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

   1
   2 #include "bcachefs.h"
   3 #include "btree_cache.h"
   4 #include "btree_iter.h"
   5 #include "btree_key_cache.h"
   6 #include "btree_locking.h"
   7 #include "btree_update.h"
   8 #include "error.h"
   9 #include "journal.h"
  10 #include "journal_reclaim.h"
  11
  12 #include <trace/events/bcachefs.h>
  13
  14 static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
  15                                        const void *obj)
  16 {
  17         const struct bkey_cached *ck = obj;
  18         const struct bkey_cached_key *key = arg->key;
  19
  20         return cmp_int(ck->key.btree_id, key->btree_id) ?:
  21                 bkey_cmp(ck->key.pos, key->pos);
  22 }
  23
  24 static const struct rhashtable_params bch2_btree_key_cache_params = {
  25         .head_offset    = offsetof(struct bkey_cached, hash),
  26         .key_offset     = offsetof(struct bkey_cached, key),
  27         .key_len        = sizeof(struct bkey_cached_key),
  28         .obj_cmpfn      = bch2_btree_key_cache_cmp_fn,
  29 };
  30
  31 __flatten
  32 static inline struct bkey_cached *
  33 btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
  34 {
  35         struct bkey_cached_key key = {
  36                 .btree_id       = btree_id,
  37                 .pos            = pos,
  38         };
  39
  40         return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
  41                                       bch2_btree_key_cache_params);
  42 }
  43
  44 static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
  45 {
  46         if (!six_trylock_intent(&ck->c.lock))
  47                 return false;
  48
  49         if (!six_trylock_write(&ck->c.lock)) {
  50                 six_unlock_intent(&ck->c.lock);
  51                 return false;
  52         }
  53
  54         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
  55                 six_unlock_write(&ck->c.lock);
  56                 six_unlock_intent(&ck->c.lock);
  57                 return false;
  58         }
  59
  60         return true;
  61 }
  62
  63 static void bkey_cached_evict(struct btree_key_cache *c,
  64                               struct bkey_cached *ck)
  65 {
  66         BUG_ON(rhashtable_remove_fast(&c->table, &ck->hash,
  67                                       bch2_btree_key_cache_params));
  68         memset(&ck->key, ~0, sizeof(ck->key));
  69 }
  70
  71 static void bkey_cached_free(struct btree_key_cache *c,
  72                              struct bkey_cached *ck)
  73 {
  74         list_move(&ck->list, &c->freed);
  75
  76         kfree(ck->k);
  77         ck->k           = NULL;
  78         ck->u64s        = 0;
  79
  80         six_unlock_write(&ck->c.lock);
  81         six_unlock_intent(&ck->c.lock);
  82 }
  83
  84 static struct bkey_cached *
  85 bkey_cached_alloc(struct btree_key_cache *c)
  86 {
  87         struct bkey_cached *ck;
  88
  89         list_for_each_entry(ck, &c->freed, list)
  90                 if (bkey_cached_lock_for_evict(ck))
  91                         return ck;
  92
  93         list_for_each_entry(ck, &c->clean, list)
  94                 if (bkey_cached_lock_for_evict(ck)) {
  95                         bkey_cached_evict(c, ck);
  96                         return ck;
  97                 }
  98
  99         ck = kzalloc(sizeof(*ck), GFP_NOFS);
 100         if (!ck)
 101                 return NULL;
 102
 103         INIT_LIST_HEAD(&ck->list);
 104         six_lock_init(&ck->c.lock);
 105         BUG_ON(!six_trylock_intent(&ck->c.lock));
 106         BUG_ON(!six_trylock_write(&ck->c.lock));
 107
 108         return ck;
 109 }
 110
 111 static struct bkey_cached *
 112 btree_key_cache_create(struct btree_key_cache *c,
 113                        enum btree_id btree_id,
 114                        struct bpos pos)
 115 {
 116         struct bkey_cached *ck;
 117
 118         ck = bkey_cached_alloc(c);
 119         if (!ck)
 120                 return ERR_PTR(-ENOMEM);
 121
 122         ck->c.level             = 0;
 123         ck->c.btree_id          = btree_id;
 124         ck->key.btree_id        = btree_id;
 125         ck->key.pos             = pos;
 126         ck->valid               = false;
 127
 128         BUG_ON(ck->flags);
 129
 130         if (rhashtable_lookup_insert_fast(&c->table,
 131                                           &ck->hash,
 132                                           bch2_btree_key_cache_params)) {
 133                 /* We raced with another fill: */
 134                 bkey_cached_free(c, ck);
 135                 return NULL;
 136         }
 137
 138         list_move(&ck->list, &c->clean);
 139         six_unlock_write(&ck->c.lock);
 140
 141         return ck;
 142 }
 143
 144 static int btree_key_cache_fill(struct btree_trans *trans,
 145                                 struct btree_iter *ck_iter,
 146                                 struct bkey_cached *ck)
 147 {
 148         struct btree_iter *iter;
 149         struct bkey_s_c k;
 150         unsigned new_u64s = 0;
 151         struct bkey_i *new_k = NULL;
 152         int ret;
 153
 154         iter = bch2_trans_get_iter(trans, ck->key.btree_id,
 155                                    ck->key.pos, BTREE_ITER_SLOTS);
 156         if (IS_ERR(iter))
 157                 return PTR_ERR(iter);
 158
 159         k = bch2_btree_iter_peek_slot(iter);
 160         ret = bkey_err(k);
 161         if (ret) {
 162                 bch2_trans_iter_put(trans, iter);
 163                 return ret;
 164         }
 165
 166         if (!bch2_btree_node_relock(ck_iter, 0)) {
 167                 bch2_trans_iter_put(trans, iter);
 168                 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 169                 return -EINTR;
 170         }
 171
 172         if (k.k->u64s > ck->u64s) {
 173                 new_u64s = roundup_pow_of_two(k.k->u64s);
 174                 new_k = kmalloc(new_u64s * sizeof(u64), GFP_NOFS);
 175                 if (!new_k) {
 176                         bch2_trans_iter_put(trans, iter);
 177                         return -ENOMEM;
 178                 }
 179         }
 180
 181         bch2_btree_node_lock_write(ck_iter->l[0].b, ck_iter);
 182         if (new_k) {
 183                 kfree(ck->k);
 184                 ck->u64s = new_u64s;
 185                 ck->k = new_k;
 186         }
 187
 188         bkey_reassemble(ck->k, k);
 189         ck->valid = true;
 190         bch2_btree_node_unlock_write(ck_iter->l[0].b, ck_iter);
 191
 192         /* We're not likely to need this iterator again: */
 193         bch2_trans_iter_free(trans, iter);
 194
 195         return 0;
 196 }
 197
 198 static int bkey_cached_check_fn(struct six_lock *lock, void *p)
 199 {
 200         struct bkey_cached *ck = container_of(lock, struct bkey_cached, c.lock);
 201         const struct btree_iter *iter = p;
 202
 203         return ck->key.btree_id == iter->btree_id &&
 204                 !bkey_cmp(ck->key.pos, iter->pos) ? 0 : -1;
 205 }
 206
 207 int bch2_btree_iter_traverse_cached(struct btree_iter *iter)
 208 {
 209         struct btree_trans *trans = iter->trans;
 210         struct bch_fs *c = trans->c;
 211         struct bkey_cached *ck;
 212         int ret = 0;
 213
 214         BUG_ON(iter->level);
 215
 216         if (btree_node_locked(iter, 0)) {
 217                 ck = (void *) iter->l[0].b;
 218                 goto fill;
 219         }
 220 retry:
 221         ck = btree_key_cache_find(c, iter->btree_id, iter->pos);
 222         if (!ck) {
 223                 if (iter->flags & BTREE_ITER_CACHED_NOCREATE) {
 224                         iter->l[0].b = NULL;
 225                         return 0;
 226                 }
 227
 228                 mutex_lock(&c->btree_key_cache.lock);
 229                 ck = btree_key_cache_create(&c->btree_key_cache,
 230                                             iter->btree_id, iter->pos);
 231                 mutex_unlock(&c->btree_key_cache.lock);
 232
 233                 ret = PTR_ERR_OR_ZERO(ck);
 234                 if (ret)
 235                         goto err;
 236                 if (!ck)
 237                         goto retry;
 238
 239                 mark_btree_node_locked(iter, 0, SIX_LOCK_intent);
 240                 iter->locks_want = 1;
 241         } else {
 242                 enum six_lock_type lock_want = __btree_lock_want(iter, 0);
 243
 244                 if (!btree_node_lock((void *) ck, iter->pos, 0, iter, lock_want,
 245                                      bkey_cached_check_fn, iter)) {
 246                         if (ck->key.btree_id != iter->btree_id ||
 247                             bkey_cmp(ck->key.pos, iter->pos)) {
 248                                 goto retry;
 249                         }
 250
 251                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 252                         ret = -EINTR;
 253                         goto err;
 254                 }
 255
 256                 if (ck->key.btree_id != iter->btree_id ||
 257                     bkey_cmp(ck->key.pos, iter->pos)) {
 258                         six_unlock_type(&ck->c.lock, lock_want);
 259                         goto retry;
 260                 }
 261
 262                 mark_btree_node_locked(iter, 0, lock_want);
 263         }
 264
 265         iter->l[0].lock_seq     = ck->c.lock.state.seq;
 266         iter->l[0].b            = (void *) ck;
 267 fill:
 268         if (!ck->valid && !(iter->flags & BTREE_ITER_CACHED_NOFILL)) {
 269                 if (!btree_node_intent_locked(iter, 0))
 270                         bch2_btree_iter_upgrade(iter, 1);
 271                 if (!btree_node_intent_locked(iter, 0)) {
 272                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
 273                         ret = -EINTR;
 274                         goto err;
 275                 }
 276
 277                 ret = btree_key_cache_fill(trans, iter, ck);
 278                 if (ret)
 279                         goto err;
 280         }
 281
 282         iter->uptodate = BTREE_ITER_NEED_PEEK;
 283         bch2_btree_iter_downgrade(iter);
 284         return ret;
 285 err:
 286         if (ret != -EINTR) {
 287                 btree_node_unlock(iter, 0);
 288                 iter->flags |= BTREE_ITER_ERROR;
 289                 iter->l[0].b = BTREE_ITER_NO_NODE_ERROR;
 290         }
 291         return ret;
 292 }
 293
 294 static int btree_key_cache_flush_pos(struct btree_trans *trans,
 295                                      struct bkey_cached_key key,
 296                                      u64 journal_seq,
 297                                      bool evict)
 298 {
 299         struct bch_fs *c = trans->c;
 300         struct journal *j = &c->journal;
 301         struct btree_iter *c_iter = NULL, *b_iter = NULL;
 302         struct bkey_cached *ck;
 303         int ret;
 304
 305         b_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 306                                      BTREE_ITER_SLOTS|
 307                                      BTREE_ITER_INTENT);
 308         ret = PTR_ERR_OR_ZERO(b_iter);
 309         if (ret)
 310                 goto out;
 311
 312         c_iter = bch2_trans_get_iter(trans, key.btree_id, key.pos,
 313                                      BTREE_ITER_CACHED|
 314                                      BTREE_ITER_CACHED_NOFILL|
 315                                      BTREE_ITER_CACHED_NOCREATE|
 316                                      BTREE_ITER_INTENT);
 317         ret = PTR_ERR_OR_ZERO(c_iter);
 318         if (ret)
 319                 goto out;
 320 retry:
 321         ret = bch2_btree_iter_traverse(c_iter);
 322         if (ret)
 323                 goto err;
 324
 325         ck = (void *) c_iter->l[0].b;
 326         if (!ck ||
 327             (journal_seq && ck->journal.seq != journal_seq))
 328                 goto out;
 329
 330         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 331                 if (!evict)
 332                         goto out;
 333                 goto evict;
 334         }
 335
 336         ret   = bch2_btree_iter_traverse(b_iter) ?:
 337                 bch2_trans_update(trans, b_iter, ck->k, BTREE_TRIGGER_NORUN) ?:
 338                 bch2_trans_commit(trans, NULL, NULL,
 339                                   BTREE_INSERT_NOUNLOCK|
 340                                   BTREE_INSERT_NOCHECK_RW|
 341                                   BTREE_INSERT_NOFAIL|
 342                                   BTREE_INSERT_USE_RESERVE|
 343                                   BTREE_INSERT_USE_ALLOC_RESERVE|
 344                                   BTREE_INSERT_JOURNAL_RESERVED|
 345                                   BTREE_INSERT_JOURNAL_RECLAIM);
 346 err:
 347         if (ret == -EINTR)
 348                 goto retry;
 349
 350         BUG_ON(ret && !bch2_journal_error(j));
 351
 352         if (ret)
 353                 goto out;
 354
 355         bch2_journal_pin_drop(j, &ck->journal);
 356         bch2_journal_preres_put(j, &ck->res);
 357         clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
 358
 359         if (!evict) {
 360                 mutex_lock(&c->btree_key_cache.lock);
 361                 list_move_tail(&ck->list, &c->btree_key_cache.clean);
 362                 mutex_unlock(&c->btree_key_cache.lock);
 363         } else {
 364 evict:
 365                 BUG_ON(!btree_node_intent_locked(c_iter, 0));
 366
 367                 mark_btree_node_unlocked(c_iter, 0);
 368                 c_iter->l[0].b = NULL;
 369
 370                 six_lock_write(&ck->c.lock, NULL, NULL);
 371
 372                 mutex_lock(&c->btree_key_cache.lock);
 373                 bkey_cached_evict(&c->btree_key_cache, ck);
 374                 bkey_cached_free(&c->btree_key_cache, ck);
 375                 mutex_unlock(&c->btree_key_cache.lock);
 376         }
 377 out:
 378         bch2_trans_iter_put(trans, b_iter);
 379         bch2_trans_iter_put(trans, c_iter);
 380         return ret;
 381 }
 382
 383 static void btree_key_cache_journal_flush(struct journal *j,
 384                                           struct journal_entry_pin *pin,
 385                                           u64 seq)
 386 {
 387         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 388         struct bkey_cached *ck =
 389                 container_of(pin, struct bkey_cached, journal);
 390         struct bkey_cached_key key;
 391         struct btree_trans trans;
 392
 393         six_lock_read(&ck->c.lock, NULL, NULL);
 394         key = ck->key;
 395
 396         if (ck->journal.seq != seq ||
 397             !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 398                 six_unlock_read(&ck->c.lock);
 399                 return;
 400         }
 401         six_unlock_read(&ck->c.lock);
 402
 403         bch2_trans_init(&trans, c, 0, 0);
 404         btree_key_cache_flush_pos(&trans, key, seq, false);
 405         bch2_trans_exit(&trans);
 406 }
 407
 408 /*
 409  * Flush and evict a key from the key cache:
 410  */
 411 int bch2_btree_key_cache_flush(struct btree_trans *trans,
 412                                enum btree_id id, struct bpos pos)
 413 {
 414         struct bch_fs *c = trans->c;
 415         struct bkey_cached_key key = { id, pos };
 416
 417         /* Fastpath - assume it won't be found: */
 418         if (!btree_key_cache_find(c, id, pos))
 419                 return 0;
 420
 421         return btree_key_cache_flush_pos(trans, key, 0, true);
 422 }
 423
 424 bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 425                                   struct btree_iter *iter,
 426                                   struct bkey_i *insert)
 427 {
 428         struct bch_fs *c = trans->c;
 429         struct bkey_cached *ck = (void *) iter->l[0].b;
 430
 431         BUG_ON(insert->u64s > ck->u64s);
 432
 433         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 434                 int difference;
 435
 436                 BUG_ON(jset_u64s(insert->u64s) > trans->journal_preres.u64s);
 437
 438                 difference = jset_u64s(insert->u64s) - ck->res.u64s;
 439                 if (difference > 0) {
 440                         trans->journal_preres.u64s      -= difference;
 441                         ck->res.u64s                    += difference;
 442                 }
 443         }
 444
 445         bkey_copy(ck->k, insert);
 446         ck->valid = true;
 447
 448         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 449                 mutex_lock(&c->btree_key_cache.lock);
 450                 list_del_init(&ck->list);
 451
 452                 set_bit(BKEY_CACHED_DIRTY, &ck->flags);
 453                 mutex_unlock(&c->btree_key_cache.lock);
 454         }
 455
 456         bch2_journal_pin_update(&c->journal, trans->journal_res.seq,
 457                                 &ck->journal, btree_key_cache_journal_flush);
 458         return true;
 459 }
 460
 461 #ifdef CONFIG_BCACHEFS_DEBUG
 462 void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
 463                                enum btree_id id, struct bpos pos)
 464 {
 465         BUG_ON(btree_key_cache_find(trans->c, id, pos));
 466 }
 467 #endif
 468
 469 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *c)
 470 {
 471         struct bkey_cached *ck, *n;
 472
 473         mutex_lock(&c->lock);
 474         list_for_each_entry_safe(ck, n, &c->clean, list) {
 475                 kfree(ck->k);
 476                 kfree(ck);
 477         }
 478         list_for_each_entry_safe(ck, n, &c->freed, list)
 479                 kfree(ck);
 480         mutex_unlock(&c->lock);
 481
 482         rhashtable_destroy(&c->table);
 483 }
 484
 485 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
 486 {
 487         mutex_init(&c->lock);
 488         INIT_LIST_HEAD(&c->freed);
 489         INIT_LIST_HEAD(&c->clean);
 490 }
 491
 492 int bch2_fs_btree_key_cache_init(struct btree_key_cache *c)
 493 {
 494         return rhashtable_init(&c->table, &bch2_btree_key_cache_params);
 495 }
 496
 497 void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *c)
 498 {
 499         struct bucket_table *tbl;
 500         struct bkey_cached *ck;
 501         struct rhash_head *pos;
 502         size_t i;
 503
 504         mutex_lock(&c->lock);
 505         tbl = rht_dereference_rcu(c->table.tbl, &c->table);
 506
 507         for (i = 0; i < tbl->size; i++) {
 508                 rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
 509                         pr_buf(out, "%s:",
 510                                bch2_btree_ids[ck->key.btree_id]);
 511                         bch2_bpos_to_text(out, ck->key.pos);
 512
 513                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags))
 514                                 pr_buf(out, " journal seq %llu", ck->journal.seq);
 515                         pr_buf(out, "\n");
 516                 }
 517         }
 518         mutex_unlock(&c->lock);
 519 }