpu;
 969 #endif
 970
 971         if (bc->shrink.list.next)
 972                 unregister_shrinker(&bc->shrink);
 973
 974         mutex_lock(&bc->lock);
 975
 976         /*
 977          * The loop is needed to guard against racing with rehash:
 978          */
 979         while (atomic_long_read(&bc->nr_keys)) {
 980                 rcu_read_lock();
 981                 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 982                 if (tbl)
 983                         for (i = 0; i < tbl->size; i++)
 984                                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 985                                         bkey_cached_evict(bc, ck);
 986                                         list_add(&ck->list, &items);
 987                                 }
 988                 rcu_read_unlock();
 989         }
 990
 991 #ifdef __KERNEL__
 992         for_each_possible_cpu(cpu) {
 993                 struct btree_key_cache_freelist *f =
 994                         per_cpu_ptr(bc->pcpu_freed, cpu);
 995
 996                 for (i = 0; i < f->nr; i++) {
 997                         ck = f->objs[i];
 998                         list_add(&ck->list, &items);
 999                 }
1000         }
1001 #endif
1002
1003         list_splice(&bc->freed_pcpu,    &items);
1004         list_splice(&bc->freed_nonpcpu, &items);
1005
1006         mutex_unlock(&bc->lock);
1007
1008         list_for_each_entry_safe(ck, n, &items, list) {
1009                 cond_resched();
1010
1011                 bch2_journal_pin_drop(&c->journal, &ck->journal);
1012                 bch2_journal_preres_put(&c->journal, &ck->res);
1013
1014                 list_del(&ck->list);
1015                 kfree(ck->k);
1016                 six_lock_pcpu_free(&ck->c.lock);
1017                 kmem_cache_free(bch2_key_cache, ck);
1018         }
1019
1020         if (atomic_long_read(&bc->nr_dirty) &&
1021             !bch2_journal_error(&c->journal) &&
1022             test_bit(BCH_FS_WAS_RW, &c->flags))
1023                 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
1024                       atomic_long_read(&bc->nr_dirty));
1025
1026         if (atomic_long_read(&bc->nr_keys))
1027                 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
1028                       atomic_long_read(&bc->nr_keys));
1029
1030         if (bc->table_init_done)
1031                 rhashtable_destroy(&bc->table);
1032
1033         free_percpu(bc->pcpu_freed);
1034 }
1035
1036 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1037 {
1038         mutex_init(&c->lock);
1039         INIT_LIST_HEAD(&c->freed_pcpu);
1040         INIT_LIST_HEAD(&c->freed_nonpcpu);
1041 }
1042
1043 static void bch2_btree_key_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
1044 {
1045         struct btree_key_cache *bc =
1046                 container_of(shrink, struct btree_key_cache, shrink);
1047         char *cbuf;
1048         size_t buflen = seq_buf_get_buf(s, &cbuf);
1049         struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
1050
1051         bch2_btree_key_cache_to_text(&out, bc);
1052         seq_buf_commit(s, out.pos);
1053 }
1054
1055 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1056 {
1057         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1058
1059 #ifdef __KERNEL__
1060         bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1061         if (!bc->pcpu_freed)
1062                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1063 #endif
1064
1065         if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1066                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1067
1068         bc->table_init_done = true;
1069
1070         bc->shrink.seeks                = 0;
1071         bc->shrink.count_objects        = bch2_btree_key_cache_count;
1072         bc->shrink.scan_objects         = bch2_btree_key_cache_scan;
1073         bc->shrink.to_text              = bch2_btree_key_cache_shrinker_to_text;
1074         if (register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name))
1075                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1076         return 0;
1077 }
1078
1079 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1080 {
1081         prt_printf(out, "nr_freed:\t%zu",       atomic_long_read(&c->nr_freed));
1082         prt_newline(out);
1083         prt_printf(out, "nr_keys:\t%lu",        atomic_long_read(&c->nr_keys));
1084         prt_newline(out);
1085         prt_printf(out, "nr_dirty:\t%lu",       atomic_long_read(&c->nr_dirty));
1086         prt_newline(out);
1087 }
1088
1089 void bch2_btree_key_cache_exit(void)
1090 {
1091         kmem_cache_destroy(bch2_key_cache);
1092 }
1093
1094 int __init bch2_btree_key_cache_init(void)
1095 {
1096         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
1097         if (!bch2_key_cache)
1098                 return -ENOMEM;
1099
1100         return 0;
1101 }