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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_buf.h"
   6 #include "btree_cache.h"
   7 #include "btree_iter.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "debug.h"
  12 #include "error.h"
  13 #include "extents.h"
  14 #include "journal.h"
  15 #include "recovery.h"
  16 #include "replicas.h"
  17 #include "subvolume.h"
  18
  19 #include <linux/prefetch.h>
  20 #include <trace/events/bcachefs.h>
  21
  22 static void btree_trans_verify_sorted(struct btree_trans *);
  23 static void btree_path_check_sort(struct btree_trans *, struct btree_path *, int);
  24
  25 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
  26 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
  27                                        struct btree_path *);
  28
  29 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
  30 {
  31 #ifdef CONFIG_BCACHEFS_DEBUG
  32         return iter->ip_allocated;
  33 #else
  34         return 0;
  35 #endif
  36 }
  37
  38 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
  39
  40 /*
  41  * Unlocks before scheduling
  42  * Note: does not revalidate iterator
  43  */
  44 static inline int bch2_trans_cond_resched(struct btree_trans *trans)
  45 {
  46         if (need_resched() || race_fault()) {
  47                 bch2_trans_unlock(trans);
  48                 schedule();
  49                 return bch2_trans_relock(trans) ? 0 : -EINTR;
  50         } else {
  51                 return 0;
  52         }
  53 }
  54
  55 static inline int __btree_path_cmp(const struct btree_path *l,
  56                                    enum btree_id        r_btree_id,
  57                                    bool                 r_cached,
  58                                    struct bpos          r_pos,
  59                                    unsigned             r_level)
  60 {
  61         return   cmp_int(l->btree_id,   r_btree_id) ?:
  62                  cmp_int((int) l->cached,       (int) r_cached) ?:
  63                  bpos_cmp(l->pos,       r_pos) ?:
  64                 -cmp_int(l->level,      r_level);
  65 }
  66
  67 static inline int btree_path_cmp(const struct btree_path *l,
  68                                  const struct btree_path *r)
  69 {
  70         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  71 }
  72
  73 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  74 {
  75         /* Are we iterating over keys in all snapshots? */
  76         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  77                 p = bpos_successor(p);
  78         } else {
  79                 p = bpos_nosnap_successor(p);
  80                 p.snapshot = iter->snapshot;
  81         }
  82
  83         return p;
  84 }
  85
  86 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  87 {
  88         /* Are we iterating over keys in all snapshots? */
  89         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  90                 p = bpos_predecessor(p);
  91         } else {
  92                 p = bpos_nosnap_predecessor(p);
  93                 p.snapshot = iter->snapshot;
  94         }
  95
  96         return p;
  97 }
  98
  99 static inline bool is_btree_node(struct btree_path *path, unsigned l)
 100 {
 101         return l < BTREE_MAX_DEPTH &&
 102                 (unsigned long) path->l[l].b >= 128;
 103 }
 104
 105 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
 106 {
 107         struct bpos pos = iter->pos;
 108
 109         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 110             bkey_cmp(pos, POS_MAX))
 111                 pos = bkey_successor(iter, pos);
 112         return pos;
 113 }
 114
 115 static inline bool btree_path_pos_before_node(struct btree_path *path,
 116                                               struct btree *b)
 117 {
 118         return bpos_cmp(path->pos, b->data->min_key) < 0;
 119 }
 120
 121 static inline bool btree_path_pos_after_node(struct btree_path *path,
 122                                              struct btree *b)
 123 {
 124         return bpos_cmp(b->key.k.p, path->pos) < 0;
 125 }
 126
 127 static inline bool btree_path_pos_in_node(struct btree_path *path,
 128                                           struct btree *b)
 129 {
 130         return path->btree_id == b->c.btree_id &&
 131                 !btree_path_pos_before_node(path, b) &&
 132                 !btree_path_pos_after_node(path, b);
 133 }
 134
 135 /* Btree node locking: */
 136
 137 void bch2_btree_node_unlock_write(struct btree_trans *trans,
 138                         struct btree_path *path, struct btree *b)
 139 {
 140         bch2_btree_node_unlock_write_inlined(trans, path, b);
 141 }
 142
 143 void __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree *b)
 144 {
 145         struct btree_path *linked;
 146         unsigned readers = 0;
 147
 148         trans_for_each_path(trans, linked)
 149                 if (linked->l[b->c.level].b == b &&
 150                     btree_node_read_locked(linked, b->c.level))
 151                         readers++;
 152
 153         /*
 154          * Must drop our read locks before calling six_lock_write() -
 155          * six_unlock() won't do wakeups until the reader count
 156          * goes to 0, and it's safe because we have the node intent
 157          * locked:
 158          */
 159         if (!b->c.lock.readers)
 160                 atomic64_sub(__SIX_VAL(read_lock, readers),
 161                              &b->c.lock.state.counter);
 162         else
 163                 this_cpu_sub(*b->c.lock.readers, readers);
 164
 165         btree_node_lock_type(trans->c, b, SIX_LOCK_write);
 166
 167         if (!b->c.lock.readers)
 168                 atomic64_add(__SIX_VAL(read_lock, readers),
 169                              &b->c.lock.state.counter);
 170         else
 171                 this_cpu_add(*b->c.lock.readers, readers);
 172 }
 173
 174 bool __bch2_btree_node_relock(struct btree_trans *trans,
 175                               struct btree_path *path, unsigned level)
 176 {
 177         struct btree *b = btree_path_node(path, level);
 178         int want = __btree_lock_want(path, level);
 179
 180         if (!is_btree_node(path, level))
 181                 goto fail;
 182
 183         if (race_fault())
 184                 goto fail;
 185
 186         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 187             (btree_node_lock_seq_matches(path, b, level) &&
 188              btree_node_lock_increment(trans, b, level, want))) {
 189                 mark_btree_node_locked(path, level, want);
 190                 return true;
 191         }
 192 fail:
 193         trace_btree_node_relock_fail(trans->fn, _RET_IP_,
 194                                      path->btree_id,
 195                                      &path->pos,
 196                                      (unsigned long) b,
 197                                      path->l[level].lock_seq,
 198                                      is_btree_node(path, level) ? b->c.lock.state.seq : 0);
 199         return false;
 200 }
 201
 202 bool bch2_btree_node_upgrade(struct btree_trans *trans,
 203                              struct btree_path *path, unsigned level)
 204 {
 205         struct btree *b = path->l[level].b;
 206
 207         if (!is_btree_node(path, level))
 208                 return false;
 209
 210         switch (btree_lock_want(path, level)) {
 211         case BTREE_NODE_UNLOCKED:
 212                 BUG_ON(btree_node_locked(path, level));
 213                 return true;
 214         case BTREE_NODE_READ_LOCKED:
 215                 BUG_ON(btree_node_intent_locked(path, level));
 216                 return bch2_btree_node_relock(trans, path, level);
 217         case BTREE_NODE_INTENT_LOCKED:
 218                 break;
 219         }
 220
 221         if (btree_node_intent_locked(path, level))
 222                 return true;
 223
 224         if (race_fault())
 225                 return false;
 226
 227         if (btree_node_locked(path, level)
 228             ? six_lock_tryupgrade(&b->c.lock)
 229             : six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 230                 goto success;
 231
 232         if (btree_node_lock_seq_matches(path, b, level) &&
 233             btree_node_lock_increment(trans, b, level, BTREE_NODE_INTENT_LOCKED)) {
 234                 btree_node_unlock(path, level);
 235                 goto success;
 236         }
 237
 238         return false;
 239 success:
 240         mark_btree_node_intent_locked(path, level);
 241         return true;
 242 }
 243
 244 static inline bool btree_path_get_locks(struct btree_trans *trans,
 245                                         struct btree_path *path,
 246                                         bool upgrade)
 247 {
 248         unsigned l = path->level;
 249         int fail_idx = -1;
 250
 251         do {
 252                 if (!btree_path_node(path, l))
 253                         break;
 254
 255                 if (!(upgrade
 256                       ? bch2_btree_node_upgrade(trans, path, l)
 257                       : bch2_btree_node_relock(trans, path, l)))
 258                         fail_idx = l;
 259
 260                 l++;
 261         } while (l < path->locks_want);
 262
 263         /*
 264          * When we fail to get a lock, we have to ensure that any child nodes
 265          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 266          * the node that we failed to relock:
 267          */
 268         if (fail_idx >= 0) {
 269                 __bch2_btree_path_unlock(path);
 270                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 271
 272                 do {
 273                         path->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS;
 274                         --fail_idx;
 275                 } while (fail_idx >= 0);
 276         }
 277
 278         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 279                 path->uptodate = BTREE_ITER_UPTODATE;
 280
 281         bch2_trans_verify_locks(trans);
 282
 283         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 284 }
 285
 286 static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b,
 287                                   bool cached)
 288 {
 289         return !cached
 290                 ? container_of(_b, struct btree, c)->key.k.p
 291                 : container_of(_b, struct bkey_cached, c)->key.pos;
 292 }
 293
 294 /* Slowpath: */
 295 bool __bch2_btree_node_lock(struct btree_trans *trans,
 296                             struct btree_path *path,
 297                             struct btree *b,
 298                             struct bpos pos, unsigned level,
 299                             enum six_lock_type type,
 300                             six_lock_should_sleep_fn should_sleep_fn, void *p,
 301                             unsigned long ip)
 302 {
 303         struct btree_path *linked, *deadlock_path = NULL;
 304         u64 start_time = local_clock();
 305         unsigned reason = 9;
 306         bool ret;
 307
 308         /* Check if it's safe to block: */
 309         trans_for_each_path(trans, linked) {
 310                 if (!linked->nodes_locked)
 311                         continue;
 312
 313                 /*
 314                  * Can't block taking an intent lock if we have _any_ nodes read
 315                  * locked:
 316                  *
 317                  * - Our read lock blocks another thread with an intent lock on
 318                  *   the same node from getting a write lock, and thus from
 319                  *   dropping its intent lock
 320                  *
 321                  * - And the other thread may have multiple nodes intent locked:
 322                  *   both the node we want to intent lock, and the node we
 323                  *   already have read locked - deadlock:
 324                  */
 325                 if (type == SIX_LOCK_intent &&
 326                     linked->nodes_locked != linked->nodes_intent_locked) {
 327                         deadlock_path = linked;
 328                         reason = 1;
 329                 }
 330
 331                 if (linked->btree_id != path->btree_id) {
 332                         if (linked->btree_id > path->btree_id) {
 333                                 deadlock_path = linked;
 334                                 reason = 3;
 335                         }
 336                         continue;
 337                 }
 338
 339                 /*
 340                  * Within the same btree, cached paths come before non
 341                  * cached paths:
 342                  */
 343                 if (linked->cached != path->cached) {
 344                         if (path->cached) {
 345                                 deadlock_path = linked;
 346                                 reason = 4;
 347                         }
 348                         continue;
 349                 }
 350
 351                 /*
 352                  * Interior nodes must be locked before their descendants: if
 353                  * another path has possible descendants locked of the node
 354                  * we're about to lock, it must have the ancestors locked too:
 355                  */
 356                 if (level > __fls(linked->nodes_locked)) {
 357                         deadlock_path = linked;
 358                         reason = 5;
 359                 }
 360
 361                 /* Must lock btree nodes in key order: */
 362                 if (btree_node_locked(linked, level) &&
 363                     bpos_cmp(pos, btree_node_pos((void *) linked->l[level].b,
 364                                                  linked->cached)) <= 0) {
 365                         deadlock_path = linked;
 366                         reason = 7;
 367                         BUG_ON(trans->in_traverse_all);
 368                 }
 369         }
 370
 371         if (unlikely(deadlock_path)) {
 372                 trace_trans_restart_would_deadlock(trans->fn, ip,
 373                                 trans->in_traverse_all, reason,
 374                                 deadlock_path->btree_id,
 375                                 deadlock_path->cached,
 376                                 &deadlock_path->pos,
 377                                 path->btree_id,
 378                                 path->cached,
 379                                 &pos);
 380                 btree_trans_restart(trans);
 381                 return false;
 382         }
 383
 384         if (six_trylock_type(&b->c.lock, type))
 385                 return true;
 386
 387         trans->locking_path_idx = path->idx;
 388         trans->locking_pos      = pos;
 389         trans->locking_btree_id = path->btree_id;
 390         trans->locking_level    = level;
 391         trans->locking          = b;
 392
 393         ret = six_lock_type(&b->c.lock, type, should_sleep_fn, p) == 0;
 394
 395         trans->locking = NULL;
 396
 397         if (ret)
 398                 bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)],
 399                                        start_time);
 400         return ret;
 401 }
 402
 403 /* Btree iterator locking: */
 404
 405 #ifdef CONFIG_BCACHEFS_DEBUG
 406
 407 static void bch2_btree_path_verify_locks(struct btree_path *path)
 408 {
 409         unsigned l;
 410
 411         if (!path->nodes_locked) {
 412                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
 413                        btree_path_node(path, path->level));
 414                 return;
 415         }
 416
 417         for (l = 0; btree_path_node(path, l); l++)
 418                 BUG_ON(btree_lock_want(path, l) !=
 419                        btree_node_locked_type(path, l));
 420 }
 421
 422 void bch2_trans_verify_locks(struct btree_trans *trans)
 423 {
 424         struct btree_path *path;
 425
 426         trans_for_each_path(trans, path)
 427                 bch2_btree_path_verify_locks(path);
 428 }
 429 #else
 430 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
 431 #endif
 432
 433 /* Btree path locking: */
 434
 435 /*
 436  * Only for btree_cache.c - only relocks intent locks
 437  */
 438 bool bch2_btree_path_relock_intent(struct btree_trans *trans,
 439                                    struct btree_path *path)
 440 {
 441         unsigned l;
 442
 443         for (l = path->level;
 444              l < path->locks_want && btree_path_node(path, l);
 445              l++) {
 446                 if (!bch2_btree_node_relock(trans, path, l)) {
 447                         __bch2_btree_path_unlock(path);
 448                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 449                         trace_trans_restart_relock_path_intent(trans->fn, _RET_IP_,
 450                                                    path->btree_id, &path->pos);
 451                         btree_trans_restart(trans);
 452                         return false;
 453                 }
 454         }
 455
 456         return true;
 457 }
 458
 459 __flatten
 460 static bool bch2_btree_path_relock(struct btree_trans *trans,
 461                         struct btree_path *path, unsigned long trace_ip)
 462 {
 463         bool ret = btree_path_get_locks(trans, path, false);
 464
 465         if (!ret) {
 466                 trace_trans_restart_relock_path(trans->fn, trace_ip,
 467                                                 path->btree_id, &path->pos);
 468                 btree_trans_restart(trans);
 469         }
 470         return ret;
 471 }
 472
 473 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 474                                struct btree_path *path,
 475                                unsigned new_locks_want)
 476 {
 477         struct btree_path *linked;
 478
 479         EBUG_ON(path->locks_want >= new_locks_want);
 480
 481         path->locks_want = new_locks_want;
 482
 483         if (btree_path_get_locks(trans, path, true))
 484                 return true;
 485
 486         /*
 487          * XXX: this is ugly - we'd prefer to not be mucking with other
 488          * iterators in the btree_trans here.
 489          *
 490          * On failure to upgrade the iterator, setting iter->locks_want and
 491          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 492          * get the locks we want on transaction restart.
 493          *
 494          * But if this iterator was a clone, on transaction restart what we did
 495          * to this iterator isn't going to be preserved.
 496          *
 497          * Possibly we could add an iterator field for the parent iterator when
 498          * an iterator is a copy - for now, we'll just upgrade any other
 499          * iterators with the same btree id.
 500          *
 501          * The code below used to be needed to ensure ancestor nodes get locked
 502          * before interior nodes - now that's handled by
 503          * bch2_btree_path_traverse_all().
 504          */
 505         trans_for_each_path(trans, linked)
 506                 if (linked != path &&
 507                     linked->cached == path->cached &&
 508                     linked->btree_id == path->btree_id &&
 509                     linked->locks_want < new_locks_want) {
 510                         linked->locks_want = new_locks_want;
 511                         btree_path_get_locks(trans, linked, true);
 512                 }
 513
 514         return false;
 515 }
 516
 517 void __bch2_btree_path_downgrade(struct btree_path *path,
 518                                  unsigned new_locks_want)
 519 {
 520         unsigned l;
 521
 522         EBUG_ON(path->locks_want < new_locks_want);
 523
 524         path->locks_want = new_locks_want;
 525
 526         while (path->nodes_locked &&
 527                (l = __fls(path->nodes_locked)) >= path->locks_want) {
 528                 if (l > path->level) {
 529                         btree_node_unlock(path, l);
 530                 } else {
 531                         if (btree_node_intent_locked(path, l)) {
 532                                 six_lock_downgrade(&path->l[l].b->c.lock);
 533                                 path->nodes_intent_locked ^= 1 << l;
 534                         }
 535                         break;
 536                 }
 537         }
 538
 539         bch2_btree_path_verify_locks(path);
 540 }
 541
 542 void bch2_trans_downgrade(struct btree_trans *trans)
 543 {
 544         struct btree_path *path;
 545
 546         trans_for_each_path(trans, path)
 547                 bch2_btree_path_downgrade(path);
 548 }
 549
 550 /* Btree transaction locking: */
 551
 552 bool bch2_trans_relock(struct btree_trans *trans)
 553 {
 554         struct btree_path *path;
 555
 556         if (unlikely(trans->restarted))
 557                 return false;
 558
 559         trans_for_each_path(trans, path)
 560                 if (path->should_be_locked &&
 561                     !bch2_btree_path_relock(trans, path, _RET_IP_)) {
 562                         trace_trans_restart_relock(trans->fn, _RET_IP_,
 563                                         path->btree_id, &path->pos);
 564                         BUG_ON(!trans->restarted);
 565                         return false;
 566                 }
 567         return true;
 568 }
 569
 570 void bch2_trans_unlock(struct btree_trans *trans)
 571 {
 572         struct btree_path *path;
 573
 574         trans_for_each_path(trans, path)
 575                 __bch2_btree_path_unlock(path);
 576
 577         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
 578 }
 579
 580 /* Btree iterator: */
 581
 582 #ifdef CONFIG_BCACHEFS_DEBUG
 583
 584 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 585                                           struct btree_path *path)
 586 {
 587         struct bkey_cached *ck;
 588         bool locked = btree_node_locked(path, 0);
 589
 590         if (!bch2_btree_node_relock(trans, path, 0))
 591                 return;
 592
 593         ck = (void *) path->l[0].b;
 594         BUG_ON(ck->key.btree_id != path->btree_id ||
 595                bkey_cmp(ck->key.pos, path->pos));
 596
 597         if (!locked)
 598                 btree_node_unlock(path, 0);
 599 }
 600
 601 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 602                                 struct btree_path *path, unsigned level)
 603 {
 604         struct btree_path_level *l;
 605         struct btree_node_iter tmp;
 606         bool locked;
 607         struct bkey_packed *p, *k;
 608         char buf1[100], buf2[100], buf3[100];
 609         const char *msg;
 610
 611         if (!bch2_debug_check_iterators)
 612                 return;
 613
 614         l       = &path->l[level];
 615         tmp     = l->iter;
 616         locked  = btree_node_locked(path, level);
 617
 618         if (path->cached) {
 619                 if (!level)
 620                         bch2_btree_path_verify_cached(trans, path);
 621                 return;
 622         }
 623
 624         if (!btree_path_node(path, level))
 625                 return;
 626
 627         if (!bch2_btree_node_relock(trans, path, level))
 628                 return;
 629
 630         BUG_ON(!btree_path_pos_in_node(path, l->b));
 631
 632         bch2_btree_node_iter_verify(&l->iter, l->b);
 633
 634         /*
 635          * For interior nodes, the iterator will have skipped past deleted keys:
 636          */
 637         p = level
 638                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 639                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 640         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 641
 642         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 643                 msg = "before";
 644                 goto err;
 645         }
 646
 647         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 648                 msg = "after";
 649                 goto err;
 650         }
 651
 652         if (!locked)
 653                 btree_node_unlock(path, level);
 654         return;
 655 err:
 656         strcpy(buf2, "(none)");
 657         strcpy(buf3, "(none)");
 658
 659         bch2_bpos_to_text(&PBUF(buf1), path->pos);
 660
 661         if (p) {
 662                 struct bkey uk = bkey_unpack_key(l->b, p);
 663                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 664         }
 665
 666         if (k) {
 667                 struct bkey uk = bkey_unpack_key(l->b, k);
 668                 bch2_bkey_to_text(&PBUF(buf3), &uk);
 669         }
 670
 671         panic("path should be %s key at level %u:\n"
 672               "path pos %s\n"
 673               "prev key %s\n"
 674               "cur  key %s\n",
 675               msg, level, buf1, buf2, buf3);
 676 }
 677
 678 static void bch2_btree_path_verify(struct btree_trans *trans,
 679                                    struct btree_path *path)
 680 {
 681         struct bch_fs *c = trans->c;
 682         unsigned i;
 683
 684         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 685
 686         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 687                 if (!path->l[i].b) {
 688                         BUG_ON(!path->cached &&
 689                                c->btree_roots[path->btree_id].b->c.level > i);
 690                         break;
 691                 }
 692
 693                 bch2_btree_path_verify_level(trans, path, i);
 694         }
 695
 696         bch2_btree_path_verify_locks(path);
 697 }
 698
 699 void bch2_trans_verify_paths(struct btree_trans *trans)
 700 {
 701         struct btree_path *path;
 702
 703         trans_for_each_path(trans, path)
 704                 bch2_btree_path_verify(trans, path);
 705 }
 706
 707 static void bch2_btree_iter_verify(struct btree_iter *iter)
 708 {
 709         struct btree_trans *trans = iter->trans;
 710
 711         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 712
 713         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
 714
 715         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 716                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 717
 718         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 719                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 720                !btree_type_has_snapshots(iter->btree_id));
 721
 722         if (iter->update_path)
 723                 bch2_btree_path_verify(trans, iter->update_path);
 724         bch2_btree_path_verify(trans, iter->path);
 725 }
 726
 727 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 728 {
 729         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 730                !iter->pos.snapshot);
 731
 732         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 733                iter->pos.snapshot != iter->snapshot);
 734
 735         BUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
 736                bkey_cmp(iter->pos, iter->k.p) > 0);
 737 }
 738
 739 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 740 {
 741         struct btree_trans *trans = iter->trans;
 742         struct btree_iter copy;
 743         struct bkey_s_c prev;
 744         int ret = 0;
 745
 746         if (!bch2_debug_check_iterators)
 747                 return 0;
 748
 749         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 750                 return 0;
 751
 752         if (bkey_err(k) || !k.k)
 753                 return 0;
 754
 755         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 756                                           iter->snapshot,
 757                                           k.k->p.snapshot));
 758
 759         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 760                              BTREE_ITER_NOPRESERVE|
 761                              BTREE_ITER_ALL_SNAPSHOTS);
 762         prev = bch2_btree_iter_prev(&copy);
 763         if (!prev.k)
 764                 goto out;
 765
 766         ret = bkey_err(prev);
 767         if (ret)
 768                 goto out;
 769
 770         if (!bkey_cmp(prev.k->p, k.k->p) &&
 771             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 772                                       prev.k->p.snapshot) > 0) {
 773                 char buf1[100], buf2[200];
 774
 775                 bch2_bkey_to_text(&PBUF(buf1), k.k);
 776                 bch2_bkey_to_text(&PBUF(buf2), prev.k);
 777
 778                 panic("iter snap %u\n"
 779                       "k    %s\n"
 780                       "prev %s\n",
 781                       iter->snapshot,
 782                       buf1, buf2);
 783         }
 784 out:
 785         bch2_trans_iter_exit(trans, &copy);
 786         return ret;
 787 }
 788
 789 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 790                             struct bpos pos, bool key_cache)
 791 {
 792         struct btree_path *path;
 793         unsigned idx;
 794         char buf[100];
 795
 796         trans_for_each_path_inorder(trans, path, idx) {
 797                 int cmp = cmp_int(path->btree_id, id) ?:
 798                         cmp_int(path->cached, key_cache);
 799
 800                 if (cmp > 0)
 801                         break;
 802                 if (cmp < 0)
 803                         continue;
 804
 805                 if (!(path->nodes_locked & 1) ||
 806                     !path->should_be_locked)
 807                         continue;
 808
 809                 if (!key_cache) {
 810                         if (bkey_cmp(pos, path->l[0].b->data->min_key) >= 0 &&
 811                             bkey_cmp(pos, path->l[0].b->key.k.p) <= 0)
 812                                 return;
 813                 } else {
 814                         if (!bkey_cmp(pos, path->pos))
 815                                 return;
 816                 }
 817         }
 818
 819         bch2_dump_trans_paths_updates(trans);
 820         panic("not locked: %s %s%s\n",
 821               bch2_btree_ids[id],
 822               (bch2_bpos_to_text(&PBUF(buf), pos), buf),
 823               key_cache ? " cached" : "");
 824 }
 825
 826 #else
 827
 828 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 829                                                 struct btree_path *path, unsigned l) {}
 830 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 831                                           struct btree_path *path) {}
 832 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 833 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 834 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 835
 836 #endif
 837
 838 /* Btree path: fixups after btree updates */
 839
 840 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 841                                         struct btree *b,
 842                                         struct bset_tree *t,
 843                                         struct bkey_packed *k)
 844 {
 845         struct btree_node_iter_set *set;
 846
 847         btree_node_iter_for_each(iter, set)
 848                 if (set->end == t->end_offset) {
 849                         set->k = __btree_node_key_to_offset(b, k);
 850                         bch2_btree_node_iter_sort(iter, b);
 851                         return;
 852                 }
 853
 854         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 855 }
 856
 857 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 858                                                struct btree *b,
 859                                                struct bkey_packed *where)
 860 {
 861         struct btree_path_level *l = &path->l[b->c.level];
 862
 863         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 864                 return;
 865
 866         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 867                 bch2_btree_node_iter_advance(&l->iter, l->b);
 868 }
 869
 870 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 871                                       struct btree *b,
 872                                       struct bkey_packed *where)
 873 {
 874         struct btree_path *path;
 875
 876         trans_for_each_path_with_node(trans, b, path) {
 877                 __bch2_btree_path_fix_key_modified(path, b, where);
 878                 bch2_btree_path_verify_level(trans, path, b->c.level);
 879         }
 880 }
 881
 882 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 883                                        struct btree *b,
 884                                        struct btree_node_iter *node_iter,
 885                                        struct bset_tree *t,
 886                                        struct bkey_packed *where,
 887                                        unsigned clobber_u64s,
 888                                        unsigned new_u64s)
 889 {
 890         const struct bkey_packed *end = btree_bkey_last(b, t);
 891         struct btree_node_iter_set *set;
 892         unsigned offset = __btree_node_key_to_offset(b, where);
 893         int shift = new_u64s - clobber_u64s;
 894         unsigned old_end = t->end_offset - shift;
 895         unsigned orig_iter_pos = node_iter->data[0].k;
 896         bool iter_current_key_modified =
 897                 orig_iter_pos >= offset &&
 898                 orig_iter_pos <= offset + clobber_u64s;
 899
 900         btree_node_iter_for_each(node_iter, set)
 901                 if (set->end == old_end)
 902                         goto found;
 903
 904         /* didn't find the bset in the iterator - might have to readd it: */
 905         if (new_u64s &&
 906             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 907                 bch2_btree_node_iter_push(node_iter, b, where, end);
 908                 goto fixup_done;
 909         } else {
 910                 /* Iterator is after key that changed */
 911                 return;
 912         }
 913 found:
 914         set->end = t->end_offset;
 915
 916         /* Iterator hasn't gotten to the key that changed yet: */
 917         if (set->k < offset)
 918                 return;
 919
 920         if (new_u64s &&
 921             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 922                 set->k = offset;
 923         } else if (set->k < offset + clobber_u64s) {
 924                 set->k = offset + new_u64s;
 925                 if (set->k == set->end)
 926                         bch2_btree_node_iter_set_drop(node_iter, set);
 927         } else {
 928                 /* Iterator is after key that changed */
 929                 set->k = (int) set->k + shift;
 930                 return;
 931         }
 932
 933         bch2_btree_node_iter_sort(node_iter, b);
 934 fixup_done:
 935         if (node_iter->data[0].k != orig_iter_pos)
 936                 iter_current_key_modified = true;
 937
 938         /*
 939          * When a new key is added, and the node iterator now points to that
 940          * key, the iterator might have skipped past deleted keys that should
 941          * come after the key the iterator now points to. We have to rewind to
 942          * before those deleted keys - otherwise
 943          * bch2_btree_node_iter_prev_all() breaks:
 944          */
 945         if (!bch2_btree_node_iter_end(node_iter) &&
 946             iter_current_key_modified &&
 947             b->c.level) {
 948                 struct bset_tree *t;
 949                 struct bkey_packed *k, *k2, *p;
 950
 951                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 952
 953                 for_each_bset(b, t) {
 954                         bool set_pos = false;
 955
 956                         if (node_iter->data[0].end == t->end_offset)
 957                                 continue;
 958
 959                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 960
 961                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 962                                bkey_iter_cmp(b, k, p) < 0) {
 963                                 k2 = p;
 964                                 set_pos = true;
 965                         }
 966
 967                         if (set_pos)
 968                                 btree_node_iter_set_set_pos(node_iter,
 969                                                             b, t, k2);
 970                 }
 971         }
 972 }
 973
 974 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 975                               struct btree_path *path,
 976                               struct btree *b,
 977                               struct btree_node_iter *node_iter,
 978                               struct bkey_packed *where,
 979                               unsigned clobber_u64s,
 980                               unsigned new_u64s)
 981 {
 982         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 983         struct btree_path *linked;
 984
 985         if (node_iter != &path->l[b->c.level].iter) {
 986                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 987                                            where, clobber_u64s, new_u64s);
 988
 989                 if (bch2_debug_check_iterators)
 990                         bch2_btree_node_iter_verify(node_iter, b);
 991         }
 992
 993         trans_for_each_path_with_node(trans, b, linked) {
 994                 __bch2_btree_node_iter_fix(linked, b,
 995                                            &linked->l[b->c.level].iter, t,
 996                                            where, clobber_u64s, new_u64s);
 997                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 998         }
 999 }
1000
1001 /* Btree path level: pointer to a particular btree node and node iter */
1002
1003 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
1004                                                   struct btree_path_level *l,
1005                                                   struct bkey *u,
1006                                                   struct bkey_packed *k)
1007 {
1008         struct bkey_s_c ret;
1009
1010         if (unlikely(!k)) {
1011                 /*
1012                  * signal to bch2_btree_iter_peek_slot() that we're currently at
1013                  * a hole
1014                  */
1015                 u->type = KEY_TYPE_deleted;
1016                 return bkey_s_c_null;
1017         }
1018
1019         ret = bkey_disassemble(l->b, k, u);
1020
1021         /*
1022          * XXX: bch2_btree_bset_insert_key() generates invalid keys when we
1023          * overwrite extents - it sets k->type = KEY_TYPE_deleted on the key
1024          * being overwritten but doesn't change k->size. But this is ok, because
1025          * those keys are never written out, we just have to avoid a spurious
1026          * assertion here:
1027          */
1028         if (bch2_debug_check_bkeys && !bkey_deleted(ret.k))
1029                 bch2_bkey_debugcheck(c, l->b, ret);
1030
1031         return ret;
1032 }
1033
1034 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
1035                                                         struct btree_path_level *l,
1036                                                         struct bkey *u)
1037 {
1038         return __btree_iter_unpack(c, l, u,
1039                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
1040 }
1041
1042 static inline struct bkey_s_c btree_path_level_peek(struct bch_fs *c,
1043                                                     struct btree_path *path,
1044                                                     struct btree_path_level *l,
1045                                                     struct bkey *u)
1046 {
1047         struct bkey_s_c k = __btree_iter_unpack(c, l, u,
1048                         bch2_btree_node_iter_peek(&l->iter, l->b));
1049
1050         path->pos = k.k ? k.k->p : l->b->key.k.p;
1051         return k;
1052 }
1053
1054 static inline struct bkey_s_c btree_path_level_prev(struct bch_fs *c,
1055                                                     struct btree_path *path,
1056                                                     struct btree_path_level *l,
1057                                                     struct bkey *u)
1058 {
1059         struct bkey_s_c k = __btree_iter_unpack(c, l, u,
1060                         bch2_btree_node_iter_prev(&l->iter, l->b));
1061
1062         path->pos = k.k ? k.k->p : l->b->data->min_key;
1063         return k;
1064 }
1065
1066 static inline bool btree_path_advance_to_pos(struct btree_path *path,
1067                                              struct btree_path_level *l,
1068                                              int max_advance)
1069 {
1070         struct bkey_packed *k;
1071         int nr_advanced = 0;
1072
1073         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
1074                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
1075                 if (max_advance > 0 && nr_advanced >= max_advance)
1076                         return false;
1077
1078                 bch2_btree_node_iter_advance(&l->iter, l->b);
1079                 nr_advanced++;
1080         }
1081
1082         return true;
1083 }
1084
1085 /*
1086  * Verify that iterator for parent node points to child node:
1087  */
1088 static void btree_path_verify_new_node(struct btree_trans *trans,
1089                                        struct btree_path *path, struct btree *b)
1090 {
1091         struct bch_fs *c = trans->c;
1092         struct btree_path_level *l;
1093         unsigned plevel;
1094         bool parent_locked;
1095         struct bkey_packed *k;
1096
1097         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
1098                 return;
1099
1100         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
1101                 return;
1102
1103         plevel = b->c.level + 1;
1104         if (!btree_path_node(path, plevel))
1105                 return;
1106
1107         parent_locked = btree_node_locked(path, plevel);
1108
1109         if (!bch2_btree_node_relock(trans, path, plevel))
1110                 return;
1111
1112         l = &path->l[plevel];
1113         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1114         if (!k ||
1115             bkey_deleted(k) ||
1116             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
1117                 char buf1[100];
1118                 char buf2[100];
1119                 char buf3[100];
1120                 char buf4[100];
1121                 struct bkey uk = bkey_unpack_key(b, k);
1122
1123                 bch2_dump_btree_node(c, l->b);
1124                 bch2_bpos_to_text(&PBUF(buf1), path->pos);
1125                 bch2_bkey_to_text(&PBUF(buf2), &uk);
1126                 bch2_bpos_to_text(&PBUF(buf3), b->data->min_key);
1127                 bch2_bpos_to_text(&PBUF(buf3), b->data->max_key);
1128                 panic("parent iter doesn't point to new node:\n"
1129                       "iter pos %s %s\n"
1130                       "iter key %s\n"
1131                       "new node %s-%s\n",
1132                       bch2_btree_ids[path->btree_id], buf1,
1133                       buf2, buf3, buf4);
1134         }
1135
1136         if (!parent_locked)
1137                 btree_node_unlock(path, plevel);
1138 }
1139
1140 static inline void __btree_path_level_init(struct btree_path *path,
1141                                            unsigned level)
1142 {
1143         struct btree_path_level *l = &path->l[level];
1144
1145         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1146
1147         /*
1148          * Iterators to interior nodes should always be pointed at the first non
1149          * whiteout:
1150          */
1151         if (level)
1152                 bch2_btree_node_iter_peek(&l->iter, l->b);
1153 }
1154
1155 static inline void btree_path_level_init(struct btree_trans *trans,
1156                                          struct btree_path *path,
1157                                          struct btree *b)
1158 {
1159         BUG_ON(path->cached);
1160
1161         btree_path_verify_new_node(trans, path, b);
1162
1163         EBUG_ON(!btree_path_pos_in_node(path, b));
1164         EBUG_ON(b->c.lock.state.seq & 1);
1165
1166         path->l[b->c.level].lock_seq = b->c.lock.state.seq;
1167         path->l[b->c.level].b = b;
1168         __btree_path_level_init(path, b->c.level);
1169 }
1170
1171 /* Btree path: fixups after btree node updates: */
1172
1173 /*
1174  * A btree node is being replaced - update the iterator to point to the new
1175  * node:
1176  */
1177 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
1178 {
1179         struct btree_path *path;
1180
1181         trans_for_each_path(trans, path)
1182                 if (!path->cached &&
1183                     btree_path_pos_in_node(path, b)) {
1184                         enum btree_node_locked_type t =
1185                                 btree_lock_want(path, b->c.level);
1186
1187                         if (path->nodes_locked &&
1188                             t != BTREE_NODE_UNLOCKED) {
1189                                 btree_node_unlock(path, b->c.level);
1190                                 six_lock_increment(&b->c.lock, t);
1191                                 mark_btree_node_locked(path, b->c.level, t);
1192                         }
1193
1194                         btree_path_level_init(trans, path, b);
1195                 }
1196 }
1197
1198 /*
1199  * A btree node has been modified in such a way as to invalidate iterators - fix
1200  * them:
1201  */
1202 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
1203 {
1204         struct btree_path *path;
1205
1206         trans_for_each_path_with_node(trans, b, path)
1207                 __btree_path_level_init(path, b->c.level);
1208 }
1209
1210 /* Btree path: traverse, set_pos: */
1211
1212 static int lock_root_check_fn(struct six_lock *lock, void *p)
1213 {
1214         struct btree *b = container_of(lock, struct btree, c.lock);
1215         struct btree **rootp = p;
1216
1217         return b == *rootp ? 0 : -1;
1218 }
1219
1220 static inline int btree_path_lock_root(struct btree_trans *trans,
1221                                        struct btree_path *path,
1222                                        unsigned depth_want,
1223                                        unsigned long trace_ip)
1224 {
1225         struct bch_fs *c = trans->c;
1226         struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
1227         enum six_lock_type lock_type;
1228         unsigned i;
1229
1230         EBUG_ON(path->nodes_locked);
1231
1232         while (1) {
1233                 b = READ_ONCE(*rootp);
1234                 path->level = READ_ONCE(b->c.level);
1235
1236                 if (unlikely(path->level < depth_want)) {
1237                         /*
1238                          * the root is at a lower depth than the depth we want:
1239                          * got to the end of the btree, or we're walking nodes
1240                          * greater than some depth and there are no nodes >=
1241                          * that depth
1242                          */
1243                         path->level = depth_want;
1244                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
1245                                 path->l[i].b = NULL;
1246                         return 1;
1247                 }
1248
1249                 lock_type = __btree_lock_want(path, path->level);
1250                 if (unlikely(!btree_node_lock(trans, path, b, SPOS_MAX,
1251                                               path->level, lock_type,
1252                                               lock_root_check_fn, rootp,
1253                                               trace_ip))) {
1254                         if (trans->restarted)
1255                                 return -EINTR;
1256                         continue;
1257                 }
1258
1259                 if (likely(b == READ_ONCE(*rootp) &&
1260                            b->c.level == path->level &&
1261                            !race_fault())) {
1262                         for (i = 0; i < path->level; i++)
1263                                 path->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
1264                         path->l[path->level].b = b;
1265                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
1266                                 path->l[i].b = NULL;
1267
1268                         mark_btree_node_locked(path, path->level, lock_type);
1269                         btree_path_level_init(trans, path, b);
1270                         return 0;
1271                 }
1272
1273                 six_unlock_type(&b->c.lock, lock_type);
1274         }
1275 }
1276
1277 noinline
1278 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
1279 {
1280         struct bch_fs *c = trans->c;
1281         struct btree_path_level *l = path_l(path);
1282         struct btree_node_iter node_iter = l->iter;
1283         struct bkey_packed *k;
1284         struct bkey_buf tmp;
1285         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1286                 ? (path->level > 1 ? 0 :  2)
1287                 : (path->level > 1 ? 1 : 16);
1288         bool was_locked = btree_node_locked(path, path->level);
1289         int ret = 0;
1290
1291         bch2_bkey_buf_init(&tmp);
1292
1293         while (nr && !ret) {
1294                 if (!bch2_btree_node_relock(trans, path, path->level))
1295                         break;
1296
1297                 bch2_btree_node_iter_advance(&node_iter, l->b);
1298                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
1299                 if (!k)
1300                         break;
1301
1302                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
1303                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
1304                                                path->level - 1);
1305         }
1306
1307         if (!was_locked)
1308                 btree_node_unlock(path, path->level);
1309
1310         bch2_bkey_buf_exit(&tmp, c);
1311         return ret;
1312 }
1313
1314 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
1315                                  struct btree_and_journal_iter *jiter)
1316 {
1317         struct bch_fs *c = trans->c;
1318         struct bkey_s_c k;
1319         struct bkey_buf tmp;
1320         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1321                 ? (path->level > 1 ? 0 :  2)
1322                 : (path->level > 1 ? 1 : 16);
1323         bool was_locked = btree_node_locked(path, path->level);
1324         int ret = 0;
1325
1326         bch2_bkey_buf_init(&tmp);
1327
1328         while (nr && !ret) {
1329                 if (!bch2_btree_node_relock(trans, path, path->level))
1330                         break;
1331
1332                 bch2_btree_and_journal_iter_advance(jiter);
1333                 k = bch2_btree_and_journal_iter_peek(jiter);
1334                 if (!k.k)
1335                         break;
1336
1337                 bch2_bkey_buf_reassemble(&tmp, c, k);
1338                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
1339                                                path->level - 1);
1340         }
1341
1342         if (!was_locked)
1343                 btree_node_unlock(path, path->level);
1344
1345         bch2_bkey_buf_exit(&tmp, c);
1346         return ret;
1347 }
1348
1349 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
1350                                             struct btree_path *path,
1351                                             unsigned plevel, struct btree *b)
1352 {
1353         struct btree_path_level *l = &path->l[plevel];
1354         bool locked = btree_node_locked(path, plevel);
1355         struct bkey_packed *k;
1356         struct bch_btree_ptr_v2 *bp;
1357
1358         if (!bch2_btree_node_relock(trans, path, plevel))
1359                 return;
1360
1361         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1362         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
1363
1364         bp = (void *) bkeyp_val(&l->b->format, k);
1365         bp->mem_ptr = (unsigned long)b;
1366
1367         if (!locked)
1368                 btree_node_unlock(path, plevel);
1369 }
1370
1371 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
1372                                                      struct btree_path *path,
1373                                                      unsigned flags,
1374                                                      struct bkey_buf *out)
1375 {
1376         struct bch_fs *c = trans->c;
1377         struct btree_path_level *l = path_l(path);
1378         struct btree_and_journal_iter jiter;
1379         struct bkey_s_c k;
1380         int ret = 0;
1381
1382         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
1383
1384         k = bch2_btree_and_journal_iter_peek(&jiter);
1385
1386         bch2_bkey_buf_reassemble(out, c, k);
1387
1388         if (flags & BTREE_ITER_PREFETCH)
1389                 ret = btree_path_prefetch_j(trans, path, &jiter);
1390
1391         bch2_btree_and_journal_iter_exit(&jiter);
1392         return ret;
1393 }
1394
1395 static __always_inline int btree_path_down(struct btree_trans *trans,
1396                                            struct btree_path *path,
1397                                            unsigned flags,
1398                                            unsigned long trace_ip)
1399 {
1400         struct bch_fs *c = trans->c;
1401         struct btree_path_level *l = path_l(path);
1402         struct btree *b;
1403         unsigned level = path->level - 1;
1404         enum six_lock_type lock_type = __btree_lock_want(path, level);
1405         bool replay_done = test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
1406         struct bkey_buf tmp;
1407         int ret;
1408
1409         EBUG_ON(!btree_node_locked(path, path->level));
1410
1411         bch2_bkey_buf_init(&tmp);
1412
1413         if (unlikely(!replay_done)) {
1414                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
1415                 if (ret)
1416                         goto err;
1417         } else {
1418                 bch2_bkey_buf_unpack(&tmp, c, l->b,
1419                                  bch2_btree_node_iter_peek(&l->iter, l->b));
1420
1421                 if (flags & BTREE_ITER_PREFETCH) {
1422                         ret = btree_path_prefetch(trans, path);
1423                         if (ret)
1424                                 goto err;
1425                 }
1426         }
1427
1428         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
1429         ret = PTR_ERR_OR_ZERO(b);
1430         if (unlikely(ret))
1431                 goto err;
1432
1433         mark_btree_node_locked(path, level, lock_type);
1434         btree_path_level_init(trans, path, b);
1435
1436         if (likely(replay_done && tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
1437             unlikely(b != btree_node_mem_ptr(tmp.k)))
1438                 btree_node_mem_ptr_set(trans, path, level + 1, b);
1439
1440         if (btree_node_read_locked(path, level + 1))
1441                 btree_node_unlock(path, level + 1);
1442         path->level = level;
1443
1444         bch2_btree_path_verify_locks(path);
1445 err:
1446         bch2_bkey_buf_exit(&tmp, c);
1447         return ret;
1448 }
1449
1450 static int btree_path_traverse_one(struct btree_trans *, struct btree_path *,
1451                                    unsigned, unsigned long);
1452
1453 static int __btree_path_traverse_all(struct btree_trans *trans, int ret,
1454                                      unsigned long trace_ip)
1455 {
1456         struct bch_fs *c = trans->c;
1457         struct btree_path *path;
1458         int i;
1459
1460         if (trans->in_traverse_all)
1461                 return -EINTR;
1462
1463         trans->in_traverse_all = true;
1464 retry_all:
1465         trans->restarted = false;
1466
1467         trans_for_each_path(trans, path)
1468                 path->should_be_locked = false;
1469
1470         btree_trans_verify_sorted(trans);
1471
1472         for (i = trans->nr_sorted - 2; i >= 0; --i) {
1473                 struct btree_path *path1 = trans->paths + trans->sorted[i];
1474                 struct btree_path *path2 = trans->paths + trans->sorted[i + 1];
1475
1476                 if (path1->btree_id == path2->btree_id &&
1477                     path1->locks_want < path2->locks_want)
1478                         __bch2_btree_path_upgrade(trans, path1, path2->locks_want);
1479                 else if (!path1->locks_want && path2->locks_want)
1480                         __bch2_btree_path_upgrade(trans, path1, 1);
1481         }
1482
1483         bch2_trans_unlock(trans);
1484         cond_resched();
1485
1486         if (unlikely(ret == -ENOMEM)) {
1487                 struct closure cl;
1488
1489                 closure_init_stack(&cl);
1490
1491                 do {
1492                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1493                         closure_sync(&cl);
1494                 } while (ret);
1495         }
1496
1497         if (unlikely(ret == -EIO))
1498                 goto out;
1499
1500         BUG_ON(ret && ret != -EINTR);
1501
1502         /* Now, redo traversals in correct order: */
1503         i = 0;
1504         while (i < trans->nr_sorted) {
1505                 path = trans->paths + trans->sorted[i];
1506
1507                 EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
1508
1509                 ret = btree_path_traverse_one(trans, path, 0, _THIS_IP_);
1510                 if (ret)
1511                         goto retry_all;
1512
1513                 EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
1514
1515                 if (path->nodes_locked ||
1516                     !btree_path_node(path, path->level))
1517                         i++;
1518         }
1519
1520         /*
1521          * BTREE_ITER_NEED_RELOCK is ok here - if we called bch2_trans_unlock()
1522          * and relock(), relock() won't relock since path->should_be_locked
1523          * isn't set yet, which is all fine
1524          */
1525         trans_for_each_path(trans, path)
1526                 BUG_ON(path->uptodate >= BTREE_ITER_NEED_TRAVERSE);
1527 out:
1528         bch2_btree_cache_cannibalize_unlock(c);
1529
1530         trans->in_traverse_all = false;
1531
1532         trace_trans_traverse_all(trans->fn, trace_ip);
1533         return ret;
1534 }
1535
1536 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
1537 {
1538         return __btree_path_traverse_all(trans, 0, _RET_IP_);
1539 }
1540
1541 static inline bool btree_path_good_node(struct btree_trans *trans,
1542                                         struct btree_path *path,
1543                                         unsigned l, int check_pos)
1544 {
1545         if (!is_btree_node(path, l) ||
1546             !bch2_btree_node_relock(trans, path, l))
1547                 return false;
1548
1549         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1550                 return false;
1551         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1552                 return false;
1553         return true;
1554 }
1555
1556 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1557                                                      struct btree_path *path,
1558                                                      int check_pos)
1559 {
1560         unsigned i, l = path->level;
1561
1562         while (btree_path_node(path, l) &&
1563                !btree_path_good_node(trans, path, l, check_pos)) {
1564                 btree_node_unlock(path, l);
1565                 path->l[l].b = BTREE_ITER_NO_NODE_UP;
1566                 l++;
1567         }
1568
1569         /* If we need intent locks, take them too: */
1570         for (i = l + 1;
1571              i < path->locks_want && btree_path_node(path, i);
1572              i++)
1573                 if (!bch2_btree_node_relock(trans, path, i))
1574                         while (l <= i) {
1575                                 btree_node_unlock(path, l);
1576                                 path->l[l].b = BTREE_ITER_NO_NODE_UP;
1577                                 l++;
1578                         }
1579
1580         return l;
1581 }
1582
1583 /*
1584  * This is the main state machine for walking down the btree - walks down to a
1585  * specified depth
1586  *
1587  * Returns 0 on success, -EIO on error (error reading in a btree node).
1588  *
1589  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1590  * stashed in the iterator and returned from bch2_trans_exit().
1591  */
1592 static int btree_path_traverse_one(struct btree_trans *trans,
1593                                    struct btree_path *path,
1594                                    unsigned flags,
1595                                    unsigned long trace_ip)
1596 {
1597         unsigned depth_want = path->level;
1598         int ret = 0;
1599
1600         if (unlikely(trans->restarted)) {
1601                 ret = -EINTR;
1602                 goto out;
1603         }
1604
1605         /*
1606          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1607          * and re-traverse the path without a transaction restart:
1608          */
1609         if (path->should_be_locked) {
1610                 ret = bch2_btree_path_relock(trans, path, trace_ip) ? 0 : -EINTR;
1611                 goto out;
1612         }
1613
1614         if (path->cached) {
1615                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1616                 goto out;
1617         }
1618
1619         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1620                 goto out;
1621
1622         path->level = btree_path_up_until_good_node(trans, path, 0);
1623
1624         /*
1625          * Note: path->nodes[path->level] may be temporarily NULL here - that
1626          * would indicate to other code that we got to the end of the btree,
1627          * here it indicates that relocking the root failed - it's critical that
1628          * btree_path_lock_root() comes next and that it can't fail
1629          */
1630         while (path->level > depth_want) {
1631                 ret = btree_path_node(path, path->level)
1632                         ? btree_path_down(trans, path, flags, trace_ip)
1633                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1634                 if (unlikely(ret)) {
1635                         if (ret == 1) {
1636                                 /*
1637                                  * No nodes at this level - got to the end of
1638                                  * the btree:
1639                                  */
1640                                 ret = 0;
1641                                 goto out;
1642                         }
1643
1644                         __bch2_btree_path_unlock(path);
1645                         path->level = depth_want;
1646
1647                         if (ret == -EIO)
1648                                 path->l[path->level].b =
1649                                         BTREE_ITER_NO_NODE_ERROR;
1650                         else
1651                                 path->l[path->level].b =
1652                                         BTREE_ITER_NO_NODE_DOWN;
1653                         goto out;
1654                 }
1655         }
1656
1657         path->uptodate = BTREE_ITER_UPTODATE;
1658 out:
1659         BUG_ON((ret == -EINTR) != !!trans->restarted);
1660         bch2_btree_path_verify(trans, path);
1661         return ret;
1662 }
1663
1664 static int __btree_path_traverse_all(struct btree_trans *, int, unsigned long);
1665
1666 int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
1667                                           struct btree_path *path, unsigned flags)
1668 {
1669         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
1670                 return 0;
1671
1672         return  bch2_trans_cond_resched(trans) ?:
1673                 btree_path_traverse_one(trans, path, flags, _RET_IP_);
1674 }
1675
1676 static void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1677                             struct btree_path *src)
1678 {
1679         unsigned i;
1680
1681         memcpy(&dst->pos, &src->pos,
1682                sizeof(struct btree_path) - offsetof(struct btree_path, pos));
1683
1684         for (i = 0; i < BTREE_MAX_DEPTH; i++)
1685                 if (btree_node_locked(dst, i))
1686                         six_lock_increment(&dst->l[i].b->c.lock,
1687                                            __btree_lock_want(dst, i));
1688
1689         btree_path_check_sort(trans, dst, 0);
1690 }
1691
1692 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1693                                            bool intent)
1694 {
1695         struct btree_path *new = btree_path_alloc(trans, src);
1696
1697         btree_path_copy(trans, new, src);
1698         __btree_path_get(new, intent);
1699         return new;
1700 }
1701
1702 inline struct btree_path * __must_check
1703 bch2_btree_path_make_mut(struct btree_trans *trans,
1704                          struct btree_path *path, bool intent,
1705                          unsigned long ip)
1706 {
1707         if (path->ref > 1 || path->preserve) {
1708                 __btree_path_put(path, intent);
1709                 path = btree_path_clone(trans, path, intent);
1710                 path->preserve = false;
1711 #ifdef CONFIG_BCACHEFS_DEBUG
1712                 path->ip_allocated = ip;
1713 #endif
1714                 btree_trans_verify_sorted(trans);
1715         }
1716
1717         return path;
1718 }
1719
1720 struct btree_path * __must_check
1721 bch2_btree_path_set_pos(struct btree_trans *trans,
1722                    struct btree_path *path, struct bpos new_pos,
1723                    bool intent, unsigned long ip)
1724 {
1725         int cmp = bpos_cmp(new_pos, path->pos);
1726         unsigned l = path->level;
1727
1728         EBUG_ON(trans->restarted);
1729         EBUG_ON(!path->ref);
1730
1731         if (!cmp)
1732                 return path;
1733
1734         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1735
1736         path->pos               = new_pos;
1737         path->should_be_locked  = false;
1738
1739         btree_path_check_sort(trans, path, cmp);
1740
1741         if (unlikely(path->cached)) {
1742                 btree_node_unlock(path, 0);
1743                 path->l[0].b = BTREE_ITER_NO_NODE_CACHED;
1744                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1745                 goto out;
1746         }
1747
1748         l = btree_path_up_until_good_node(trans, path, cmp);
1749
1750         if (btree_path_node(path, l)) {
1751                 /*
1752                  * We might have to skip over many keys, or just a few: try
1753                  * advancing the node iterator, and if we have to skip over too
1754                  * many keys just reinit it (or if we're rewinding, since that
1755                  * is expensive).
1756                  */
1757                 if (cmp < 0 ||
1758                     !btree_path_advance_to_pos(path, &path->l[l], 8))
1759                         __btree_path_level_init(path, l);
1760         }
1761
1762         if (l != path->level) {
1763                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1764                 __bch2_btree_path_unlock(path);
1765         }
1766 out:
1767         bch2_btree_path_verify(trans, path);
1768         return path;
1769 }
1770
1771 /* Btree path: main interface: */
1772
1773 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1774 {
1775         struct btree_path *next;
1776
1777         next = prev_btree_path(trans, path);
1778         if (next && !btree_path_cmp(next, path))
1779                 return next;
1780
1781         next = next_btree_path(trans, path);
1782         if (next && !btree_path_cmp(next, path))
1783                 return next;
1784
1785         return NULL;
1786 }
1787
1788 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1789 {
1790         struct btree_path *next;
1791
1792         next = prev_btree_path(trans, path);
1793         if (next && next->level == path->level && path_l(next)->b == path_l(path)->b)
1794                 return next;
1795
1796         next = next_btree_path(trans, path);
1797         if (next && next->level == path->level && path_l(next)->b == path_l(path)->b)
1798                 return next;
1799
1800         return NULL;
1801 }
1802
1803 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1804 {
1805         __bch2_btree_path_unlock(path);
1806         btree_path_list_remove(trans, path);
1807         trans->paths_allocated &= ~(1ULL << path->idx);
1808 }
1809
1810 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1811 {
1812         struct btree_path *dup;
1813
1814         EBUG_ON(trans->paths + path->idx != path);
1815         EBUG_ON(!path->ref);
1816
1817         if (!__btree_path_put(path, intent))
1818                 return;
1819
1820         /*
1821          * Perhaps instead we should check for duplicate paths in traverse_all:
1822          */
1823         if (path->preserve &&
1824             (dup = have_path_at_pos(trans, path))) {
1825                 dup->preserve = true;
1826                 path->preserve = false;
1827                 goto free;
1828         }
1829
1830         if (!path->preserve &&
1831             (dup = have_node_at_pos(trans, path)))
1832                 goto free;
1833         return;
1834 free:
1835         if (path->should_be_locked &&
1836             !btree_node_locked(dup, path->level))
1837                 return;
1838
1839         dup->should_be_locked |= path->should_be_locked;
1840         __bch2_path_free(trans, path);
1841 }
1842
1843 noinline __cold
1844 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1845 {
1846         struct btree_path *path;
1847         struct btree_insert_entry *i;
1848         unsigned idx;
1849         char buf1[300], buf2[300];
1850
1851         btree_trans_verify_sorted(trans);
1852
1853         trans_for_each_path_inorder(trans, path, idx)
1854                 printk(KERN_ERR "path: idx %u ref %u:%u%s%s btree %s pos %s locks %u %pS\n",
1855                        path->idx, path->ref, path->intent_ref,
1856                        path->should_be_locked ? " S" : "",
1857                        path->preserve ? " P" : "",
1858                        bch2_btree_ids[path->btree_id],
1859                        (bch2_bpos_to_text(&PBUF(buf1), path->pos), buf1),
1860                        path->nodes_locked,
1861 #ifdef CONFIG_BCACHEFS_DEBUG
1862                        (void *) path->ip_allocated
1863 #else
1864                        NULL
1865 #endif
1866                        );
1867
1868         trans_for_each_update(trans, i) {
1869                 struct bkey u;
1870                 struct bkey_s_c old = bch2_btree_path_peek_slot(i->path, &u);
1871
1872                 printk(KERN_ERR "update: btree %s %pS\n  old %s\n  new %s",
1873                        bch2_btree_ids[i->btree_id],
1874                        (void *) i->ip_allocated,
1875                        (bch2_bkey_val_to_text(&PBUF(buf1), trans->c, old), buf1),
1876                        (bch2_bkey_val_to_text(&PBUF(buf2), trans->c, bkey_i_to_s_c(i->k)), buf2));
1877         }
1878 }
1879
1880 static struct btree_path *btree_path_alloc(struct btree_trans *trans,
1881                                            struct btree_path *pos)
1882 {
1883         struct btree_path *path;
1884         unsigned idx;
1885
1886         if (unlikely(trans->paths_allocated ==
1887                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1)))) {
1888                 bch2_dump_trans_paths_updates(trans);
1889                 panic("trans path oveflow\n");
1890         }
1891
1892         idx = __ffs64(~trans->paths_allocated);
1893         trans->paths_allocated |= 1ULL << idx;
1894
1895         path = &trans->paths[idx];
1896
1897         path->idx               = idx;
1898         path->ref               = 0;
1899         path->intent_ref        = 0;
1900         path->nodes_locked      = 0;
1901         path->nodes_intent_locked = 0;
1902
1903         btree_path_list_add(trans, pos, path);
1904         return path;
1905 }
1906
1907 struct btree_path *bch2_path_get(struct btree_trans *trans,
1908                                  enum btree_id btree_id, struct bpos pos,
1909                                  unsigned locks_want, unsigned level,
1910                                  unsigned flags, unsigned long ip)
1911 {
1912         struct btree_path *path, *path_pos = NULL;
1913         bool cached = flags & BTREE_ITER_CACHED;
1914         bool intent = flags & BTREE_ITER_INTENT;
1915         int i;
1916
1917         BUG_ON(trans->restarted);
1918
1919         trans_for_each_path_inorder(trans, path, i) {
1920                 if (__btree_path_cmp(path,
1921                                      btree_id,
1922                                      cached,
1923                                      pos,
1924                                      level) > 0)
1925                         break;
1926
1927                 path_pos = path;
1928         }
1929
1930         if (path_pos &&
1931             path_pos->cached    == cached &&
1932             path_pos->btree_id  == btree_id &&
1933             path_pos->level     == level) {
1934                 __btree_path_get(path_pos, intent);
1935                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1936         } else {
1937                 path = btree_path_alloc(trans, path_pos);
1938                 path_pos = NULL;
1939
1940                 __btree_path_get(path, intent);
1941                 path->pos                       = pos;
1942                 path->btree_id                  = btree_id;
1943                 path->cached                    = cached;
1944                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1945                 path->should_be_locked          = false;
1946                 path->level                     = level;
1947                 path->locks_want                = locks_want;
1948                 path->nodes_locked              = 0;
1949                 path->nodes_intent_locked       = 0;
1950                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1951                         path->l[i].b            = BTREE_ITER_NO_NODE_INIT;
1952 #ifdef CONFIG_BCACHEFS_DEBUG
1953                 path->ip_allocated              = ip;
1954 #endif
1955                 btree_trans_verify_sorted(trans);
1956         }
1957
1958         if (!(flags & BTREE_ITER_NOPRESERVE))
1959                 path->preserve = true;
1960
1961         if (path->intent_ref)
1962                 locks_want = max(locks_want, level + 1);
1963
1964         /*
1965          * If the path has locks_want greater than requested, we don't downgrade
1966          * it here - on transaction restart because btree node split needs to
1967          * upgrade locks, we might be putting/getting the iterator again.
1968          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1969          * a successful transaction commit.
1970          */
1971
1972         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1973         if (locks_want > path->locks_want) {
1974                 path->locks_want = locks_want;
1975                 btree_path_get_locks(trans, path, true);
1976         }
1977
1978         return path;
1979 }
1980
1981 inline struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1982 {
1983
1984         struct bkey_s_c k;
1985
1986         if (!path->cached) {
1987                 struct btree_path_level *l = path_l(path);
1988                 struct bkey_packed *_k;
1989
1990                 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1991
1992                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1993                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1994
1995                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_cmp(k.k->p, path->pos) == 0);
1996
1997                 if (!k.k || bpos_cmp(path->pos, k.k->p))
1998                         goto hole;
1999         } else {
2000                 struct bkey_cached *ck = (void *) path->l[0].b;
2001
2002                 EBUG_ON(ck &&
2003                         (path->btree_id != ck->key.btree_id ||
2004                          bkey_cmp(path->pos, ck->key.pos)));
2005
2006                 /* BTREE_ITER_CACHED_NOFILL|BTREE_ITER_CACHED_NOCREATE? */
2007                 if (unlikely(!ck || !ck->valid))
2008                         return bkey_s_c_null;
2009
2010                 EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
2011
2012                 k = bkey_i_to_s_c(ck->k);
2013         }
2014
2015         return k;
2016 hole:
2017         bkey_init(u);
2018         u->p = path->pos;
2019         return (struct bkey_s_c) { u, NULL };
2020 }
2021
2022 /* Btree iterators: */
2023
2024 int __must_check
2025 __bch2_btree_iter_traverse(struct btree_iter *iter)
2026 {
2027         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
2028 }
2029
2030 int __must_check
2031 bch2_btree_iter_traverse(struct btree_iter *iter)
2032 {
2033         int ret;
2034
2035         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
2036                                         btree_iter_search_key(iter),
2037                                         iter->flags & BTREE_ITER_INTENT,
2038                                         btree_iter_ip_allocated(iter));
2039
2040         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
2041         if (ret)
2042                 return ret;
2043
2044         iter->path->should_be_locked = true;
2045         return 0;
2046 }
2047
2048 /* Iterate across nodes (leaf and interior nodes) */
2049
2050 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
2051 {
2052         struct btree_trans *trans = iter->trans;
2053         struct btree *b = NULL;
2054         int ret;
2055
2056         EBUG_ON(iter->path->cached);
2057         bch2_btree_iter_verify(iter);
2058
2059         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2060         if (ret)
2061                 goto err;
2062
2063         b = btree_path_node(iter->path, iter->path->level);
2064         if (!b)
2065                 goto out;
2066
2067         BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0);
2068
2069         bkey_init(&iter->k);
2070         iter->k.p = iter->pos = b->key.k.p;
2071
2072         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
2073                                         iter->flags & BTREE_ITER_INTENT,
2074                                         btree_iter_ip_allocated(iter));
2075         iter->path->should_be_locked = true;
2076         BUG_ON(iter->path->uptodate);
2077 out:
2078         bch2_btree_iter_verify_entry_exit(iter);
2079         bch2_btree_iter_verify(iter);
2080
2081         return b;
2082 err:
2083         b = ERR_PTR(ret);
2084         goto out;
2085 }
2086
2087 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
2088 {
2089         struct btree_trans *trans = iter->trans;
2090         struct btree_path *path = iter->path;
2091         struct btree *b = NULL;
2092         unsigned l;
2093         int ret;
2094
2095         BUG_ON(trans->restarted);
2096         EBUG_ON(iter->path->cached);
2097         bch2_btree_iter_verify(iter);
2098
2099         /* already at end? */
2100         if (!btree_path_node(path, path->level))
2101                 return NULL;
2102
2103         /* got to end? */
2104         if (!btree_path_node(path, path->level + 1)) {
2105                 btree_node_unlock(path, path->level);
2106                 path->l[path->level].b = BTREE_ITER_NO_NODE_UP;
2107                 path->level++;
2108                 return NULL;
2109         }
2110
2111         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
2112                 __bch2_btree_path_unlock(path);
2113                 path->l[path->level].b = BTREE_ITER_NO_NODE_GET_LOCKS;
2114                 path->l[path->level + 1].b = BTREE_ITER_NO_NODE_GET_LOCKS;
2115                 trace_trans_restart_relock_next_node(trans->fn, _THIS_IP_,
2116                                            path->btree_id, &path->pos);
2117                 btree_trans_restart(trans);
2118                 ret = -EINTR;
2119                 goto err;
2120         }
2121
2122         b = btree_path_node(path, path->level + 1);
2123
2124         if (!bpos_cmp(iter->pos, b->key.k.p)) {
2125                 btree_node_unlock(path, path->level);
2126                 path->l[path->level].b = BTREE_ITER_NO_NODE_UP;
2127                 path->level++;
2128         } else {
2129                 /*
2130                  * Haven't gotten to the end of the parent node: go back down to
2131                  * the next child node
2132                  */
2133                 path = iter->path =
2134                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
2135                                            iter->flags & BTREE_ITER_INTENT,
2136                                            btree_iter_ip_allocated(iter));
2137
2138                 path->level = iter->min_depth;
2139
2140                 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
2141                         if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
2142                                 btree_node_unlock(path, l);
2143
2144                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
2145                 bch2_btree_iter_verify(iter);
2146
2147                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
2148                 if (ret)
2149                         goto err;
2150
2151                 b = path->l[path->level].b;
2152         }
2153
2154         bkey_init(&iter->k);
2155         iter->k.p = iter->pos = b->key.k.p;
2156
2157         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
2158                                         iter->flags & BTREE_ITER_INTENT,
2159                                         btree_iter_ip_allocated(iter));
2160         iter->path->should_be_locked = true;
2161         BUG_ON(iter->path->uptodate);
2162 out:
2163         bch2_btree_iter_verify_entry_exit(iter);
2164         bch2_btree_iter_verify(iter);
2165
2166         return b;
2167 err:
2168         b = ERR_PTR(ret);
2169         goto out;
2170 }
2171
2172 /* Iterate across keys (in leaf nodes only) */
2173
2174 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
2175 {
2176         struct bpos pos = iter->k.p;
2177         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
2178                     ? bpos_cmp(pos, SPOS_MAX)
2179                     : bkey_cmp(pos, SPOS_MAX)) != 0;
2180
2181         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
2182                 pos = bkey_successor(iter, pos);
2183         bch2_btree_iter_set_pos(iter, pos);
2184         return ret;
2185 }
2186
2187 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
2188 {
2189         struct bpos pos = bkey_start_pos(&iter->k);
2190         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
2191                     ? bpos_cmp(pos, POS_MIN)
2192                     : bkey_cmp(pos, POS_MIN)) != 0;
2193
2194         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
2195                 pos = bkey_predecessor(iter, pos);
2196         bch2_btree_iter_set_pos(iter, pos);
2197         return ret;
2198 }
2199
2200 static inline struct bkey_i *btree_trans_peek_updates(struct btree_trans *trans,
2201                                                       enum btree_id btree_id,
2202                                                       struct bpos pos)
2203 {
2204         struct btree_insert_entry *i;
2205
2206         trans_for_each_update(trans, i)
2207                 if ((cmp_int(btree_id,  i->btree_id) ?:
2208                      bpos_cmp(pos,      i->k->k.p)) <= 0) {
2209                         if (btree_id == i->btree_id)
2210                                 return i->k;
2211                         break;
2212                 }
2213
2214         return NULL;
2215 }
2216
2217 static noinline
2218 struct bkey_i *__btree_trans_peek_journal(struct btree_trans *trans,
2219                                           struct btree_path *path)
2220 {
2221         struct journal_keys *keys = &trans->c->journal_keys;
2222         size_t idx = bch2_journal_key_search(keys, path->btree_id,
2223                                              path->level, path->pos);
2224
2225         while (idx < keys->nr && keys->d[idx].overwritten)
2226                 idx++;
2227
2228         return (idx < keys->nr &&
2229                 keys->d[idx].btree_id   == path->btree_id &&
2230                 keys->d[idx].level      == path->level)
2231                 ? keys->d[idx].k
2232                 : NULL;
2233 }
2234
2235 static noinline
2236 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
2237                                          struct btree_iter *iter,
2238                                          struct bkey_s_c k)
2239 {
2240         struct bkey_i *next_journal =
2241                 __btree_trans_peek_journal(trans, iter->path);
2242
2243         if (next_journal &&
2244             bpos_cmp(next_journal->k.p,
2245                      k.k ? k.k->p : iter->path->l[0].b->key.k.p) <= 0) {
2246                 iter->k = next_journal->k;
2247                 k = bkey_i_to_s_c(next_journal);
2248         }
2249
2250         return k;
2251 }
2252
2253 /*
2254  * Checks btree key cache for key at iter->pos and returns it if present, or
2255  * bkey_s_c_null:
2256  */
2257 static noinline
2258 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
2259 {
2260         struct btree_trans *trans = iter->trans;
2261         struct bch_fs *c = trans->c;
2262         struct bkey u;
2263         int ret;
2264
2265         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
2266                 return bkey_s_c_null;
2267
2268         if (!iter->key_cache_path)
2269                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
2270                                                      iter->flags & BTREE_ITER_INTENT, 0,
2271                                                      iter->flags|BTREE_ITER_CACHED,
2272                                                      _THIS_IP_);
2273
2274         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
2275                                         iter->flags & BTREE_ITER_INTENT,
2276                                         btree_iter_ip_allocated(iter));
2277
2278         ret = bch2_btree_path_traverse(trans, iter->key_cache_path, iter->flags|BTREE_ITER_CACHED);
2279         if (unlikely(ret))
2280                 return bkey_s_c_err(ret);
2281
2282         iter->key_cache_path->should_be_locked = true;
2283
2284         return bch2_btree_path_peek_slot(iter->key_cache_path, &u);
2285 }
2286
2287 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
2288 {
2289         struct btree_trans *trans = iter->trans;
2290         struct bkey_i *next_update;
2291         struct bkey_s_c k, k2;
2292         int ret;
2293
2294         EBUG_ON(iter->path->cached || iter->path->level);
2295         bch2_btree_iter_verify(iter);
2296
2297         while (1) {
2298                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2299                                         iter->flags & BTREE_ITER_INTENT,
2300                                         btree_iter_ip_allocated(iter));
2301
2302                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2303                 if (unlikely(ret)) {
2304                         /* ensure that iter->k is consistent with iter->pos: */
2305                         bch2_btree_iter_set_pos(iter, iter->pos);
2306                         k = bkey_s_c_err(ret);
2307                         goto out;
2308                 }
2309
2310                 iter->path->should_be_locked = true;
2311
2312                 k = btree_path_level_peek_all(trans->c, &iter->path->l[0], &iter->k);
2313
2314                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2315                     k.k &&
2316                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2317                         ret = bkey_err(k2);
2318                         if (ret) {
2319                                 k = k2;
2320                                 bch2_btree_iter_set_pos(iter, iter->pos);
2321                                 goto out;
2322                         }
2323
2324                         k = k2;
2325                         iter->k = *k.k;
2326                 }
2327
2328                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2329                         k = btree_trans_peek_journal(trans, iter, k);
2330
2331                 next_update = iter->flags & BTREE_ITER_WITH_UPDATES
2332                         ? btree_trans_peek_updates(trans, iter->btree_id, search_key)
2333                         : NULL;
2334                 if (next_update &&
2335                     bpos_cmp(next_update->k.p,
2336                              k.k ? k.k->p : iter->path->l[0].b->key.k.p) <= 0) {
2337                         iter->k = next_update->k;
2338                         k = bkey_i_to_s_c(next_update);
2339                 }
2340
2341                 if (k.k && bkey_deleted(k.k)) {
2342                         /*
2343                          * If we've got a whiteout, and it's after the search
2344                          * key, advance the search key to the whiteout instead
2345                          * of just after the whiteout - it might be a btree
2346                          * whiteout, with a real key at the same position, since
2347                          * in the btree deleted keys sort before non deleted.
2348                          */
2349                         search_key = bpos_cmp(search_key, k.k->p)
2350                                 ? k.k->p
2351                                 : bpos_successor(k.k->p);
2352                         continue;
2353                 }
2354
2355                 if (likely(k.k)) {
2356                         break;
2357                 } else if (likely(bpos_cmp(iter->path->l[0].b->key.k.p, SPOS_MAX))) {
2358                         /* Advance to next leaf node: */
2359                         search_key = bpos_successor(iter->path->l[0].b->key.k.p);
2360                 } else {
2361                         /* End of btree: */
2362                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2363                         k = bkey_s_c_null;
2364                         goto out;
2365                 }
2366         }
2367 out:
2368         bch2_btree_iter_verify(iter);
2369
2370         return k;
2371 }
2372
2373 /**
2374  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2375  * current position
2376  */
2377 struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
2378 {
2379         struct btree_trans *trans = iter->trans;
2380         struct bpos search_key = btree_iter_search_key(iter);
2381         struct bkey_s_c k;
2382         int ret;
2383
2384         if (iter->update_path) {
2385                 bch2_path_put(trans, iter->update_path,
2386                               iter->flags & BTREE_ITER_INTENT);
2387                 iter->update_path = NULL;
2388         }
2389
2390         bch2_btree_iter_verify_entry_exit(iter);
2391
2392         while (1) {
2393                 k = __bch2_btree_iter_peek(iter, search_key);
2394                 if (!k.k || bkey_err(k))
2395                         goto out;
2396
2397                 if (iter->update_path &&
2398                     bkey_cmp(iter->update_path->pos, k.k->p)) {
2399                         bch2_path_put(trans, iter->update_path,
2400                                       iter->flags & BTREE_ITER_INTENT);
2401                         iter->update_path = NULL;
2402                 }
2403
2404                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2405                     (iter->flags & BTREE_ITER_INTENT) &&
2406                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2407                     !iter->update_path) {
2408                         struct bpos pos = k.k->p;
2409
2410                         if (pos.snapshot < iter->snapshot) {
2411                                 search_key = bpos_successor(k.k->p);
2412                                 continue;
2413                         }
2414
2415                         pos.snapshot = iter->snapshot;
2416
2417                         /*
2418                          * advance, same as on exit for iter->path, but only up
2419                          * to snapshot
2420                          */
2421                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2422                         iter->update_path = iter->path;
2423
2424                         iter->update_path = bch2_btree_path_set_pos(trans,
2425                                                 iter->update_path, pos,
2426                                                 iter->flags & BTREE_ITER_INTENT,
2427                                                 btree_iter_ip_allocated(iter));
2428
2429                         BUG_ON(!(iter->update_path->nodes_locked & 1));
2430                         iter->update_path->should_be_locked = true;
2431                 }
2432
2433                 /*
2434                  * We can never have a key in a leaf node at POS_MAX, so
2435                  * we don't have to check these successor() calls:
2436                  */
2437                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2438                     !bch2_snapshot_is_ancestor(trans->c,
2439                                                iter->snapshot,
2440                                                k.k->p.snapshot)) {
2441                         search_key = bpos_successor(k.k->p);
2442                         continue;
2443                 }
2444
2445                 if (bkey_whiteout(k.k) &&
2446                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2447                         search_key = bkey_successor(iter, k.k->p);
2448                         continue;
2449                 }
2450
2451                 break;
2452         }
2453
2454         /*
2455          * iter->pos should be mononotically increasing, and always be equal to
2456          * the key we just returned - except extents can straddle iter->pos:
2457          */
2458         if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2459                 iter->pos = k.k->p;
2460         else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
2461                 iter->pos = bkey_start_pos(k.k);
2462
2463         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2464                                 iter->flags & BTREE_ITER_INTENT,
2465                                 btree_iter_ip_allocated(iter));
2466         BUG_ON(!iter->path->nodes_locked);
2467 out:
2468         if (iter->update_path) {
2469                 BUG_ON(!(iter->update_path->nodes_locked & 1));
2470                 iter->update_path->should_be_locked = true;
2471         }
2472         iter->path->should_be_locked = true;
2473
2474         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2475                 iter->pos.snapshot = iter->snapshot;
2476
2477         ret = bch2_btree_iter_verify_ret(iter, k);
2478         if (unlikely(ret)) {
2479                 bch2_btree_iter_set_pos(iter, iter->pos);
2480                 k = bkey_s_c_err(ret);
2481         }
2482
2483         bch2_btree_iter_verify_entry_exit(iter);
2484
2485         return k;
2486 }
2487
2488 /**
2489  * bch2_btree_iter_next: returns first key greater than iterator's current
2490  * position
2491  */
2492 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2493 {
2494         if (!bch2_btree_iter_advance(iter))
2495                 return bkey_s_c_null;
2496
2497         return bch2_btree_iter_peek(iter);
2498 }
2499
2500 /**
2501  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2502  * iterator's current position
2503  */
2504 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2505 {
2506         struct btree_trans *trans = iter->trans;
2507         struct bpos search_key = iter->pos;
2508         struct btree_path *saved_path = NULL;
2509         struct bkey_s_c k;
2510         struct bkey saved_k;
2511         const struct bch_val *saved_v;
2512         int ret;
2513
2514         EBUG_ON(iter->path->cached || iter->path->level);
2515         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2516
2517         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2518                 return bkey_s_c_err(-EIO);
2519
2520         bch2_btree_iter_verify(iter);
2521         bch2_btree_iter_verify_entry_exit(iter);
2522
2523         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2524                 search_key.snapshot = U32_MAX;
2525
2526         while (1) {
2527                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2528                                                 iter->flags & BTREE_ITER_INTENT,
2529                                                 btree_iter_ip_allocated(iter));
2530
2531                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2532                 if (unlikely(ret)) {
2533                         /* ensure that iter->k is consistent with iter->pos: */
2534                         bch2_btree_iter_set_pos(iter, iter->pos);
2535                         k = bkey_s_c_err(ret);
2536                         goto out;
2537                 }
2538
2539                 k = btree_path_level_peek(trans->c, iter->path,
2540                                           &iter->path->l[0], &iter->k);
2541                 if (!k.k ||
2542                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2543                      ? bpos_cmp(bkey_start_pos(k.k), search_key) >= 0
2544                      : bpos_cmp(k.k->p, search_key) > 0))
2545                         k = btree_path_level_prev(trans->c, iter->path,
2546                                                   &iter->path->l[0], &iter->k);
2547
2548                 btree_path_check_sort(trans, iter->path, 0);
2549
2550                 if (likely(k.k)) {
2551                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2552                                 if (k.k->p.snapshot == iter->snapshot)
2553                                         goto got_key;
2554
2555                                 /*
2556                                  * If we have a saved candidate, and we're no
2557                                  * longer at the same _key_ (not pos), return
2558                                  * that candidate
2559                                  */
2560                                 if (saved_path && bkey_cmp(k.k->p, saved_k.p)) {
2561                                         bch2_path_put(trans, iter->path,
2562                                                       iter->flags & BTREE_ITER_INTENT);
2563                                         iter->path = saved_path;
2564                                         saved_path = NULL;
2565                                         iter->k = saved_k;
2566                                         k.v     = saved_v;
2567                                         goto got_key;
2568                                 }
2569
2570                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2571                                                               iter->snapshot,
2572                                                               k.k->p.snapshot)) {
2573                                         if (saved_path)
2574                                                 bch2_path_put(trans, saved_path,
2575                                                       iter->flags & BTREE_ITER_INTENT);
2576                                         saved_path = btree_path_clone(trans, iter->path,
2577                                                                 iter->flags & BTREE_ITER_INTENT);
2578                                         saved_k = *k.k;
2579                                         saved_v = k.v;
2580                                 }
2581
2582                                 search_key = bpos_predecessor(k.k->p);
2583                                 continue;
2584                         }
2585 got_key:
2586                         if (bkey_whiteout(k.k) &&
2587                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2588                                 search_key = bkey_predecessor(iter, k.k->p);
2589                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2590                                         search_key.snapshot = U32_MAX;
2591                                 continue;
2592                         }
2593
2594                         break;
2595                 } else if (likely(bpos_cmp(iter->path->l[0].b->data->min_key, POS_MIN))) {
2596                         /* Advance to previous leaf node: */
2597                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2598                 } else {
2599                         /* Start of btree: */
2600                         bch2_btree_iter_set_pos(iter, POS_MIN);
2601                         k = bkey_s_c_null;
2602                         goto out;
2603                 }
2604         }
2605
2606         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
2607
2608         /* Extents can straddle iter->pos: */
2609         if (bkey_cmp(k.k->p, iter->pos) < 0)
2610                 iter->pos = k.k->p;
2611
2612         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2613                 iter->pos.snapshot = iter->snapshot;
2614 out:
2615         if (saved_path)
2616                 bch2_path_put(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2617         iter->path->should_be_locked = true;
2618
2619         bch2_btree_iter_verify_entry_exit(iter);
2620         bch2_btree_iter_verify(iter);
2621
2622         return k;
2623 }
2624
2625 /**
2626  * bch2_btree_iter_prev: returns first key less than iterator's current
2627  * position
2628  */
2629 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2630 {
2631         if (!bch2_btree_iter_rewind(iter))
2632                 return bkey_s_c_null;
2633
2634         return bch2_btree_iter_peek_prev(iter);
2635 }
2636
2637 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2638 {
2639         struct btree_trans *trans = iter->trans;
2640         struct bpos search_key;
2641         struct bkey_s_c k;
2642         int ret;
2643
2644         EBUG_ON(iter->path->level);
2645         bch2_btree_iter_verify(iter);
2646         bch2_btree_iter_verify_entry_exit(iter);
2647
2648         /* extents can't span inode numbers: */
2649         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2650             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2651                 if (iter->pos.inode == KEY_INODE_MAX)
2652                         return bkey_s_c_null;
2653
2654                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2655         }
2656
2657         search_key = btree_iter_search_key(iter);
2658         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2659                                         iter->flags & BTREE_ITER_INTENT,
2660                                         btree_iter_ip_allocated(iter));
2661
2662         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2663         if (unlikely(ret))
2664                 return bkey_s_c_err(ret);
2665
2666         if ((iter->flags & BTREE_ITER_CACHED) ||
2667             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2668                 struct bkey_i *next_update;
2669
2670                 if ((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2671                     (next_update = btree_trans_peek_updates(trans,
2672                                                 iter->btree_id, search_key)) &&
2673                     !bpos_cmp(next_update->k.p, iter->pos)) {
2674                         iter->k = next_update->k;
2675                         k = bkey_i_to_s_c(next_update);
2676                         goto out;
2677                 }
2678
2679                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2680                     (next_update = __btree_trans_peek_journal(trans, iter->path)) &&
2681                     !bpos_cmp(next_update->k.p, iter->pos)) {
2682                         iter->k = next_update->k;
2683                         k = bkey_i_to_s_c(next_update);
2684                         goto out;
2685                 }
2686
2687                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2688                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2689                         if (!bkey_err(k))
2690                                 iter->k = *k.k;
2691                         goto out;
2692                 }
2693
2694                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2695         } else {
2696                 struct bpos next;
2697
2698                 if (iter->flags & BTREE_ITER_INTENT) {
2699                         struct btree_iter iter2;
2700
2701                         bch2_trans_copy_iter(&iter2, iter);
2702                         k = bch2_btree_iter_peek(&iter2);
2703
2704                         if (k.k && !bkey_err(k)) {
2705                                 iter->k = iter2.k;
2706                                 k.k = &iter->k;
2707                         }
2708                         bch2_trans_iter_exit(trans, &iter2);
2709                 } else {
2710                         struct bpos pos = iter->pos;
2711
2712                         k = bch2_btree_iter_peek(iter);
2713                         iter->pos = pos;
2714                 }
2715
2716                 if (unlikely(bkey_err(k)))
2717                         return k;
2718
2719                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2720
2721                 if (bkey_cmp(iter->pos, next) < 0) {
2722                         bkey_init(&iter->k);
2723                         iter->k.p = iter->pos;
2724
2725                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2726                                 bch2_key_resize(&iter->k,
2727                                                 min_t(u64, KEY_SIZE_MAX,
2728                                                       (next.inode == iter->pos.inode
2729                                                        ? next.offset
2730                                                        : KEY_OFFSET_MAX) -
2731                                                       iter->pos.offset));
2732                                 EBUG_ON(!iter->k.size);
2733                         }
2734
2735                         k = (struct bkey_s_c) { &iter->k, NULL };
2736                 }
2737         }
2738 out:
2739         iter->path->should_be_locked = true;
2740
2741         bch2_btree_iter_verify_entry_exit(iter);
2742         bch2_btree_iter_verify(iter);
2743         ret = bch2_btree_iter_verify_ret(iter, k);
2744         if (unlikely(ret))
2745                 return bkey_s_c_err(ret);
2746
2747         return k;
2748 }
2749
2750 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2751 {
2752         if (!bch2_btree_iter_advance(iter))
2753                 return bkey_s_c_null;
2754
2755         return bch2_btree_iter_peek_slot(iter);
2756 }
2757
2758 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2759 {
2760         if (!bch2_btree_iter_rewind(iter))
2761                 return bkey_s_c_null;
2762
2763         return bch2_btree_iter_peek_slot(iter);
2764 }
2765
2766 /* new transactional stuff: */
2767
2768 static inline void btree_path_verify_sorted_ref(struct btree_trans *trans,
2769                                                 struct btree_path *path)
2770 {
2771         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2772         EBUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2773         EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
2774 }
2775
2776 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2777 {
2778 #ifdef CONFIG_BCACHEFS_DEBUG
2779         unsigned i;
2780
2781         for (i = 0; i < trans->nr_sorted; i++)
2782                 btree_path_verify_sorted_ref(trans, trans->paths + trans->sorted[i]);
2783 #endif
2784 }
2785
2786 static void btree_trans_verify_sorted(struct btree_trans *trans)
2787 {
2788 #ifdef CONFIG_BCACHEFS_DEBUG
2789         struct btree_path *path, *prev = NULL;
2790         unsigned i;
2791
2792         trans_for_each_path_inorder(trans, path, i) {
2793                 BUG_ON(prev && btree_path_cmp(prev, path) > 0);
2794                 prev = path;
2795         }
2796 #endif
2797 }
2798
2799 static inline void btree_path_swap(struct btree_trans *trans,
2800                                    struct btree_path *l, struct btree_path *r)
2801 {
2802         swap(l->sorted_idx, r->sorted_idx);
2803         swap(trans->sorted[l->sorted_idx],
2804              trans->sorted[r->sorted_idx]);
2805
2806         btree_path_verify_sorted_ref(trans, l);
2807         btree_path_verify_sorted_ref(trans, r);
2808 }
2809
2810 static void btree_path_check_sort(struct btree_trans *trans, struct btree_path *path,
2811                                   int cmp)
2812 {
2813         struct btree_path *n;
2814
2815         if (cmp <= 0) {
2816                 n = prev_btree_path(trans, path);
2817                 if (n && btree_path_cmp(n, path) > 0) {
2818                         do {
2819                                 btree_path_swap(trans, n, path);
2820                                 n = prev_btree_path(trans, path);
2821                         } while (n && btree_path_cmp(n, path) > 0);
2822
2823                         goto out;
2824                 }
2825         }
2826
2827         if (cmp >= 0) {
2828                 n = next_btree_path(trans, path);
2829                 if (n && btree_path_cmp(path, n) > 0) {
2830                         do {
2831                                 btree_path_swap(trans, path, n);
2832                                 n = next_btree_path(trans, path);
2833                         } while (n && btree_path_cmp(path, n) > 0);
2834                 }
2835         }
2836 out:
2837         btree_trans_verify_sorted(trans);
2838 }
2839
2840 static inline void btree_path_list_remove(struct btree_trans *trans,
2841                                           struct btree_path *path)
2842 {
2843         unsigned i;
2844
2845         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2846
2847         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2848
2849         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2850                 trans->paths[trans->sorted[i]].sorted_idx = i;
2851
2852         path->sorted_idx = U8_MAX;
2853
2854         btree_trans_verify_sorted_refs(trans);
2855 }
2856
2857 static inline void btree_path_list_add(struct btree_trans *trans,
2858                                        struct btree_path *pos,
2859                                        struct btree_path *path)
2860 {
2861         unsigned i;
2862
2863         btree_trans_verify_sorted_refs(trans);
2864
2865         path->sorted_idx = pos ? pos->sorted_idx + 1 : 0;
2866
2867         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2868
2869         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2870                 trans->paths[trans->sorted[i]].sorted_idx = i;
2871
2872         btree_trans_verify_sorted_refs(trans);
2873 }
2874
2875 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2876 {
2877         if (iter->path)
2878                 bch2_path_put(trans, iter->path,
2879                               iter->flags & BTREE_ITER_INTENT);
2880         if (iter->update_path)
2881                 bch2_path_put(trans, iter->update_path,
2882                               iter->flags & BTREE_ITER_INTENT);
2883         if (iter->key_cache_path)
2884                 bch2_path_put(trans, iter->key_cache_path,
2885                               iter->flags & BTREE_ITER_INTENT);
2886         iter->path = NULL;
2887         iter->update_path = NULL;
2888         iter->key_cache_path = NULL;
2889 }
2890
2891 static void __bch2_trans_iter_init(struct btree_trans *trans,
2892                                    struct btree_iter *iter,
2893                                    unsigned btree_id, struct bpos pos,
2894                                    unsigned locks_want,
2895                                    unsigned depth,
2896                                    unsigned flags,
2897                                    unsigned long ip)
2898 {
2899         EBUG_ON(trans->restarted);
2900
2901         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
2902             btree_node_type_is_extents(btree_id))
2903                 flags |= BTREE_ITER_IS_EXTENTS;
2904
2905         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
2906             !btree_type_has_snapshots(btree_id))
2907                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
2908
2909         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
2910             btree_type_has_snapshots(btree_id))
2911                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
2912
2913         if (!test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags))
2914                 flags |= BTREE_ITER_WITH_JOURNAL;
2915
2916         if (!btree_id_cached(trans->c, btree_id)) {
2917                 flags &= ~BTREE_ITER_CACHED;
2918                 flags &= ~BTREE_ITER_WITH_KEY_CACHE;
2919         } else if (!(flags & BTREE_ITER_CACHED))
2920                 flags |= BTREE_ITER_WITH_KEY_CACHE;
2921
2922         iter->trans     = trans;
2923         iter->path      = NULL;
2924         iter->update_path = NULL;
2925         iter->key_cache_path = NULL;
2926         iter->btree_id  = btree_id;
2927         iter->min_depth = depth;
2928         iter->flags     = flags;
2929         iter->snapshot  = pos.snapshot;
2930         iter->pos       = pos;
2931         iter->k.type    = KEY_TYPE_deleted;
2932         iter->k.p       = pos;
2933         iter->k.size    = 0;
2934 #ifdef CONFIG_BCACHEFS_DEBUG
2935         iter->ip_allocated = ip;
2936 #endif
2937
2938         iter->path = bch2_path_get(trans, btree_id, iter->pos,
2939                                    locks_want, depth, flags, ip);
2940 }
2941
2942 void bch2_trans_iter_init(struct btree_trans *trans,
2943                           struct btree_iter *iter,
2944                           unsigned btree_id, struct bpos pos,
2945                           unsigned flags)
2946 {
2947         __bch2_trans_iter_init(trans, iter, btree_id, pos,
2948                                0, 0, flags, _RET_IP_);
2949 }
2950
2951 void bch2_trans_node_iter_init(struct btree_trans *trans,
2952                                struct btree_iter *iter,
2953                                enum btree_id btree_id,
2954                                struct bpos pos,
2955                                unsigned locks_want,
2956                                unsigned depth,
2957                                unsigned flags)
2958 {
2959         __bch2_trans_iter_init(trans, iter, btree_id, pos, locks_want, depth,
2960                                BTREE_ITER_NOT_EXTENTS|
2961                                __BTREE_ITER_ALL_SNAPSHOTS|
2962                                BTREE_ITER_ALL_SNAPSHOTS|
2963                                flags, _RET_IP_);
2964         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2965         BUG_ON(iter->path->level        != depth);
2966         BUG_ON(iter->min_depth          != depth);
2967 }
2968
2969 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2970 {
2971         *dst = *src;
2972         if (src->path)
2973                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2974         if (src->update_path)
2975                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2976         dst->key_cache_path = NULL;
2977 }
2978
2979 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2980 {
2981         size_t new_top = trans->mem_top + size;
2982         void *p;
2983
2984         if (new_top > trans->mem_bytes) {
2985                 size_t old_bytes = trans->mem_bytes;
2986                 size_t new_bytes = roundup_pow_of_two(new_top);
2987                 void *new_mem;
2988
2989                 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2990
2991                 new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2992                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2993                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2994                         new_bytes = BTREE_TRANS_MEM_MAX;
2995                         kfree(trans->mem);
2996                 }
2997
2998                 if (!new_mem)
2999                         return ERR_PTR(-ENOMEM);
3000
3001                 trans->mem = new_mem;
3002                 trans->mem_bytes = new_bytes;
3003
3004                 if (old_bytes) {
3005                         trace_trans_restart_mem_realloced(trans->fn, _RET_IP_, new_bytes);
3006                         btree_trans_restart(trans);
3007                         return ERR_PTR(-EINTR);
3008                 }
3009         }
3010
3011         p = trans->mem + trans->mem_top;
3012         trans->mem_top += size;
3013         memset(p, 0, size);
3014         return p;
3015 }
3016
3017 /**
3018  * bch2_trans_begin() - reset a transaction after a interrupted attempt
3019  * @trans: transaction to reset
3020  *
3021  * While iterating over nodes or updating nodes a attempt to lock a btree
3022  * node may return EINTR when the trylock fails. When this occurs
3023  * bch2_trans_begin() should be called and the transaction retried.
3024  */
3025 void bch2_trans_begin(struct btree_trans *trans)
3026 {
3027         struct btree_insert_entry *i;
3028         struct btree_path *path;
3029
3030         trans_for_each_update(trans, i)
3031                 __btree_path_put(i->path, true);
3032
3033         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
3034         trans->extra_journal_res        = 0;
3035         trans->nr_updates               = 0;
3036         trans->mem_top                  = 0;
3037
3038         trans->hooks                    = NULL;
3039         trans->extra_journal_entries    = NULL;
3040         trans->extra_journal_entry_u64s = 0;
3041
3042         if (trans->fs_usage_deltas) {
3043                 trans->fs_usage_deltas->used = 0;
3044                 memset(&trans->fs_usage_deltas->memset_start, 0,
3045                        (void *) &trans->fs_usage_deltas->memset_end -
3046                        (void *) &trans->fs_usage_deltas->memset_start);
3047         }
3048
3049         trans_for_each_path(trans, path) {
3050                 path->should_be_locked = false;
3051
3052                 /*
3053                  * XXX: we probably shouldn't be doing this if the transaction
3054                  * was restarted, but currently we still overflow transaction
3055                  * iterators if we do that
3056                  */
3057                 if (!path->ref && !path->preserve)
3058                         __bch2_path_free(trans, path);
3059                 else if (!path->ref)
3060                         path->preserve = false;
3061         }
3062
3063         bch2_trans_cond_resched(trans);
3064
3065         if (trans->restarted)
3066                 bch2_btree_path_traverse_all(trans);
3067
3068         trans->restarted = false;
3069 }
3070
3071 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
3072 {
3073         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
3074         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
3075         void *p = NULL;
3076
3077         BUG_ON(trans->used_mempool);
3078
3079 #ifdef __KERNEL__
3080         p = this_cpu_xchg(c->btree_paths_bufs->path , NULL);
3081 #endif
3082         if (!p)
3083                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
3084
3085         trans->paths            = p; p += paths_bytes;
3086         trans->updates          = p; p += updates_bytes;
3087 }
3088
3089 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
3090                        unsigned expected_nr_iters,
3091                        size_t expected_mem_bytes,
3092                        const char *fn)
3093         __acquires(&c->btree_trans_barrier)
3094 {
3095         memset(trans, 0, sizeof(*trans));
3096         trans->c                = c;
3097         trans->fn               = fn;
3098
3099         bch2_trans_alloc_paths(trans, c);
3100
3101         if (expected_mem_bytes) {
3102                 trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes);
3103                 trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL);
3104
3105                 if (!unlikely(trans->mem)) {
3106                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
3107                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
3108                 }
3109         }
3110
3111         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
3112
3113         trans->pid = current->pid;
3114         mutex_lock(&c->btree_trans_lock);
3115         list_add(&trans->list, &c->btree_trans_list);
3116         mutex_unlock(&c->btree_trans_lock);
3117 }
3118
3119 static void check_btree_paths_leaked(struct btree_trans *trans)
3120 {
3121 #ifdef CONFIG_BCACHEFS_DEBUG
3122         struct bch_fs *c = trans->c;
3123         struct btree_path *path;
3124
3125         trans_for_each_path(trans, path)
3126                 if (path->ref)
3127                         goto leaked;
3128         return;
3129 leaked:
3130         bch_err(c, "btree paths leaked from %s!", trans->fn);
3131         trans_for_each_path(trans, path)
3132                 if (path->ref)
3133                         printk(KERN_ERR "  btree %s %pS\n",
3134                                bch2_btree_ids[path->btree_id],
3135                                (void *) path->ip_allocated);
3136         /* Be noisy about this: */
3137         bch2_fatal_error(c);
3138 #endif
3139 }
3140
3141 void bch2_trans_exit(struct btree_trans *trans)
3142         __releases(&c->btree_trans_barrier)
3143 {
3144         struct btree_insert_entry *i;
3145         struct bch_fs *c = trans->c;
3146
3147         bch2_trans_unlock(trans);
3148
3149         trans_for_each_update(trans, i)
3150                 __btree_path_put(i->path, true);
3151         trans->nr_updates               = 0;
3152
3153         check_btree_paths_leaked(trans);
3154
3155         mutex_lock(&c->btree_trans_lock);
3156         list_del(&trans->list);
3157         mutex_unlock(&c->btree_trans_lock);
3158
3159         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3160
3161         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3162
3163         if (trans->fs_usage_deltas) {
3164                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3165                     REPLICAS_DELTA_LIST_MAX)
3166                         mempool_free(trans->fs_usage_deltas,
3167                                      &c->replicas_delta_pool);
3168                 else
3169                         kfree(trans->fs_usage_deltas);
3170         }
3171
3172         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3173                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3174         else
3175                 kfree(trans->mem);
3176
3177 #ifdef __KERNEL__
3178         /*
3179          * Userspace doesn't have a real percpu implementation:
3180          */
3181         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3182 #endif
3183
3184         if (trans->paths)
3185                 mempool_free(trans->paths, &c->btree_paths_pool);
3186
3187         trans->mem      = (void *) 0x1;
3188         trans->paths    = (void *) 0x1;
3189 }
3190
3191 static void __maybe_unused
3192 bch2_btree_path_node_to_text(struct printbuf *out,
3193                              struct btree_bkey_cached_common *_b,
3194                              bool cached)
3195 {
3196         pr_buf(out, "    l=%u %s:",
3197                _b->level, bch2_btree_ids[_b->btree_id]);
3198         bch2_bpos_to_text(out, btree_node_pos(_b, cached));
3199 }
3200
3201 static bool trans_has_locks(struct btree_trans *trans)
3202 {
3203         struct btree_path *path;
3204
3205         trans_for_each_path(trans, path)
3206                 if (path->nodes_locked)
3207                         return true;
3208         return false;
3209 }
3210
3211 void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
3212 {
3213         struct btree_trans *trans;
3214         struct btree_path *path;
3215         struct btree *b;
3216         unsigned l;
3217
3218         mutex_lock(&c->btree_trans_lock);
3219         list_for_each_entry(trans, &c->btree_trans_list, list) {
3220                 if (!trans_has_locks(trans))
3221                         continue;
3222
3223                 pr_buf(out, "%i %s\n", trans->pid, trans->fn);
3224
3225                 trans_for_each_path(trans, path) {
3226                         if (!path->nodes_locked)
3227                                 continue;
3228
3229                         pr_buf(out, "  path %u %c l=%u %s:",
3230                                path->idx,
3231                                path->cached ? 'c' : 'b',
3232                                path->level,
3233                                bch2_btree_ids[path->btree_id]);
3234                         bch2_bpos_to_text(out, path->pos);
3235                         pr_buf(out, "\n");
3236
3237                         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3238                                 if (btree_node_locked(path, l)) {
3239                                         pr_buf(out, "    %s l=%u ",
3240                                                btree_node_intent_locked(path, l) ? "i" : "r", l);
3241                                         bch2_btree_path_node_to_text(out,
3242                                                         (void *) path->l[l].b,
3243                                                         path->cached);
3244                                         pr_buf(out, "\n");
3245                                 }
3246                         }
3247                 }
3248
3249                 b = READ_ONCE(trans->locking);
3250                 if (b) {
3251                         path = &trans->paths[trans->locking_path_idx];
3252                         pr_buf(out, "  locking path %u %c l=%u %s:",
3253                                trans->locking_path_idx,
3254                                path->cached ? 'c' : 'b',
3255                                trans->locking_level,
3256                                bch2_btree_ids[trans->locking_btree_id]);
3257                         bch2_bpos_to_text(out, trans->locking_pos);
3258
3259                         pr_buf(out, " node ");
3260                         bch2_btree_path_node_to_text(out,
3261                                         (void *) b, path->cached);
3262                         pr_buf(out, "\n");
3263                 }
3264         }
3265         mutex_unlock(&c->btree_trans_lock);
3266 }
3267
3268 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3269 {
3270         if (c->btree_trans_barrier_initialized)
3271                 cleanup_srcu_struct(&c->btree_trans_barrier);
3272         mempool_exit(&c->btree_trans_mem_pool);
3273         mempool_exit(&c->btree_paths_pool);
3274 }
3275
3276 int bch2_fs_btree_iter_init(struct bch_fs *c)
3277 {
3278         unsigned nr = BTREE_ITER_MAX;
3279         int ret;
3280
3281         INIT_LIST_HEAD(&c->btree_trans_list);
3282         mutex_init(&c->btree_trans_lock);
3283
3284         ret   = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3285                         sizeof(struct btree_path) * nr +
3286                         sizeof(struct btree_insert_entry) * nr) ?:
3287                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3288                                           BTREE_TRANS_MEM_MAX) ?:
3289                 init_srcu_struct(&c->btree_trans_barrier);
3290         if (!ret)
3291                 c->btree_trans_barrier_initialized = true;
3292         return ret;
3293 }