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 <linux/seq_buf.h>
  16 #include <trace/events/bcachefs.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         bch2_trans_iter_init(trans, &iter, ck->key.btree_id, ck->key.pos,
 391                              BTREE_ITER_KEY_CACHE_FILL|
 392                              BTREE_ITER_CACHED_NOFILL);
 393         k = bch2_btree_iter_peek_slot(&iter);
 394         ret = bkey_err(k);
 395         if (ret)
 396                 goto err;
 397
 398         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 399                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 400                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 401                 goto err;
 402         }
 403
 404         /*
 405          * bch2_varint_decode can read past the end of the buffer by at
 406          * most 7 bytes (it won't be used):
 407          */
 408         new_u64s = k.k->u64s + 1;
 409
 410         /*
 411          * Allocate some extra space so that the transaction commit path is less
 412          * likely to have to reallocate, since that requires a transaction
 413          * restart:
 414          */
 415         new_u64s = min(256U, (new_u64s * 3) / 2);
 416
 417         if (new_u64s > ck->u64s) {
 418                 new_u64s = roundup_pow_of_two(new_u64s);
 419                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 420                 if (!new_k) {
 421                         bch2_trans_unlock(trans);
 422
 423                         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 424                         if (!new_k) {
 425                                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 426                                         bch2_btree_ids[ck->key.btree_id], new_u64s);
 427                                 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
 428                                 goto err;
 429                         }
 430
 431                         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 432                                 kfree(new_k);
 433                                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 434                                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 435                                 goto err;
 436                         }
 437
 438                         ret = bch2_trans_relock(trans);
 439                         if (ret) {
 440                                 kfree(new_k);
 441                                 goto err;
 442                         }
 443                 }
 444         }
 445
 446         ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
 447         if (ret) {
 448                 kfree(new_k);
 449                 goto err;
 450         }
 451
 452         if (new_k) {
 453                 kfree(ck->k);
 454                 ck->u64s = new_u64s;
 455                 ck->k = new_k;
 456         }
 457
 458         bkey_reassemble(ck->k, k);
 459         ck->valid = true;
 460         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 461
 462         /* We're not likely to need this iterator again: */
 463         set_btree_iter_dontneed(&iter);
 464 err:
 465         bch2_trans_iter_exit(trans, &iter);
 466         return ret;
 467 }
 468
 469 static noinline int
 470 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
 471                                          unsigned flags)
 472 {
 473         struct bch_fs *c = trans->c;
 474         struct bkey_cached *ck;
 475         int ret = 0;
 476
 477         BUG_ON(path->level);
 478
 479         path->l[1].b = NULL;
 480
 481         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 482                 ck = (void *) path->l[0].b;
 483                 goto fill;
 484         }
 485 retry:
 486         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 487         if (!ck) {
 488                 ck = btree_key_cache_create(trans, path);
 489                 ret = PTR_ERR_OR_ZERO(ck);
 490                 if (ret)
 491                         goto err;
 492                 if (!ck)
 493                         goto retry;
 494
 495                 mark_btree_node_locked(trans, path, 0, SIX_LOCK_intent);
 496                 path->locks_want = 1;
 497         } else {
 498                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 499
 500                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 501                                       lock_want, _THIS_IP_);
 502                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 503                         goto err;
 504
 505                 BUG_ON(ret);
 506
 507                 if (ck->key.btree_id != path->btree_id ||
 508                     !bpos_eq(ck->key.pos, path->pos)) {
 509                         six_unlock_type(&ck->c.lock, lock_want);
 510                         goto retry;
 511                 }
 512
 513                 mark_btree_node_locked(trans, path, 0, lock_want);
 514         }
 515
 516         path->l[0].lock_seq     = ck->c.lock.state.seq;
 517         path->l[0].b            = (void *) ck;
 518 fill:
 519         path->uptodate = BTREE_ITER_UPTODATE;
 520
 521         if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
 522                 /*
 523                  * Using the underscore version because we haven't set
 524                  * path->uptodate yet:
 525                  */
 526                 if (!path->locks_want &&
 527                     !__bch2_btree_path_upgrade(trans, path, 1)) {
 528                         trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
 529                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
 530                         goto err;
 531                 }
 532
 533                 ret = btree_key_cache_fill(trans, path, ck);
 534                 if (ret)
 535                         goto err;
 536
 537                 ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
 538                 if (ret)
 539                         goto err;
 540
 541                 path->uptodate = BTREE_ITER_UPTODATE;
 542         }
 543
 544         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 545                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 546
 547         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 548         BUG_ON(path->uptodate);
 549
 550         return ret;
 551 err:
 552         path->uptodate = BTREE_ITER_NEED_TRAVERSE;
 553         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 554                 btree_node_unlock(trans, path, 0);
 555                 path->l[0].b = ERR_PTR(ret);
 556         }
 557         return ret;
 558 }
 559
 560 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 561                                     unsigned flags)
 562 {
 563         struct bch_fs *c = trans->c;
 564         struct bkey_cached *ck;
 565         int ret = 0;
 566
 567         EBUG_ON(path->level);
 568
 569         path->l[1].b = NULL;
 570
 571         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 572                 ck = (void *) path->l[0].b;
 573                 goto fill;
 574         }
 575 retry:
 576         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 577         if (!ck) {
 578                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 579         } else {
 580                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 581
 582                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 583                                       lock_want, _THIS_IP_);
 584                 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
 585
 586                 if (ret)
 587                         return ret;
 588
 589                 if (ck->key.btree_id != path->btree_id ||
 590                     !bpos_eq(ck->key.pos, path->pos)) {
 591                         six_unlock_type(&ck->c.lock, lock_want);
 592                         goto retry;
 593                 }
 594
 595                 mark_btree_node_locked(trans, path, 0, lock_want);
 596         }
 597
 598         path->l[0].lock_seq     = ck->c.lock.state.seq;
 599         path->l[0].b            = (void *) ck;
 600 fill:
 601         if (!ck->valid)
 602                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 603
 604         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 605                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 606
 607         path->uptodate = BTREE_ITER_UPTODATE;
 608         EBUG_ON(!ck->valid);
 609         EBUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 610
 611         return ret;
 612 }
 613
 614 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 615                                      struct bkey_cached_key key,
 616                                      u64 journal_seq,
 617                                      unsigned commit_flags,
 618                                      bool evict)
 619 {
 620         struct bch_fs *c = trans->c;
 621         struct journal *j = &c->journal;
 622         struct btree_iter c_iter, b_iter;
 623         struct bkey_cached *ck = NULL;
 624         int ret;
 625
 626         bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
 627                              BTREE_ITER_SLOTS|
 628                              BTREE_ITER_INTENT|
 629                              BTREE_ITER_ALL_SNAPSHOTS);
 630         bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
 631                              BTREE_ITER_CACHED|
 632                              BTREE_ITER_INTENT);
 633         b_iter.flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 634
 635         ret = bch2_btree_iter_traverse(&c_iter);
 636         if (ret)
 637                 goto out;
 638
 639         ck = (void *) c_iter.path->l[0].b;
 640         if (!ck)
 641                 goto out;
 642
 643         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 644                 if (evict)
 645                         goto evict;
 646                 goto out;
 647         }
 648
 649         BUG_ON(!ck->valid);
 650
 651         if (journal_seq && ck->journal.seq != journal_seq)
 652                 goto out;
 653
 654         /*
 655          * Since journal reclaim depends on us making progress here, and the
 656          * allocator/copygc depend on journal reclaim making progress, we need
 657          * to be using alloc reserves:
 658          */
 659         ret   = bch2_btree_iter_traverse(&b_iter) ?:
 660                 bch2_trans_update(trans, &b_iter, ck->k,
 661                                   BTREE_UPDATE_KEY_CACHE_RECLAIM|
 662                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 663                                   BTREE_TRIGGER_NORUN) ?:
 664                 bch2_trans_commit(trans, NULL, NULL,
 665                                   BTREE_INSERT_NOCHECK_RW|
 666                                   BTREE_INSERT_NOFAIL|
 667                                   BTREE_INSERT_USE_RESERVE|
 668                                   (ck->journal.seq == journal_last_seq(j)
 669                                    ? JOURNAL_WATERMARK_reserved
 670                                    : 0)|
 671                                   commit_flags);
 672
 673         bch2_fs_fatal_err_on(ret &&
 674                              !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 675                              !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
 676                              !bch2_journal_error(j), c,
 677                              "error flushing key cache: %s", bch2_err_str(ret));
 678         if (ret)
 679                 goto out;
 680
 681         bch2_journal_pin_drop(j, &ck->journal);
 682         bch2_journal_preres_put(j, &ck->res);
 683
 684         BUG_ON(!btree_node_locked(c_iter.path, 0));
 685
 686         if (!evict) {
 687                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 688                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 689                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 690                 }
 691         } else {
 692                 struct btree_path *path2;
 693 evict:
 694                 trans_for_each_path(trans, path2)
 695                         if (path2 != c_iter.path)
 696                                 __bch2_btree_path_unlock(trans, path2);
 697
 698                 bch2_btree_node_lock_write_nofail(trans, c_iter.path, &ck->c);
 699
 700                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 701                         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 702                         atomic_long_dec(&c->btree_key_cache.nr_dirty);
 703                 }
 704
 705                 mark_btree_node_locked_noreset(c_iter.path, 0, BTREE_NODE_UNLOCKED);
 706                 bkey_cached_evict(&c->btree_key_cache, ck);
 707                 bkey_cached_free_fast(&c->btree_key_cache, ck);
 708         }
 709 out:
 710         bch2_trans_iter_exit(trans, &b_iter);
 711         bch2_trans_iter_exit(trans, &c_iter);
 712         return ret;
 713 }
 714
 715 int bch2_btree_key_cache_journal_flush(struct journal *j,
 716                                 struct journal_entry_pin *pin, u64 seq)
 717 {
 718         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 719         struct bkey_cached *ck =
 720                 container_of(pin, struct bkey_cached, journal);
 721         struct bkey_cached_key key;
 722         struct btree_trans trans;
 723         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 724         int ret = 0;
 725
 726         bch2_trans_init(&trans, c, 0, 0);
 727
 728         btree_node_lock_nopath_nofail(&trans, &ck->c, SIX_LOCK_read);
 729         key = ck->key;
 730
 731         if (ck->journal.seq != seq ||
 732             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 733                 six_unlock_read(&ck->c.lock);
 734                 goto unlock;
 735         }
 736
 737         if (ck->seq != seq) {
 738                 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
 739                                         bch2_btree_key_cache_journal_flush);
 740                 six_unlock_read(&ck->c.lock);
 741                 goto unlock;
 742         }
 743         six_unlock_read(&ck->c.lock);
 744
 745         ret = commit_do(&trans, NULL, NULL, 0,
 746                 btree_key_cache_flush_pos(&trans, key, seq,
 747                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 748 unlock:
 749         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 750
 751         bch2_trans_exit(&trans);
 752         return ret;
 753 }
 754
 755 /*
 756  * Flush and evict a key from the key cache:
 757  */
 758 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 759                                enum btree_id id, struct bpos pos)
 760 {
 761         struct bch_fs *c = trans->c;
 762         struct bkey_cached_key key = { id, pos };
 763
 764         /* Fastpath - assume it won't be found: */
 765         if (!bch2_btree_key_cache_find(c, id, pos))
 766                 return 0;
 767
 768         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 769 }
 770
 771 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 772                                   unsigned flags,
 773                                   struct btree_insert_entry *insert_entry)
 774 {
 775         struct bch_fs *c = trans->c;
 776         struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
 777         struct bkey_i *insert = insert_entry->k;
 778         bool kick_reclaim = false;
 779
 780         BUG_ON(insert->k.u64s > ck->u64s);
 781
 782         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 783                 int difference;
 784
 785                 BUG_ON(jset_u64s(insert->k.u64s) > trans->journal_preres.u64s);
 786
 787                 difference = jset_u64s(insert->k.u64s) - ck->res.u64s;
 788                 if (difference > 0) {
 789                         trans->journal_preres.u64s      -= difference;
 790                         ck->res.u64s                    += difference;
 791                 }
 792         }
 793
 794         bkey_copy(ck->k, insert);
 795         ck->valid = true;
 796
 797         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 798                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 799                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 800
 801                 if (bch2_nr_btree_keys_need_flush(c))
 802                         kick_reclaim = true;
 803         }
 804
 805         /*
 806          * To minimize lock contention, we only add the journal pin here and
 807          * defer pin updates to the flush callback via ->seq. Be careful not to
 808          * update ->seq on nojournal commits because we don't want to update the
 809          * pin to a seq that doesn't include journal updates on disk. Otherwise
 810          * we risk losing the update after a crash.
 811          *
 812          * The only exception is if the pin is not active in the first place. We
 813          * have to add the pin because journal reclaim drives key cache
 814          * flushing. The flush callback will not proceed unless ->seq matches
 815          * the latest pin, so make sure it starts with a consistent value.
 816          */
 817         if (!(insert_entry->flags & BTREE_UPDATE_NOJOURNAL) ||
 818             !journal_pin_active(&ck->journal)) {
 819                 ck->seq = trans->journal_res.seq;
 820         }
 821         bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 822                              &ck->journal, bch2_btree_key_cache_journal_flush);
 823
 824         if (kick_reclaim)
 825                 journal_reclaim_kick(&c->journal);
 826         return true;
 827 }
 828
 829 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 830                                struct btree_path *path)
 831 {
 832         struct bch_fs *c = trans->c;
 833         struct bkey_cached *ck = (void *) path->l[0].b;
 834
 835         BUG_ON(!ck->valid);
 836
 837         /*
 838          * We just did an update to the btree, bypassing the key cache: the key
 839          * cache key is now stale and must be dropped, even if dirty:
 840          */
 841         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 842                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 843                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 844                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 845         }
 846
 847         ck->valid = false;
 848 }
 849
 850 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 851                                            struct shrink_control *sc)
 852 {
 853         struct bch_fs *c = container_of(shrink, struct bch_fs,
 854                                         btree_key_cache.shrink);
 855         struct btree_key_cache *bc = &c->btree_key_cache;
 856         struct bucket_table *tbl;
 857         struct bkey_cached *ck, *t;
 858         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 859         unsigned start, flags;
 860         int srcu_idx;
 861
 862         mutex_lock(&bc->lock);
 863         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 864         flags = memalloc_nofs_save();
 865
 866         /*
 867          * Newest freed entries are at the end of the list - once we hit one
 868          * that's too new to be freed, we can bail out:
 869          */
 870         list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
 871                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 872                                                  ck->btree_trans_barrier_seq))
 873                         break;
 874
 875                 list_del(&ck->list);
 876                 six_lock_pcpu_free(&ck->c.lock);
 877                 kmem_cache_free(bch2_key_cache, ck);
 878                 atomic_long_dec(&bc->nr_freed);
 879                 scanned++;
 880                 freed++;
 881         }
 882
 883         if (scanned >= nr)
 884                 goto out;
 885
 886         list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
 887                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 888                                                  ck->btree_trans_barrier_seq))
 889                         break;
 890
 891                 list_del(&ck->list);
 892                 six_lock_pcpu_free(&ck->c.lock);
 893                 kmem_cache_free(bch2_key_cache, ck);
 894                 atomic_long_dec(&bc->nr_freed);
 895                 scanned++;
 896                 freed++;
 897         }
 898
 899         if (scanned >= nr)
 900                 goto out;
 901
 902         rcu_read_lock();
 903         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 904         if (bc->shrink_iter >= tbl->size)
 905                 bc->shrink_iter = 0;
 906         start = bc->shrink_iter;
 907
 908         do {
 909                 struct rhash_head *pos, *next;
 910
 911                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 912
 913                 while (!rht_is_a_nulls(pos)) {
 914                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 915                         ck = container_of(pos, struct bkey_cached, hash);
 916
 917                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 918                                 goto next;
 919
 920                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 921                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 922                         else if (bkey_cached_lock_for_evict(ck)) {
 923                                 bkey_cached_evict(bc, ck);
 924                                 bkey_cached_free(bc, ck);
 925                         }
 926
 927                         scanned++;
 928                         if (scanned >= nr)
 929                                 break;
 930 next:
 931                         pos = next;
 932                 }
 933
 934                 bc->shrink_iter++;
 935                 if (bc->shrink_iter >= tbl->size)
 936                         bc->shrink_iter = 0;
 937         } while (scanned < nr && bc->shrink_iter != start);
 938
 939         rcu_read_unlock();
 940 out:
 941         memalloc_nofs_restore(flags);
 942         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 943         mutex_unlock(&bc->lock);
 944
 945         return freed;
 946 }
 947
 948 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 949                                             struct shrink_control *sc)
 950 {
 951         struct bch_fs *c = container_of(shrink, struct bch_fs,
 952                                         btree_key_cache.shrink);
 953         struct btree_key_cache *bc = &c->btree_key_cache;
 954         long nr = atomic_long_read(&bc->nr_keys) -
 955                 atomic_long_read(&bc->nr_dirty);
 956
 957         return max(0L, nr);
 958 }
 959
 960 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 961 {
 962         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 963         struct bucket_table *tbl;
 964         struct bkey_cached *ck, *n;
 965         struct rhash_head *pos;
 966         LIST_HEAD(items);
 967         unsigned i;
 968 #ifdef __KERNEL__
 969         int cpu;
 970 #endif
 971
 972         if (bc->shrink.list.next)
 973                 unregister_shrinker(&bc->shrink);
 974
 975         mutex_lock(&bc->lock);
 976
 977         /*
 978          * The loop is needed to guard against racing with rehash:
 979          */
 980         while (atomic_long_read(&bc->nr_keys)) {
 981                 rcu_read_lock();
 982                 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 983                 if (tbl)
 984                         for (i = 0; i < tbl->size; i++)
 985                                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 986                                         bkey_cached_evict(bc, ck);
 987                                         list_add(&ck->list, &items);
 988                                 }
 989                 rcu_read_unlock();
 990         }
 991
 992 #ifdef __KERNEL__
 993         for_each_possible_cpu(cpu) {
 994                 struct btree_key_cache_freelist *f =
 995                         per_cpu_ptr(bc->pcpu_freed, cpu);
 996
 997                 for (i = 0; i < f->nr; i++) {
 998                         ck = f->objs[i];
 999                         list_add(&ck->list, &items);
1000                 }
1001         }
1002 #endif
1003
1004         list_splice(&bc->freed_pcpu,    &items);
1005         list_splice(&bc->freed_nonpcpu, &items);
1006
1007         mutex_unlock(&bc->lock);
1008
1009         list_for_each_entry_safe(ck, n, &items, list) {
1010                 cond_resched();
1011
1012                 bch2_journal_pin_drop(&c->journal, &ck->journal);
1013                 bch2_journal_preres_put(&c->journal, &ck->res);
1014
1015                 list_del(&ck->list);
1016                 kfree(ck->k);
1017                 six_lock_pcpu_free(&ck->c.lock);
1018                 kmem_cache_free(bch2_key_cache, ck);
1019         }
1020
1021         if (atomic_long_read(&bc->nr_dirty) &&
1022             !bch2_journal_error(&c->journal) &&
1023             test_bit(BCH_FS_WAS_RW, &c->flags))
1024                 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
1025                       atomic_long_read(&bc->nr_dirty));
1026
1027         if (atomic_long_read(&bc->nr_keys))
1028                 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
1029                       atomic_long_read(&bc->nr_keys));
1030
1031         if (bc->table_init_done)
1032                 rhashtable_destroy(&bc->table);
1033
1034         free_percpu(bc->pcpu_freed);
1035 }
1036
1037 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1038 {
1039         mutex_init(&c->lock);
1040         INIT_LIST_HEAD(&c->freed_pcpu);
1041         INIT_LIST_HEAD(&c->freed_nonpcpu);
1042 }
1043
1044 static void bch2_btree_key_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
1045 {
1046         struct btree_key_cache *bc =
1047                 container_of(shrink, struct btree_key_cache, shrink);
1048         char *cbuf;
1049         size_t buflen = seq_buf_get_buf(s, &cbuf);
1050         struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
1051
1052         bch2_btree_key_cache_to_text(&out, bc);
1053         seq_buf_commit(s, out.pos);
1054 }
1055
1056 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1057 {
1058         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1059
1060 #ifdef __KERNEL__
1061         bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1062         if (!bc->pcpu_freed)
1063                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1064 #endif
1065
1066         if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1067                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1068
1069         bc->table_init_done = true;
1070
1071         bc->shrink.seeks                = 0;
1072         bc->shrink.count_objects        = bch2_btree_key_cache_count;
1073         bc->shrink.scan_objects         = bch2_btree_key_cache_scan;
1074         bc->shrink.to_text              = bch2_btree_key_cache_shrinker_to_text;
1075         if (register_shrinker(&bc->shrink, "%s/btree_key_cache", c->name))
1076                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1077         return 0;
1078 }
1079
1080 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1081 {
1082         prt_printf(out, "nr_freed:\t%zu",       atomic_long_read(&c->nr_freed));
1083         prt_newline(out);
1084         prt_printf(out, "nr_keys:\t%lu",        atomic_long_read(&c->nr_keys));
1085         prt_newline(out);
1086         prt_printf(out, "nr_dirty:\t%lu",       atomic_long_read(&c->nr_dirty));
1087         prt_newline(out);
1088 }
1089
1090 void bch2_btree_key_cache_exit(void)
1091 {
1092         kmem_cache_destroy(bch2_key_cache);
1093 }
1094
1095 int __init bch2_btree_key_cache_init(void)
1096 {
1097         bch2_key_cache = KMEM_CACHE(bkey_cached, 0);
1098         if (!bch2_key_cache)
1099                 return -ENOMEM;
1100
1101         return 0;
1102 }