git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_key_cache.c

   1
   2 #include "bcachefs.h"
   3 #include "btree_iter.h"
   4 #include "btree_key_cache.h"
   5 #include "btree_locking.h"
   6 #include "btree_update.h"
   7 #include "error.h"
   8 #include "journal.h"
   9 #include "journal_reclaim.h"
  10
  11 #include <trace/events/bcachefs.h>
  12
  13 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  14                                        const void *obj)
  15 {
  16         const struct bkey_cached *ck = obj;
  17         const struct bkey_cached_key *key = arg->key;
  18
  19         return cmp_int(ck->key.btree_id, key->btree_id) ?:
  20                 bkey_cmp(ck->key.pos, key->pos);
  21 }
  22
  23 static const struct rhashtable_params bch2_btree_key_cache_params = {
  24         .head_offset    = offsetof(struct bkey_cached, hash),
  25         .key_offset     = offsetof(struct bkey_cached, key),
  26         .key_len        = sizeof(struct bkey_cached_key),
  27         .obj_cmpfn      = bch2_btree_key_cache_cmp_fn,
  28 };
  29
  30 __flatten
  31 static inline struct bkey_cached *
  32 btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
  33 {
  34         struct bkey_cached_key key = {
  35                 .btree_id       = btree_id,
  36                 .pos            = pos,
  37         };
  38
  39         return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
  40                                       bch2_btree_key_cache_params);
  41 }
  42
  43 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
  44 {
  45         if (!six_trylock_intent(&ck->c.lock))
  46                 return false;
  47
  48         if (!six_trylock_write(&ck->c.lock)) {
  49                 six_unlock_intent(&ck->c.lock);
  50                 return false;
  51         }
  52
  53         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
  54                 six_unlock_write(&ck->c.lock);
  55                 six_unlock_intent(&ck->c.lock);
  56                 return false;
  57         }
  58
  59         return true;
  60 }
  61
  62 static void bkey_cached_evict(struct btree_key_cache *c,
  63                               struct bkey_cached *ck)
  64 {
  65         BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
  66                                       bch2_btree_key_cache_params));
  67         memset(&ck->key, ~0, sizeof(ck->key));
  68 }
  69
  70 static void bkey_cached_free(struct btree_key_cache *c,
  71                              struct bkey_cached *ck)
  72 {
  73         list_move(&ck->list, &c->freed);
  74
  75         kfree(ck->k);
  76         ck->k           = NULL;
  77         ck->u64s        = 0;
  78
  79         six_unlock_write(&ck->c.lock);
  80         six_unlock_intent(&ck->c.lock);
  81 }
  82
  83 static struct bkey_cached *
  84 bkey_cached_alloc(struct btree_key_cache *c)
  85 {
  86         struct bkey_cached *ck;
  87
  88         list_for_each_entry(ck, &c->freed, list)
  89                 if (bkey_cached_lock_for_evict(ck))
  90                         return ck;
  91
  92         list_for_each_entry(ck, &c->clean, list)
  93                 if (bkey_cached_lock_for_evict(ck)) {
  94                         bkey_cached_evict(c, ck);
  95                         return ck;
  96                 }
  97
  98         ck = kzalloc(sizeof(*ck), GFP_NOFS);
  99         if (!ck)
 100                 return NULL;
 101
 102         INIT_LIST_HEAD(&ck->list);
 103         six_lock_init(&ck->c.lock);
 104         BUG_ON(!six_trylock_intent(&ck->c.lock));
 105         BUG_ON(!six_trylock_write(&ck->c.lock));
 106
 107         return ck;
 108 }
 109
 110 static struct bkey_cached *
 111 btree_key_cache_create(struct btree_key_cache *c,
 112                        enum btree_id btree_id,
 113                        struct bpos pos)
 114 {
 115         struct bkey_cached *ck;
 116
 117         ck = bkey_cached_alloc(c);
 118         if (!ck)
 119                 return ERR_PTR(-ENOMEM);
 120
 121         ck->c.level             = 0;
 122         ck->c.btree_id          = btree_id;
 123         ck->key.btree_id        = btree_id;
 124         ck->key.pos             = pos;
 125         ck->valid               = false;
 126
 127         BUG_ON(ck->flags);
 128
 129         if (rhashtable_lookup_insert_fast(&c->table,
 130                                           &ck->hash,
 131                                           bch2_btree_key_cache_params)) {
 132                 /* We raced with another fill: */
 133                 bkey_cached_free(c, ck);
 134                 return NULL;
 135         }
 136
 137         list_move(&ck->list, &c->clean);
 138         six_unlock_write(&ck->c.lock);
 139
 140         return ck;
 141 }
 142
 143 static int btree_key_cache_fill(struct btree_trans *trans,
 144                                 struct btree_iter *ck_iter,
 145                                 struct bkey_cached *ck)
 146 {
 147         struct btree_iter *iter;
 148         struct bkey_s_c k;
 149         unsigned new_u64s = 0;
 150         struct bkey_i *new_k = NULL;
 151         int ret;
 152
 153         iter = bch2_trans_get_iter(trans, ck->key.btree_id,
 154                                    ck->key.pos, BTREE_ITER_SLOTS);
 155         if (IS_ERR(iter))
 156                 return PTR_ERR(iter);
 157
 158         k = bch2_btree_iter_peek_slot(iter);
 159         ret = bkey_err(k);
 160         if (ret) {
 161                 bch2_trans_iter_put(trans, iter);
 162                 return ret;
 163         }
 164
 165         if (!bch2_btree_node_relock(ck_iter, 0)) {
 166                 bch2_trans_iter_put(trans, iter);
 167                 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 168                 return -EINTR;
 169         }
 170
 171         if (k.k->u64s > ck->u64s) {
 172                 new_u64s = roundup_pow_of_two(k.k->u64s);
 173                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 174                 if (!new_k) {
 175                         bch2_trans_iter_put(trans, iter);
 176                         return -ENOMEM;
 177                 }
 178         }
 179
 180         bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
 181         if (new_k) {
 182                 kfree(ck->k);
 183                 ck->u64s = new_u64s;
 184                 ck->k = new_k;
 185         }
 186
 187         bkey_reassemble(ck->k, k);
 188         ck->valid = true;
 189         bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
 190
 191         /* We're not likely to need this iterator again: */
 192         bch2_trans_iter_free(trans, iter);
 193
 194         return 0;
 195 }
 196
 197 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 198 {
 199         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 200         const struct btree_iter *iter = p;
 201
 202         return ck->key.btree_id == iter->btree_id &&
 203                 !bkey_cmp(ck->key.pos, iter->pos) ? 0 : -1;
 204 }
 205
 206 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
 207 {
 208         struct btree_trans *trans = iter->trans;
 209         struct bch_fs *c = trans->c;
 210         struct bkey_cached *ck;
 211         int ret = 0;
 212
 213         BUG_ON(iter->level);
 214
 215         if (btree_node_locked(iter, 0)) {
 216                 ck = (void *) iter->l[0].b;
 217                 goto fill;
 218         }
 219 retry:
 220         ck = btree_key_cache_find(c, iter->btree_id, iter->pos);
 221         if (!ck) {
 222                 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
 223                         iter->l[0].b = NULL;
 224                         return 0;
 225                 }
 226
 227                 mutex_lock(&c->btree_key_cache.lock);
 228                 ck = btree_key_cache_create(&c->btree_key_cache,
 229                                             iter->btree_id, iter->pos);
 230                 mutex_unlock(&c->btree_key_cache.lock);
 231
 232                 ret = PTR_ERR_OR_ZERO(ck);
 233                 if (ret)
 234                         goto err;
 235                 if (!ck)
 236                         goto retry;
 237
 238                 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
 239                 iter->locks_want = 1;
 240         } else {
 241                 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
 242
 243                 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
 244                                      bkey_cached_check_fn, iter)) {
 245                         if (ck->key.btree_id != iter->btree_id ||
 246                             bkey_cmp(ck->key.pos, iter->pos)) {
 247                                 goto retry;
 248                         }
 249
 250                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 251                         ret = -EINTR;
 252                         goto err;
 253                 }
 254
 255                 if (ck->key.btree_id != iter->btree_id ||
 256                     bkey_cmp(ck->key.pos, iter->pos)) {
 257                         six_unlock_type(&ck->c.lock, lock_want);
 258                         goto retry;
 259                 }
 260
 261                 mark_btree_node_locked(iter, 0, lock_want);
 262         }
 263
 264         iter->l[0].lock_seq     = ck->c.lock.state.seq;
 265         iter->l[0].b            = (void *) ck;
 266 fill:
 267         if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
 268                 if (!btree_node_intent_locked(iter, 0))
 269                         bch2_btree_iter_upgrade(iter, 1);
 270                 if (!btree_node_intent_locked(iter, 0)) {
 271                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 272                         ret = -EINTR;
 273                         goto err;
 274                 }
 275
 276                 ret = btree_key_cache_fill(trans, iter, ck);
 277                 if (ret)
 278                         goto err;
 279         }
 280
 281         iter->uptodate = BTREE_ITER_NEED_PEEK;
 282         bch2_btree_iter_downgrade(iter);
 283         return ret;
 284 err:
 285         if (ret != -EINTR) {
 286                 btree_node_unlock(iter, 0);
 287                 iter->flags |= BTREE_ITER_ERROR;
 288                 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
 289         }
 290         return ret;
 291 }
 292
 293 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 294                                      struct bkey_cached_key key,
 295                                      u64 journal_seq,
 296                                      bool evict)
 297 {
 298         struct bch_fs *c = trans->c;
 299         struct journal *j = &c->journal;
 300         struct btree_iter *c_iter = NULL, *b_iter = NULL;
 301         struct bkey_cached *ck;
 302         int ret;
 303
 304         b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 305                                      BTREE_ITER_SLOTS|
 306                                      BTREE_ITER_INTENT);
 307         ret = PTR_ERR_OR_ZERO(b_iter);
 308         if (ret)
 309                 goto out;
 310
 311         c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 312                                      BTREE_ITER_CACHED|
 313                                      BTREE_ITER_CACHED_NOFILL|
 314                                      BTREE_ITER_CACHED_NOCREATE|
 315                                      BTREE_ITER_INTENT);
 316         ret = PTR_ERR_OR_ZERO(c_iter);
 317         if (ret)
 318                 goto out;
 319 retry:
 320         ret = bch2_btree_iter_traverse(c_iter);
 321         if (ret)
 322                 goto err;
 323
 324         ck = (void *) c_iter->l[0].b;
 325         if (!ck ||
 326             (journal_seq && ck->journal.seq != journal_seq))
 327                 goto out;
 328
 329         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 330                 if (!evict)
 331                         goto out;
 332                 goto evict;
 333         }
 334
 335         ret   = bch2_btree_iter_traverse(b_iter) ?:
 336                 bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
 337                 bch2_trans_commit(trans, NULL, NULL,
 338                                   BTREE_INSERT_NOUNLOCK|
 339                                   BTREE_INSERT_NOCHECK_RW|
 340                                   BTREE_INSERT_NOFAIL|
 341                                   BTREE_INSERT_USE_RESERVE|
 342                                   BTREE_INSERT_USE_ALLOC_RESERVE|
 343                                   BTREE_INSERT_JOURNAL_RESERVED|
 344                                   BTREE_INSERT_JOURNAL_RECLAIM);
 345 err:
 346         if (ret == -EINTR)
 347                 goto retry;
 348
 349         BUG_ON(ret && !bch2_journal_error(j));
 350
 351         if (ret)
 352                 goto out;
 353
 354         bch2_journal_pin_drop(j, &ck->journal);
 355         bch2_journal_preres_put(j, &ck->res);
 356         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 357
 358         if (!evict) {
 359                 mutex_lock(&c->btree_key_cache.lock);
 360                 list_move_tail(&ck->list, &c->btree_key_cache.clean);
 361                 mutex_unlock(&c->btree_key_cache.lock);
 362         } else {
 363 evict:
 364                 BUG_ON(!btree_node_intent_locked(c_iter, 0));
 365
 366                 mark_btree_node_unlocked(c_iter, 0);
 367                 c_iter->l[0].b = NULL;
 368
 369                 six_lock_write(&ck->c.lock, NULL, NULL);
 370
 371                 mutex_lock(&c->btree_key_cache.lock);
 372                 bkey_cached_evict(&c->btree_key_cache, ck);
 373                 bkey_cached_free(&c->btree_key_cache, ck);
 374                 mutex_unlock(&c->btree_key_cache.lock);
 375         }
 376 out:
 377         bch2_trans_iter_put(trans, b_iter);
 378         bch2_trans_iter_put(trans, c_iter);
 379         return ret;
 380 }
 381
 382 static void btree_key_cache_journal_flush(struct journal *j,
 383                                           struct journal_entry_pin *pin,
 384                                           u64 seq)
 385 {
 386         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 387         struct bkey_cached *ck =
 388                 container_of(pin, struct bkey_cached, journal);
 389         struct bkey_cached_key key;
 390         struct btree_trans trans;
 391
 392         six_lock_read(&ck->c.lock, NULL, NULL);
 393         key = READ_ONCE(ck->key);
 394
 395         if (ck->journal.seq != seq ||
 396             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 397                 six_unlock_read(&ck->c.lock);
 398                 return;
 399         }
 400         six_unlock_read(&ck->c.lock);
 401
 402         bch2_trans_init(&trans, c, 0, 0);
 403         btree_key_cache_flush_pos(&trans, key, seq, false);
 404         bch2_trans_exit(&trans);
 405 }
 406
 407 /*
 408  * Flush and evict a key from the key cache:
 409  */
 410 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 411                                enum btree_id id, struct bpos pos)
 412 {
 413         struct bch_fs *c = trans->c;
 414         struct bkey_cached_key key = { id, pos };
 415
 416         /* Fastpath - assume it won't be found: */
 417         if (!btree_key_cache_find(c, id, pos))
 418                 return 0;
 419
 420         return btree_key_cache_flush_pos(trans, key, 0, true);
 421 }
 422
 423 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 424                                   struct btree_iter *iter,
 425                                   struct bkey_i *insert)
 426 {
 427         struct bch_fs *c = trans->c;
 428         struct bkey_cached *ck = (void *) iter->l[0].b;
 429
 430         BUG_ON(insert->u64s > ck->u64s);
 431
 432         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 433                 int difference;
 434
 435                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 436
 437                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 438                 if (difference > 0) {
 439                         trans->journal_preres.u64s      -= difference;
 440                         ck->res.u64s                    += difference;
 441                 }
 442         }
 443
 444         bkey_copy(ck->k, insert);
 445         ck->valid = true;
 446
 447         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 448                 mutex_lock(&c->btree_key_cache.lock);
 449                 list_del_init(&ck->list);
 450
 451                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 452                 mutex_unlock(&c->btree_key_cache.lock);
 453         }
 454
 455         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 456                                 &ck->journal, btree_key_cache_journal_flush);
 457         return true;
 458 }
 459
 460 #ifdef CONFIG_BCACHEFS_DEBUG
 461 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
 462                                enum btree_id id, struct bpos pos)
 463 {
 464         BUG_ON(btree_key_cache_find(trans->c, id, pos));
 465 }
 466 #endif
 467
 468 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *c)
 469 {
 470         struct bkey_cached *ck, *n;
 471
 472         mutex_lock(&c->lock);
 473         list_for_each_entry_safe(ck, n, &c->clean, list) {
 474                 kfree(ck->k);
 475                 kfree(ck);
 476         }
 477         list_for_each_entry_safe(ck, n, &c->freed, list)
 478                 kfree(ck);
 479         mutex_unlock(&c->lock);
 480
 481         rhashtable_destroy(&c->table);
 482 }
 483
 484 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 485 {
 486         mutex_init(&c->lock);
 487         INIT_LIST_HEAD(&c->freed);
 488         INIT_LIST_HEAD(&c->clean);
 489 }
 490
 491 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
 492 {
 493         return rhashtable_init(&c->table, &bch2_btree_key_cache_params);
 494 }