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 = six_lock_seq(&ck->c.lock);
 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 = allocate_dropping_locks(trans, ret,
 269                         kmem_cache_zalloc(bch2_key_cache, _gfp));
 270         if (ret) {
 271                 kmem_cache_free(bch2_key_cache, ck);
 272                 return ERR_PTR(ret);
 273         }
 274
 275         if (!ck)
 276                 return NULL;
 277
 278         INIT_LIST_HEAD(&ck->list);
 279         bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
 280
 281         ck->c.cached = true;
 282         BUG_ON(!six_trylock_intent(&ck->c.lock));
 283         BUG_ON(!six_trylock_write(&ck->c.lock));
 284         *was_new = true;
 285         return ck;
 286 }
 287
 288 static struct bkey_cached *
 289 bkey_cached_reuse(struct btree_key_cache *c)
 290 {
 291         struct bucket_table *tbl;
 292         struct rhash_head *pos;
 293         struct bkey_cached *ck;
 294         unsigned i;
 295
 296         mutex_lock(&c->lock);
 297         rcu_read_lock();
 298         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 299         for (i = 0; i < tbl->size; i++)
 300                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 301                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 302                             bkey_cached_lock_for_evict(ck)) {
 303                                 bkey_cached_evict(c, ck);
 304                                 goto out;
 305                         }
 306                 }
 307         ck = NULL;
 308 out:
 309         rcu_read_unlock();
 310         mutex_unlock(&c->lock);
 311         return ck;
 312 }
 313
 314 static struct bkey_cached *
 315 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
 316 {
 317         struct bch_fs *c = trans->c;
 318         struct btree_key_cache *bc = &c->btree_key_cache;
 319         struct bkey_cached *ck;
 320         bool was_new = false;
 321
 322         ck = bkey_cached_alloc(trans, path, &was_new);
 323         if (IS_ERR(ck))
 324                 return ck;
 325
 326         if (unlikely(!ck)) {
 327                 ck = bkey_cached_reuse(bc);
 328                 if (unlikely(!ck)) {
 329                         bch_err(c, "error allocating memory for key cache item, btree %s",
 330                                 bch2_btree_ids[path->btree_id]);
 331                         return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
 332                 }
 333
 334                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 335         }
 336
 337         ck->c.level             = 0;
 338         ck->c.btree_id          = path->btree_id;
 339         ck->key.btree_id        = path->btree_id;
 340         ck->key.pos             = path->pos;
 341         ck->valid               = false;
 342         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 343
 344         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 345                                           &ck->hash,
 346                                           bch2_btree_key_cache_params))) {
 347                 /* We raced with another fill: */
 348
 349                 if (likely(was_new)) {
 350                         six_unlock_write(&ck->c.lock);
 351                         six_unlock_intent(&ck->c.lock);
 352                         kfree(ck);
 353                 } else {
 354                         bkey_cached_free_fast(bc, ck);
 355                 }
 356
 357                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
 358                 return NULL;
 359         }
 360
 361         atomic_long_inc(&bc->nr_keys);
 362
 363         six_unlock_write(&ck->c.lock);
 364
 365         return ck;
 366 }
 367
 368 static int btree_key_cache_fill(struct btree_trans *trans,
 369                                 struct btree_path *ck_path,
 370                                 struct bkey_cached *ck)
 371 {
 372         struct btree_iter iter;
 373         struct bkey_s_c k;
 374         unsigned new_u64s = 0;
 375         struct bkey_i *new_k = NULL;
 376         int ret;
 377
 378         k = bch2_bkey_get_iter(trans, &iter, ck->key.btree_id, ck->key.pos,
 379                                BTREE_ITER_KEY_CACHE_FILL|
 380                                BTREE_ITER_CACHED_NOFILL);
 381         ret = bkey_err(k);
 382         if (ret)
 383                 goto err;
 384
 385         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 386                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 387                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 388                 goto err;
 389         }
 390
 391         /*
 392          * bch2_varint_decode can read past the end of the buffer by at
 393          * most 7 bytes (it won't be used):
 394          */
 395         new_u64s = k.k->u64s + 1;
 396
 397         /*
 398          * Allocate some extra space so that the transaction commit path is less
 399          * likely to have to reallocate, since that requires a transaction
 400          * restart:
 401          */
 402         new_u64s = min(256U, (new_u64s * 3) / 2);
 403
 404         if (new_u64s > ck->u64s) {
 405                 new_u64s = roundup_pow_of_two(new_u64s);
 406                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 407                 if (!new_k) {
 408                         bch2_trans_unlock(trans);
 409
 410                         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 411                         if (!new_k) {
 412                                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 413                                         bch2_btree_ids[ck->key.btree_id], new_u64s);
 414                                 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
 415                                 goto err;
 416                         }
 417
 418                         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 419                                 kfree(new_k);
 420                                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 421                                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 422                                 goto err;
 423                         }
 424
 425                         ret = bch2_trans_relock(trans);
 426                         if (ret) {
 427                                 kfree(new_k);
 428                                 goto err;
 429                         }
 430                 }
 431         }
 432
 433         ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
 434         if (ret) {
 435                 kfree(new_k);
 436                 goto err;
 437         }
 438
 439         if (new_k) {
 440                 kfree(ck->k);
 441                 ck->u64s = new_u64s;
 442                 ck->k = new_k;
 443         }
 444
 445         bkey_reassemble(ck->k, k);
 446         ck->valid = true;
 447         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 448
 449         /* We're not likely to need this iterator again: */
 450         set_btree_iter_dontneed(&iter);
 451 err:
 452         bch2_trans_iter_exit(trans, &iter);
 453         return ret;
 454 }
 455
 456 static noinline int
 457 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
 458                                          unsigned flags)
 459 {
 460         struct bch_fs *c = trans->c;
 461         struct bkey_cached *ck;
 462         int ret = 0;
 463
 464         BUG_ON(path->level);
 465
 466         path->l[1].b = NULL;
 467
 468         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 469                 ck = (void *) path->l[0].b;
 470                 goto fill;
 471         }
 472 retry:
 473         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 474         if (!ck) {
 475                 ck = btree_key_cache_create(trans, path);
 476                 ret = PTR_ERR_OR_ZERO(ck);
 477                 if (ret)
 478                         goto err;
 479                 if (!ck)
 480                         goto retry;
 481
 482                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 483                 path->locks_want = 1;
 484         } else {
 485                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 486
 487                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 488                                       lock_want, _THIS_IP_);
 489                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 490                         goto err;
 491
 492                 BUG_ON(ret);
 493
 494                 if (ck->key.btree_id != path->btree_id ||
 495                     !bpos_eq(ck->key.pos, path->pos)) {
 496                         six_unlock_type(&ck->c.lock, lock_want);
 497                         goto retry;
 498                 }
 499
 500                 mark_btree_node_locked(trans, path, 0, lock_want);
 501         }
 502
 503         path->l[0].lock_seq     = six_lock_seq(&ck->c.lock);
 504         path->l[0].b            = (void *) ck;
 505 fill:
 506         path->uptodate = BTREE_ITER_UPTODATE;
 507
 508         if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
 509                 /*
 510                  * Using the underscore version because we haven't set
 511                  * path->uptodate yet:
 512                  */
 513                 if (!path->locks_want &&
 514                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 515                         trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
 516                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
 517                         goto err;
 518                 }
 519
 520                 ret = btree_key_cache_fill(trans, path, ck);
 521                 if (ret)
 522                         goto err;
 523
 524                 ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
 525                 if (ret)
 526                         goto err;
 527
 528                 path->uptodate = BTREE_ITER_UPTODATE;
 529         }
 530
 531         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 532                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 533
 534         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 535         BUG_ON(path->uptodate);
 536
 537         return ret;
 538 err:
 539         path->uptodate = BTREE_ITER_NEED_TRAVERSE;
 540         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 541                 btree_node_unlock(trans, path, 0);
 542                 path->l[0].b = ERR_PTR(ret);
 543         }
 544         return ret;
 545 }
 546
 547 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 548                                     unsigned flags)
 549 {
 550         struct bch_fs *c = trans->c;
 551         struct bkey_cached *ck;
 552         int ret = 0;
 553
 554         EBUG_ON(path->level);
 555
 556         path->l[1].b = NULL;
 557
 558         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 559                 ck = (void *) path->l[0].b;
 560                 goto fill;
 561         }
 562 retry:
 563         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 564         if (!ck) {
 565                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 566         } else {
 567                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 568
 569                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 570                                       lock_want, _THIS_IP_);
 571                 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
 572
 573                 if (ret)
 574                         return ret;
 575
 576                 if (ck->key.btree_id != path->btree_id ||
 577                     !bpos_eq(ck->key.pos, path->pos)) {
 578                         six_unlock_type(&ck->c.lock, lock_want);
 579                         goto retry;
 580                 }
 581
 582                 mark_btree_node_locked(trans, path, 0, lock_want);
 583         }
 584
 585         path->l[0].lock_seq     = six_lock_seq(&ck->c.lock);
 586         path->l[0].b            = (void *) ck;
 587 fill:
 588         if (!ck->valid)
 589                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 590
 591         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 592                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 593
 594         path->uptodate = BTREE_ITER_UPTODATE;
 595         EBUG_ON(!ck->valid);
 596         EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 597
 598         return ret;
 599 }
 600
 601 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 602                                      struct bkey_cached_key key,
 603                                      u64 journal_seq,
 604                                      unsigned commit_flags,
 605                                      bool evict)
 606 {
 607         struct bch_fs *c = trans->c;
 608         struct journal *j = &c->journal;
 609         struct btree_iter c_iter, b_iter;
 610         struct bkey_cached *ck = NULL;
 611         int ret;
 612
 613         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 614                              BTREE_ITER_SLOTS|
 615                              BTREE_ITER_INTENT|
 616                              BTREE_ITER_ALL_SNAPSHOTS);
 617         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 618                              BTREE_ITER_CACHED|
 619                              BTREE_ITER_INTENT);
 620         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 621
 622         ret = bch2_btree_iter_traverse(&c_iter);
 623         if (ret)
 624                 goto out;
 625
 626         ck = (void *) c_iter.path->l[0].b;
 627         if (!ck)
 628                 goto out;
 629
 630         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 631                 if (evict)
 632                         goto evict;
 633                 goto out;
 634         }
 635
 636         BUG_ON(!ck->valid);
 637
 638         if (journal_seq && ck->journal.seq != journal_seq)
 639                 goto out;
 640
 641         /*
 642          * Since journal reclaim depends on us making progress here, and the
 643          * allocator/copygc depend on journal reclaim making progress, we need
 644          * to be using alloc reserves:
 645          */
 646         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 647                 bch2_trans_update(trans, &b_iter, ck->k,
 648                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 649                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 650                                   BTREE_TRIGGER_NORUN) ?:
 651                 bch2_trans_commit(trans, NULL, NULL,
 652                                   BTREE_INSERT_NOCHECK_RW|
 653                                   BTREE_INSERT_NOFAIL|
 654                                   (ck->journal.seq == journal_last_seq(j)
 655                                    ? BCH_WATERMARK_reclaim
 656                                    : 0)|
 657                                   commit_flags);
 658
 659         bch2_fs_fatal_err_on(ret &&
 660                              !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 661                              !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
 662                              !bch2_journal_error(j), c,
 663                              "error flushing key cache: %s", bch2_err_str(ret));
 664         if (ret)
 665                 goto out;
 666
 667         bch2_journal_pin_drop(j, &ck->journal);
 668         bch2_journal_preres_put(j, &ck->res);
 669
 670         BUG_ON(!btree_node_locked(c_iter.path, 0));
 671
 672         if (!evict) {
 673                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 674                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 675                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 676                 }
 677         } else {
 678                 struct btree_path *path2;
 679 evict:
 680                 trans_for_each_path(trans, path2)
 681                         if (path2 != c_iter.path)
 682                                 __bch2_btree_path_unlock(trans, path2);
 683
 684                 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
 685
 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
 691                 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
 692                 bkey_cached_evict(&c->btree_key_cache, ck);
 693                 bkey_cached_free_fast(&c->btree_key_cache, ck);
 694         }
 695 out:
 696         bch2_trans_iter_exit(trans, &b_iter);
 697         bch2_trans_iter_exit(trans, &c_iter);
 698         return ret;
 699 }
 700
 701 int bch2_btree_key_cache_journal_flush(struct journal *j,
 702                                 struct journal_entry_pin *pin, u64 seq)
 703 {
 704         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 705         struct bkey_cached *ck =
 706                 container_of(pin, struct bkey_cached, journal);
 707         struct bkey_cached_key key;
 708         struct btree_trans trans;
 709         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 710         int ret = 0;
 711
 712         bch2_trans_init(&trans, c, 0, 0);
 713
 714         btree_node_lock_nopath_nofail(&trans, &ck->c, SIX_LOCK_read);
 715         key = ck->key;
 716
 717         if (ck->journal.seq != seq ||
 718             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 719                 six_unlock_read(&ck->c.lock);
 720                 goto unlock;
 721         }
 722
 723         if (ck->seq != seq) {
 724                 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
 725                                         bch2_btree_key_cache_journal_flush);
 726                 six_unlock_read(&ck->c.lock);
 727                 goto unlock;
 728         }
 729         six_unlock_read(&ck->c.lock);
 730
 731         ret = commit_do(&trans, NULL, NULL, 0,
 732                 btree_key_cache_flush_pos(&trans, key, seq,
 733                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 734 unlock:
 735         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 736
 737         bch2_trans_exit(&trans);
 738         return ret;
 739 }
 740
 741 /*
 742  * Flush and evict a key from the key cache:
 743  */
 744 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 745                                enum btree_id id, struct bpos pos)
 746 {
 747         struct bch_fs *c = trans->c;
 748         struct bkey_cached_key key = { id, pos };
 749
 750         /* Fastpath - assume it won't be found: */
 751         if (!bch2_btree_key_cache_find(c, id, pos))
 752                 return 0;
 753
 754         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 755 }
 756
 757 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 758                                   unsigned flags,
 759                                   struct btree_insert_entry *insert_entry)
 760 {
 761         struct bch_fs *c = trans->c;
 762         struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
 763         struct bkey_i *insert = insert_entry->k;
 764         bool kick_reclaim = false;
 765
 766         BUG_ON(insert->k.u64s > ck->u64s);
 767
 768         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 769                 int difference;
 770
 771                 BUG_ON(jset_u64s(insert->k.u64s) > trans->journal_preres.u64s);
 772
 773                 difference = jset_u64s(insert->k.u64s) - ck->res.u64s;
 774                 if (difference > 0) {
 775                         trans->journal_preres.u64s      -= difference;
 776                         ck->res.u64s                    += difference;
 777                 }
 778         }
 779
 780         bkey_copy(ck->k, insert);
 781         ck->valid = true;
 782
 783         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 784                 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
 785                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 786                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 787
 788                 if (bch2_nr_btree_keys_need_flush(c))
 789                         kick_reclaim = true;
 790         }
 791
 792         /*
 793          * To minimize lock contention, we only add the journal pin here and
 794          * defer pin updates to the flush callback via ->seq. Be careful not to
 795          * update ->seq on nojournal commits because we don't want to update the
 796          * pin to a seq that doesn't include journal updates on disk. Otherwise
 797          * we risk losing the update after a crash.
 798          *
 799          * The only exception is if the pin is not active in the first place. We
 800          * have to add the pin because journal reclaim drives key cache
 801          * flushing. The flush callback will not proceed unless ->seq matches
 802          * the latest pin, so make sure it starts with a consistent value.
 803          */
 804         if (!(insert_entry->flags & BTREE_UPDATE_NOJOURNAL) ||
 805             !journal_pin_active(&ck->journal)) {
 806                 ck->seq = trans->journal_res.seq;
 807         }
 808         bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 809                              &ck->journal, bch2_btree_key_cache_journal_flush);
 810
 811         if (kick_reclaim)
 812                 journal_reclaim_kick(&c->journal);
 813         return true;
 814 }
 815
 816 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 817                                struct btree_path *path)
 818 {
 819         struct bch_fs *c = trans->c;
 820         struct bkey_cached *ck = (void *) path->l[0].b;
 821
 822         BUG_ON(!ck->valid);
 823
 824         /*
 825          * We just did an update to the btree, bypassing the key cache: the key
 826          * cache key is now stale and must be dropped, even if dirty:
 827          */
 828         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 829                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 830                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 831                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 832         }
 833
 834         ck->valid = false;
 835 }
 836
 837 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 838                                            struct shrink_control *sc)
 839 {
 840         struct bch_fs *c = container_of(shrink, struct bch_fs,
 841                                         btree_key_cache.shrink);
 842         struct btree_key_cache *bc = &c->btree_key_cache;
 843         struct bucket_table *tbl;
 844         struct bkey_cached *ck, *t;
 845         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 846         unsigned start, flags;
 847         int srcu_idx;
 848
 849         mutex_lock(&bc->lock);
 850         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 851         flags = memalloc_nofs_save();
 852
 853         /*
 854          * Newest freed entries are at the end of the list - once we hit one
 855          * that's too new to be freed, we can bail out:
 856          */
 857         list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
 858                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 859                                                  ck->btree_trans_barrier_seq))
 860                         break;
 861
 862                 list_del(&ck->list);
 863                 six_lock_exit(&ck->c.lock);
 864                 kmem_cache_free(bch2_key_cache, ck);
 865                 atomic_long_dec(&bc->nr_freed);
 866                 scanned++;
 867                 freed++;
 868         }
 869
 870         if (scanned >= nr)
 871                 goto out;
 872
 873         list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
 874                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 875                                                  ck->btree_trans_barrier_seq))
 876                         break;
 877
 878                 list_del(&ck->list);
 879                 six_lock_exit(&ck->c.lock);
 880                 kmem_cache_free(bch2_key_cache, ck);
 881                 atomic_long_dec(&bc->nr_freed);
 882                 scanned++;
 883                 freed++;
 884         }
 885
 886         if (scanned >= nr)
 887                 goto out;
 888
 889         rcu_read_lock();
 890         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 891         if (bc->shrink_iter >= tbl->size)
 892                 bc->shrink_iter = 0;
 893         start = bc->shrink_iter;
 894
 895         do {
 896                 struct rhash_head *pos, *next;
 897
 898                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 899
 900                 while (!rht_is_a_nulls(pos)) {
 901                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 902                         ck = container_of(pos, struct bkey_cached, hash);
 903
 904                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 905                                 goto next;
 906
 907                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 908                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 909                         else if (bkey_cached_lock_for_evict(ck)) {
 910                                 bkey_cached_evict(bc, ck);
 911                                 bkey_cached_free(bc, ck);
 912                         }
 913
 914                         scanned++;
 915                         if (scanned >= nr)
 916                                 break;
 917 next:
 918                         pos = next;
 919                 }
 920
 921                 bc->shrink_iter++;
 922                 if (bc->shrink_iter >= tbl->size)
 923                         bc->shrink_iter = 0;
 924         } while (scanned < nr && bc->shrink_iter != start);
 925
 926         rcu_read_unlock();
 927 out:
 928         memalloc_nofs_restore(flags);
 929         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 930         mutex_unlock(&bc->lock);
 931
 932         return freed;
 933 }
 934
 935 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 936                                             struct shrink_control *sc)
 937 {
 938         struct bch_fs *c = container_of(shrink, struct bch_fs,
 939                                         btree_key_cache.shrink);
 940         struct btree_key_cache *bc = &c->btree_key_cache;
 941         long nr = atomic_long_read(&bc->nr_keys) -
 942                 atomic_long_read(&bc->nr_dirty);
 943
 944         return max(0L, nr);
 945 }
 946
 947 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 948 {
 949         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 950         struct bucket_table *tbl;
 951         struct bkey_cached *ck, *n;
 952         struct rhash_head *pos;
 953         LIST_HEAD(items);
 954         unsigned i;
 955 #ifdef __KERNEL__
 956         int cpu;
 957 #endif
 958
 959         unregister_shrinker(&bc->shrink);
 960
 961         mutex_lock(&bc->lock);
 962
 963         /*
 964          * The loop is needed to guard against racing with rehash:
 965          */
 966         while (atomic_long_read(&bc->nr_keys)) {
 967                 rcu_read_lock();
 968                 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 969                 if (tbl)
 970                         for (i = 0; i < tbl->size; i++)
 971                                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 972                                         bkey_cached_evict(bc, ck);
 973                                         list_add(&ck->list, &items);
 974                                 }
 975                 rcu_read_unlock();
 976         }
 977
 978 #ifdef __KERNEL__
 979         for_each_possible_cpu(cpu) {
 980                 struct btree_key_cache_freelist *f =
 981                         per_cpu_ptr(bc->pcpu_freed, cpu);
 982
 983                 for (i = 0; i < f->nr; i++) {
 984                         ck = f->objs[i];
 985                         list_add(&ck->list, &items);
 986                 }
 987         }
 988 #endif
 989
 990         list_splice(&bc->freed_pcpu,    &items);
 991         list_splice(&bc->freed_nonpcpu, &items);
 992
 993         mutex_unlock(&bc->lock);
 994
 995         list_for_each_entry_safe(ck, n, &items, list) {
 996                 cond_resched();
 997
 998                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 999                 bch2_journal_preres_put(&c->journal, &ck->res);
1000
1001                 list_del(&ck->list);
1002                 kfree(ck->k);
1003                 six_lock_exit(&ck->c.lock);
1004                 kmem_cache_free(bch2_key_cache, ck);
1005         }
1006
1007         if (atomic_long_read(&bc->nr_dirty) &&
1008             !bch2_journal_error(&c->journal) &&
1009             test_bit(BCH_FS_WAS_RW, &c->flags))
1010                 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
1011                       atomic_long_read(&bc->nr_dirty));
1012
1013         if (atomic_long_read(&bc->nr_keys))
1014                 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
1015                       atomic_long_read(&bc->nr_keys));
1016
1017         if (bc->table_init_done)
1018                 rhashtable_destroy(&bc->table);
1019
1020         free_percpu(bc->pcpu_freed);
1021 }
1022
1023 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1024 {
1025         mutex_init(&c->lock);
1026         INIT_LIST_HEAD(&c->freed_pcpu);
1027         INIT_LIST_HEAD(&c->freed_nonpcpu);
1028 }
1029
1030 static void bch2_btree_key_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
1031 {
1032         struct btree_key_cache *bc =
1033                 container_of(shrink, struct btree_key_cache, shrink);
1034         char *cbuf;
1035         size_t buflen = seq_buf_get_buf(s, &cbuf);
1036         struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
1037
1038         bch2_btree_key_cache_to_text(&out, bc);
1039         seq_buf_commit(s, out.pos);
1040 }
1041
1042 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1043 {
1044         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1045
1046 #ifdef __KERNEL__
1047         bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1048         if (!bc->pcpu_freed)
1049                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1050 #endif
1051
1052         if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1053                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1054
1055         bc->table_init_done = true;
1056
1057         bc->shrink.seeks                = 0;
1058         bc->shrink.count_objects        = bch2_btree_key_cache_count;
1059         bc->shrink.scan_objects         = bch2_btree_key_cache_scan;
1060         bc->shrink.to_text              = bch2_btree_key_cache_shrinker_to_text;
1061         if (register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name))
1062                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1063         return 0;
1064 }
1065
1066 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1067 {
1068         prt_printf(out, "nr_freed:\t%zu",       atomic_long_read(&c->nr_freed));
1069         prt_newline(out);
1070         prt_printf(out, "nr_keys:\t%lu",        atomic_long_read(&c->nr_keys));
1071         prt_newline(out);
1072         prt_printf(out, "nr_dirty:\t%lu",       atomic_long_read(&c->nr_dirty));
1073         prt_newline(out);
1074 }
1075
1076 void bch2_btree_key_cache_exit(void)
1077 {
1078         kmem_cache_destroy(bch2_key_cache);
1079 }
1080
1081 int __init bch2_btree_key_cache_init(void)
1082 {
1083         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
1084         if (!bch2_key_cache)
1085                 return -ENOMEM;
1086
1087         return 0;
1088 }