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                 ret = btree_node_lock_nopath(trans, &ck->c, SIX_LOCK_intent, _THIS_IP_);
 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 = six_lock_seq(&ck->c.lock);
 254                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
 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         ck = allocate_dropping_locks(trans, ret,
 267                         kmem_cache_zalloc(bch2_key_cache, _gfp));
 268         if (ret) {
 269                 kmem_cache_free(bch2_key_cache, ck);
 270                 return ERR_PTR(ret);
 271         }
 272
 273         if (!ck)
 274                 return NULL;
 275
 276         INIT_LIST_HEAD(&ck->list);
 277         bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
 278
 279         ck->c.cached = true;
 280         BUG_ON(!six_trylock_intent(&ck->c.lock));
 281         BUG_ON(!six_trylock_write(&ck->c.lock));
 282         *was_new = true;
 283         return ck;
 284 }
 285
 286 static struct bkey_cached *
 287 bkey_cached_reuse(struct btree_key_cache *c)
 288 {
 289         struct bucket_table *tbl;
 290         struct rhash_head *pos;
 291         struct bkey_cached *ck;
 292         unsigned i;
 293
 294         mutex_lock(&c->lock);
 295         rcu_read_lock();
 296         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 297         for (i = 0; i < tbl->size; i++)
 298                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 299                         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 300                             bkey_cached_lock_for_evict(ck)) {
 301                                 bkey_cached_evict(c, ck);
 302                                 goto out;
 303                         }
 304                 }
 305         ck = NULL;
 306 out:
 307         rcu_read_unlock();
 308         mutex_unlock(&c->lock);
 309         return ck;
 310 }
 311
 312 static struct bkey_cached *
 313 btree_key_cache_create(struct btree_trans *trans, struct btree_path *path)
 314 {
 315         struct bch_fs *c = trans->c;
 316         struct btree_key_cache *bc = &c->btree_key_cache;
 317         struct bkey_cached *ck;
 318         bool was_new = false;
 319
 320         ck = bkey_cached_alloc(trans, path, &was_new);
 321         if (IS_ERR(ck))
 322                 return ck;
 323
 324         if (unlikely(!ck)) {
 325                 ck = bkey_cached_reuse(bc);
 326                 if (unlikely(!ck)) {
 327                         bch_err(c, "error allocating memory for key cache item, btree %s",
 328                                 bch2_btree_id_str(path->btree_id));
 329                         return ERR_PTR(-BCH_ERR_ENOMEM_btree_key_cache_create);
 330                 }
 331
 332                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
 333         }
 334
 335         ck->c.level             = 0;
 336         ck->c.btree_id          = path->btree_id;
 337         ck->key.btree_id        = path->btree_id;
 338         ck->key.pos             = path->pos;
 339         ck->valid               = false;
 340         ck->flags               = 1U << BKEY_CACHED_ACCESSED;
 341
 342         if (unlikely(rhashtable_lookup_insert_fast(&bc->table,
 343                                           &ck->hash,
 344                                           bch2_btree_key_cache_params))) {
 345                 /* We raced with another fill: */
 346
 347                 if (likely(was_new)) {
 348                         six_unlock_write(&ck->c.lock);
 349                         six_unlock_intent(&ck->c.lock);
 350                         kfree(ck);
 351                 } else {
 352                         bkey_cached_free_fast(bc, ck);
 353                 }
 354
 355                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
 356                 return NULL;
 357         }
 358
 359         atomic_long_inc(&bc->nr_keys);
 360
 361         six_unlock_write(&ck->c.lock);
 362
 363         return ck;
 364 }
 365
 366 static int btree_key_cache_fill(struct btree_trans *trans,
 367                                 struct btree_path *ck_path,
 368                                 struct bkey_cached *ck)
 369 {
 370         struct btree_iter iter;
 371         struct bkey_s_c k;
 372         unsigned new_u64s = 0;
 373         struct bkey_i *new_k = NULL;
 374         int ret;
 375
 376         k = bch2_bkey_get_iter(trans, &iter, ck->key.btree_id, ck->key.pos,
 377                                BTREE_ITER_KEY_CACHE_FILL|
 378                                BTREE_ITER_CACHED_NOFILL);
 379         ret = bkey_err(k);
 380         if (ret)
 381                 goto err;
 382
 383         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 384                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 385                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 386                 goto err;
 387         }
 388
 389         /*
 390          * bch2_varint_decode can read past the end of the buffer by at
 391          * most 7 bytes (it won't be used):
 392          */
 393         new_u64s = k.k->u64s + 1;
 394
 395         /*
 396          * Allocate some extra space so that the transaction commit path is less
 397          * likely to have to reallocate, since that requires a transaction
 398          * restart:
 399          */
 400         new_u64s = min(256U, (new_u64s * 3) / 2);
 401
 402         if (new_u64s > ck->u64s) {
 403                 new_u64s = roundup_pow_of_two(new_u64s);
 404                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 405                 if (!new_k) {
 406                         bch2_trans_unlock(trans);
 407
 408                         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 409                         if (!new_k) {
 410                                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 411                                         bch2_btree_id_str(ck->key.btree_id), new_u64s);
 412                                 ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
 413                                 goto err;
 414                         }
 415
 416                         if (!bch2_btree_node_relock(trans, ck_path, 0)) {
 417                                 kfree(new_k);
 418                                 trace_and_count(trans->c, trans_restart_relock_key_cache_fill, trans, _THIS_IP_, ck_path);
 419                                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_fill);
 420                                 goto err;
 421                         }
 422
 423                         ret = bch2_trans_relock(trans);
 424                         if (ret) {
 425                                 kfree(new_k);
 426                                 goto err;
 427                         }
 428                 }
 429         }
 430
 431         ret = bch2_btree_node_lock_write(trans, ck_path, &ck_path->l[0].b->c);
 432         if (ret) {
 433                 kfree(new_k);
 434                 goto err;
 435         }
 436
 437         if (new_k) {
 438                 kfree(ck->k);
 439                 ck->u64s = new_u64s;
 440                 ck->k = new_k;
 441         }
 442
 443         bkey_reassemble(ck->k, k);
 444         ck->valid = true;
 445         bch2_btree_node_unlock_write(trans, ck_path, ck_path->l[0].b);
 446
 447         /* We're not likely to need this iterator again: */
 448         set_btree_iter_dontneed(&iter);
 449 err:
 450         bch2_trans_iter_exit(trans, &iter);
 451         return ret;
 452 }
 453
 454 static noinline int
 455 bch2_btree_path_traverse_cached_slowpath(struct btree_trans *trans, struct btree_path *path,
 456                                          unsigned flags)
 457 {
 458         struct bch_fs *c = trans->c;
 459         struct bkey_cached *ck;
 460         int ret = 0;
 461
 462         BUG_ON(path->level);
 463
 464         path->l[1].b = NULL;
 465
 466         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 467                 ck = (void *) path->l[0].b;
 468                 goto fill;
 469         }
 470 retry:
 471         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 472         if (!ck) {
 473                 ck = btree_key_cache_create(trans, path);
 474                 ret = PTR_ERR_OR_ZERO(ck);
 475                 if (ret)
 476                         goto err;
 477                 if (!ck)
 478                         goto retry;
 479
 480                 mark_btree_node_locked(trans, path, 0, BTREE_NODE_INTENT_LOCKED);
 481                 path->locks_want = 1;
 482         } else {
 483                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 484
 485                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 486                                       lock_want, _THIS_IP_);
 487                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 488                         goto err;
 489
 490                 BUG_ON(ret);
 491
 492                 if (ck->key.btree_id != path->btree_id ||
 493                     !bpos_eq(ck->key.pos, path->pos)) {
 494                         six_unlock_type(&ck->c.lock, lock_want);
 495                         goto retry;
 496                 }
 497
 498                 mark_btree_node_locked(trans, path, 0,
 499                                        (enum btree_node_locked_type) lock_want);
 500         }
 501
 502         path->l[0].lock_seq     = six_lock_seq(&ck->c.lock);
 503         path->l[0].b            = (void *) ck;
 504 fill:
 505         path->uptodate = BTREE_ITER_UPTODATE;
 506
 507         if (!ck->valid && !(flags & BTREE_ITER_CACHED_NOFILL)) {
 508                 /*
 509                  * Using the underscore version because we haven't set
 510                  * path->uptodate yet:
 511                  */
 512                 if (!path->locks_want &&
 513                     !__bch2_btree_path_upgrade(trans, path, 1, NULL)) {
 514                         trace_and_count(trans->c, trans_restart_key_cache_upgrade, trans, _THIS_IP_);
 515                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_upgrade);
 516                         goto err;
 517                 }
 518
 519                 ret = btree_key_cache_fill(trans, path, ck);
 520                 if (ret)
 521                         goto err;
 522
 523                 ret = bch2_btree_path_relock(trans, path, _THIS_IP_);
 524                 if (ret)
 525                         goto err;
 526
 527                 path->uptodate = BTREE_ITER_UPTODATE;
 528         }
 529
 530         if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 531                 set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 532
 533         BUG_ON(btree_node_locked_type(path, 0) != btree_lock_want(path, 0));
 534         BUG_ON(path->uptodate);
 535
 536         return ret;
 537 err:
 538         path->uptodate = BTREE_ITER_NEED_TRAVERSE;
 539         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 540                 btree_node_unlock(trans, path, 0);
 541                 path->l[0].b = ERR_PTR(ret);
 542         }
 543         return ret;
 544 }
 545
 546 int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
 547                                     unsigned flags)
 548 {
 549         struct bch_fs *c = trans->c;
 550         struct bkey_cached *ck;
 551         int ret = 0;
 552
 553         EBUG_ON(path->level);
 554
 555         path->l[1].b = NULL;
 556
 557         if (bch2_btree_node_relock_notrace(trans, path, 0)) {
 558                 ck = (void *) path->l[0].b;
 559                 goto fill;
 560         }
 561 retry:
 562         ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
 563         if (!ck) {
 564                 return bch2_btree_path_traverse_cached_slowpath(trans, path, flags);
 565         } else {
 566                 enum six_lock_type lock_want = __btree_lock_want(path, 0);
 567
 568                 ret = btree_node_lock(trans, path, (void *) ck, 0,
 569                                       lock_want, _THIS_IP_);
 570                 EBUG_ON(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart));
 571
 572                 if (ret)
 573                         return ret;
 574
 575                 if (ck->key.btree_id != path->btree_id ||
 576                     !bpos_eq(ck->key.pos, path->pos)) {
 577                         six_unlock_type(&ck->c.lock, lock_want);
 578                         goto retry;
 579                 }
 580
 581                 mark_btree_node_locked(trans, path, 0,
 582                                        (enum btree_node_locked_type) 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 = bch2_trans_get(c);
 709         int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 710         int ret = 0;
 711
 712         btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
 713         key = ck->key;
 714
 715         if (ck->journal.seq != seq ||
 716             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 717                 six_unlock_read(&ck->c.lock);
 718                 goto unlock;
 719         }
 720
 721         if (ck->seq != seq) {
 722                 bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
 723                                         bch2_btree_key_cache_journal_flush);
 724                 six_unlock_read(&ck->c.lock);
 725                 goto unlock;
 726         }
 727         six_unlock_read(&ck->c.lock);
 728
 729         ret = commit_do(trans, NULL, NULL, 0,
 730                 btree_key_cache_flush_pos(trans, key, seq,
 731                                 BTREE_INSERT_JOURNAL_RECLAIM, false));
 732 unlock:
 733         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 734
 735         bch2_trans_put(trans);
 736         return ret;
 737 }
 738
 739 /*
 740  * Flush and evict a key from the key cache:
 741  */
 742 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 743                                enum btree_id id, struct bpos pos)
 744 {
 745         struct bch_fs *c = trans->c;
 746         struct bkey_cached_key key = { id, pos };
 747
 748         /* Fastpath - assume it won't be found: */
 749         if (!bch2_btree_key_cache_find(c, id, pos))
 750                 return 0;
 751
 752         return btree_key_cache_flush_pos(trans, key, 0, 0, true);
 753 }
 754
 755 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 756                                   unsigned flags,
 757                                   struct btree_insert_entry *insert_entry)
 758 {
 759         struct bch_fs *c = trans->c;
 760         struct bkey_cached *ck = (void *) insert_entry->path->l[0].b;
 761         struct bkey_i *insert = insert_entry->k;
 762         bool kick_reclaim = false;
 763
 764         BUG_ON(insert->k.u64s > ck->u64s);
 765
 766         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 767                 int difference;
 768
 769                 BUG_ON(jset_u64s(insert->k.u64s) > trans->journal_preres.u64s);
 770
 771                 difference = jset_u64s(insert->k.u64s) - ck->res.u64s;
 772                 if (difference > 0) {
 773                         trans->journal_preres.u64s      -= difference;
 774                         ck->res.u64s                    += difference;
 775                 }
 776         }
 777
 778         bkey_copy(ck->k, insert);
 779         ck->valid = true;
 780
 781         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 782                 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
 783                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 784                 atomic_long_inc(&c->btree_key_cache.nr_dirty);
 785
 786                 if (bch2_nr_btree_keys_need_flush(c))
 787                         kick_reclaim = true;
 788         }
 789
 790         /*
 791          * To minimize lock contention, we only add the journal pin here and
 792          * defer pin updates to the flush callback via ->seq. Be careful not to
 793          * update ->seq on nojournal commits because we don't want to update the
 794          * pin to a seq that doesn't include journal updates on disk. Otherwise
 795          * we risk losing the update after a crash.
 796          *
 797          * The only exception is if the pin is not active in the first place. We
 798          * have to add the pin because journal reclaim drives key cache
 799          * flushing. The flush callback will not proceed unless ->seq matches
 800          * the latest pin, so make sure it starts with a consistent value.
 801          */
 802         if (!(insert_entry->flags & BTREE_UPDATE_NOJOURNAL) ||
 803             !journal_pin_active(&ck->journal)) {
 804                 ck->seq = trans->journal_res.seq;
 805         }
 806         bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 807                              &ck->journal, bch2_btree_key_cache_journal_flush);
 808
 809         if (kick_reclaim)
 810                 journal_reclaim_kick(&c->journal);
 811         return true;
 812 }
 813
 814 void bch2_btree_key_cache_drop(struct btree_trans *trans,
 815                                struct btree_path *path)
 816 {
 817         struct bch_fs *c = trans->c;
 818         struct bkey_cached *ck = (void *) path->l[0].b;
 819
 820         BUG_ON(!ck->valid);
 821
 822         /*
 823          * We just did an update to the btree, bypassing the key cache: the key
 824          * cache key is now stale and must be dropped, even if dirty:
 825          */
 826         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 827                 clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 828                 atomic_long_dec(&c->btree_key_cache.nr_dirty);
 829                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 830         }
 831
 832         ck->valid = false;
 833 }
 834
 835 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 836                                            struct shrink_control *sc)
 837 {
 838         struct bch_fs *c = container_of(shrink, struct bch_fs,
 839                                         btree_key_cache.shrink);
 840         struct btree_key_cache *bc = &c->btree_key_cache;
 841         struct bucket_table *tbl;
 842         struct bkey_cached *ck, *t;
 843         size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
 844         unsigned start, flags;
 845         int srcu_idx;
 846
 847         mutex_lock(&bc->lock);
 848         srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
 849         flags = memalloc_nofs_save();
 850
 851         /*
 852          * Newest freed entries are at the end of the list - once we hit one
 853          * that's too new to be freed, we can bail out:
 854          */
 855         list_for_each_entry_safe(ck, t, &bc->freed_nonpcpu, list) {
 856                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 857                                                  ck->btree_trans_barrier_seq))
 858                         break;
 859
 860                 list_del(&ck->list);
 861                 six_lock_exit(&ck->c.lock);
 862                 kmem_cache_free(bch2_key_cache, ck);
 863                 atomic_long_dec(&bc->nr_freed);
 864                 scanned++;
 865                 freed++;
 866         }
 867
 868         if (scanned >= nr)
 869                 goto out;
 870
 871         list_for_each_entry_safe(ck, t, &bc->freed_pcpu, list) {
 872                 if (!poll_state_synchronize_srcu(&c->btree_trans_barrier,
 873                                                  ck->btree_trans_barrier_seq))
 874                         break;
 875
 876                 list_del(&ck->list);
 877                 six_lock_exit(&ck->c.lock);
 878                 kmem_cache_free(bch2_key_cache, ck);
 879                 atomic_long_dec(&bc->nr_freed);
 880                 scanned++;
 881                 freed++;
 882         }
 883
 884         if (scanned >= nr)
 885                 goto out;
 886
 887         rcu_read_lock();
 888         tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 889         if (bc->shrink_iter >= tbl->size)
 890                 bc->shrink_iter = 0;
 891         start = bc->shrink_iter;
 892
 893         do {
 894                 struct rhash_head *pos, *next;
 895
 896                 pos = rht_ptr_rcu(rht_bucket(tbl, bc->shrink_iter));
 897
 898                 while (!rht_is_a_nulls(pos)) {
 899                         next = rht_dereference_bucket_rcu(pos->next, tbl, bc->shrink_iter);
 900                         ck = container_of(pos, struct bkey_cached, hash);
 901
 902                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 903                                 goto next;
 904
 905                         if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
 906                                 clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
 907                         else if (bkey_cached_lock_for_evict(ck)) {
 908                                 bkey_cached_evict(bc, ck);
 909                                 bkey_cached_free(bc, ck);
 910                         }
 911
 912                         scanned++;
 913                         if (scanned >= nr)
 914                                 break;
 915 next:
 916                         pos = next;
 917                 }
 918
 919                 bc->shrink_iter++;
 920                 if (bc->shrink_iter >= tbl->size)
 921                         bc->shrink_iter = 0;
 922         } while (scanned < nr && bc->shrink_iter != start);
 923
 924         rcu_read_unlock();
 925 out:
 926         memalloc_nofs_restore(flags);
 927         srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
 928         mutex_unlock(&bc->lock);
 929
 930         return freed;
 931 }
 932
 933 static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
 934                                             struct shrink_control *sc)
 935 {
 936         struct bch_fs *c = container_of(shrink, struct bch_fs,
 937                                         btree_key_cache.shrink);
 938         struct btree_key_cache *bc = &c->btree_key_cache;
 939         long nr = atomic_long_read(&bc->nr_keys) -
 940                 atomic_long_read(&bc->nr_dirty);
 941
 942         return max(0L, nr);
 943 }
 944
 945 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
 946 {
 947         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
 948         struct bucket_table *tbl;
 949         struct bkey_cached *ck, *n;
 950         struct rhash_head *pos;
 951         LIST_HEAD(items);
 952         unsigned i;
 953 #ifdef __KERNEL__
 954         int cpu;
 955 #endif
 956
 957         unregister_shrinker(&bc->shrink);
 958
 959         mutex_lock(&bc->lock);
 960
 961         /*
 962          * The loop is needed to guard against racing with rehash:
 963          */
 964         while (atomic_long_read(&bc->nr_keys)) {
 965                 rcu_read_lock();
 966                 tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
 967                 if (tbl)
 968                         for (i = 0; i < tbl->size; i++)
 969                                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 970                                         bkey_cached_evict(bc, ck);
 971                                         list_add(&ck->list, &items);
 972                                 }
 973                 rcu_read_unlock();
 974         }
 975
 976 #ifdef __KERNEL__
 977         for_each_possible_cpu(cpu) {
 978                 struct btree_key_cache_freelist *f =
 979                         per_cpu_ptr(bc->pcpu_freed, cpu);
 980
 981                 for (i = 0; i < f->nr; i++) {
 982                         ck = f->objs[i];
 983                         list_add(&ck->list, &items);
 984                 }
 985         }
 986 #endif
 987
 988         list_splice(&bc->freed_pcpu,    &items);
 989         list_splice(&bc->freed_nonpcpu, &items);
 990
 991         mutex_unlock(&bc->lock);
 992
 993         list_for_each_entry_safe(ck, n, &items, list) {
 994                 cond_resched();
 995
 996                 bch2_journal_pin_drop(&c->journal, &ck->journal);
 997                 bch2_journal_preres_put(&c->journal, &ck->res);
 998
 999                 list_del(&ck->list);
1000                 kfree(ck->k);
1001                 six_lock_exit(&ck->c.lock);
1002                 kmem_cache_free(bch2_key_cache, ck);
1003         }
1004
1005         if (atomic_long_read(&bc->nr_dirty) &&
1006             !bch2_journal_error(&c->journal) &&
1007             test_bit(BCH_FS_WAS_RW, &c->flags))
1008                 panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
1009                       atomic_long_read(&bc->nr_dirty));
1010
1011         if (atomic_long_read(&bc->nr_keys))
1012                 panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
1013                       atomic_long_read(&bc->nr_keys));
1014
1015         if (bc->table_init_done)
1016                 rhashtable_destroy(&bc->table);
1017
1018         free_percpu(bc->pcpu_freed);
1019 }
1020
1021 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
1022 {
1023         mutex_init(&c->lock);
1024         INIT_LIST_HEAD(&c->freed_pcpu);
1025         INIT_LIST_HEAD(&c->freed_nonpcpu);
1026 }
1027
1028 static void bch2_btree_key_cache_shrinker_to_text(struct seq_buf *s, struct shrinker *shrink)
1029 {
1030         struct btree_key_cache *bc =
1031                 container_of(shrink, struct btree_key_cache, shrink);
1032         char *cbuf;
1033         size_t buflen = seq_buf_get_buf(s, &cbuf);
1034         struct printbuf out = PRINTBUF_EXTERN(cbuf, buflen);
1035
1036         bch2_btree_key_cache_to_text(&out, bc);
1037         seq_buf_commit(s, out.pos);
1038 }
1039
1040 int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
1041 {
1042         struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
1043
1044 #ifdef __KERNEL__
1045         bc->pcpu_freed = alloc_percpu(struct btree_key_cache_freelist);
1046         if (!bc->pcpu_freed)
1047                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1048 #endif
1049
1050         if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
1051                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1052
1053         bc->table_init_done = true;
1054
1055         bc->shrink.seeks                = 0;
1056         bc->shrink.count_objects        = bch2_btree_key_cache_count;
1057         bc->shrink.scan_objects         = bch2_btree_key_cache_scan;
1058         bc->shrink.to_text              = bch2_btree_key_cache_shrinker_to_text;
1059         if (register_shrinker(&bc->shrink, "%s-btree_key_cache", c->name))
1060                 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
1061         return 0;
1062 }
1063
1064 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
1065 {
1066         prt_printf(out, "nr_freed:\t%lu",       atomic_long_read(&c->nr_freed));
1067         prt_newline(out);
1068         prt_printf(out, "nr_keys:\t%lu",        atomic_long_read(&c->nr_keys));
1069         prt_newline(out);
1070         prt_printf(out, "nr_dirty:\t%lu",       atomic_long_read(&c->nr_dirty));
1071         prt_newline(out);
1072 }
1073
1074 void bch2_btree_key_cache_exit(void)
1075 {
1076         kmem_cache_destroy(bch2_key_cache);
1077 }
1078
1079 int __init bch2_btree_key_cache_init(void)
1080 {
1081         bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
1082         if (!bch2_key_cache)
1083                 return -ENOMEM;
1084
1085         return 0;
1086 }