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 "replicas.h"
  16
  17 #include <linux/prefetch.h>
  18 #include <trace/events/bcachefs.h>
  19
  20 static void btree_iter_set_search_pos(struct btree_iter *, struct bpos);
  21 static struct btree_iter *btree_iter_child_alloc(struct btree_iter *, unsigned long);
  22 static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *);
  23 static void btree_iter_copy(struct btree_iter *, struct btree_iter *);
  24
  25 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  26 {
  27         EBUG_ON(btree_iter_type(iter) == BTREE_ITER_NODES);
  28
  29         /* Are we iterating over keys in all snapshots? */
  30         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  31                 p = bpos_successor(p);
  32         } else {
  33                 p = bpos_nosnap_successor(p);
  34                 p.snapshot = iter->snapshot;
  35         }
  36
  37         return p;
  38 }
  39
  40 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  41 {
  42         EBUG_ON(btree_iter_type(iter) == BTREE_ITER_NODES);
  43
  44         /* Are we iterating over keys in all snapshots? */
  45         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  46                 p = bpos_predecessor(p);
  47         } else {
  48                 p = bpos_nosnap_predecessor(p);
  49                 p.snapshot = iter->snapshot;
  50         }
  51
  52         return p;
  53 }
  54
  55 static inline bool is_btree_node(struct btree_iter *iter, unsigned l)
  56 {
  57         return l < BTREE_MAX_DEPTH &&
  58                 (unsigned long) iter->l[l].b >= 128;
  59 }
  60
  61 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
  62 {
  63         struct bpos pos = iter->pos;
  64
  65         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
  66             bkey_cmp(pos, POS_MAX))
  67                 pos = bkey_successor(iter, pos);
  68         return pos;
  69 }
  70
  71 static inline bool btree_iter_pos_before_node(struct btree_iter *iter,
  72                                               struct btree *b)
  73 {
  74         return bpos_cmp(iter->real_pos, b->data->min_key) < 0;
  75 }
  76
  77 static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
  78                                              struct btree *b)
  79 {
  80         return bpos_cmp(b->key.k.p, iter->real_pos) < 0;
  81 }
  82
  83 static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
  84                                           struct btree *b)
  85 {
  86         return iter->btree_id == b->c.btree_id &&
  87                 !btree_iter_pos_before_node(iter, b) &&
  88                 !btree_iter_pos_after_node(iter, b);
  89 }
  90
  91 /* Btree node locking: */
  92
  93 void bch2_btree_node_unlock_write(struct btree *b, struct btree_iter *iter)
  94 {
  95         bch2_btree_node_unlock_write_inlined(b, iter);
  96 }
  97
  98 void __bch2_btree_node_lock_write(struct btree *b, struct btree_iter *iter)
  99 {
 100         struct btree_iter *linked;
 101         unsigned readers = 0;
 102
 103         EBUG_ON(!btree_node_intent_locked(iter, b->c.level));
 104
 105         trans_for_each_iter(iter->trans, linked)
 106                 if (linked->l[b->c.level].b == b &&
 107                     btree_node_read_locked(linked, b->c.level))
 108                         readers++;
 109
 110         /*
 111          * Must drop our read locks before calling six_lock_write() -
 112          * six_unlock() won't do wakeups until the reader count
 113          * goes to 0, and it's safe because we have the node intent
 114          * locked:
 115          */
 116         atomic64_sub(__SIX_VAL(read_lock, readers),
 117                      &b->c.lock.state.counter);
 118         btree_node_lock_type(iter->trans->c, b, SIX_LOCK_write);
 119         atomic64_add(__SIX_VAL(read_lock, readers),
 120                      &b->c.lock.state.counter);
 121 }
 122
 123 bool __bch2_btree_node_relock(struct btree_iter *iter, unsigned level)
 124 {
 125         struct btree *b = btree_iter_node(iter, level);
 126         int want = __btree_lock_want(iter, level);
 127
 128         if (!is_btree_node(iter, level))
 129                 return false;
 130
 131         if (race_fault())
 132                 return false;
 133
 134         if (six_relock_type(&b->c.lock, want, iter->l[level].lock_seq) ||
 135             (btree_node_lock_seq_matches(iter, b, level) &&
 136              btree_node_lock_increment(iter->trans, b, level, want))) {
 137                 mark_btree_node_locked(iter, level, want);
 138                 return true;
 139         } else {
 140                 return false;
 141         }
 142 }
 143
 144 static bool bch2_btree_node_upgrade(struct btree_iter *iter, unsigned level)
 145 {
 146         struct btree *b = iter->l[level].b;
 147
 148         EBUG_ON(btree_lock_want(iter, level) != BTREE_NODE_INTENT_LOCKED);
 149
 150         if (!is_btree_node(iter, level))
 151                 return false;
 152
 153         if (btree_node_intent_locked(iter, level))
 154                 return true;
 155
 156         if (race_fault())
 157                 return false;
 158
 159         if (btree_node_locked(iter, level)
 160             ? six_lock_tryupgrade(&b->c.lock)
 161             : six_relock_type(&b->c.lock, SIX_LOCK_intent, iter->l[level].lock_seq))
 162                 goto success;
 163
 164         if (btree_node_lock_seq_matches(iter, b, level) &&
 165             btree_node_lock_increment(iter->trans, b, level, BTREE_NODE_INTENT_LOCKED)) {
 166                 btree_node_unlock(iter, level);
 167                 goto success;
 168         }
 169
 170         return false;
 171 success:
 172         mark_btree_node_intent_locked(iter, level);
 173         return true;
 174 }
 175
 176 static inline bool btree_iter_get_locks(struct btree_iter *iter, bool upgrade,
 177                                         unsigned long trace_ip)
 178 {
 179         unsigned l = iter->level;
 180         int fail_idx = -1;
 181
 182         do {
 183                 if (!btree_iter_node(iter, l))
 184                         break;
 185
 186                 if (!(upgrade
 187                       ? bch2_btree_node_upgrade(iter, l)
 188                       : bch2_btree_node_relock(iter, l))) {
 189                         (upgrade
 190                          ? trace_node_upgrade_fail
 191                          : trace_node_relock_fail)(iter->trans->ip, trace_ip,
 192                                         iter->btree_id, &iter->real_pos,
 193                                         l, iter->l[l].lock_seq,
 194                                         is_btree_node(iter, l)
 195                                         ? 0
 196                                         : (unsigned long) iter->l[l].b,
 197                                         is_btree_node(iter, l)
 198                                         ? iter->l[l].b->c.lock.state.seq
 199                                         : 0);
 200
 201                         fail_idx = l;
 202                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
 203                 }
 204
 205                 l++;
 206         } while (l < iter->locks_want);
 207
 208         /*
 209          * When we fail to get a lock, we have to ensure that any child nodes
 210          * can't be relocked so bch2_btree_iter_traverse has to walk back up to
 211          * the node that we failed to relock:
 212          */
 213         while (fail_idx >= 0) {
 214                 btree_node_unlock(iter, fail_idx);
 215                 iter->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS;
 216                 --fail_idx;
 217         }
 218
 219         if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
 220                 iter->uptodate = BTREE_ITER_NEED_PEEK;
 221
 222         bch2_btree_trans_verify_locks(iter->trans);
 223
 224         return iter->uptodate < BTREE_ITER_NEED_RELOCK;
 225 }
 226
 227 static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b,
 228                                   enum btree_iter_type type)
 229 {
 230         return  type != BTREE_ITER_CACHED
 231                 ? container_of(_b, struct btree, c)->key.k.p
 232                 : container_of(_b, struct bkey_cached, c)->key.pos;
 233 }
 234
 235 /* Slowpath: */
 236 bool __bch2_btree_node_lock(struct btree *b, struct bpos pos,
 237                             unsigned level, struct btree_iter *iter,
 238                             enum six_lock_type type,
 239                             six_lock_should_sleep_fn should_sleep_fn, void *p,
 240                             unsigned long ip)
 241 {
 242         struct btree_trans *trans = iter->trans;
 243         struct btree_iter *linked, *deadlock_iter = NULL;
 244         u64 start_time = local_clock();
 245         unsigned reason = 9;
 246         bool ret;
 247
 248         /* Check if it's safe to block: */
 249         trans_for_each_iter(trans, linked) {
 250                 if (!linked->nodes_locked)
 251                         continue;
 252
 253                 /*
 254                  * Can't block taking an intent lock if we have _any_ nodes read
 255                  * locked:
 256                  *
 257                  * - Our read lock blocks another thread with an intent lock on
 258                  *   the same node from getting a write lock, and thus from
 259                  *   dropping its intent lock
 260                  *
 261                  * - And the other thread may have multiple nodes intent locked:
 262                  *   both the node we want to intent lock, and the node we
 263                  *   already have read locked - deadlock:
 264                  */
 265                 if (type == SIX_LOCK_intent &&
 266                     linked->nodes_locked != linked->nodes_intent_locked) {
 267                         deadlock_iter = linked;
 268                         reason = 1;
 269                 }
 270
 271                 if (linked->btree_id != iter->btree_id) {
 272                         if (linked->btree_id > iter->btree_id) {
 273                                 deadlock_iter = linked;
 274                                 reason = 3;
 275                         }
 276                         continue;
 277                 }
 278
 279                 /*
 280                  * Within the same btree, cached iterators come before non
 281                  * cached iterators:
 282                  */
 283                 if (btree_iter_is_cached(linked) != btree_iter_is_cached(iter)) {
 284                         if (btree_iter_is_cached(iter)) {
 285                                 deadlock_iter = linked;
 286                                 reason = 4;
 287                         }
 288                         continue;
 289                 }
 290
 291                 /*
 292                  * Interior nodes must be locked before their descendants: if
 293                  * another iterator has possible descendants locked of the node
 294                  * we're about to lock, it must have the ancestors locked too:
 295                  */
 296                 if (level > __fls(linked->nodes_locked)) {
 297                         deadlock_iter = linked;
 298                         reason = 5;
 299                 }
 300
 301                 /* Must lock btree nodes in key order: */
 302                 if (btree_node_locked(linked, level) &&
 303                     bpos_cmp(pos, btree_node_pos((void *) linked->l[level].b,
 304                                                  btree_iter_type(linked))) <= 0) {
 305                         deadlock_iter = linked;
 306                         reason = 7;
 307                         BUG_ON(trans->in_traverse_all);
 308                 }
 309         }
 310
 311         if (unlikely(deadlock_iter)) {
 312                 trace_trans_restart_would_deadlock(iter->trans->ip, ip,
 313                                 trans->in_traverse_all, reason,
 314                                 deadlock_iter->btree_id,
 315                                 btree_iter_type(deadlock_iter),
 316                                 &deadlock_iter->real_pos,
 317                                 iter->btree_id,
 318                                 btree_iter_type(iter),
 319                                 &pos);
 320                 return false;
 321         }
 322
 323         if (six_trylock_type(&b->c.lock, type))
 324                 return true;
 325
 326 #ifdef CONFIG_BCACHEFS_DEBUG
 327         trans->locking_iter_idx = iter->idx;
 328         trans->locking_pos      = pos;
 329         trans->locking_btree_id = iter->btree_id;
 330         trans->locking_level    = level;
 331         trans->locking          = b;
 332 #endif
 333
 334         ret = six_lock_type(&b->c.lock, type, should_sleep_fn, p) == 0;
 335
 336 #ifdef CONFIG_BCACHEFS_DEBUG
 337         trans->locking = NULL;
 338 #endif
 339         if (ret)
 340                 bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)],
 341                                        start_time);
 342         return ret;
 343 }
 344
 345 /* Btree iterator locking: */
 346
 347 #ifdef CONFIG_BCACHEFS_DEBUG
 348 static void bch2_btree_iter_verify_locks(struct btree_iter *iter)
 349 {
 350         unsigned l;
 351
 352         if (!(iter->trans->iters_linked & (1ULL << iter->idx))) {
 353                 BUG_ON(iter->nodes_locked);
 354                 return;
 355         }
 356
 357         for (l = 0; btree_iter_node(iter, l); l++) {
 358                 if (iter->uptodate >= BTREE_ITER_NEED_RELOCK &&
 359                     !btree_node_locked(iter, l))
 360                         continue;
 361
 362                 BUG_ON(btree_lock_want(iter, l) !=
 363                        btree_node_locked_type(iter, l));
 364         }
 365 }
 366
 367 void bch2_btree_trans_verify_locks(struct btree_trans *trans)
 368 {
 369         struct btree_iter *iter;
 370
 371         trans_for_each_iter(trans, iter)
 372                 bch2_btree_iter_verify_locks(iter);
 373 }
 374 #else
 375 static inline void bch2_btree_iter_verify_locks(struct btree_iter *iter) {}
 376 #endif
 377
 378 __flatten
 379 bool bch2_btree_iter_relock(struct btree_iter *iter, unsigned long trace_ip)
 380 {
 381         return btree_iter_get_locks(iter, false, trace_ip);
 382 }
 383
 384 bool __bch2_btree_iter_upgrade(struct btree_iter *iter,
 385                                unsigned new_locks_want)
 386 {
 387         struct btree_iter *linked;
 388
 389         EBUG_ON(iter->locks_want >= new_locks_want);
 390
 391         iter->locks_want = new_locks_want;
 392
 393         if (btree_iter_get_locks(iter, true, _THIS_IP_))
 394                 return true;
 395
 396         /*
 397          * XXX: this is ugly - we'd prefer to not be mucking with other
 398          * iterators in the btree_trans here.
 399          *
 400          * On failure to upgrade the iterator, setting iter->locks_want and
 401          * calling get_locks() is sufficient to make bch2_btree_iter_traverse()
 402          * get the locks we want on transaction restart.
 403          *
 404          * But if this iterator was a clone, on transaction restart what we did
 405          * to this iterator isn't going to be preserved.
 406          *
 407          * Possibly we could add an iterator field for the parent iterator when
 408          * an iterator is a copy - for now, we'll just upgrade any other
 409          * iterators with the same btree id.
 410          *
 411          * The code below used to be needed to ensure ancestor nodes get locked
 412          * before interior nodes - now that's handled by
 413          * bch2_btree_iter_traverse_all().
 414          */
 415         trans_for_each_iter(iter->trans, linked)
 416                 if (linked != iter &&
 417                     btree_iter_type(linked) == btree_iter_type(iter) &&
 418                     linked->btree_id == iter->btree_id &&
 419                     linked->locks_want < new_locks_want) {
 420                         linked->locks_want = new_locks_want;
 421                         btree_iter_get_locks(linked, true, _THIS_IP_);
 422                 }
 423
 424         return false;
 425 }
 426
 427 void __bch2_btree_iter_downgrade(struct btree_iter *iter,
 428                                  unsigned new_locks_want)
 429 {
 430         unsigned l;
 431
 432         EBUG_ON(iter->locks_want < new_locks_want);
 433
 434         iter->locks_want = new_locks_want;
 435
 436         while (iter->nodes_locked &&
 437                (l = __fls(iter->nodes_locked)) >= iter->locks_want) {
 438                 if (l > iter->level) {
 439                         btree_node_unlock(iter, l);
 440                 } else {
 441                         if (btree_node_intent_locked(iter, l)) {
 442                                 six_lock_downgrade(&iter->l[l].b->c.lock);
 443                                 iter->nodes_intent_locked ^= 1 << l;
 444                         }
 445                         break;
 446                 }
 447         }
 448
 449         bch2_btree_trans_verify_locks(iter->trans);
 450 }
 451
 452 void bch2_trans_downgrade(struct btree_trans *trans)
 453 {
 454         struct btree_iter *iter;
 455
 456         trans_for_each_iter(trans, iter)
 457                 bch2_btree_iter_downgrade(iter);
 458 }
 459
 460 /* Btree transaction locking: */
 461
 462 static inline bool btree_iter_should_be_locked(struct btree_trans *trans,
 463                                                struct btree_iter *iter)
 464 {
 465         return (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) ||
 466                 iter->should_be_locked;
 467 }
 468
 469 bool bch2_trans_relock(struct btree_trans *trans)
 470 {
 471         struct btree_iter *iter;
 472
 473         trans_for_each_iter(trans, iter)
 474                 if (!bch2_btree_iter_relock(iter, _RET_IP_) &&
 475                     btree_iter_should_be_locked(trans, iter)) {
 476                         trace_trans_restart_relock(trans->ip, _RET_IP_,
 477                                         iter->btree_id, &iter->real_pos);
 478                         return false;
 479                 }
 480         return true;
 481 }
 482
 483 void bch2_trans_unlock(struct btree_trans *trans)
 484 {
 485         struct btree_iter *iter;
 486
 487         trans_for_each_iter(trans, iter)
 488                 __bch2_btree_iter_unlock(iter);
 489
 490         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
 491 }
 492
 493 /* Btree iterator: */
 494
 495 #ifdef CONFIG_BCACHEFS_DEBUG
 496
 497 static void bch2_btree_iter_verify_cached(struct btree_iter *iter)
 498 {
 499         struct bkey_cached *ck;
 500         bool locked = btree_node_locked(iter, 0);
 501
 502         if (!bch2_btree_node_relock(iter, 0))
 503                 return;
 504
 505         ck = (void *) iter->l[0].b;
 506         BUG_ON(ck->key.btree_id != iter->btree_id ||
 507                bkey_cmp(ck->key.pos, iter->pos));
 508
 509         if (!locked)
 510                 btree_node_unlock(iter, 0);
 511 }
 512
 513 static void bch2_btree_iter_verify_level(struct btree_iter *iter,
 514                                          unsigned level)
 515 {
 516         struct btree_iter_level *l;
 517         struct btree_node_iter tmp;
 518         bool locked;
 519         struct bkey_packed *p, *k;
 520         char buf1[100], buf2[100], buf3[100];
 521         const char *msg;
 522
 523         if (!bch2_debug_check_iterators)
 524                 return;
 525
 526         l       = &iter->l[level];
 527         tmp     = l->iter;
 528         locked  = btree_node_locked(iter, level);
 529
 530         if (btree_iter_type(iter) == BTREE_ITER_CACHED) {
 531                 if (!level)
 532                         bch2_btree_iter_verify_cached(iter);
 533                 return;
 534         }
 535
 536         BUG_ON(iter->level < iter->min_depth);
 537
 538         if (!btree_iter_node(iter, level))
 539                 return;
 540
 541         if (!bch2_btree_node_relock(iter, level))
 542                 return;
 543
 544         BUG_ON(!btree_iter_pos_in_node(iter, l->b));
 545
 546         /*
 547          * node iterators don't use leaf node iterator:
 548          */
 549         if (btree_iter_type(iter) == BTREE_ITER_NODES &&
 550             level <= iter->min_depth)
 551                 goto unlock;
 552
 553         bch2_btree_node_iter_verify(&l->iter, l->b);
 554
 555         /*
 556          * For interior nodes, the iterator will have skipped past
 557          * deleted keys:
 558          *
 559          * For extents, the iterator may have skipped past deleted keys (but not
 560          * whiteouts)
 561          */
 562         p = level || btree_node_type_is_extents(iter->btree_id)
 563                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 564                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 565         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 566
 567         if (p && bkey_iter_pos_cmp(l->b, p, &iter->real_pos) >= 0) {
 568                 msg = "before";
 569                 goto err;
 570         }
 571
 572         if (k && bkey_iter_pos_cmp(l->b, k, &iter->real_pos) < 0) {
 573                 msg = "after";
 574                 goto err;
 575         }
 576 unlock:
 577         if (!locked)
 578                 btree_node_unlock(iter, level);
 579         return;
 580 err:
 581         strcpy(buf2, "(none)");
 582         strcpy(buf3, "(none)");
 583
 584         bch2_bpos_to_text(&PBUF(buf1), iter->real_pos);
 585
 586         if (p) {
 587                 struct bkey uk = bkey_unpack_key(l->b, p);
 588                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 589         }
 590
 591         if (k) {
 592                 struct bkey uk = bkey_unpack_key(l->b, k);
 593                 bch2_bkey_to_text(&PBUF(buf3), &uk);
 594         }
 595
 596         panic("iterator should be %s key at level %u:\n"
 597               "iter pos %s\n"
 598               "prev key %s\n"
 599               "cur  key %s\n",
 600               msg, level, buf1, buf2, buf3);
 601 }
 602
 603 static void bch2_btree_iter_verify(struct btree_iter *iter)
 604 {
 605         struct btree_trans *trans = iter->trans;
 606         struct bch_fs *c = trans->c;
 607         enum btree_iter_type type = btree_iter_type(iter);
 608         unsigned i;
 609
 610         EBUG_ON(iter->btree_id >= BTREE_ID_NR);
 611
 612         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 613                iter->pos.snapshot != iter->snapshot);
 614
 615         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 616                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 617
 618         BUG_ON(type == BTREE_ITER_NODES &&
 619                !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 620
 621         BUG_ON(type != BTREE_ITER_NODES &&
 622                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 623                !btree_type_has_snapshots(iter->btree_id));
 624
 625         for (i = 0; i < (type != BTREE_ITER_CACHED ? BTREE_MAX_DEPTH : 1); i++) {
 626                 if (!iter->l[i].b) {
 627                         BUG_ON(c->btree_roots[iter->btree_id].b->c.level > i);
 628                         break;
 629                 }
 630
 631                 bch2_btree_iter_verify_level(iter, i);
 632         }
 633
 634         bch2_btree_iter_verify_locks(iter);
 635 }
 636
 637 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 638 {
 639         enum btree_iter_type type = btree_iter_type(iter);
 640
 641         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 642                iter->pos.snapshot != iter->snapshot);
 643
 644         BUG_ON((type == BTREE_ITER_KEYS ||
 645                 type == BTREE_ITER_CACHED) &&
 646                (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
 647                 bkey_cmp(iter->pos, iter->k.p) > 0));
 648 }
 649
 650 void bch2_btree_trans_verify_iters(struct btree_trans *trans, struct btree *b)
 651 {
 652         struct btree_iter *iter;
 653
 654         if (!bch2_debug_check_iterators)
 655                 return;
 656
 657         trans_for_each_iter_with_node(trans, b, iter)
 658                 bch2_btree_iter_verify_level(iter, b->c.level);
 659 }
 660
 661 #else
 662
 663 static inline void bch2_btree_iter_verify_level(struct btree_iter *iter, unsigned l) {}
 664 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 665 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 666
 667 #endif
 668
 669 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 670                                         struct btree *b,
 671                                         struct bset_tree *t,
 672                                         struct bkey_packed *k)
 673 {
 674         struct btree_node_iter_set *set;
 675
 676         btree_node_iter_for_each(iter, set)
 677                 if (set->end == t->end_offset) {
 678                         set->k = __btree_node_key_to_offset(b, k);
 679                         bch2_btree_node_iter_sort(iter, b);
 680                         return;
 681                 }
 682
 683         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 684 }
 685
 686 static void __bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 687                                                struct btree *b,
 688                                                struct bkey_packed *where)
 689 {
 690         struct btree_iter_level *l = &iter->l[b->c.level];
 691
 692         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 693                 return;
 694
 695         if (bkey_iter_pos_cmp(l->b, where, &iter->real_pos) < 0)
 696                 bch2_btree_node_iter_advance(&l->iter, l->b);
 697
 698         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 699 }
 700
 701 void bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 702                                       struct btree *b,
 703                                       struct bkey_packed *where)
 704 {
 705         struct btree_iter *linked;
 706
 707         trans_for_each_iter_with_node(iter->trans, b, linked) {
 708                 __bch2_btree_iter_fix_key_modified(linked, b, where);
 709                 bch2_btree_iter_verify_level(linked, b->c.level);
 710         }
 711 }
 712
 713 static void __bch2_btree_node_iter_fix(struct btree_iter *iter,
 714                                       struct btree *b,
 715                                       struct btree_node_iter *node_iter,
 716                                       struct bset_tree *t,
 717                                       struct bkey_packed *where,
 718                                       unsigned clobber_u64s,
 719                                       unsigned new_u64s)
 720 {
 721         const struct bkey_packed *end = btree_bkey_last(b, t);
 722         struct btree_node_iter_set *set;
 723         unsigned offset = __btree_node_key_to_offset(b, where);
 724         int shift = new_u64s - clobber_u64s;
 725         unsigned old_end = t->end_offset - shift;
 726         unsigned orig_iter_pos = node_iter->data[0].k;
 727         bool iter_current_key_modified =
 728                 orig_iter_pos >= offset &&
 729                 orig_iter_pos <= offset + clobber_u64s;
 730
 731         btree_node_iter_for_each(node_iter, set)
 732                 if (set->end == old_end)
 733                         goto found;
 734
 735         /* didn't find the bset in the iterator - might have to readd it: */
 736         if (new_u64s &&
 737             bkey_iter_pos_cmp(b, where, &iter->real_pos) >= 0) {
 738                 bch2_btree_node_iter_push(node_iter, b, where, end);
 739                 goto fixup_done;
 740         } else {
 741                 /* Iterator is after key that changed */
 742                 return;
 743         }
 744 found:
 745         set->end = t->end_offset;
 746
 747         /* Iterator hasn't gotten to the key that changed yet: */
 748         if (set->k < offset)
 749                 return;
 750
 751         if (new_u64s &&
 752             bkey_iter_pos_cmp(b, where, &iter->real_pos) >= 0) {
 753                 set->k = offset;
 754         } else if (set->k < offset + clobber_u64s) {
 755                 set->k = offset + new_u64s;
 756                 if (set->k == set->end)
 757                         bch2_btree_node_iter_set_drop(node_iter, set);
 758         } else {
 759                 /* Iterator is after key that changed */
 760                 set->k = (int) set->k + shift;
 761                 return;
 762         }
 763
 764         bch2_btree_node_iter_sort(node_iter, b);
 765 fixup_done:
 766         if (node_iter->data[0].k != orig_iter_pos)
 767                 iter_current_key_modified = true;
 768
 769         /*
 770          * When a new key is added, and the node iterator now points to that
 771          * key, the iterator might have skipped past deleted keys that should
 772          * come after the key the iterator now points to. We have to rewind to
 773          * before those deleted keys - otherwise
 774          * bch2_btree_node_iter_prev_all() breaks:
 775          */
 776         if (!bch2_btree_node_iter_end(node_iter) &&
 777             iter_current_key_modified &&
 778             (b->c.level ||
 779              btree_node_type_is_extents(iter->btree_id))) {
 780                 struct bset_tree *t;
 781                 struct bkey_packed *k, *k2, *p;
 782
 783                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 784
 785                 for_each_bset(b, t) {
 786                         bool set_pos = false;
 787
 788                         if (node_iter->data[0].end == t->end_offset)
 789                                 continue;
 790
 791                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 792
 793                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 794                                bkey_iter_cmp(b, k, p) < 0) {
 795                                 k2 = p;
 796                                 set_pos = true;
 797                         }
 798
 799                         if (set_pos)
 800                                 btree_node_iter_set_set_pos(node_iter,
 801                                                             b, t, k2);
 802                 }
 803         }
 804
 805         if (!b->c.level &&
 806             node_iter == &iter->l[0].iter &&
 807             iter_current_key_modified)
 808                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 809 }
 810
 811 void bch2_btree_node_iter_fix(struct btree_iter *iter,
 812                               struct btree *b,
 813                               struct btree_node_iter *node_iter,
 814                               struct bkey_packed *where,
 815                               unsigned clobber_u64s,
 816                               unsigned new_u64s)
 817 {
 818         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 819         struct btree_iter *linked;
 820
 821         if (node_iter != &iter->l[b->c.level].iter) {
 822                 __bch2_btree_node_iter_fix(iter, b, node_iter, t,
 823                                            where, clobber_u64s, new_u64s);
 824
 825                 if (bch2_debug_check_iterators)
 826                         bch2_btree_node_iter_verify(node_iter, b);
 827         }
 828
 829         trans_for_each_iter_with_node(iter->trans, b, linked) {
 830                 __bch2_btree_node_iter_fix(linked, b,
 831                                            &linked->l[b->c.level].iter, t,
 832                                            where, clobber_u64s, new_u64s);
 833                 bch2_btree_iter_verify_level(linked, b->c.level);
 834         }
 835 }
 836
 837 static inline struct bkey_s_c __btree_iter_unpack(struct btree_iter *iter,
 838                                                   struct btree_iter_level *l,
 839                                                   struct bkey *u,
 840                                                   struct bkey_packed *k)
 841 {
 842         struct bkey_s_c ret;
 843
 844         if (unlikely(!k)) {
 845                 /*
 846                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 847                  * a hole
 848                  */
 849                 u->type = KEY_TYPE_deleted;
 850                 return bkey_s_c_null;
 851         }
 852
 853         ret = bkey_disassemble(l->b, k, u);
 854
 855         /*
 856          * XXX: bch2_btree_bset_insert_key() generates invalid keys when we
 857          * overwrite extents - it sets k->type = KEY_TYPE_deleted on the key
 858          * being overwritten but doesn't change k->size. But this is ok, because
 859          * those keys are never written out, we just have to avoid a spurious
 860          * assertion here:
 861          */
 862         if (bch2_debug_check_bkeys && !bkey_deleted(ret.k))
 863                 bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
 864
 865         return ret;
 866 }
 867
 868 /* peek_all() doesn't skip deleted keys */
 869 static inline struct bkey_s_c btree_iter_level_peek_all(struct btree_iter *iter,
 870                                                         struct btree_iter_level *l)
 871 {
 872         return __btree_iter_unpack(iter, l, &iter->k,
 873                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 874 }
 875
 876 static inline struct bkey_s_c btree_iter_level_peek(struct btree_iter *iter,
 877                                                     struct btree_iter_level *l)
 878 {
 879         struct bkey_s_c k = __btree_iter_unpack(iter, l, &iter->k,
 880                         bch2_btree_node_iter_peek(&l->iter, l->b));
 881
 882         iter->real_pos = k.k ? k.k->p : l->b->key.k.p;
 883         return k;
 884 }
 885
 886 static inline struct bkey_s_c btree_iter_level_prev(struct btree_iter *iter,
 887                                                     struct btree_iter_level *l)
 888 {
 889         struct bkey_s_c k = __btree_iter_unpack(iter, l, &iter->k,
 890                         bch2_btree_node_iter_prev(&l->iter, l->b));
 891
 892         iter->real_pos = k.k ? k.k->p : l->b->data->min_key;
 893         return k;
 894 }
 895
 896 static inline bool btree_iter_advance_to_pos(struct btree_iter *iter,
 897                                              struct btree_iter_level *l,
 898                                              int max_advance)
 899 {
 900         struct bkey_packed *k;
 901         int nr_advanced = 0;
 902
 903         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 904                bkey_iter_pos_cmp(l->b, k, &iter->real_pos) < 0) {
 905                 if (max_advance > 0 && nr_advanced >= max_advance)
 906                         return false;
 907
 908                 bch2_btree_node_iter_advance(&l->iter, l->b);
 909                 nr_advanced++;
 910         }
 911
 912         return true;
 913 }
 914
 915 /*
 916  * Verify that iterator for parent node points to child node:
 917  */
 918 static void btree_iter_verify_new_node(struct btree_iter *iter, struct btree *b)
 919 {
 920         struct btree_iter_level *l;
 921         unsigned plevel;
 922         bool parent_locked;
 923         struct bkey_packed *k;
 924
 925         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 926                 return;
 927
 928         plevel = b->c.level + 1;
 929         if (!btree_iter_node(iter, plevel))
 930                 return;
 931
 932         parent_locked = btree_node_locked(iter, plevel);
 933
 934         if (!bch2_btree_node_relock(iter, plevel))
 935                 return;
 936
 937         l = &iter->l[plevel];
 938         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 939         if (!k ||
 940             bkey_deleted(k) ||
 941             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
 942                 char buf1[100];
 943                 char buf2[100];
 944                 char buf3[100];
 945                 char buf4[100];
 946                 struct bkey uk = bkey_unpack_key(b, k);
 947
 948                 bch2_dump_btree_node(iter->trans->c, l->b);
 949                 bch2_bpos_to_text(&PBUF(buf1), iter->real_pos);
 950                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 951                 bch2_bpos_to_text(&PBUF(buf3), b->data->min_key);
 952                 bch2_bpos_to_text(&PBUF(buf3), b->data->max_key);
 953                 panic("parent iter doesn't point to new node:\n"
 954                       "iter pos %s %s\n"
 955                       "iter key %s\n"
 956                       "new node %s-%s\n",
 957                       bch2_btree_ids[iter->btree_id], buf1,
 958                       buf2, buf3, buf4);
 959         }
 960
 961         if (!parent_locked)
 962                 btree_node_unlock(iter, b->c.level + 1);
 963 }
 964
 965 static inline void __btree_iter_init(struct btree_iter *iter,
 966                                      unsigned level)
 967 {
 968         struct btree_iter_level *l = &iter->l[level];
 969
 970         bch2_btree_node_iter_init(&l->iter, l->b, &iter->real_pos);
 971
 972         /*
 973          * Iterators to interior nodes should always be pointed at the first non
 974          * whiteout:
 975          */
 976         if (level)
 977                 bch2_btree_node_iter_peek(&l->iter, l->b);
 978
 979         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 980 }
 981
 982 static inline void btree_iter_node_set(struct btree_iter *iter,
 983                                        struct btree *b)
 984 {
 985         BUG_ON(btree_iter_type(iter) == BTREE_ITER_CACHED);
 986
 987         btree_iter_verify_new_node(iter, b);
 988
 989         EBUG_ON(!btree_iter_pos_in_node(iter, b));
 990         EBUG_ON(b->c.lock.state.seq & 1);
 991
 992         iter->l[b->c.level].lock_seq = b->c.lock.state.seq;
 993         iter->l[b->c.level].b = b;
 994         __btree_iter_init(iter, b->c.level);
 995 }
 996
 997 /*
 998  * A btree node is being replaced - update the iterator to point to the new
 999  * node:
1000  */
1001 void bch2_btree_iter_node_replace(struct btree_iter *iter, struct btree *b)
1002 {
1003         enum btree_node_locked_type t;
1004         struct btree_iter *linked;
1005
1006         trans_for_each_iter(iter->trans, linked)
1007                 if (btree_iter_type(linked) != BTREE_ITER_CACHED &&
1008                     btree_iter_pos_in_node(linked, b)) {
1009                         /*
1010                          * bch2_btree_iter_node_drop() has already been called -
1011                          * the old node we're replacing has already been
1012                          * unlocked and the pointer invalidated
1013                          */
1014                         BUG_ON(btree_node_locked(linked, b->c.level));
1015
1016                         t = btree_lock_want(linked, b->c.level);
1017                         if (t != BTREE_NODE_UNLOCKED) {
1018                                 six_lock_increment(&b->c.lock, t);
1019                                 mark_btree_node_locked(linked, b->c.level, t);
1020                         }
1021
1022                         btree_iter_node_set(linked, b);
1023                 }
1024 }
1025
1026 void bch2_btree_iter_node_drop(struct btree_iter *iter, struct btree *b)
1027 {
1028         struct btree_iter *linked;
1029         unsigned level = b->c.level;
1030
1031         trans_for_each_iter(iter->trans, linked)
1032                 if (linked->l[level].b == b) {
1033                         btree_node_unlock(linked, level);
1034                         linked->l[level].b = BTREE_ITER_NO_NODE_DROP;
1035                 }
1036 }
1037
1038 /*
1039  * A btree node has been modified in such a way as to invalidate iterators - fix
1040  * them:
1041  */
1042 void bch2_btree_iter_reinit_node(struct btree_iter *iter, struct btree *b)
1043 {
1044         struct btree_iter *linked;
1045
1046         trans_for_each_iter_with_node(iter->trans, b, linked)
1047                 __btree_iter_init(linked, b->c.level);
1048 }
1049
1050 static int lock_root_check_fn(struct six_lock *lock, void *p)
1051 {
1052         struct btree *b = container_of(lock, struct btree, c.lock);
1053         struct btree **rootp = p;
1054
1055         return b == *rootp ? 0 : -1;
1056 }
1057
1058 static inline int btree_iter_lock_root(struct btree_iter *iter,
1059                                        unsigned depth_want,
1060                                        unsigned long trace_ip)
1061 {
1062         struct bch_fs *c = iter->trans->c;
1063         struct btree *b, **rootp = &c->btree_roots[iter->btree_id].b;
1064         enum six_lock_type lock_type;
1065         unsigned i;
1066
1067         EBUG_ON(iter->nodes_locked);
1068
1069         while (1) {
1070                 b = READ_ONCE(*rootp);
1071                 iter->level = READ_ONCE(b->c.level);
1072
1073                 if (unlikely(iter->level < depth_want)) {
1074                         /*
1075                          * the root is at a lower depth than the depth we want:
1076                          * got to the end of the btree, or we're walking nodes
1077                          * greater than some depth and there are no nodes >=
1078                          * that depth
1079                          */
1080                         iter->level = depth_want;
1081                         for (i = iter->level; i < BTREE_MAX_DEPTH; i++)
1082                                 iter->l[i].b = NULL;
1083                         return 1;
1084                 }
1085
1086                 lock_type = __btree_lock_want(iter, iter->level);
1087                 if (unlikely(!btree_node_lock(b, SPOS_MAX, iter->level,
1088                                               iter, lock_type,
1089                                               lock_root_check_fn, rootp,
1090                                               trace_ip)))
1091                         return -EINTR;
1092
1093                 if (likely(b == READ_ONCE(*rootp) &&
1094                            b->c.level == iter->level &&
1095                            !race_fault())) {
1096                         for (i = 0; i < iter->level; i++)
1097                                 iter->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
1098                         iter->l[iter->level].b = b;
1099                         for (i = iter->level + 1; i < BTREE_MAX_DEPTH; i++)
1100                                 iter->l[i].b = NULL;
1101
1102                         mark_btree_node_locked(iter, iter->level, lock_type);
1103                         btree_iter_node_set(iter, b);
1104                         return 0;
1105                 }
1106
1107                 six_unlock_type(&b->c.lock, lock_type);
1108         }
1109 }
1110
1111 noinline
1112 static void btree_iter_prefetch(struct btree_iter *iter)
1113 {
1114         struct bch_fs *c = iter->trans->c;
1115         struct btree_iter_level *l = &iter->l[iter->level];
1116         struct btree_node_iter node_iter = l->iter;
1117         struct bkey_packed *k;
1118         struct bkey_buf tmp;
1119         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1120                 ? (iter->level > 1 ? 0 :  2)
1121                 : (iter->level > 1 ? 1 : 16);
1122         bool was_locked = btree_node_locked(iter, iter->level);
1123
1124         bch2_bkey_buf_init(&tmp);
1125
1126         while (nr) {
1127                 if (!bch2_btree_node_relock(iter, iter->level))
1128                         break;
1129
1130                 bch2_btree_node_iter_advance(&node_iter, l->b);
1131                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
1132                 if (!k)
1133                         break;
1134
1135                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
1136                 bch2_btree_node_prefetch(c, iter, tmp.k, iter->btree_id,
1137                                          iter->level - 1);
1138         }
1139
1140         if (!was_locked)
1141                 btree_node_unlock(iter, iter->level);
1142
1143         bch2_bkey_buf_exit(&tmp, c);
1144 }
1145
1146 static noinline void btree_node_mem_ptr_set(struct btree_iter *iter,
1147                                             unsigned plevel, struct btree *b)
1148 {
1149         struct btree_iter_level *l = &iter->l[plevel];
1150         bool locked = btree_node_locked(iter, plevel);
1151         struct bkey_packed *k;
1152         struct bch_btree_ptr_v2 *bp;
1153
1154         if (!bch2_btree_node_relock(iter, plevel))
1155                 return;
1156
1157         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1158         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
1159
1160         bp = (void *) bkeyp_val(&l->b->format, k);
1161         bp->mem_ptr = (unsigned long)b;
1162
1163         if (!locked)
1164                 btree_node_unlock(iter, plevel);
1165 }
1166
1167 static __always_inline int btree_iter_down(struct btree_iter *iter,
1168                                            unsigned long trace_ip)
1169 {
1170         struct bch_fs *c = iter->trans->c;
1171         struct btree_iter_level *l = &iter->l[iter->level];
1172         struct btree *b;
1173         unsigned level = iter->level - 1;
1174         enum six_lock_type lock_type = __btree_lock_want(iter, level);
1175         struct bkey_buf tmp;
1176         int ret;
1177
1178         EBUG_ON(!btree_node_locked(iter, iter->level));
1179
1180         bch2_bkey_buf_init(&tmp);
1181         bch2_bkey_buf_unpack(&tmp, c, l->b,
1182                          bch2_btree_node_iter_peek(&l->iter, l->b));
1183
1184         b = bch2_btree_node_get(c, iter, tmp.k, level, lock_type, trace_ip);
1185         ret = PTR_ERR_OR_ZERO(b);
1186         if (unlikely(ret))
1187                 goto err;
1188
1189         mark_btree_node_locked(iter, level, lock_type);
1190         btree_iter_node_set(iter, b);
1191
1192         if (tmp.k->k.type == KEY_TYPE_btree_ptr_v2 &&
1193             unlikely(b != btree_node_mem_ptr(tmp.k)))
1194                 btree_node_mem_ptr_set(iter, level + 1, b);
1195
1196         if (iter->flags & BTREE_ITER_PREFETCH)
1197                 btree_iter_prefetch(iter);
1198
1199         if (btree_node_read_locked(iter, level + 1))
1200                 btree_node_unlock(iter, level + 1);
1201         iter->level = level;
1202
1203         bch2_btree_iter_verify_locks(iter);
1204 err:
1205         bch2_bkey_buf_exit(&tmp, c);
1206         return ret;
1207 }
1208
1209 static int btree_iter_traverse_one(struct btree_iter *, unsigned long);
1210
1211 static int __btree_iter_traverse_all(struct btree_trans *trans, int ret,
1212                                      unsigned long trace_ip)
1213 {
1214         struct bch_fs *c = trans->c;
1215         struct btree_iter *iter;
1216         u8 sorted[BTREE_ITER_MAX];
1217         int i, nr_sorted = 0;
1218         bool relock_fail;
1219
1220         if (trans->in_traverse_all)
1221                 return -EINTR;
1222
1223         trans->in_traverse_all = true;
1224 retry_all:
1225         nr_sorted = 0;
1226         relock_fail = false;
1227
1228         trans_for_each_iter(trans, iter) {
1229                 if (!bch2_btree_iter_relock(iter, _THIS_IP_))
1230                         relock_fail = true;
1231                 sorted[nr_sorted++] = iter->idx;
1232         }
1233
1234         if (!relock_fail) {
1235                 trans->in_traverse_all = false;
1236                 return 0;
1237         }
1238
1239 #define btree_iter_cmp_by_idx(_l, _r)                           \
1240                 btree_iter_lock_cmp(&trans->iters[_l], &trans->iters[_r])
1241
1242         bubble_sort(sorted, nr_sorted, btree_iter_cmp_by_idx);
1243 #undef btree_iter_cmp_by_idx
1244
1245         for (i = nr_sorted - 2; i >= 0; --i) {
1246                 struct btree_iter *iter1 = trans->iters + sorted[i];
1247                 struct btree_iter *iter2 = trans->iters + sorted[i + 1];
1248
1249                 if (iter1->btree_id == iter2->btree_id &&
1250                     iter1->locks_want < iter2->locks_want)
1251                         __bch2_btree_iter_upgrade(iter1, iter2->locks_want);
1252                 else if (!iter1->locks_want && iter2->locks_want)
1253                         __bch2_btree_iter_upgrade(iter1, 1);
1254         }
1255
1256         bch2_trans_unlock(trans);
1257         cond_resched();
1258
1259         if (unlikely(ret == -ENOMEM)) {
1260                 struct closure cl;
1261
1262                 closure_init_stack(&cl);
1263
1264                 do {
1265                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1266                         closure_sync(&cl);
1267                 } while (ret);
1268         }
1269
1270         if (unlikely(ret == -EIO)) {
1271                 trans->error = true;
1272                 goto out;
1273         }
1274
1275         BUG_ON(ret && ret != -EINTR);
1276
1277         /* Now, redo traversals in correct order: */
1278         for (i = 0; i < nr_sorted; i++) {
1279                 unsigned idx = sorted[i];
1280
1281                 /*
1282                  * sucessfully traversing one iterator can cause another to be
1283                  * unlinked, in btree_key_cache_fill()
1284                  */
1285                 if (!(trans->iters_linked & (1ULL << idx)))
1286                         continue;
1287
1288                 ret = btree_iter_traverse_one(&trans->iters[idx], _THIS_IP_);
1289                 if (ret)
1290                         goto retry_all;
1291         }
1292
1293         if (hweight64(trans->iters_live) > 1)
1294                 ret = -EINTR;
1295         else
1296                 trans_for_each_iter(trans, iter)
1297                         if (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) {
1298                                 ret = -EINTR;
1299                                 break;
1300                         }
1301 out:
1302         bch2_btree_cache_cannibalize_unlock(c);
1303
1304         trans->in_traverse_all = false;
1305
1306         trace_trans_traverse_all(trans->ip, trace_ip);
1307         return ret;
1308 }
1309
1310 int bch2_btree_iter_traverse_all(struct btree_trans *trans)
1311 {
1312         return __btree_iter_traverse_all(trans, 0, _RET_IP_);
1313 }
1314
1315 static inline bool btree_iter_good_node(struct btree_iter *iter,
1316                                         unsigned l, int check_pos)
1317 {
1318         if (!is_btree_node(iter, l) ||
1319             !bch2_btree_node_relock(iter, l))
1320                 return false;
1321
1322         if (check_pos < 0 && btree_iter_pos_before_node(iter, iter->l[l].b))
1323                 return false;
1324         if (check_pos > 0 && btree_iter_pos_after_node(iter, iter->l[l].b))
1325                 return false;
1326         return true;
1327 }
1328
1329 static inline unsigned btree_iter_up_until_good_node(struct btree_iter *iter,
1330                                                      int check_pos)
1331 {
1332         unsigned l = iter->level;
1333
1334         while (btree_iter_node(iter, l) &&
1335                !btree_iter_good_node(iter, l, check_pos)) {
1336                 btree_node_unlock(iter, l);
1337                 iter->l[l].b = BTREE_ITER_NO_NODE_UP;
1338                 l++;
1339         }
1340
1341         return l;
1342 }
1343
1344 /*
1345  * This is the main state machine for walking down the btree - walks down to a
1346  * specified depth
1347  *
1348  * Returns 0 on success, -EIO on error (error reading in a btree node).
1349  *
1350  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1351  * stashed in the iterator and returned from bch2_trans_exit().
1352  */
1353 static int btree_iter_traverse_one(struct btree_iter *iter,
1354                                    unsigned long trace_ip)
1355 {
1356         unsigned l, depth_want = iter->level;
1357         int ret = 0;
1358
1359         if (btree_iter_type(iter) == BTREE_ITER_CACHED) {
1360                 ret = bch2_btree_iter_traverse_cached(iter);
1361                 goto out;
1362         }
1363
1364         if (unlikely(iter->level >= BTREE_MAX_DEPTH))
1365                 goto out;
1366
1367         iter->level = btree_iter_up_until_good_node(iter, 0);
1368
1369         /* If we need intent locks, take them too: */
1370         for (l = iter->level + 1;
1371              l < iter->locks_want && btree_iter_node(iter, l);
1372              l++)
1373                 if (!bch2_btree_node_relock(iter, l))
1374                         while (iter->level <= l) {
1375                                 btree_node_unlock(iter, iter->level);
1376                                 iter->l[iter->level].b = BTREE_ITER_NO_NODE_UP;
1377                                 iter->level++;
1378                         }
1379
1380         /*
1381          * Note: iter->nodes[iter->level] may be temporarily NULL here - that
1382          * would indicate to other code that we got to the end of the btree,
1383          * here it indicates that relocking the root failed - it's critical that
1384          * btree_iter_lock_root() comes next and that it can't fail
1385          */
1386         while (iter->level > depth_want) {
1387                 ret = btree_iter_node(iter, iter->level)
1388                         ? btree_iter_down(iter, trace_ip)
1389                         : btree_iter_lock_root(iter, depth_want, trace_ip);
1390                 if (unlikely(ret)) {
1391                         if (ret == 1) {
1392                                 /*
1393                                  * Got to the end of the btree (in
1394                                  * BTREE_ITER_NODES mode)
1395                                  */
1396                                 ret = 0;
1397                                 goto out;
1398                         }
1399
1400                         __bch2_btree_iter_unlock(iter);
1401                         iter->level = depth_want;
1402
1403                         if (ret == -EIO) {
1404                                 iter->flags |= BTREE_ITER_ERROR;
1405                                 iter->l[iter->level].b =
1406                                         BTREE_ITER_NO_NODE_ERROR;
1407                         } else {
1408                                 iter->l[iter->level].b =
1409                                         BTREE_ITER_NO_NODE_DOWN;
1410                         }
1411                         goto out;
1412                 }
1413         }
1414
1415         iter->uptodate = BTREE_ITER_NEED_PEEK;
1416 out:
1417         trace_iter_traverse(iter->trans->ip, trace_ip,
1418                             iter->btree_id, &iter->real_pos, ret);
1419         bch2_btree_iter_verify(iter);
1420         return ret;
1421 }
1422
1423 static int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
1424 {
1425         struct btree_trans *trans = iter->trans;
1426         int ret;
1427
1428         ret =   bch2_trans_cond_resched(trans) ?:
1429                 btree_iter_traverse_one(iter, _RET_IP_);
1430         if (unlikely(ret))
1431                 ret = __btree_iter_traverse_all(trans, ret, _RET_IP_);
1432
1433         return ret;
1434 }
1435
1436 /*
1437  * Note:
1438  * bch2_btree_iter_traverse() is for external users, btree_iter_traverse() is
1439  * for internal btree iterator users
1440  *
1441  * bch2_btree_iter_traverse sets iter->real_pos to iter->pos,
1442  * btree_iter_traverse() does not:
1443  */
1444 static inline int __must_check
1445 btree_iter_traverse(struct btree_iter *iter)
1446 {
1447         return iter->uptodate >= BTREE_ITER_NEED_RELOCK
1448                 ? __bch2_btree_iter_traverse(iter)
1449                 : 0;
1450 }
1451
1452 int __must_check
1453 bch2_btree_iter_traverse(struct btree_iter *iter)
1454 {
1455         int ret;
1456
1457         btree_iter_set_search_pos(iter, btree_iter_search_key(iter));
1458
1459         ret = btree_iter_traverse(iter);
1460         if (ret)
1461                 return ret;
1462
1463         iter->should_be_locked = true;
1464         return 0;
1465 }
1466
1467 /* Iterate across nodes (leaf and interior nodes) */
1468
1469 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1470 {
1471         struct btree *b;
1472         int ret;
1473
1474         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
1475         bch2_btree_iter_verify(iter);
1476
1477         ret = btree_iter_traverse(iter);
1478         if (ret)
1479                 return NULL;
1480
1481         b = btree_iter_node(iter, iter->level);
1482         if (!b)
1483                 return NULL;
1484
1485         BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0);
1486
1487         iter->pos = iter->real_pos = b->key.k.p;
1488
1489         bch2_btree_iter_verify(iter);
1490         iter->should_be_locked = true;
1491
1492         return b;
1493 }
1494
1495 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1496 {
1497         struct btree *b;
1498         int ret;
1499
1500         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
1501         bch2_btree_iter_verify(iter);
1502
1503         /* already got to end? */
1504         if (!btree_iter_node(iter, iter->level))
1505                 return NULL;
1506
1507         bch2_trans_cond_resched(iter->trans);
1508
1509         btree_node_unlock(iter, iter->level);
1510         iter->l[iter->level].b = BTREE_ITER_NO_NODE_UP;
1511         iter->level++;
1512
1513         btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1514         ret = btree_iter_traverse(iter);
1515         if (ret)
1516                 return NULL;
1517
1518         /* got to end? */
1519         b = btree_iter_node(iter, iter->level);
1520         if (!b)
1521                 return NULL;
1522
1523         if (bpos_cmp(iter->pos, b->key.k.p) < 0) {
1524                 /*
1525                  * Haven't gotten to the end of the parent node: go back down to
1526                  * the next child node
1527                  */
1528                 btree_iter_set_search_pos(iter, bpos_successor(iter->pos));
1529
1530                 /* Unlock to avoid screwing up our lock invariants: */
1531                 btree_node_unlock(iter, iter->level);
1532
1533                 iter->level = iter->min_depth;
1534                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1535                 bch2_btree_iter_verify(iter);
1536
1537                 ret = btree_iter_traverse(iter);
1538                 if (ret)
1539                         return NULL;
1540
1541                 b = iter->l[iter->level].b;
1542         }
1543
1544         iter->pos = iter->real_pos = b->key.k.p;
1545
1546         bch2_btree_iter_verify(iter);
1547         iter->should_be_locked = true;
1548
1549         return b;
1550 }
1551
1552 /* Iterate across keys (in leaf nodes only) */
1553
1554 static void btree_iter_set_search_pos(struct btree_iter *iter, struct bpos new_pos)
1555 {
1556 #ifdef CONFIG_BCACHEFS_DEBUG
1557         struct bpos old_pos = iter->real_pos;
1558 #endif
1559         int cmp = bpos_cmp(new_pos, iter->real_pos);
1560         unsigned l = iter->level;
1561
1562         if (!cmp)
1563                 goto out;
1564
1565         iter->real_pos = new_pos;
1566         iter->should_be_locked = false;
1567
1568         if (unlikely(btree_iter_type(iter) == BTREE_ITER_CACHED)) {
1569                 btree_node_unlock(iter, 0);
1570                 iter->l[0].b = BTREE_ITER_NO_NODE_CACHED;
1571                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1572                 return;
1573         }
1574
1575         l = btree_iter_up_until_good_node(iter, cmp);
1576
1577         if (btree_iter_node(iter, l)) {
1578                 /*
1579                  * We might have to skip over many keys, or just a few: try
1580                  * advancing the node iterator, and if we have to skip over too
1581                  * many keys just reinit it (or if we're rewinding, since that
1582                  * is expensive).
1583                  */
1584                 if (cmp < 0 ||
1585                     !btree_iter_advance_to_pos(iter, &iter->l[l], 8))
1586                         __btree_iter_init(iter, l);
1587
1588                 /* Don't leave it locked if we're not supposed to: */
1589                 if (btree_lock_want(iter, l) == BTREE_NODE_UNLOCKED)
1590                         btree_node_unlock(iter, l);
1591         }
1592 out:
1593         if (l != iter->level)
1594                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1595         else
1596                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
1597
1598         bch2_btree_iter_verify(iter);
1599 #ifdef CONFIG_BCACHEFS_DEBUG
1600         trace_iter_set_search_pos(iter->trans->ip, _RET_IP_,
1601                                   iter->btree_id,
1602                                   &old_pos, &new_pos, l);
1603 #endif
1604 }
1605
1606 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1607 {
1608         struct bpos pos = iter->k.p;
1609         bool ret = bpos_cmp(pos, SPOS_MAX) != 0;
1610
1611         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1612                 pos = bkey_successor(iter, pos);
1613         bch2_btree_iter_set_pos(iter, pos);
1614         return ret;
1615 }
1616
1617 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1618 {
1619         struct bpos pos = bkey_start_pos(&iter->k);
1620         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1621                     ? bpos_cmp(pos, POS_MIN)
1622                     : bkey_cmp(pos, POS_MIN)) != 0;
1623
1624         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1625                 pos = bkey_predecessor(iter, pos);
1626         bch2_btree_iter_set_pos(iter, pos);
1627         return ret;
1628 }
1629
1630 static inline bool btree_iter_set_pos_to_next_leaf(struct btree_iter *iter)
1631 {
1632         struct bpos next_pos = iter->l[0].b->key.k.p;
1633         bool ret = bpos_cmp(next_pos, SPOS_MAX) != 0;
1634
1635         /*
1636          * Typically, we don't want to modify iter->pos here, since that
1637          * indicates where we searched from - unless we got to the end of the
1638          * btree, in that case we want iter->pos to reflect that:
1639          */
1640         if (ret)
1641                 btree_iter_set_search_pos(iter, bpos_successor(next_pos));
1642         else
1643                 bch2_btree_iter_set_pos(iter, SPOS_MAX);
1644
1645         return ret;
1646 }
1647
1648 static inline bool btree_iter_set_pos_to_prev_leaf(struct btree_iter *iter)
1649 {
1650         struct bpos next_pos = iter->l[0].b->data->min_key;
1651         bool ret = bpos_cmp(next_pos, POS_MIN) != 0;
1652
1653         if (ret)
1654                 btree_iter_set_search_pos(iter, bpos_predecessor(next_pos));
1655         else
1656                 bch2_btree_iter_set_pos(iter, POS_MIN);
1657
1658         return ret;
1659 }
1660
1661 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter,
1662                                                       struct bpos pos)
1663 {
1664         struct btree_insert_entry *i;
1665
1666         if (!(iter->flags & BTREE_ITER_WITH_UPDATES))
1667                 return NULL;
1668
1669         trans_for_each_update(iter->trans, i)
1670                 if ((cmp_int(iter->btree_id,    i->iter->btree_id) ?:
1671                      bkey_cmp(pos,              i->k->k.p)) <= 0) {
1672                         if (iter->btree_id ==   i->iter->btree_id)
1673                                 return i->k;
1674                         break;
1675                 }
1676
1677         return NULL;
1678 }
1679
1680 /**
1681  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
1682  * current position
1683  */
1684 struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
1685 {
1686         struct bpos search_key = btree_iter_search_key(iter);
1687         struct bkey_i *next_update;
1688         struct bkey_s_c k;
1689         int ret;
1690
1691         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1692         bch2_btree_iter_verify(iter);
1693         bch2_btree_iter_verify_entry_exit(iter);
1694 start:
1695         next_update = btree_trans_peek_updates(iter, search_key);
1696         btree_iter_set_search_pos(iter, search_key);
1697
1698         while (1) {
1699                 ret = btree_iter_traverse(iter);
1700                 if (unlikely(ret))
1701                         return bkey_s_c_err(ret);
1702
1703                 k = btree_iter_level_peek(iter, &iter->l[0]);
1704
1705                 if (next_update &&
1706                     bpos_cmp(next_update->k.p, iter->real_pos) <= 0) {
1707                         iter->k = next_update->k;
1708                         k = bkey_i_to_s_c(next_update);
1709                 }
1710
1711                 if (likely(k.k)) {
1712                         if (bkey_deleted(k.k)) {
1713                                 search_key = bkey_successor(iter, k.k->p);
1714                                 goto start;
1715                         }
1716
1717                         break;
1718                 }
1719
1720                 if (!btree_iter_set_pos_to_next_leaf(iter))
1721                         return bkey_s_c_null;
1722         }
1723
1724         /*
1725          * iter->pos should be mononotically increasing, and always be equal to
1726          * the key we just returned - except extents can straddle iter->pos:
1727          */
1728         if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
1729                 iter->pos = k.k->p;
1730         else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
1731                 iter->pos = bkey_start_pos(k.k);
1732
1733         bch2_btree_iter_verify_entry_exit(iter);
1734         bch2_btree_iter_verify(iter);
1735         iter->should_be_locked = true;
1736         return k;
1737 }
1738
1739 /**
1740  * bch2_btree_iter_next: returns first key greater than iterator's current
1741  * position
1742  */
1743 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
1744 {
1745         if (!bch2_btree_iter_advance(iter))
1746                 return bkey_s_c_null;
1747
1748         return bch2_btree_iter_peek(iter);
1749 }
1750
1751 /**
1752  * bch2_btree_iter_peek_prev: returns first key less than or equal to
1753  * iterator's current position
1754  */
1755 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
1756 {
1757         struct btree_iter_level *l = &iter->l[0];
1758         struct bkey_s_c k;
1759         int ret;
1760
1761         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1762         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
1763         bch2_btree_iter_verify(iter);
1764         bch2_btree_iter_verify_entry_exit(iter);
1765
1766         btree_iter_set_search_pos(iter, iter->pos);
1767
1768         while (1) {
1769                 ret = btree_iter_traverse(iter);
1770                 if (unlikely(ret)) {
1771                         k = bkey_s_c_err(ret);
1772                         goto no_key;
1773                 }
1774
1775                 k = btree_iter_level_peek(iter, l);
1776                 if (!k.k ||
1777                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
1778                      ? bkey_cmp(bkey_start_pos(k.k), iter->pos) >= 0
1779                      : bkey_cmp(k.k->p, iter->pos) > 0))
1780                         k = btree_iter_level_prev(iter, l);
1781
1782                 if (likely(k.k))
1783                         break;
1784
1785                 if (!btree_iter_set_pos_to_prev_leaf(iter)) {
1786                         k = bkey_s_c_null;
1787                         goto no_key;
1788                 }
1789         }
1790
1791         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
1792
1793         /* Extents can straddle iter->pos: */
1794         if (bkey_cmp(k.k->p, iter->pos) < 0)
1795                 iter->pos = k.k->p;
1796 out:
1797         bch2_btree_iter_verify_entry_exit(iter);
1798         bch2_btree_iter_verify(iter);
1799         iter->should_be_locked = true;
1800         return k;
1801 no_key:
1802         /*
1803          * btree_iter_level_peek() may have set iter->k to a key we didn't want, and
1804          * then we errored going to the previous leaf - make sure it's
1805          * consistent with iter->pos:
1806          */
1807         bkey_init(&iter->k);
1808         iter->k.p = iter->pos;
1809         goto out;
1810 }
1811
1812 /**
1813  * bch2_btree_iter_prev: returns first key less than iterator's current
1814  * position
1815  */
1816 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
1817 {
1818         if (!bch2_btree_iter_rewind(iter))
1819                 return bkey_s_c_null;
1820
1821         return bch2_btree_iter_peek_prev(iter);
1822 }
1823
1824 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
1825 {
1826         struct bpos search_key;
1827         struct bkey_s_c k;
1828         int ret;
1829
1830         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS &&
1831                 btree_iter_type(iter) != BTREE_ITER_CACHED);
1832         bch2_btree_iter_verify(iter);
1833         bch2_btree_iter_verify_entry_exit(iter);
1834
1835         /* extents can't span inode numbers: */
1836         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
1837             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
1838                 if (iter->pos.inode == KEY_INODE_MAX)
1839                         return bkey_s_c_null;
1840
1841                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
1842         }
1843
1844         search_key = btree_iter_search_key(iter);
1845         btree_iter_set_search_pos(iter, search_key);
1846
1847         ret = btree_iter_traverse(iter);
1848         if (unlikely(ret))
1849                 return bkey_s_c_err(ret);
1850
1851         if (btree_iter_type(iter) == BTREE_ITER_CACHED ||
1852             !(iter->flags & BTREE_ITER_IS_EXTENTS)) {
1853                 struct bkey_i *next_update;
1854                 struct bkey_cached *ck;
1855
1856                 switch (btree_iter_type(iter)) {
1857                 case BTREE_ITER_KEYS:
1858                         k = btree_iter_level_peek_all(iter, &iter->l[0]);
1859                         EBUG_ON(k.k && bkey_deleted(k.k) && bpos_cmp(k.k->p, iter->pos) == 0);
1860                         break;
1861                 case BTREE_ITER_CACHED:
1862                         ck = (void *) iter->l[0].b;
1863                         EBUG_ON(iter->btree_id != ck->key.btree_id ||
1864                                 bkey_cmp(iter->pos, ck->key.pos));
1865                         BUG_ON(!ck->valid);
1866
1867                         k = bkey_i_to_s_c(ck->k);
1868                         break;
1869                 case BTREE_ITER_NODES:
1870                         BUG();
1871                 }
1872
1873                 next_update = btree_trans_peek_updates(iter, search_key);
1874                 if (next_update &&
1875                     (!k.k || bpos_cmp(next_update->k.p, k.k->p) <= 0)) {
1876                         iter->k = next_update->k;
1877                         k = bkey_i_to_s_c(next_update);
1878                 }
1879         } else {
1880                 if ((iter->flags & BTREE_ITER_INTENT)) {
1881                         struct btree_iter *child =
1882                                 btree_iter_child_alloc(iter, _THIS_IP_);
1883
1884                         btree_iter_copy(child, iter);
1885                         k = bch2_btree_iter_peek(child);
1886
1887                         if (k.k && !bkey_err(k))
1888                                 iter->k = child->k;
1889                 } else {
1890                         struct bpos pos = iter->pos;
1891
1892                         k = bch2_btree_iter_peek(iter);
1893                         iter->pos = pos;
1894                 }
1895
1896                 if (unlikely(bkey_err(k)))
1897                         return k;
1898         }
1899
1900         if (!(iter->flags & BTREE_ITER_IS_EXTENTS)) {
1901                 if (!k.k ||
1902                     ((iter->flags & BTREE_ITER_ALL_SNAPSHOTS)
1903                      ? bpos_cmp(iter->pos, k.k->p)
1904                      : bkey_cmp(iter->pos, k.k->p))) {
1905                         bkey_init(&iter->k);
1906                         iter->k.p = iter->pos;
1907                         k = (struct bkey_s_c) { &iter->k, NULL };
1908                 }
1909         } else {
1910                 struct bpos next = k.k ? bkey_start_pos(k.k) : POS_MAX;
1911
1912                 if (bkey_cmp(iter->pos, next) < 0) {
1913                         bkey_init(&iter->k);
1914                         iter->k.p = iter->pos;
1915                         bch2_key_resize(&iter->k,
1916                                         min_t(u64, KEY_SIZE_MAX,
1917                                               (next.inode == iter->pos.inode
1918                                                ? next.offset
1919                                                : KEY_OFFSET_MAX) -
1920                                               iter->pos.offset));
1921
1922                         k = (struct bkey_s_c) { &iter->k, NULL };
1923                         EBUG_ON(!k.k->size);
1924                 }
1925         }
1926
1927         bch2_btree_iter_verify_entry_exit(iter);
1928         bch2_btree_iter_verify(iter);
1929         iter->should_be_locked = true;
1930
1931         return k;
1932 }
1933
1934 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
1935 {
1936         if (!bch2_btree_iter_advance(iter))
1937                 return bkey_s_c_null;
1938
1939         return bch2_btree_iter_peek_slot(iter);
1940 }
1941
1942 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
1943 {
1944         if (!bch2_btree_iter_rewind(iter))
1945                 return bkey_s_c_null;
1946
1947         return bch2_btree_iter_peek_slot(iter);
1948 }
1949
1950 static inline void bch2_btree_iter_init(struct btree_trans *trans,
1951                         struct btree_iter *iter, enum btree_id btree_id)
1952 {
1953         struct bch_fs *c = trans->c;
1954         unsigned i;
1955
1956         iter->trans                     = trans;
1957         iter->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1958         iter->btree_id                  = btree_id;
1959         iter->real_pos                  = POS_MIN;
1960         iter->level                     = 0;
1961         iter->min_depth                 = 0;
1962         iter->locks_want                = 0;
1963         iter->nodes_locked              = 0;
1964         iter->nodes_intent_locked       = 0;
1965         for (i = 0; i < ARRAY_SIZE(iter->l); i++)
1966                 iter->l[i].b            = BTREE_ITER_NO_NODE_INIT;
1967
1968         prefetch(c->btree_roots[btree_id].b);
1969 }
1970
1971 /* new transactional stuff: */
1972
1973 static void btree_iter_child_free(struct btree_iter *iter)
1974 {
1975         struct btree_iter *child = btree_iter_child(iter);
1976
1977         if (child) {
1978                 bch2_trans_iter_free(iter->trans, child);
1979                 iter->child_idx = U8_MAX;
1980         }
1981 }
1982
1983 static struct btree_iter *btree_iter_child_alloc(struct btree_iter *iter,
1984                                                  unsigned long ip)
1985 {
1986         struct btree_trans *trans = iter->trans;
1987         struct btree_iter *child = btree_iter_child(iter);
1988
1989         if (!child) {
1990                 child = btree_trans_iter_alloc(trans);
1991                 child->ip_allocated     = ip;
1992                 iter->child_idx         = child->idx;
1993
1994                 trans->iters_live       |= 1ULL << child->idx;
1995                 trans->iters_touched    |= 1ULL << child->idx;
1996         }
1997
1998         return child;
1999 }
2000
2001 static inline void __bch2_trans_iter_free(struct btree_trans *trans,
2002                                           unsigned idx)
2003 {
2004         btree_iter_child_free(&trans->iters[idx]);
2005
2006         __bch2_btree_iter_unlock(&trans->iters[idx]);
2007         trans->iters_linked             &= ~(1ULL << idx);
2008         trans->iters_live               &= ~(1ULL << idx);
2009         trans->iters_touched            &= ~(1ULL << idx);
2010 }
2011
2012 int bch2_trans_iter_put(struct btree_trans *trans,
2013                         struct btree_iter *iter)
2014 {
2015         int ret;
2016
2017         if (IS_ERR_OR_NULL(iter))
2018                 return 0;
2019
2020         BUG_ON(trans->iters + iter->idx != iter);
2021         BUG_ON(!btree_iter_live(trans, iter));
2022
2023         ret = btree_iter_err(iter);
2024
2025         if (!(trans->iters_touched & (1ULL << iter->idx)) &&
2026             !(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT))
2027                 __bch2_trans_iter_free(trans, iter->idx);
2028
2029         trans->iters_live       &= ~(1ULL << iter->idx);
2030         return ret;
2031 }
2032
2033 int bch2_trans_iter_free(struct btree_trans *trans,
2034                          struct btree_iter *iter)
2035 {
2036         if (IS_ERR_OR_NULL(iter))
2037                 return 0;
2038
2039         set_btree_iter_dontneed(trans, iter);
2040
2041         return bch2_trans_iter_put(trans, iter);
2042 }
2043
2044 noinline __cold
2045 static void btree_trans_iter_alloc_fail(struct btree_trans *trans)
2046 {
2047
2048         struct btree_iter *iter;
2049         struct btree_insert_entry *i;
2050         char buf[100];
2051
2052         trans_for_each_iter(trans, iter)
2053                 printk(KERN_ERR "iter: btree %s pos %s%s%s%s %pS\n",
2054                        bch2_btree_ids[iter->btree_id],
2055                        (bch2_bpos_to_text(&PBUF(buf), iter->pos), buf),
2056                        btree_iter_live(trans, iter) ? " live" : "",
2057                        (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "",
2058                        iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "",
2059                        (void *) iter->ip_allocated);
2060
2061         trans_for_each_update(trans, i) {
2062                 char buf[300];
2063
2064                 bch2_bkey_val_to_text(&PBUF(buf), trans->c, bkey_i_to_s_c(i->k));
2065                 printk(KERN_ERR "update: btree %s %s\n",
2066                        bch2_btree_ids[i->iter->btree_id], buf);
2067         }
2068         panic("trans iter oveflow\n");
2069 }
2070
2071 static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans)
2072 {
2073         struct btree_iter *iter;
2074         unsigned idx;
2075
2076         if (unlikely(trans->iters_linked ==
2077                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
2078                 btree_trans_iter_alloc_fail(trans);
2079
2080         idx = __ffs64(~trans->iters_linked);
2081         iter = &trans->iters[idx];
2082
2083         iter->trans             = trans;
2084         iter->idx               = idx;
2085         iter->child_idx         = U8_MAX;
2086         iter->flags             = 0;
2087         iter->nodes_locked      = 0;
2088         iter->nodes_intent_locked = 0;
2089         trans->iters_linked     |= 1ULL << idx;
2090         return iter;
2091 }
2092
2093 static void btree_iter_copy(struct btree_iter *dst, struct btree_iter *src)
2094 {
2095         unsigned i;
2096
2097         __bch2_btree_iter_unlock(dst);
2098         btree_iter_child_free(dst);
2099
2100         memcpy(&dst->flags, &src->flags,
2101                sizeof(struct btree_iter) - offsetof(struct btree_iter, flags));
2102
2103         for (i = 0; i < BTREE_MAX_DEPTH; i++)
2104                 if (btree_node_locked(dst, i))
2105                         six_lock_increment(&dst->l[i].b->c.lock,
2106                                            __btree_lock_want(dst, i));
2107
2108         dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
2109         dst->flags &= ~BTREE_ITER_SET_POS_AFTER_COMMIT;
2110 }
2111
2112 struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans,
2113                                          unsigned btree_id, struct bpos pos,
2114                                          unsigned locks_want,
2115                                          unsigned depth,
2116                                          unsigned flags)
2117 {
2118         struct btree_iter *iter, *best = NULL;
2119         struct bpos real_pos, pos_min = POS_MIN;
2120
2121         if ((flags & BTREE_ITER_TYPE) != BTREE_ITER_NODES &&
2122             btree_node_type_is_extents(btree_id) &&
2123             !(flags & BTREE_ITER_NOT_EXTENTS) &&
2124             !(flags & BTREE_ITER_ALL_SNAPSHOTS))
2125                 flags |= BTREE_ITER_IS_EXTENTS;
2126
2127         if ((flags & BTREE_ITER_TYPE) != BTREE_ITER_NODES &&
2128             !btree_type_has_snapshots(btree_id))
2129                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
2130
2131         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS))
2132                 pos.snapshot = btree_type_has_snapshots(btree_id)
2133                         ? U32_MAX : 0;
2134
2135         real_pos = pos;
2136
2137         if ((flags & BTREE_ITER_IS_EXTENTS) &&
2138             bkey_cmp(pos, POS_MAX))
2139                 real_pos = bpos_nosnap_successor(pos);
2140
2141         trans_for_each_iter(trans, iter) {
2142                 if (btree_iter_type(iter) != (flags & BTREE_ITER_TYPE))
2143                         continue;
2144
2145                 if (iter->btree_id != btree_id)
2146                         continue;
2147
2148                 if (best) {
2149                         int cmp = bkey_cmp(bpos_diff(best->real_pos, real_pos),
2150                                            bpos_diff(iter->real_pos, real_pos));
2151
2152                         if (cmp < 0 ||
2153                             ((cmp == 0 && btree_iter_keep(trans, iter))))
2154                                 continue;
2155                 }
2156
2157                 best = iter;
2158         }
2159
2160         if (!best) {
2161                 iter = btree_trans_iter_alloc(trans);
2162                 bch2_btree_iter_init(trans, iter, btree_id);
2163         } else if (btree_iter_keep(trans, best)) {
2164                 iter = btree_trans_iter_alloc(trans);
2165                 btree_iter_copy(iter, best);
2166         } else {
2167                 iter = best;
2168         }
2169
2170         trans->iters_live       |= 1ULL << iter->idx;
2171         trans->iters_touched    |= 1ULL << iter->idx;
2172
2173         iter->flags = flags;
2174
2175         iter->snapshot = pos.snapshot;
2176
2177         /*
2178          * If the iterator has locks_want greater than requested, we explicitly
2179          * do not downgrade it here - on transaction restart because btree node
2180          * split needs to upgrade locks, we might be putting/getting the
2181          * iterator again. Downgrading iterators only happens via an explicit
2182          * bch2_trans_downgrade().
2183          */
2184
2185         locks_want = min(locks_want, BTREE_MAX_DEPTH);
2186         if (locks_want > iter->locks_want) {
2187                 iter->locks_want = locks_want;
2188                 btree_iter_get_locks(iter, true, _THIS_IP_);
2189         }
2190
2191         while (iter->level != depth) {
2192                 btree_node_unlock(iter, iter->level);
2193                 iter->l[iter->level].b = BTREE_ITER_NO_NODE_INIT;
2194                 iter->uptodate = BTREE_ITER_NEED_TRAVERSE;
2195                 if (iter->level < depth)
2196                         iter->level++;
2197                 else
2198                         iter->level--;
2199         }
2200
2201         iter->min_depth = depth;
2202
2203         bch2_btree_iter_set_pos(iter, pos);
2204         btree_iter_set_search_pos(iter, real_pos);
2205
2206         trace_trans_get_iter(_RET_IP_, trans->ip,
2207                              btree_id,
2208                              &real_pos, locks_want, iter->uptodate,
2209                              best ? &best->real_pos     : &pos_min,
2210                              best ? best->locks_want    : U8_MAX,
2211                              best ? best->uptodate      : U8_MAX);
2212
2213         return iter;
2214 }
2215
2216 struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *trans,
2217                                             enum btree_id btree_id,
2218                                             struct bpos pos,
2219                                             unsigned locks_want,
2220                                             unsigned depth,
2221                                             unsigned flags)
2222 {
2223         struct btree_iter *iter =
2224                 __bch2_trans_get_iter(trans, btree_id, pos,
2225                                       locks_want, depth,
2226                                       BTREE_ITER_NODES|
2227                                       BTREE_ITER_NOT_EXTENTS|
2228                                       BTREE_ITER_ALL_SNAPSHOTS|
2229                                       flags);
2230
2231         BUG_ON(bkey_cmp(iter->pos, pos));
2232         BUG_ON(iter->locks_want != min(locks_want, BTREE_MAX_DEPTH));
2233         BUG_ON(iter->level      != depth);
2234         BUG_ON(iter->min_depth  != depth);
2235         iter->ip_allocated = _RET_IP_;
2236
2237         return iter;
2238 }
2239
2240 struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *trans,
2241                                         struct btree_iter *src)
2242 {
2243         struct btree_iter *iter;
2244
2245         iter = btree_trans_iter_alloc(trans);
2246         btree_iter_copy(iter, src);
2247
2248         trans->iters_live |= 1ULL << iter->idx;
2249         /*
2250          * We don't need to preserve this iter since it's cheap to copy it
2251          * again - this will cause trans_iter_put() to free it right away:
2252          */
2253         set_btree_iter_dontneed(trans, iter);
2254
2255         return iter;
2256 }
2257
2258 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2259 {
2260         size_t new_top = trans->mem_top + size;
2261         void *p;
2262
2263         if (new_top > trans->mem_bytes) {
2264                 size_t old_bytes = trans->mem_bytes;
2265                 size_t new_bytes = roundup_pow_of_two(new_top);
2266                 void *new_mem;
2267
2268                 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2269
2270                 new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2271                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2272                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2273                         new_bytes = BTREE_TRANS_MEM_MAX;
2274                         kfree(trans->mem);
2275                 }
2276
2277                 if (!new_mem)
2278                         return ERR_PTR(-ENOMEM);
2279
2280                 trans->mem = new_mem;
2281                 trans->mem_bytes = new_bytes;
2282
2283                 if (old_bytes) {
2284                         trace_trans_restart_mem_realloced(trans->ip, _RET_IP_, new_bytes);
2285                         return ERR_PTR(-EINTR);
2286                 }
2287         }
2288
2289         p = trans->mem + trans->mem_top;
2290         trans->mem_top += size;
2291         memset(p, 0, size);
2292         return p;
2293 }
2294
2295 inline void bch2_trans_unlink_iters(struct btree_trans *trans)
2296 {
2297         u64 iters = trans->iters_linked &
2298                 ~trans->iters_touched &
2299                 ~trans->iters_live;
2300
2301         while (iters) {
2302                 unsigned idx = __ffs64(iters);
2303
2304                 iters &= ~(1ULL << idx);
2305                 __bch2_trans_iter_free(trans, idx);
2306         }
2307 }
2308
2309 /**
2310  * bch2_trans_reset() - reset a transaction after a interrupted attempt
2311  * @trans: transaction to reset
2312  * @flags: transaction reset flags.
2313  *
2314  * While iterating over nodes or updating nodes a attempt to lock a btree
2315  * node may return EINTR when the trylock fails. When this occurs
2316  * bch2_trans_reset() or bch2_trans_begin() should be called and the
2317  * transaction retried.
2318  *
2319  * Transaction reset flags include:
2320  *
2321  *  - TRANS_RESET_NOUNLOCK   - Do not attempt to unlock and reschedule the
2322  *                             transaction.
2323  *  - TRANS_RESET_NOTRAVERSE - Do not traverse all linked iters.
2324  */
2325 void bch2_trans_reset(struct btree_trans *trans, unsigned flags)
2326 {
2327         struct btree_iter *iter;
2328
2329         trans_for_each_iter(trans, iter) {
2330                 iter->flags &= ~(BTREE_ITER_KEEP_UNTIL_COMMIT|
2331                                  BTREE_ITER_SET_POS_AFTER_COMMIT);
2332                 iter->should_be_locked = false;
2333         }
2334
2335         bch2_trans_unlink_iters(trans);
2336
2337         trans->iters_touched &= trans->iters_live;
2338
2339         trans->extra_journal_res        = 0;
2340         trans->nr_updates               = 0;
2341         trans->mem_top                  = 0;
2342
2343         trans->hooks                    = NULL;
2344         trans->extra_journal_entries    = NULL;
2345         trans->extra_journal_entry_u64s = 0;
2346
2347         if (trans->fs_usage_deltas) {
2348                 trans->fs_usage_deltas->used = 0;
2349                 memset(&trans->fs_usage_deltas->memset_start, 0,
2350                        (void *) &trans->fs_usage_deltas->memset_end -
2351                        (void *) &trans->fs_usage_deltas->memset_start);
2352         }
2353
2354         if (!(flags & TRANS_RESET_NOUNLOCK))
2355                 bch2_trans_cond_resched(trans);
2356
2357         if (!(flags & TRANS_RESET_NOTRAVERSE) &&
2358             trans->iters_linked)
2359                 bch2_btree_iter_traverse_all(trans);
2360 }
2361
2362 static void bch2_trans_alloc_iters(struct btree_trans *trans, struct bch_fs *c)
2363 {
2364         size_t iters_bytes      = sizeof(struct btree_iter) * BTREE_ITER_MAX;
2365         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2366         void *p = NULL;
2367
2368         BUG_ON(trans->used_mempool);
2369
2370 #ifdef __KERNEL__
2371         p = this_cpu_xchg(c->btree_iters_bufs->iter, NULL);
2372 #endif
2373         if (!p)
2374                 p = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
2375
2376         trans->iters            = p; p += iters_bytes;
2377         trans->updates          = p; p += updates_bytes;
2378 }
2379
2380 void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
2381                      unsigned expected_nr_iters,
2382                      size_t expected_mem_bytes)
2383         __acquires(&c->btree_trans_barrier)
2384 {
2385         memset(trans, 0, sizeof(*trans));
2386         trans->c                = c;
2387         trans->ip               = _RET_IP_;
2388
2389         /*
2390          * reallocating iterators currently completely breaks
2391          * bch2_trans_iter_put(), we always allocate the max:
2392          */
2393         bch2_trans_alloc_iters(trans, c);
2394
2395         if (expected_mem_bytes) {
2396                 trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes);
2397                 trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL);
2398
2399                 if (!unlikely(trans->mem)) {
2400                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2401                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2402                 }
2403         }
2404
2405         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2406
2407 #ifdef CONFIG_BCACHEFS_DEBUG
2408         trans->pid = current->pid;
2409         mutex_lock(&c->btree_trans_lock);
2410         list_add(&trans->list, &c->btree_trans_list);
2411         mutex_unlock(&c->btree_trans_lock);
2412 #endif
2413 }
2414
2415 int bch2_trans_exit(struct btree_trans *trans)
2416         __releases(&c->btree_trans_barrier)
2417 {
2418         struct bch_fs *c = trans->c;
2419
2420         bch2_trans_unlock(trans);
2421
2422 #ifdef CONFIG_BCACHEFS_DEBUG
2423         if (trans->iters_live) {
2424                 struct btree_iter *iter;
2425
2426                 trans_for_each_iter(trans, iter)
2427                         btree_iter_child_free(iter);
2428         }
2429
2430         if (trans->iters_live) {
2431                 struct btree_iter *iter;
2432
2433                 bch_err(c, "btree iterators leaked!");
2434                 trans_for_each_iter(trans, iter)
2435                         if (btree_iter_live(trans, iter))
2436                                 printk(KERN_ERR "  btree %s allocated at %pS\n",
2437                                        bch2_btree_ids[iter->btree_id],
2438                                        (void *) iter->ip_allocated);
2439                 /* Be noisy about this: */
2440                 bch2_fatal_error(c);
2441         }
2442
2443         mutex_lock(&trans->c->btree_trans_lock);
2444         list_del(&trans->list);
2445         mutex_unlock(&trans->c->btree_trans_lock);
2446 #endif
2447
2448         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2449
2450         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
2451
2452         if (trans->fs_usage_deltas) {
2453                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
2454                     REPLICAS_DELTA_LIST_MAX)
2455                         mempool_free(trans->fs_usage_deltas,
2456                                      &trans->c->replicas_delta_pool);
2457                 else
2458                         kfree(trans->fs_usage_deltas);
2459         }
2460
2461         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
2462                 mempool_free(trans->mem, &trans->c->btree_trans_mem_pool);
2463         else
2464                 kfree(trans->mem);
2465
2466 #ifdef __KERNEL__
2467         /*
2468          * Userspace doesn't have a real percpu implementation:
2469          */
2470         trans->iters = this_cpu_xchg(c->btree_iters_bufs->iter, trans->iters);
2471 #endif
2472
2473         if (trans->iters)
2474                 mempool_free(trans->iters, &trans->c->btree_iters_pool);
2475
2476         trans->mem      = (void *) 0x1;
2477         trans->iters    = (void *) 0x1;
2478
2479         return trans->error ? -EIO : 0;
2480 }
2481
2482 static void __maybe_unused
2483 bch2_btree_iter_node_to_text(struct printbuf *out,
2484                              struct btree_bkey_cached_common *_b,
2485                              enum btree_iter_type type)
2486 {
2487         pr_buf(out, "    l=%u %s:",
2488                _b->level, bch2_btree_ids[_b->btree_id]);
2489         bch2_bpos_to_text(out, btree_node_pos(_b, type));
2490 }
2491
2492 #ifdef CONFIG_BCACHEFS_DEBUG
2493 static bool trans_has_btree_nodes_locked(struct btree_trans *trans)
2494 {
2495         struct btree_iter *iter;
2496
2497         trans_for_each_iter(trans, iter)
2498                 if (btree_iter_type(iter) != BTREE_ITER_CACHED &&
2499                     iter->nodes_locked)
2500                         return true;
2501         return false;
2502 }
2503 #endif
2504
2505 void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
2506 {
2507 #ifdef CONFIG_BCACHEFS_DEBUG
2508         struct btree_trans *trans;
2509         struct btree_iter *iter;
2510         struct btree *b;
2511         unsigned l;
2512
2513         mutex_lock(&c->btree_trans_lock);
2514         list_for_each_entry(trans, &c->btree_trans_list, list) {
2515                 if (!trans_has_btree_nodes_locked(trans))
2516                         continue;
2517
2518                 pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip);
2519
2520                 trans_for_each_iter(trans, iter) {
2521                         if (!iter->nodes_locked)
2522                                 continue;
2523
2524                         pr_buf(out, "  iter %u %c %s:",
2525                                iter->idx,
2526                                btree_iter_type(iter) == BTREE_ITER_CACHED ? 'c' : 'b',
2527                                bch2_btree_ids[iter->btree_id]);
2528                         bch2_bpos_to_text(out, iter->pos);
2529                         pr_buf(out, "\n");
2530
2531                         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
2532                                 if (btree_node_locked(iter, l)) {
2533                                         pr_buf(out, "    %s l=%u ",
2534                                                btree_node_intent_locked(iter, l) ? "i" : "r", l);
2535                                         bch2_btree_iter_node_to_text(out,
2536                                                         (void *) iter->l[l].b,
2537                                                         btree_iter_type(iter));
2538                                         pr_buf(out, "\n");
2539                                 }
2540                         }
2541                 }
2542
2543                 b = READ_ONCE(trans->locking);
2544                 if (b) {
2545                         iter = &trans->iters[trans->locking_iter_idx];
2546                         pr_buf(out, "  locking iter %u %c l=%u %s:",
2547                                trans->locking_iter_idx,
2548                                btree_iter_type(iter) == BTREE_ITER_CACHED ? 'c' : 'b',
2549                                trans->locking_level,
2550                                bch2_btree_ids[trans->locking_btree_id]);
2551                         bch2_bpos_to_text(out, trans->locking_pos);
2552
2553                         pr_buf(out, " node ");
2554                         bch2_btree_iter_node_to_text(out,
2555                                         (void *) b,
2556                                         btree_iter_type(iter));
2557                         pr_buf(out, "\n");
2558                 }
2559         }
2560         mutex_unlock(&c->btree_trans_lock);
2561 #endif
2562 }
2563
2564 void bch2_fs_btree_iter_exit(struct bch_fs *c)
2565 {
2566         mempool_exit(&c->btree_trans_mem_pool);
2567         mempool_exit(&c->btree_iters_pool);
2568         cleanup_srcu_struct(&c->btree_trans_barrier);
2569 }
2570
2571 int bch2_fs_btree_iter_init(struct bch_fs *c)
2572 {
2573         unsigned nr = BTREE_ITER_MAX;
2574
2575         INIT_LIST_HEAD(&c->btree_trans_list);
2576         mutex_init(&c->btree_trans_lock);
2577
2578         return  init_srcu_struct(&c->btree_trans_barrier) ?:
2579                 mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
2580                         sizeof(struct btree_iter) * nr +
2581                         sizeof(struct btree_insert_entry) * nr) ?:
2582                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
2583                                           BTREE_TRANS_MEM_MAX);
2584 }