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
  14 #include <linux/sched/mm.h>
  15 #include <trace/events/bcachefs.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 cmp_int(ck->key.btree_id, key->btree_id) ?:
  31                 bpos_cmp(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 (!six_trylock_write(&ck->c.lock)) {
  60                 six_unlock_intent(&ck->c.lock);
  61                 return false;
  62         }
  63
  64         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
  65                 six_unlock_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 {
 195         struct bch_fs *c = trans->c;
 196         struct btree_key_cache *bc = &c->btree_key_cache;
 197         struct bkey_cached *ck = NULL;
 198         bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
 199
 200         if (!pcpu_readers) {
 201 #ifdef __KERNEL__
 202                 struct btree_key_cache_freelist *f;
 203
 204                 preempt_disable();
 205                 f = this_cpu_ptr(bc->pcpu_freed);
 206                 if (f->nr)
 207                         ck = f->objs[--f->nr];
 208                 preempt_enable();
 209
 210                 if (!ck) {
 211                         mutex_lock(&bc->lock);
 212                         preempt_disable();
 213                         f = this_cpu_ptr(bc->pcpu_freed);
 214
 215                         while (!list_empty(&bc->freed_nonpcpu) &&
 216                                f->nr < ARRAY_SIZE(f->objs) / 2) {
 217                                 ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 218                                 list_del_init(&ck->list);
 219                                 f->objs[f->nr++] = ck;
 220                         }
 221
 222                         ck = f->nr ? f->objs[--f->nr] : NULL;
 223                         preempt_enable();
 224                         mutex_unlock(&bc->lock);
 225                 }
 226 #else
 227                 mutex_lock(&bc->lock);
 228                 if (!list_empty(&bc->freed_nonpcpu)) {
 229                         ck = list_last_entry(&bc->freed_nonpcpu, struct bkey_cached, list);
 230                         list_del_init(&ck->list);
 231                 }
 232                 mutex_unlock(&bc->lock);
 233 #endif
 234         } else {
 235                 mutex_lock(&bc->lock);
 236                 if (!list_empty(&bc->freed_pcpu)) {
 237                         ck = list_last_entry(&bc->freed_pcpu, struct bkey_cached, list);
 238                         list_del_init(&ck->list);
 239                 }
 240                 mutex_unlock(&bc->lock);
 241         }
 242
 243         if (ck) {
 244                 int ret;
 245
 246                 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent);
 247                 if (unlikely(ret)) {
 248                         bkey_cached_move_to_freelist(bc, ck);
 249                         return ERR_PTR(ret);
 250                 }
 251
 252                 path->l[0].b = (void *) ck;
 253                 path->l[0].lock_seq = ck->c.lock.state.seq;
 254                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 255
 256                 ret = bch2_btree_node_lock_write(trans, path, &ck->c);
 257                 if (unlikely(ret)) {
 258                         btree_node_unlock(trans, path, 0);
 259                         bkey_cached_move_to_freelist(bc, ck);
 260                         return ERR_PTR(ret);
 261                 }
 262
 263                 return ck;
 264         }
 265
 266         /* GFP_NOFS because we're holding btree locks: */
 267         ck = kmem_cache_alloc(bch2_key_cache, GFP_NOFS|__GFP_ZERO);
 268         if (likely(ck)) {
 269                 INIT_LIST_HEAD(&ck->list);
 270                 __six_lock_init(&ck->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
 271                 if (pcpu_readers)
 272                         six_lock_pcpu_alloc(&ck->c.lock);
 273
 274                 ck->c.cached = true;
 275                 BUG_ON(!six_trylock_intent(&ck->c.lock));
 276                 BUG_ON(!six_trylock_write(&ck->c.lock));
 277                 return ck;
 278         }
 279
 280         return NULL;
 281 }
 282
 283 static struct bkey_cached *
 284 bkey_cached_reuse(struct btree_key_cache *c)
 285 {
 286         struct bucket_table *tbl;
 287         struct rhash_head *pos;
 288         struct bkey_cached *ck;
 289         unsigned i;
 290
 291         mutex_lock(&c->lock);
 292         rcu_read_lock();
 293         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 294         for (i = 0; i < tbl->size; i++)
 295                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 296                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 297                             bkey_cached_lock_for_evict(ck)) {
 298                                 bkey_cached_evict(c, ck);
 299                                 goto out;
 300                         }
 301                 }
 302         ck = NULL;
 303 out:
 304         rcu_read_unlock();
 305         mutex_unlock(&c->lock);
 306         return ck;
 307 }
 308
 309 static struct bkey_cached *
 310 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
 311 {
 312         struct bch_fs *c = trans->c;
 313         struct btree_key_cache *bc = &c->btree_key_cache;
 314         struct bkey_cached *ck;
 315         bool was_new = true;
 316
 317         ck = bkey_cached_alloc(trans, path);
 318         if (IS_ERR(ck))
 319                 return ck;
 320
 321         if (unlikely(!ck)) {
 322                 ck = bkey_cached_reuse(bc);
 323                 if (unlikely(!ck)) {
 324                         bch_err(c, "error allocating memory for key cache item, btree %s",
 325                                 bch2_btree_ids[path->btree_id]);
 326                         return ERR_PTR(-ENOMEM);
 327                 }
 328
 329                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 330                 was_new = false;
 331         } else {
 332                 if (path->btree_id == BTREE_ID_subvolumes)
 333                         six_lock_pcpu_alloc(&ck->c.lock);
 334         }
 335
 336         ck->c.level             = 0;
 337         ck->c.btree_id          = path->btree_id;
 338         ck->key.btree_id        = path->btree_id;
 339         ck->key.pos             = path->pos;
 340         ck->valid               = false;
 341         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 342
 343         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 344                                           &ck->hash,
 345                                           bch2_btree_key_cache_params))) {
 346                 /* We raced with another fill: */
 347
 348                 if (likely(was_new)) {
 349                         six_unlock_write(&ck->c.lock);
 350                         six_unlock_intent(&ck->c.lock);
 351                         mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
 352                         kfree(ck);
 353                 } else {
 354                         bkey_cached_free_fast(bc, ck);
 355                 }
 356
 357                 return NULL;
 358         }
 359
 360         atomic_long_inc(&bc->nr_keys);
 361
 362         six_unlock_write(&ck->c.lock);
 363
 364         return ck;
 365 }
 366
 367 static int btree_key_cache_fill(struct btree_trans *trans,
 368                                 struct btree_path *ck_path,
 369                                 struct bkey_cached *ck)
 370 {
 371         struct btree_path *path;
 372         struct bkey_s_c k;
 373         unsigned new_u64s = 0;
 374         struct bkey_i *new_k = NULL;
 375         struct bkey u;
 376         int ret;
 377
 378         path = bch2_path_get(trans, ck->key.btree_id,
 379                              ck->key.pos, 0, 0, 0, _THIS_IP_);
 380         ret = bch2_btree_path_traverse(trans, path, 0);
 381         if (ret)
 382                 goto err;
 383
 384         k = bch2_btree_path_peek_slot(path, &u);
 385
 386         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 387                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 388                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
 389                 goto err;
 390         }
 391
 392         /*
 393          * bch2_varint_decode can read past the end of the buffer by at
 394          * most 7 bytes (it won't be used):
 395          */
 396         new_u64s = k.k->u64s + 1;
 397
 398         /*
 399          * Allocate some extra space so that the transaction commit path is less
 400          * likely to have to reallocate, since that requires a transaction
 401          * restart:
 402          */
 403         new_u64s = min(256U, (new_u64s * 3) / 2);
 404
 405         if (new_u64s > ck->u64s) {
 406                 new_u64s = roundup_pow_of_two(new_u64s);
 407                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 408                 if (!new_k) {
 409                         bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 410                                 bch2_btree_ids[ck->key.btree_id], new_u64s);
 411                         ret = -ENOMEM;
 412                         goto err;
 413                 }
 414         }
 415
 416         ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
 417         if (ret) {
 418                 kfree(new_k);
 419                 goto err;
 420         }
 421
 422         if (new_k) {
 423                 kfree(ck->k);
 424                 ck->u64s = new_u64s;
 425                 ck->k = new_k;
 426         }
 427
 428         bkey_reassemble(ck->k, k);
 429         ck->valid = true;
 430         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 431
 432         /* We're not likely to need this iterator again: */
 433         path->preserve = false;
 434 err:
 435         bch2_path_put(trans, path, 0);
 436         return ret;
 437 }
 438
 439 static noinline int
 440 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
 441                                          unsigned flags)
 442 {
 443         struct bch_fs *c = trans->c;
 444         struct bkey_cached *ck;
 445         int ret = 0;
 446
 447         BUG_ON(path->level);
 448
 449         path->l[1].b = NULL;
 450
 451         if (bch2_btree_node_relock(trans, path, 0)) {
 452                 ck = (void *) path->l[0].b;
 453                 goto fill;
 454         }
 455 retry:
 456         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 457         if (!ck) {
 458                 ck = btree_key_cache_create(trans, path);
 459                 ret = PTR_ERR_OR_ZERO(ck);
 460                 if (ret)
 461                         goto err;
 462                 if (!ck)
 463                         goto retry;
 464
 465                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 466                 path->locks_want = 1;
 467         } else {
 468                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 469
 470                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 471                                       lock_want, _THIS_IP_);
 472                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 473                         goto err;
 474
 475                 BUG_ON(ret);
 476
 477                 if (ck->key.btree_id != path->btree_id ||
 478                     bpos_cmp(ck->key.pos, path->pos)) {
 479                         six_unlock_type(&ck->c.lock, lock_want);
 480                         goto retry;
 481                 }
 482
 483                 mark_btree_node_locked(trans, path, 0, lock_want);
 484         }
 485
 486         path->l[0].lock_seq     = ck->c.lock.state.seq;
 487         path->l[0].b            = (void *) ck;
 488 fill:
 489         if (!ck->valid) {
 490                 /*
 491                  * Using the underscore version because we haven't set
 492                  * path->uptodate yet:
 493                  */
 494                 if (!path->locks_want &&
 495                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 496                         trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
 497                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
 498                         goto err;
 499                 }
 500
 501                 ret = btree_key_cache_fill(trans, path, ck);
 502                 if (ret)
 503                         goto err;
 504         }
 505
 506         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 507                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 508
 509         path->uptodate = BTREE_ITER_UPTODATE;
 510         BUG_ON(!ck->valid);
 511         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 512
 513         return ret;
 514 err:
 515         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 516                 btree_node_unlock(trans, path, 0);
 517                 path->l[0].b = ERR_PTR(ret);
 518         }
 519         return ret;
 520 }
 521
 522 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 523                                     unsigned flags)
 524 {
 525         struct bch_fs *c = trans->c;
 526         struct bkey_cached *ck;
 527         int ret = 0;
 528
 529         EBUG_ON(path->level);
 530
 531         path->l[1].b = NULL;
 532
 533         if (bch2_btree_node_relock(trans, path, 0)) {
 534                 ck = (void *) path->l[0].b;
 535                 goto fill;
 536         }
 537 retry:
 538         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 539         if (!ck) {
 540                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 541         } else {
 542                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 543
 544                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 545                                       lock_want, _THIS_IP_);
 546                 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
 547
 548                 if (ret)
 549                         return ret;
 550
 551                 if (ck->key.btree_id != path->btree_id ||
 552                     bpos_cmp(ck->key.pos, path->pos)) {
 553                         six_unlock_type(&ck->c.lock, lock_want);
 554                         goto retry;
 555                 }
 556
 557                 mark_btree_node_locked(trans, path, 0, lock_want);
 558         }
 559
 560         path->l[0].lock_seq     = ck->c.lock.state.seq;
 561         path->l[0].b            = (void *) ck;
 562 fill:
 563         if (!ck->valid)
 564                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 565
 566         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 567                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 568
 569         path->uptodate = BTREE_ITER_UPTODATE;
 570         EBUG_ON(!ck->valid);
 571         EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 572
 573         return ret;
 574 }
 575
 576 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 577                                      struct bkey_cached_key key,
 578                                      u64 journal_seq,
 579                                      unsigned commit_flags,
 580                                      bool evict)
 581 {
 582         struct bch_fs *c = trans->c;
 583         struct journal *j = &c->journal;
 584         struct btree_iter c_iter, b_iter;
 585         struct bkey_cached *ck = NULL;
 586         int ret;
 587
 588         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 589                              BTREE_ITER_SLOTS|
 590                              BTREE_ITER_INTENT|
 591                              BTREE_ITER_ALL_SNAPSHOTS);
 592         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 593                              BTREE_ITER_CACHED|
 594                              BTREE_ITER_INTENT);
 595         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 596
 597         ret = bch2_btree_iter_traverse(&c_iter);
 598         if (ret)
 599                 goto out;
 600
 601         ck = (void *) c_iter.path->l[0].b;
 602         if (!ck)
 603                 goto out;
 604
 605         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 606                 if (evict)
 607                         goto evict;
 608                 goto out;
 609         }
 610
 611         BUG_ON(!ck->valid);
 612
 613         if (journal_seq && ck->journal.seq != journal_seq)
 614                 goto out;
 615
 616         /*
 617          * Since journal reclaim depends on us making progress here, and the
 618          * allocator/copygc depend on journal reclaim making progress, we need
 619          * to be using alloc reserves:
 620          */
 621         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 622                 bch2_trans_update(trans, &b_iter, ck->k,
 623                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 624                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 625                                   BTREE_TRIGGER_NORUN) ?:
 626                 bch2_trans_commit(trans, NULL, NULL,
 627                                   BTREE_INSERT_NOCHECK_RW|
 628                                   BTREE_INSERT_NOFAIL|
 629                                   BTREE_INSERT_USE_RESERVE|
 630                                   (ck->journal.seq == journal_last_seq(j)
 631                                    ? JOURNAL_WATERMARK_reserved
 632                                    : 0)|
 633                                   commit_flags);
 634
 635         bch2_fs_fatal_err_on(ret &&
 636                              !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 637                              !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
 638                              !bch2_journal_error(j), c,
 639                              "error flushing key cache: %s", bch2_err_str(ret));
 640         if (ret)
 641                 goto out;
 642
 643         bch2_journal_pin_drop(j, &ck->journal);
 644         bch2_journal_preres_put(j, &ck->res);
 645
 646         BUG_ON(!btree_node_locked(c_iter.path, 0));
 647
 648         if (!evict) {
 649                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 650                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 651                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 652                 }
 653         } else {
 654                 struct btree_path *path2;
 655 evict:
 656                 trans_for_each_path(trans, path2)
 657                         if (path2 != c_iter.path)
 658                                 __bch2_btree_path_unlock(trans, path2);
 659
 660                 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
 661
 662                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 663                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 664                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 665                 }
 666
 667                 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
 668                 bkey_cached_evict(&c->btree_key_cache, ck);
 669                 bkey_cached_free_fast(&c->btree_key_cache, ck);
 670         }
 671 out:
 672         bch2_trans_iter_exit(trans, &b_iter);
 673         bch2_trans_iter_exit(trans, &c_iter);
 674         return ret;
 675 }
 676
 677 int bch2_btree_key_cache_journal_flush(struct journal *j,
 678                                 struct journal_entry_pin *pin, u64 seq)
 679 {
 680         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 681         struct bkey_cached *ck =
 682                 container_of(pin, struct bkey_cached, journal);
 683         struct bkey_cached_key key;
 684         struct btree_trans trans;
 685         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 686         int ret = 0;
 687
 688         bch2_trans_init(&trans, c, 0, 0);
 689
 690         btree_node_lock_nopath_nofail(&trans, &ck->c, SIX_LOCK_read);
 691         key = ck->key;
 692
 693         if (ck->journal.seq != seq ||
 694             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 695                 six_unlock_read(&ck->c.lock);
 696                 goto unlock;
 697         }
 698         six_unlock_read(&ck->c.lock);
 699
 700         ret = commit_do(&trans, NULL, NULL, 0,
 701                 btree_key_cache_flush_pos(&trans, key, seq,
 702                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 703 unlock:
 704         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 705
 706         bch2_trans_exit(&trans);
 707         return ret;
 708 }
 709
 710 /*
 711  * Flush and evict a key from the key cache:
 712  */
 713 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 714                                enum btree_id id, struct bpos pos)
 715 {
 716         struct bch_fs *c = trans->c;
 717         struct bkey_cached_key key = { id, pos };
 718
 719         /* Fastpath - assume it won't be found: */
 720         if (!bch2_btree_key_cache_find(c, id, pos))
 721                 return 0;
 722
 723         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 724 }
 725
 726 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 727                                   struct btree_path *path,
 728                                   struct bkey_i *insert)
 729 {
 730         struct bch_fs *c = trans->c;
 731         struct bkey_cached *ck = (void *) path->l[0].b;
 732         bool kick_reclaim = false;
 733
 734         BUG_ON(insert->u64s > ck->u64s);
 735
 736         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 737                 int difference;
 738
 739                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 740
 741                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 742                 if (difference > 0) {
 743                         trans->journal_preres.u64s      -= difference;
 744                         ck->res.u64s                    += difference;
 745                 }
 746         }
 747
 748         bkey_copy(ck->k, insert);
 749         ck->valid = true;
 750
 751         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 752                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 753                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 754
 755                 if (bch2_nr_btree_keys_need_flush(c))
 756                         kick_reclaim = true;
 757         }
 758
 759         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 760                                 &ck->journal, bch2_btree_key_cache_journal_flush);
 761
 762         if (kick_reclaim)
 763                 journal_reclaim_kick(&c->journal);
 764         return true;
 765 }
 766
 767 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 768                                struct btree_path *path)
 769 {
 770         struct bch_fs *c = trans->c;
 771         struct bkey_cached *ck = (void *) path->l[0].b;
 772
 773         BUG_ON(!ck->valid);
 774
 775         /*
 776          * We just did an update to the btree, bypassing the key cache: the key
 777          * cache key is now stale and must be dropped, even if dirty:
 778          */
 779         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 780                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 781                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 782                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 783         }
 784
 785         ck->valid = false;
 786 }
 787
 788 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 789                                            struct shrink_control *sc)
 790 {
 791         struct bch_fs *c = container_of(shrink, struct bch_fs,
 792                                         btree_key_cache.shrink);
 793         struct btree_key_cache *bc = &c->btree_key_cache;
 794         struct bucket_table *tbl;
 795         struct bkey_cached *ck, *t;
 796         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 797         unsigned start, flags;
 798         int srcu_idx;
 799
 800         mutex_lock(&bc->lock);
 801         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 802         flags = memalloc_nofs_save();
 803
 804         /*
 805          * Newest freed entries are at the end of the list - once we hit one
 806          * that's too new to be freed, we can bail out:
 807          */
 808         list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
 809                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 810                                                  ck->btree_trans_barrier_seq))
 811                         break;
 812
 813                 list_del(&ck->list);
 814                 six_lock_pcpu_free(&ck->c.lock);
 815                 kmem_cache_free(bch2_key_cache, ck);
 816                 atomic_long_dec(&bc->nr_freed);
 817                 scanned++;
 818                 freed++;
 819         }
 820
 821         if (scanned >= nr)
 822                 goto out;
 823
 824         list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
 825                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 826                                                  ck->btree_trans_barrier_seq))
 827                         break;
 828
 829                 list_del(&ck->list);
 830                 six_lock_pcpu_free(&ck->c.lock);
 831                 kmem_cache_free(bch2_key_cache, ck);
 832                 atomic_long_dec(&bc->nr_freed);
 833                 scanned++;
 834                 freed++;
 835         }
 836
 837         if (scanned >= nr)
 838                 goto out;
 839
 840         rcu_read_lock();
 841         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 842         if (bc->shrink_iter >= tbl->size)
 843                 bc->shrink_iter = 0;
 844         start = bc->shrink_iter;
 845
 846         do {
 847                 struct rhash_head *pos, *next;
 848
 849                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 850
 851                 while (!rht_is_a_nulls(pos)) {
 852                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 853                         ck = container_of(pos, struct bkey_cached, hash);
 854
 855                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 856                                 goto next;
 857
 858                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 859                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 860                         else if (bkey_cached_lock_for_evict(ck)) {
 861                                 bkey_cached_evict(bc, ck);
 862                                 bkey_cached_free(bc, ck);
 863                         }
 864
 865                         scanned++;
 866                         if (scanned >= nr)
 867                                 break;
 868 next:
 869                         pos = next;
 870                 }
 871
 872                 bc->shrink_iter++;
 873                 if (bc->shrink_iter >= tbl->size)
 874                         bc->shrink_iter = 0;
 875         } while (scanned < nr && bc->shrink_iter != start);
 876
 877         rcu_read_unlock();
 878 out:
 879         memalloc_nofs_restore(flags);
 880         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 881         mutex_unlock(&bc->lock);
 882
 883         return freed;
 884 }
 885
 886 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 887                                             struct shrink_control *sc)
 888 {
 889         struct bch_fs *c = container_of(shrink, struct bch_fs,
 890                                         btree_key_cache.shrink);
 891         struct btree_key_cache *bc = &c->btree_key_cache;
 892         long nr = atomic_long_read(&bc->nr_keys) -
 893                 atomic_long_read(&bc->nr_dirty);
 894
 895         return max(0L, nr);
 896 }
 897
 898 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 899 {
 900         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 901         struct bucket_table *tbl;
 902         struct bkey_cached *ck, *n;
 903         struct rhash_head *pos;
 904         unsigned i;
 905 #ifdef __KERNEL__
 906         int cpu;
 907 #endif
 908
 909         if (bc->shrink.list.next)
 910                 unregister_shrinker(&bc->shrink);
 911
 912         mutex_lock(&bc->lock);
 913
 914         /*
 915          * The loop is needed to guard against racing with rehash:
 916          */
 917         while (atomic_long_read(&bc->nr_keys)) {
 918                 rcu_read_lock();
 919                 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 920                 if (tbl)
 921                         for (i = 0; i < tbl->size; i++)
 922                                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 923                                         bkey_cached_evict(bc, ck);
 924                                         list_add(&ck->list, &bc->freed_nonpcpu);
 925                                 }
 926                 rcu_read_unlock();
 927         }
 928
 929 #ifdef __KERNEL__
 930         for_each_possible_cpu(cpu) {
 931                 struct btree_key_cache_freelist *f =
 932                         per_cpu_ptr(bc->pcpu_freed, cpu);
 933
 934                 for (i = 0; i < f->nr; i++) {
 935                         ck = f->objs[i];
 936                         list_add(&ck->list, &bc->freed_nonpcpu);
 937                 }
 938         }
 939 #endif
 940
 941         list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
 942
 943         list_for_each_entry_safe(ck, n, &bc->freed_nonpcpu, list) {
 944                 cond_resched();
 945
 946                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 947                 bch2_journal_preres_put(&c->journal, &ck->res);
 948
 949                 list_del(&ck->list);
 950                 kfree(ck->k);
 951                 six_lock_pcpu_free(&ck->c.lock);
 952                 kmem_cache_free(bch2_key_cache, ck);
 953         }
 954
 955         if (atomic_long_read(&bc->nr_dirty) &&
 956             !bch2_journal_error(&c->journal) &&
 957             test_bit(BCH_FS_WAS_RW, &c->flags))
 958                 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
 959                       atomic_long_read(&bc->nr_dirty));
 960
 961         if (atomic_long_read(&bc->nr_keys))
 962                 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
 963                       atomic_long_read(&bc->nr_keys));
 964
 965         mutex_unlock(&bc->lock);
 966
 967         if (bc->table_init_done)
 968                 rhashtable_destroy(&bc->table);
 969
 970         free_percpu(bc->pcpu_freed);
 971 }
 972
 973 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 974 {
 975         mutex_init(&c->lock);
 976         INIT_LIST_HEAD(&c->freed_pcpu);
 977         INIT_LIST_HEAD(&c->freed_nonpcpu);
 978 }
 979
 980 static void bch2_btree_key_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
 981 {
 982         struct btree_key_cache *bc =
 983                 container_of(shrink, struct btree_key_cache, shrink);
 984
 985         bch2_btree_key_cache_to_text(out, bc);
 986 }
 987
 988 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
 989 {
 990         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 991         int ret;
 992
 993 #ifdef __KERNEL__
 994         bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
 995         if (!bc->pcpu_freed)
 996                 return -ENOMEM;
 997 #endif
 998
 999         ret = rhashtable_init(&bc->table, &bch2_btree_key_cache_params);
1000         if (ret)
1001                 return ret;
1002
1003         bc->table_init_done = true;
1004
1005         bc->shrink.seeks                = 0;
1006         bc->shrink.count_objects        = bch2_btree_key_cache_count;
1007         bc->shrink.scan_objects         = bch2_btree_key_cache_scan;
1008         bc->shrink.to_text              = bch2_btree_key_cache_shrinker_to_text;
1009         return register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name);
1010 }
1011
1012 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1013 {
1014         prt_printf(out, "nr_freed:\t%zu",       atomic_long_read(&c->nr_freed));
1015         prt_newline(out);
1016         prt_printf(out, "nr_keys:\t%lu",        atomic_long_read(&c->nr_keys));
1017         prt_newline(out);
1018         prt_printf(out, "nr_dirty:\t%lu",       atomic_long_read(&c->nr_dirty));
1019         prt_newline(out);
1020 }
1021
1022 void bch2_btree_key_cache_exit(void)
1023 {
1024         kmem_cache_destroy(bch2_key_cache);
1025 }
1026
1027 int __init bch2_btree_key_cache_init(void)
1028 {
1029         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
1030         if (!bch2_key_cache)
1031                 return -ENOMEM;
1032
1033         return 0;
1034 }