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