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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_buf.h"
   6 #include "btree_cache.h"
   7 #include "btree_iter.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "debug.h"
  12 #include "error.h"
  13 #include "extents.h"
  14 #include "journal.h"
  15 #include "replicas.h"
  16
  17 #include <linux/prefetch.h>
  18 #include <trace/events/bcachefs.h>
  19
  20 static void btree_iter_set_search_pos(struct btree_iter *, struct bpos);
  21 static struct btree_iter *btree_iter_child_alloc(struct btree_iter *, unsigned long);
  22 static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *);
  23 static void btree_iter_copy(struct btree_iter *, struct btree_iter *);
  24
  25 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  26 {
  27         EBUG_ON(btree_iter_type(iter) == BTREE_ITER_NODES);
  28
  29         /* Are we iterating over keys in all snapshots? */
  30         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  31                 p = bpos_successor(p);
  32         } else {
  33                 p = bpos_nosnap_successor(p);
  34                 p.snapshot = iter->snapshot;
  35         }
  36
  37         return p;
  38 }
  39
  40 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  41 {
  42         EBUG_ON(btree_iter_type(iter) == BTREE_ITER_NODES);
  43
  44         /* Are we iterating over keys in all snapshots? */
  45         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  46                 p = bpos_predecessor(p);
  47         } else {
  48                 p = bpos_nosnap_predecessor(p);
  49                 p.snapshot = iter->snapshot;
  50         }
  51
  52         return p;
  53 }
  54
  55 static inline bool is_btree_node(struct btree_iter *iter, unsigned l)
  56 {
  57         return l < BTREE_MAX_DEPTH &&
  58                 (unsigned long) iter->l[l].b >= 128;
  59 }
  60
  61 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
  62 {
  63         struct bpos pos = iter->pos;
  64
  65         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
  66             bkey_cmp(pos, POS_MAX))
  67                 pos = bkey_successor(iter, pos);
  68         return pos;
  69 }
  70
  71 static inline bool btree_iter_pos_before_node(struct btree_iter *iter,
  72                                               struct btree *b)
  73 {
  74         return bpos_cmp(iter->real_pos, b->data->min_key) < 0;
  75 }
  76
  77 static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
  78                                              struct btree *b)
  79 {
  80         return bpos_cmp(b->key.k.p, iter->real_pos) < 0;
  81 }
  82
  83 static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
  84                                           struct btree *b)
  85 {
  86         return iter->btree_id == b->c.btree_id &&
  87                 !btree_iter_pos_before_node(iter, b) &&
  88                 !btree_iter_pos_after_node(iter, b);
  89 }
  90
  91 /* Btree node locking: */
  92
  93 void bch2_btree_node_unlock_write(struct btree *b, struct btree_iter *iter)
  94 {
  95         bch2_btree_node_unlock_write_inlined(b, iter);
  96 }
  97
  98 void __bch2_btree_node_lock_write(struct btree *b, struct btree_iter *iter)
  99 {
 100         struct btree_iter *linked;
 101         unsigned readers = 0;
 102
 103         EBUG_ON(!btree_node_intent_locked(iter, b->c.level));
 104
 105         trans_for_each_iter(iter->trans, linked)
 106                 if (linked->l[b->c.level].b == b &&
 107                     btree_node_read_locked(linked, b->c.level))
 108                         readers++;
 109
 110         /*
 111          * Must drop our read locks before calling six_lock_write() -
 112          * six_unlock() won't do wakeups until the reader count
 113          * goes to 0, and it's safe because we have the node intent
 114          * locked:
 115          */
 116         atomic64_sub(__SIX_VAL(read_lock, readers),
 117                      &b->c.lock.state.counter);
 118         btree_node_lock_type(iter->trans->c, b, SIX_LOCK_write);
 119         atomic64_add(__SIX_VAL(read_lock, readers),
 120                      &b->c.lock.state.counter);
 121 }
 122
 123 bool __bch2_btree_node_relock(struct btree_iter *iter, unsigned level)
 124 {
 125         struct btree *b = btree_iter_node(iter, level);
 126         int want = __btree_lock_want(iter, level);
 127
 128         if (!is_btree_node(iter, level))
 129                 return false;
 130
 131         if (race_fault())
 132                 return false;
 133
 134         if (six_relock_type(&b->c.lock, want, iter->l[level].lock_seq) ||
 135             (btree_node_lock_seq_matches(iter, b, level) &&
 136              btree_node_lock_increment(iter->trans, b, level, want))) {
 137                 mark_btree_node_locked(iter, level, want);
 138                 return true;
 139         } else {
 140                 return false;
 141         }
 142 }
 143
 144 static bool bch2_btree_node_upgrade(struct btree_iter *iter, unsigned level)
 145 {
 146         struct btree *b = iter->l[level].b;
 147
 148         EBUG_ON(btree_lock_want(iter, level) != BTREE_NODE_INTENT_LOCKED);
 149
 150         if (!is_btree_node(iter, level))
 151                 return false;
 152
 153         if (btree_node_intent_locked(iter, level))
 154                 return true;
 155
 156         if (race_fault())
 157                 return false;
 158
 159         if (btree_node_locked(iter, level)
 160             ? six_lock_tryupgrade(&b->c.lock)
 161             : six_relock_type(&b->c.lock, SIX_LOCK_intent, iter->l[level].lock_seq))
 162                 goto success;
 163
 164         if (btree_node_lock_seq_matches(iter, b, level) &&
 165             btree_node_lock_increment(iter->trans, b, level, BTREE_NODE_INTENT_LOCKED)) {
 166                 btree_node_unlock(iter, level);
 167                 goto success;
 168         }
 169
 170         return false;
 171 success:
 172         mark_btree_node_intent_locked(iter, level);
 173         return true;
 174 }
 175
 176 static inline bool btree_iter_get_locks(struct btree_iter *iter, bool upgrade,
 177                                         unsigned long trace_ip)
 178 {
 179         unsigned l = iter->level;
 180         int fail_idx = -1;
 181
 182         do {
 183                 if (!btree_iter_node(iter, l))
 184                         break;
 185
 186                 if (!(upgrade
 187                       ? bch2_btree_node_upgrade(iter, l)
 188                       : bch2_btree_node_relock(iter, l))) {
 189                         (upgrade
 190                          ? trace_node_upgrade_fail
 191                          : trace_node_relock_fail)(iter->trans->ip, trace_ip,
 192                                         iter->btree_id, &iter->real_pos,
 193                                         l, iter->l[l].lock_seq,
 194                                         is_btree_node(iter, l)
 195                                         ? 0
 196                                         : (unsigned long) iter->l[l].b,
 197                                         is_btree_node(iter, l)
 198                                         ? iter->l[l].b->c.lock.state.seq
 199                                         : 0);
 200
 201                         fail_idx = l;
 202                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
 203                 }
 204
 205                 l++;
 206         } while (l < iter->locks_want);
 207
 208         /*
 209          * When we fail to get a lock, we have to ensure that any child nodes
 210          * can't be relocked so bch2_btree_iter_traverse has to walk back up to
 211          * the node that we failed to relock:
 212          */
 213         while (fail_idx >= 0) {
 214                 btree_node_unlock(iter, fail_idx);
 215                 iter->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS;
 216                 --fail_idx;
 217         }
 218
 219         if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
 220                 iter->uptodate = BTREE_ITER_NEED_PEEK;
 221
 222         bch2_btree_trans_verify_locks(iter->trans);
 223
 224         return iter->uptodate < BTREE_ITER_NEED_RELOCK;
 225 }
 226
 227 static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b,
 228                                   enum btree_iter_type type)
 229 {
 230         return  type != BTREE_ITER_CACHED
 231                 ? container_of(_b, struct btree, c)->key.k.p
 232                 : container_of(_b, struct bkey_cached, c)->key.pos;
 233 }
 234
 235 /* Slowpath: */
 236 bool __bch2_btree_node_lock(struct btree *b, struct bpos pos,
 237                             unsigned level, struct btree_iter *iter,
 238                             enum six_lock_type type,
 239                             six_lock_should_sleep_fn should_sleep_fn, void *p,
 240                             unsigned long ip)
 241 {
 242         struct btree_trans *trans = iter->trans;
 243         struct btree_iter *linked, *deadlock_iter = NULL;
 244         u64 start_time = local_clock();
 245         unsigned reason = 9;
 246         bool ret;
 247
 248         /* Check if it's safe to block: */
 249         trans_for_each_iter(trans, linked) {
 250                 if (!linked->nodes_locked)
 251                         continue;
 252
 253                 /*
 254                  * Can't block taking an intent lock if we have _any_ nodes read
 255                  * locked:
 256                  *
 257                  * - Our read lock blocks another thread with an intent lock on
 258                  *   the same node from getting a write lock, and thus from
 259                  *   dropping its intent lock
 260                  *
 261                  * - And the other thread may have multiple nodes intent locked:
 262                  *   both the node we want to intent lock, and the node we
 263                  *   already have read locked - deadlock:
 264                  */
 265                 if (type == SIX_LOCK_intent &&
 266                     linked->nodes_locked != linked->nodes_intent_locked) {
 267                         deadlock_iter = linked;
 268                         reason = 1;
 269                 }
 270
 271                 if (linked->btree_id != iter->btree_id) {
 272                         if (linked->btree_id > iter->btree_id) {
 273                                 deadlock_iter = linked;
 274                                 reason = 3;
 275                         }
 276                         continue;
 277                 }
 278
 279                 /*
 280                  * Within the same btree, cached iterators come before non
 281                  * cached iterators:
 282                  */
 283                 if (btree_iter_is_cached(linked) != btree_iter_is_cached(iter)) {
 284                         if (btree_iter_is_cached(iter)) {
 285                                 deadlock_iter = linked;
 286                                 reason = 4;
 287                         }
 288                         continue;
 289                 }
 290
 291                 /*
 292                  * Interior nodes must be locked before their descendants: if
 293                  * another iterator has possible descendants locked of the node
 294                  * we're about to lock, it must have the ancestors locked too:
 295                  */
 296                 if (level > __fls(linked->nodes_locked)) {
 297                         deadlock_iter = linked;
 298                         reason = 5;
 299                 }
 300
 301                 /* Must lock btree nodes in key order: */
 302                 if (btree_node_locked(linked, level) &&
 303                     bpos_cmp(pos, btree_node_pos((void *) linked->l[level].b,
 304                                                  btree_iter_type(linked))) <= 0) {
 305                         deadlock_iter = linked;
 306                         reason = 7;
 307                         BUG_ON(trans->in_traverse_all);
 308                 }
 309         }
 310
 311         if (unlikely(deadlock_iter)) {
 312                 trace_trans_restart_would_deadlock(iter->trans->ip, ip,
 313                                 trans->in_traverse_all, reason,
 314                                 deadlock_iter->btree_id,
 315                                 btree_iter_type(deadlock_iter),
 316                                 &deadlock_iter->real_pos,
 317                                 iter->btree_id,
 318                                 btree_iter_type(iter),
 319                                 &pos);
 320                 return false;
 321         }
 322
 323         if (six_trylock_type(&b->c.lock, type))
 324                 return true;
 325
 326 #ifdef CONFIG_BCACHEFS_DEBUG
 327         trans->locking_iter_idx = iter->idx;
 328         trans->locking_pos      = pos;
 329         trans->locking_btree_id = iter->btree_id;
 330         trans->locking_level    = level;
 331         trans->locking          = b;
 332 #endif
 333
 334         ret = six_lock_type(&b->c.lock, type, should_sleep_fn, p) == 0;
 335
 336 #ifdef CONFIG_BCACHEFS_DEBUG
 337         trans->locking = NULL;
 338 #endif
 339         if (ret)
 340                 bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)],
 341                                        start_time);
 342         return ret;
 343 }
 344
 345 /* Btree iterator locking: */
 346
 347 #ifdef CONFIG_BCACHEFS_DEBUG
 348 static void bch2_btree_iter_verify_locks(struct btree_iter *iter)
 349 {
 350         unsigned l;
 351
 352         if (!(iter->trans->iters_linked & (1ULL << iter->idx))) {
 353                 BUG_ON(iter->nodes_locked);
 354                 return;
 355         }
 356
 357         for (l = 0; is_btree_node(iter, l); l++) {
 358                 if (iter->uptodate >= BTREE_ITER_NEED_RELOCK &&
 359                     !btree_node_locked(iter, l))
 360                         continue;
 361
 362                 BUG_ON(btree_lock_want(iter, l) !=
 363                        btree_node_locked_type(iter, l));
 364         }
 365 }
 366
 367 void bch2_btree_trans_verify_locks(struct btree_trans *trans)
 368 {
 369         struct btree_iter *iter;
 370
 371         trans_for_each_iter(trans, iter)
 372                 bch2_btree_iter_verify_locks(iter);
 373 }
 374 #else
 375 static inline void bch2_btree_iter_verify_locks(struct btree_iter *iter) {}
 376 #endif
 377
 378 __flatten
 379 static bool bch2_btree_iter_relock(struct btree_iter *iter, unsigned long trace_ip)
 380 {
 381         return btree_iter_get_locks(iter, false, trace_ip);
 382 }
 383
 384 bool __bch2_btree_iter_upgrade(struct btree_iter *iter,
 385                                unsigned new_locks_want)
 386 {
 387         struct btree_iter *linked;
 388
 389         EBUG_ON(iter->locks_want >= new_locks_want);
 390
 391         iter->locks_want = new_locks_want;
 392
 393         if (btree_iter_get_locks(iter, true, _THIS_IP_))
 394                 return true;
 395
 396         /*
 397          * XXX: this is ugly - we'd prefer to not be mucking with other
 398          * iterators in the btree_trans here.
 399          *
 400          * On failure to upgrade the iterator, setting iter->locks_want and
 401          * calling get_locks() is sufficient to make bch2_btree_iter_traverse()
 402          * get the locks we want on transaction restart.
 403          *
 404          * But if this iterator was a clone, on transaction restart what we did
 405          * to this iterator isn't going to be preserved.
 406          *
 407          * Possibly we could add an iterator field for the parent iterator when
 408          * an iterator is a copy - for now, we'll just upgrade any other
 409          * iterators with the same btree id.
 410          *
 411          * The code below used to be needed to ensure ancestor nodes get locked
 412          * before interior nodes - now that's handled by
 413          * bch2_btree_iter_traverse_all().
 414          */
 415         trans_for_each_iter(iter->trans, linked)
 416                 if (linked != iter &&
 417                     btree_iter_type(linked) == btree_iter_type(iter) &&
 418                     linked->btree_id == iter->btree_id &&
 419                     linked->locks_want < new_locks_want) {
 420                         linked->locks_want = new_locks_want;
 421                         btree_iter_get_locks(linked, true, _THIS_IP_);
 422                 }
 423
 424         return false;
 425 }
 426
 427 void __bch2_btree_iter_downgrade(struct btree_iter *iter,
 428                                  unsigned new_locks_want)
 429 {
 430         unsigned l;
 431
 432         EBUG_ON(iter->locks_want < new_locks_want);
 433
 434         iter->locks_want = new_locks_want;
 435
 436         while (iter->nodes_locked &&
 437                (l = __fls(iter->nodes_locked)) >= iter->locks_want) {
 438                 if (l > iter->level) {
 439                         btree_node_unlock(iter, l);
 440                 } else {
 441                         if (btree_node_intent_locked(iter, l)) {
 442                                 six_lock_downgrade(&iter->l[l].b->c.lock);
 443                                 iter->nodes_intent_locked ^= 1 << l;
 444                         }
 445                         break;
 446                 }
 447         }
 448
 449         bch2_btree_trans_verify_locks(iter->trans);
 450 }
 451
 452 void bch2_trans_downgrade(struct btree_trans *trans)
 453 {
 454         struct btree_iter *iter;
 455
 456         trans_for_each_iter(trans, iter)
 457                 bch2_btree_iter_downgrade(iter);
 458 }
 459
 460 /* Btree transaction locking: */
 461
 462 static inline bool btree_iter_should_be_locked(struct btree_trans *trans,
 463                                                struct btree_iter *iter)
 464 {
 465         return (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) ||
 466                 iter->should_be_locked;
 467 }
 468
 469 bool bch2_trans_relock(struct btree_trans *trans)
 470 {
 471         struct btree_iter *iter;
 472
 473         trans_for_each_iter(trans, iter)
 474                 if (!bch2_btree_iter_relock(iter, _RET_IP_) &&
 475                     btree_iter_should_be_locked(trans, iter)) {
 476                         trace_trans_restart_relock(trans->ip, _RET_IP_,
 477                                         iter->btree_id, &iter->real_pos);
 478                         return false;
 479                 }
 480         return true;
 481 }
 482
 483 void bch2_trans_unlock(struct btree_trans *trans)
 484 {
 485         struct btree_iter *iter;
 486
 487         trans_for_each_iter(trans, iter)
 488                 __bch2_btree_iter_unlock(iter);
 489 }
 490
 491 /* Btree iterator: */
 492
 493 #ifdef CONFIG_BCACHEFS_DEBUG
 494
 495 static void bch2_btree_iter_verify_cached(struct btree_iter *iter)
 496 {
 497         struct bkey_cached *ck;
 498         bool locked = btree_node_locked(iter, 0);
 499
 500         if (!bch2_btree_node_relock(iter, 0))
 501                 return;
 502
 503         ck = (void *) iter->l[0].b;
 504         BUG_ON(ck->key.btree_id != iter->btree_id ||
 505                bkey_cmp(ck->key.pos, iter->pos));
 506
 507         if (!locked)
 508                 btree_node_unlock(iter, 0);
 509 }
 510
 511 static void bch2_btree_iter_verify_level(struct btree_iter *iter,
 512                                          unsigned level)
 513 {
 514         struct btree_iter_level *l;
 515         struct btree_node_iter tmp;
 516         bool locked;
 517         struct bkey_packed *p, *k;
 518         char buf1[100], buf2[100], buf3[100];
 519         const char *msg;
 520
 521         if (!bch2_debug_check_iterators)
 522                 return;
 523
 524         l       = &iter->l[level];
 525         tmp     = l->iter;
 526         locked  = btree_node_locked(iter, level);
 527
 528         if (btree_iter_type(iter) == BTREE_ITER_CACHED) {
 529                 if (!level)
 530                         bch2_btree_iter_verify_cached(iter);
 531                 return;
 532         }
 533
 534         BUG_ON(iter->level < iter->min_depth);
 535
 536         if (!btree_iter_node(iter, level))
 537                 return;
 538
 539         if (!bch2_btree_node_relock(iter, level))
 540                 return;
 541
 542         BUG_ON(!btree_iter_pos_in_node(iter, l->b));
 543
 544         /*
 545          * node iterators don't use leaf node iterator:
 546          */
 547         if (btree_iter_type(iter) == BTREE_ITER_NODES &&
 548             level <= iter->min_depth)
 549                 goto unlock;
 550
 551         bch2_btree_node_iter_verify(&l->iter, l->b);
 552
 553         /*
 554          * For interior nodes, the iterator will have skipped past
 555          * deleted keys:
 556          *
 557          * For extents, the iterator may have skipped past deleted keys (but not
 558          * whiteouts)
 559          */
 560         p = level || btree_node_type_is_extents(iter->btree_id)
 561                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 562                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 563         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 564
 565         if (p && bkey_iter_pos_cmp(l->b, p, &iter->real_pos) >= 0) {
 566                 msg = "before";
 567                 goto err;
 568         }
 569
 570         if (k && bkey_iter_pos_cmp(l->b, k, &iter->real_pos) < 0) {
 571                 msg = "after";
 572                 goto err;
 573         }
 574 unlock:
 575         if (!locked)
 576                 btree_node_unlock(iter, level);
 577         return;
 578 err:
 579         strcpy(buf2, "(none)");
 580         strcpy(buf3, "(none)");
 581
 582         bch2_bpos_to_text(&PBUF(buf1), iter->real_pos);
 583
 584         if (p) {
 585                 struct bkey uk = bkey_unpack_key(l->b, p);
 586                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 587         }
 588
 589         if (k) {
 590                 struct bkey uk = bkey_unpack_key(l->b, k);
 591                 bch2_bkey_to_text(&PBUF(buf3), &uk);
 592         }
 593
 594         panic("iterator should be %s key at level %u:\n"
 595               "iter pos %s\n"
 596               "prev key %s\n"
 597               "cur  key %s\n",
 598               msg, level, buf1, buf2, buf3);
 599 }
 600
 601 static void bch2_btree_iter_verify(struct btree_iter *iter)
 602 {
 603         enum btree_iter_type type = btree_iter_type(iter);
 604         unsigned i;
 605
 606         EBUG_ON(iter->btree_id >= BTREE_ID_NR);
 607
 608         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 609                iter->pos.snapshot != iter->snapshot);
 610
 611         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 612                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 613
 614         BUG_ON(type == BTREE_ITER_NODES &&
 615                !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 616
 617         BUG_ON(type != BTREE_ITER_NODES &&
 618                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 619                !btree_type_has_snapshots(iter->btree_id));
 620
 621         bch2_btree_iter_verify_locks(iter);
 622
 623         for (i = 0; i < BTREE_MAX_DEPTH; i++)
 624                 bch2_btree_iter_verify_level(iter, i);
 625 }
 626
 627 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 628 {
 629         enum btree_iter_type type = btree_iter_type(iter);
 630
 631         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 632                iter->pos.snapshot != iter->snapshot);
 633
 634         BUG_ON((type == BTREE_ITER_KEYS ||
 635                 type == BTREE_ITER_CACHED) &&
 636                (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
 637                 bkey_cmp(iter->pos, iter->k.p) > 0));
 638 }
 639
 640 void bch2_btree_trans_verify_iters(struct btree_trans *trans, struct btree *b)
 641 {
 642         struct btree_iter *iter;
 643
 644         if (!bch2_debug_check_iterators)
 645                 return;
 646
 647         trans_for_each_iter_with_node(trans, b, iter)
 648                 bch2_btree_iter_verify_level(iter, b->c.level);
 649 }
 650
 651 #else
 652
 653 static inline void bch2_btree_iter_verify_level(struct btree_iter *iter, unsigned l) {}
 654 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 655 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 656
 657 #endif
 658
 659 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 660                                         struct btree *b,
 661                                         struct bset_tree *t,
 662                                         struct bkey_packed *k)
 663 {
 664         struct btree_node_iter_set *set;
 665
 666         btree_node_iter_for_each(iter, set)
 667                 if (set->end == t->end_offset) {
 668                         set->k = __btree_node_key_to_offset(b, k);
 669                         bch2_btree_node_iter_sort(iter, b);
 670                         return;
 671                 }
 672
 673         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 674 }
 675
 676 static void __bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 677                                                struct btree *b,
 678                                                struct bkey_packed *where)
 679 {
 680         struct btree_iter_level *l = &iter->l[b->c.level];
 681
 682         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 683                 return;
 684
 685         if (bkey_iter_pos_cmp(l->b, where, &iter->real_pos) < 0)
 686                 bch2_btree_node_iter_advance(&l->iter, l->b);
 687
 688         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 689 }
 690
 691 void bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 692                                       struct btree *b,
 693                                       struct bkey_packed *where)
 694 {
 695         struct btree_iter *linked;
 696
 697         trans_for_each_iter_with_node(iter->trans, b, linked) {
 698                 __bch2_btree_iter_fix_key_modified(linked, b, where);
 699                 bch2_btree_iter_verify_level(linked, b->c.level);
 700         }
 701 }
 702
 703 static void __bch2_btree_node_iter_fix(struct btree_iter *iter,
 704                                       struct btree *b,
 705                                       struct btree_node_iter *node_iter,
 706                                       struct bset_tree *t,
 707                                       struct bkey_packed *where,
 708                                       unsigned clobber_u64s,
 709                                       unsigned new_u64s)
 710 {
 711         const struct bkey_packed *end = btree_bkey_last(b, t);
 712         struct btree_node_iter_set *set;
 713         unsigned offset = __btree_node_key_to_offset(b, where);
 714         int shift = new_u64s - clobber_u64s;
 715         unsigned old_end = t->end_offset - shift;
 716         unsigned orig_iter_pos = node_iter->data[0].k;
 717         bool iter_current_key_modified =
 718                 orig_iter_pos >= offset &&
 719                 orig_iter_pos <= offset + clobber_u64s;
 720
 721         btree_node_iter_for_each(node_iter, set)
 722                 if (set->end == old_end)
 723                         goto found;
 724
 725         /* didn't find the bset in the iterator - might have to readd it: */
 726         if (new_u64s &&
 727             bkey_iter_pos_cmp(b, where, &iter->real_pos) >= 0) {
 728                 bch2_btree_node_iter_push(node_iter, b, where, end);
 729                 goto fixup_done;
 730         } else {
 731                 /* Iterator is after key that changed */
 732                 return;
 733         }
 734 found:
 735         set->end = t->end_offset;
 736
 737         /* Iterator hasn't gotten to the key that changed yet: */
 738         if (set->k < offset)
 739                 return;
 740
 741         if (new_u64s &&
 742             bkey_iter_pos_cmp(b, where, &iter->real_pos) >= 0) {
 743                 set->k = offset;
 744         } else if (set->k < offset + clobber_u64s) {
 745                 set->k = offset + new_u64s;
 746                 if (set->k == set->end)
 747                         bch2_btree_node_iter_set_drop(node_iter, set);
 748         } else {
 749                 /* Iterator is after key that changed */
 750                 set->k = (int) set->k + shift;
 751                 return;
 752         }
 753
 754         bch2_btree_node_iter_sort(node_iter, b);
 755 fixup_done:
 756         if (node_iter->data[0].k != orig_iter_pos)
 757                 iter_current_key_modified = true;
 758
 759         /*
 760          * When a new key is added, and the node iterator now points to that
 761          * key, the iterator might have skipped past deleted keys that should
 762          * come after the key the iterator now points to. We have to rewind to
 763          * before those deleted keys - otherwise
 764          * bch2_btree_node_iter_prev_all() breaks:
 765          */
 766         if (!bch2_btree_node_iter_end(node_iter) &&
 767             iter_current_key_modified &&
 768             (b->c.level ||
 769              btree_node_type_is_extents(iter->btree_id))) {
 770                 struct bset_tree *t;
 771                 struct bkey_packed *k, *k2, *p;
 772
 773                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 774
 775                 for_each_bset(b, t) {
 776                         bool set_pos = false;
 777
 778                         if (node_iter->data[0].end == t->end_offset)
 779                                 continue;
 780
 781                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 782
 783                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 784                                bkey_iter_cmp(b, k, p) < 0) {
 785                                 k2 = p;
 786                                 set_pos = true;
 787                         }
 788
 789                         if (set_pos)
 790                                 btree_node_iter_set_set_pos(node_iter,
 791                                                             b, t, k2);
 792                 }
 793         }
 794
 795         if (!b->c.level &&
 796             node_iter == &iter->l[0].iter &&
 797             iter_current_key_modified)
 798                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 799 }
 800
 801 void bch2_btree_node_iter_fix(struct btree_iter *iter,
 802                               struct btree *b,
 803                               struct btree_node_iter *node_iter,
 804                               struct bkey_packed *where,
 805                               unsigned clobber_u64s,
 806                               unsigned new_u64s)
 807 {
 808         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 809         struct btree_iter *linked;
 810
 811         if (node_iter != &iter->l[b->c.level].iter) {
 812                 __bch2_btree_node_iter_fix(iter, b, node_iter, t,
 813                                            where, clobber_u64s, new_u64s);
 814
 815                 if (bch2_debug_check_iterators)
 816                         bch2_btree_node_iter_verify(node_iter, b);
 817         }
 818
 819         trans_for_each_iter_with_node(iter->trans, b, linked) {
 820                 __bch2_btree_node_iter_fix(linked, b,
 821                                            &linked->l[b->c.level].iter, t,
 822                                            where, clobber_u64s, new_u64s);
 823                 bch2_btree_iter_verify_level(linked, b->c.level);
 824         }
 825 }
 826
 827 static inline struct bkey_s_c __btree_iter_unpack(struct btree_iter *iter,
 828                                                   struct btree_iter_level *l,
 829                                                   struct bkey *u,
 830                                                   struct bkey_packed *k)
 831 {
 832         struct bkey_s_c ret;
 833
 834         if (unlikely(!k)) {
 835                 /*
 836                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 837                  * a hole
 838                  */
 839                 u->type = KEY_TYPE_deleted;
 840                 return bkey_s_c_null;
 841         }
 842
 843         ret = bkey_disassemble(l->b, k, u);
 844
 845         /*
 846          * XXX: bch2_btree_bset_insert_key() generates invalid keys when we
 847          * overwrite extents - it sets k->type = KEY_TYPE_deleted on the key
 848          * being overwritten but doesn't change k->size. But this is ok, because
 849          * those keys are never written out, we just have to avoid a spurious
 850          * assertion here:
 851          */
 852         if (bch2_debug_check_bkeys && !bkey_deleted(ret.k))
 853                 bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
 854
 855         return ret;
 856 }
 857
 858 /* peek_all() doesn't skip deleted keys */
 859 static inline struct bkey_s_c btree_iter_level_peek_all(struct btree_iter *iter,
 860                                                         struct btree_iter_level *l)
 861 {
 862         return __btree_iter_unpack(iter, l, &iter->k,
 863                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 864 }
 865
 866 static inline struct bkey_s_c btree_iter_level_peek(struct btree_iter *iter,
 867                                                     struct btree_iter_level *l)
 868 {
 869         struct bkey_s_c k = __btree_iter_unpack(iter, l, &iter->k,
 870                         bch2_btree_node_iter_peek(&l->iter, l->b));
 871
 872         iter->real_pos = k.k ? k.k->p : l->b->key.k.p;
 873         return k;
 874 }
 875
 876 static inline struct bkey_s_c btree_iter_level_prev(struct btree_iter *iter,
 877                                                     struct btree_iter_level *l)
 878 {
 879         struct bkey_s_c k = __btree_iter_unpack(iter, l, &iter->k,
 880                         bch2_btree_node_iter_prev(&l->iter, l->b));
 881
 882         iter->real_pos = k.k ? k.k->p : l->b->data->min_key;
 883         return k;
 884 }
 885
 886 static inline bool btree_iter_advance_to_pos(struct btree_iter *iter,
 887                                              struct btree_iter_level *l,
 888                                              int max_advance)
 889 {
 890         struct bkey_packed *k;
 891         int nr_advanced = 0;
 892
 893         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 894                bkey_iter_pos_cmp(l->b, k, &iter->real_pos) < 0) {
 895                 if (max_advance > 0 && nr_advanced >= max_advance)
 896                         return false;
 897
 898                 bch2_btree_node_iter_advance(&l->iter, l->b);
 899                 nr_advanced++;
 900         }
 901
 902         return true;
 903 }
 904
 905 /*
 906  * Verify that iterator for parent node points to child node:
 907  */
 908 static void btree_iter_verify_new_node(struct btree_iter *iter, struct btree *b)
 909 {
 910         struct btree_iter_level *l;
 911         unsigned plevel;
 912         bool parent_locked;
 913         struct bkey_packed *k;
 914
 915         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 916                 return;
 917
 918         plevel = b->c.level + 1;
 919         if (!btree_iter_node(iter, plevel))
 920                 return;
 921
 922         parent_locked = btree_node_locked(iter, plevel);
 923
 924         if (!bch2_btree_node_relock(iter, plevel))
 925                 return;
 926
 927         l = &iter->l[plevel];
 928         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 929         if (!k ||
 930             bkey_deleted(k) ||
 931             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
 932                 char buf1[100];
 933                 char buf2[100];
 934                 char buf3[100];
 935                 char buf4[100];
 936                 struct bkey uk = bkey_unpack_key(b, k);
 937
 938                 bch2_dump_btree_node(iter->trans->c, l->b);
 939                 bch2_bpos_to_text(&PBUF(buf1), iter->real_pos);
 940                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 941                 bch2_bpos_to_text(&PBUF(buf3), b->data->min_key);
 942                 bch2_bpos_to_text(&PBUF(buf3), b->data->max_key);
 943                 panic("parent iter doesn't point to new node:\n"
 944                       "iter pos %s %s\n"
 945                       "iter key %s\n"
 946                       "new node %s-%s\n",
 947                       bch2_btree_ids[iter->btree_id], buf1,
 948                       buf2, buf3, buf4);
 949         }
 950
 951         if (!parent_locked)
 952                 btree_node_unlock(iter, b->c.level + 1);
 953 }
 954
 955 static inline void __btree_iter_init(struct btree_iter *iter,
 956                                      unsigned level)
 957 {
 958         struct btree_iter_level *l = &iter->l[level];
 959
 960         bch2_btree_node_iter_init(&l->iter, l->b, &iter->real_pos);
 961
 962         /*
 963          * Iterators to interior nodes should always be pointed at the first non
 964          * whiteout:
 965          */
 966         if (level)
 967                 bch2_btree_node_iter_peek(&l->iter, l->b);
 968
 969         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 970 }
 971
 972 static inline void btree_iter_node_set(struct btree_iter *iter,
 973                                        struct btree *b)
 974 {
 975         BUG_ON(btree_iter_type(iter) == BTREE_ITER_CACHED);
 976
 977         btree_iter_verify_new_node(iter, b);
 978
 979         EBUG_ON(!btree_iter_pos_in_node(iter, b));
 980         EBUG_ON(b->c.lock.state.seq & 1);
 981
 982         iter->l[b->c.level].lock_seq = b->c.lock.state.seq;
 983         iter->l[b->c.level].b = b;
 984         __btree_iter_init(iter, b->c.level);
 985 }
 986
 987 /*
 988  * A btree node is being replaced - update the iterator to point to the new
 989  * node:
 990  */
 991 void bch2_btree_iter_node_replace(struct btree_iter *iter, struct btree *b)
 992 {
 993         enum btree_node_locked_type t;
 994         struct btree_iter *linked;
 995
 996         trans_for_each_iter(iter->trans, linked)
 997                 if (btree_iter_type(linked) != BTREE_ITER_CACHED &&
 998                     btree_iter_pos_in_node(linked, b)) {
 999                         /*
1000                          * bch2_btree_iter_node_drop() has already been called -
1001                          * the old node we're replacing has already been
1002                          * unlocked and the pointer invalidated
1003                          */
1004                         BUG_ON(btree_node_locked(linked, b->c.level));
1005
1006                         t = btree_lock_want(linked, b->c.level);
1007                         if (t != BTREE_NODE_UNLOCKED) {
1008                                 six_lock_increment(&b->c.lock, t);
1009                                 mark_btree_node_locked(linked, b->c.level, t);
1010                         }
1011
1012                         btree_iter_node_set(linked, b);
1013                 }
1014 }
1015
1016 void bch2_btree_iter_node_drop(struct btree_iter *iter, struct btree *b)
1017 {
1018         struct btree_iter *linked;
1019         unsigned level = b->c.level;
1020
1021         trans_for_each_iter(iter->trans, linked)
1022                 if (linked->l[level].b == b) {
1023                         btree_node_unlock(linked, level);
1024                         linked->l[level].b = BTREE_ITER_NO_NODE_DROP;
1025                 }
1026 }
1027
1028 /*
1029  * A btree node has been modified in such a way as to invalidate iterators - fix
1030  * them:
1031  */
1032 void bch2_btree_iter_reinit_node(struct btree_iter *iter, struct btree *b)
1033 {
1034         struct btree_iter *linked;
1035
1036         trans_for_each_iter_with_node(iter->trans, b, linked)
1037                 __btree_iter_init(linked, b->c.level);
1038 }
1039
1040 static int lock_root_check_fn(struct six_lock *lock, void *p)
1041 {
1042         struct btree *b = container_of(lock, struct btree, c.lock);
1043         struct btree **rootp = p;
1044
1045         return b == *rootp ? 0 : -1;
1046 }
1047
1048 static inline int btree_iter_lock_root(struct btree_iter *iter,
1049                                        unsigned depth_want,
1050                                        unsigned long trace_ip)
1051 {
1052         struct bch_fs *c = iter->trans->c;
1053         struct btree *b, **rootp = &c->btree_roots[iter->btree_id].b;
1054         enum six_lock_type lock_type;
1055         unsigned i;
1056
1057         EBUG_ON(iter->nodes_locked);
1058
1059         while (1) {
1060                 b = READ_ONCE(*rootp);
1061                 iter->level = READ_ONCE(b->c.level);
1062
1063                 if (unlikely(iter->level < depth_want)) {
1064                         /*
1065                          * the root is at a lower depth than the depth we want:
1066                          * got to the end of the btree, or we're walking nodes
1067                          * greater than some depth and there are no nodes >=
1068                          * that depth
1069                          */
1070                         iter->level = depth_want;
1071                         for (i = iter->level; i < BTREE_MAX_DEPTH; i++)
1072                                 iter->l[i].b = NULL;
1073                         return 1;
1074                 }
1075
1076                 lock_type = __btree_lock_want(iter, iter->level);
1077                 if (unlikely(!btree_node_lock(b, POS_MAX, iter->level,
1078                                               iter, lock_type,
1079                                               lock_root_check_fn, rootp,
1080                                               trace_ip)))
1081                         return -EINTR;
1082
1083                 if (likely(b == READ_ONCE(*rootp) &&
1084                            b->c.level == iter->level &&
1085                            !race_fault())) {
1086                         for (i = 0; i < iter->level; i++)
1087                                 iter->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
1088                         iter->l[iter->level].b = b;
1089                         for (i = iter->level + 1; i < BTREE_MAX_DEPTH; i++)
1090                                 iter->l[i].b = NULL;
1091
1092                         mark_btree_node_locked(iter, iter->level, lock_type);
1093                         btree_iter_node_set(iter, b);
1094                         return 0;
1095                 }
1096
1097                 six_unlock_type(&b->c.lock, lock_type);
1098         }
1099 }
1100
1101 noinline
1102 static void btree_iter_prefetch(struct btree_iter *iter)
1103 {
1104         struct bch_fs *c = iter->trans->c;
1105         struct btree_iter_level *l = &iter->l[iter->level];
1106         struct btree_node_iter node_iter = l->iter;
1107         struct bkey_packed *k;
1108         struct bkey_buf tmp;
1109         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1110                 ? (iter->level > 1 ? 0 :  2)
1111                 : (iter->level > 1 ? 1 : 16);
1112         bool was_locked = btree_node_locked(iter, iter->level);
1113
1114         bch2_bkey_buf_init(&tmp);
1115
1116         while (nr) {
1117                 if (!bch2_btree_node_relock(iter, iter->level))
1118                         break;
1119
1120                 bch2_btree_node_iter_advance(&node_iter, l->b);
1121                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
1122                 if (!k)
1123                         break;
1124
1125                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
1126                 bch2_btree_node_prefetch(c, iter, tmp.k, iter->btree_id,
1127                                          iter->level - 1);
1128         }
1129
1130         if (!was_locked)
1131                 btree_node_unlock(iter, iter->level);
1132
1133         bch2_bkey_buf_exit(&tmp, c);
1134 }
1135
1136 static noinline void btree_node_mem_ptr_set(struct btree_iter *iter,
1137                                             unsigned plevel, struct btree *b)
1138 {
1139         struct btree_iter_level *l = &iter->l[plevel];
1140         bool locked = btree_node_locked(iter, plevel);
1141         struct bkey_packed *k;
1142         struct bch_btree_ptr_v2 *bp;
1143
1144         if (!bch2_btree_node_relock(iter, plevel))
1145                 return;
1146
1147         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1148         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
1149
1150         bp = (void *) bkeyp_val(&l->b->format, k);
1151         bp->mem_ptr = (unsigned long)b;
1152
1153         if (!locked)
1154                 btree_node_unlock(iter, plevel);
1155 }
1156
1157 static __always_inline int btree_iter_down(struct btree_iter *iter,
1158                                            unsigned long trace_ip)
1159 {
1160         struct bch_fs *c = iter->trans->c;
1161         struct btree_iter_level *l = &iter->l[iter->level];
1162         struct btree *b;
1163         unsigned level = iter->level - 1;
1164         enum six_lock_type lock_type = __btree_lock_want(iter, level);
1165         struct bkey_buf tmp;
1166         int ret;
1167
1168         EBUG_ON(!btree_node_locked(iter, iter->level));
1169
1170         bch2_bkey_buf_init(&tmp);
1171         bch2_bkey_buf_unpack(&tmp, c, l->b,
1172                          bch2_btree_node_iter_peek(&l->iter, l->b));
1173
1174         b = bch2_btree_node_get(c, iter, tmp.k, level, lock_type, trace_ip);
1175         ret = PTR_ERR_OR_ZERO(b);
1176         if (unlikely(ret))
1177                 goto err;
1178
1179         mark_btree_node_locked(iter, level, lock_type);
1180         btree_iter_node_set(iter, b);
1181
1182         if (tmp.k->k.type == KEY_TYPE_btree_ptr_v2 &&
1183             unlikely(b != btree_node_mem_ptr(tmp.k)))
1184                 btree_node_mem_ptr_set(iter, level + 1, b);
1185
1186         if (iter->flags & BTREE_ITER_PREFETCH)
1187                 btree_iter_prefetch(iter);
1188
1189         if (btree_node_read_locked(iter, level + 1))
1190                 btree_node_unlock(iter, level + 1);
1191         iter->level = level;
1192
1193         bch2_btree_iter_verify_locks(iter);
1194 err:
1195         bch2_bkey_buf_exit(&tmp, c);
1196         return ret;
1197 }
1198
1199 static int btree_iter_traverse_one(struct btree_iter *, unsigned long);
1200
1201 static int __btree_iter_traverse_all(struct btree_trans *trans, int ret,
1202                                      unsigned long trace_ip)
1203 {
1204         struct bch_fs *c = trans->c;
1205         struct btree_iter *iter;
1206         u8 sorted[BTREE_ITER_MAX];
1207         int i, nr_sorted = 0;
1208         bool relock_fail;
1209
1210         if (trans->in_traverse_all)
1211                 return -EINTR;
1212
1213         trans->in_traverse_all = true;
1214 retry_all:
1215         nr_sorted = 0;
1216         relock_fail = false;
1217
1218         trans_for_each_iter(trans, iter) {
1219                 if (!bch2_btree_iter_relock(iter, _THIS_IP_))
1220                         relock_fail = true;
1221                 sorted[nr_sorted++] = iter->idx;
1222         }
1223
1224         if (!relock_fail) {
1225                 trans->in_traverse_all = false;
1226                 return 0;
1227         }
1228
1229 #define btree_iter_cmp_by_idx(_l, _r)                           \
1230                 btree_iter_lock_cmp(&trans->iters[_l], &trans->iters[_r])
1231
1232         bubble_sort(sorted, nr_sorted, btree_iter_cmp_by_idx);
1233 #undef btree_iter_cmp_by_idx
1234
1235         for (i = nr_sorted - 2; i >= 0; --i) {
1236                 struct btree_iter *iter1 = trans->iters + sorted[i];
1237                 struct btree_iter *iter2 = trans->iters + sorted[i + 1];
1238
1239                 if (iter1->btree_id == iter2->btree_id &&
1240                     iter1->locks_want < iter2->locks_want)
1241                         __bch2_btree_iter_upgrade(iter1, iter2->locks_want);
1242                 else if (!iter1->locks_want && iter2->locks_want)
1243                         __bch2_btree_iter_upgrade(iter1, 1);
1244         }
1245
1246         bch2_trans_unlock(trans);
1247         cond_resched();
1248
1249         if (unlikely(ret == -ENOMEM)) {
1250                 struct closure cl;
1251
1252                 closure_init_stack(&cl);
1253
1254                 do {
1255                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1256                         closure_sync(&cl);
1257                 } while (ret);
1258         }
1259
1260         if (unlikely(ret == -EIO)) {
1261                 trans->error = true;
1262                 goto out;
1263         }
1264
1265         BUG_ON(ret && ret != -EINTR);
1266
1267         /* Now, redo traversals in correct order: */
1268         for (i = 0; i < nr_sorted; i++) {
1269                 unsigned idx = sorted[i];
1270
1271                 /*
1272                  * sucessfully traversing one iterator can cause another to be
1273                  * unlinked, in btree_key_cache_fill()
1274                  */
1275                 if (!(trans->iters_linked & (1ULL << idx)))
1276                         continue;
1277
1278                 ret = btree_iter_traverse_one(&trans->iters[idx], _THIS_IP_);
1279                 if (ret)
1280                         goto retry_all;
1281         }
1282
1283         if (hweight64(trans->iters_live) > 1)
1284                 ret = -EINTR;
1285         else
1286                 trans_for_each_iter(trans, iter)
1287                         if (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) {
1288                                 ret = -EINTR;
1289                                 break;
1290                         }
1291 out:
1292         bch2_btree_cache_cannibalize_unlock(c);
1293
1294         trans->in_traverse_all = false;
1295
1296         trace_trans_traverse_all(trans->ip, trace_ip);
1297         return ret;
1298 }
1299
1300 int bch2_btree_iter_traverse_all(struct btree_trans *trans)
1301 {
1302         return __btree_iter_traverse_all(trans, 0, _RET_IP_);
1303 }
1304
1305 static inline bool btree_iter_good_node(struct btree_iter *iter,
1306                                         unsigned l, int check_pos)
1307 {
1308         if (!is_btree_node(iter, l) ||
1309             !bch2_btree_node_relock(iter, l))
1310                 return false;
1311
1312         if (check_pos < 0 && btree_iter_pos_before_node(iter, iter->l[l].b))
1313                 return false;
1314         if (check_pos > 0 && btree_iter_pos_after_node(iter, iter->l[l].b))
1315                 return false;
1316         return true;
1317 }
1318
1319 static inline unsigned btree_iter_up_until_good_node(struct btree_iter *iter,
1320                                                      int check_pos)
1321 {
1322         unsigned l = iter->level;
1323
1324         while (btree_iter_node(iter, l) &&
1325                !btree_iter_good_node(iter, l, check_pos)) {
1326                 btree_node_unlock(iter, l);
1327                 iter->l[l].b = BTREE_ITER_NO_NODE_UP;
1328                 l++;
1329         }
1330
1331         return l;
1332 }
1333
1334 /*
1335  * This is the main state machine for walking down the btree - walks down to a
1336  * specified depth
1337  *
1338  * Returns 0 on success, -EIO on error (error reading in a btree node).
1339  *
1340  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1341  * stashed in the iterator and returned from bch2_trans_exit().
1342  */
1343 static int btree_iter_traverse_one(struct btree_iter *iter,
1344                                    unsigned long trace_ip)
1345 {
1346         unsigned depth_want = iter->level;
1347         int ret = 0;
1348
1349         /*
1350          * if we need interior nodes locked, call btree_iter_relock() to make
1351          * sure we walk back up enough that we lock them:
1352          */
1353         if (iter->uptodate == BTREE_ITER_NEED_RELOCK ||
1354             iter->locks_want > 1)
1355                 bch2_btree_iter_relock(iter, _THIS_IP_);
1356
1357         if (btree_iter_type(iter) == BTREE_ITER_CACHED) {
1358                 ret = bch2_btree_iter_traverse_cached(iter);
1359                 goto out;
1360         }
1361
1362         if (iter->uptodate < BTREE_ITER_NEED_RELOCK)
1363                 goto out;
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         /*
1371          * Note: iter->nodes[iter->level] may be temporarily NULL here - that
1372          * would indicate to other code that we got to the end of the btree,
1373          * here it indicates that relocking the root failed - it's critical that
1374          * btree_iter_lock_root() comes next and that it can't fail
1375          */
1376         while (iter->level > depth_want) {
1377                 ret = btree_iter_node(iter, iter->level)
1378                         ? btree_iter_down(iter, trace_ip)
1379                         : btree_iter_lock_root(iter, depth_want, trace_ip);
1380                 if (unlikely(ret)) {
1381                         if (ret == 1) {
1382                                 /*
1383                                  * Got to the end of the btree (in
1384                                  * BTREE_ITER_NODES mode)
1385                                  */
1386                                 ret = 0;
1387                                 goto out;
1388                         }
1389
1390                         iter->level = depth_want;
1391
1392                         if (ret == -EIO) {
1393                                 iter->flags |= BTREE_ITER_ERROR;
1394                                 iter->l[iter->level].b =
1395                                         BTREE_ITER_NO_NODE_ERROR;
1396                         } else {
1397                                 iter->l[iter->level].b =
1398                                         BTREE_ITER_NO_NODE_DOWN;
1399                         }
1400                         goto out;
1401                 }
1402         }
1403
1404         iter->uptodate = BTREE_ITER_NEED_PEEK;
1405 out:
1406         trace_iter_traverse(iter->trans->ip, trace_ip,
1407                             iter->btree_id, &iter->real_pos, ret);
1408         bch2_btree_iter_verify(iter);
1409         return ret;
1410 }
1411
1412 static int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
1413 {
1414         struct btree_trans *trans = iter->trans;
1415         int ret;
1416
1417         ret =   bch2_trans_cond_resched(trans) ?:
1418                 btree_iter_traverse_one(iter, _RET_IP_);
1419         if (unlikely(ret))
1420                 ret = __btree_iter_traverse_all(trans, ret, _RET_IP_);
1421
1422         return ret;
1423 }
1424
1425 /*
1426  * Note:
1427  * bch2_btree_iter_traverse() is for external users, btree_iter_traverse() is
1428  * for internal btree iterator users
1429  *
1430  * bch2_btree_iter_traverse sets iter->real_pos to iter->pos,
1431  * btree_iter_traverse() does not:
1432  */
1433 static inline int __must_check
1434 btree_iter_traverse(struct btree_iter *iter)
1435 {
1436         return iter->uptodate >= BTREE_ITER_NEED_RELOCK
1437                 ? __bch2_btree_iter_traverse(iter)
1438                 : 0;
1439 }
1440
1441 int __must_check
1442 bch2_btree_iter_traverse(struct btree_iter *iter)
1443 {
1444         int ret;
1445
1446         btree_iter_set_search_pos(iter, btree_iter_search_key(iter));
1447
1448         ret = btree_iter_traverse(iter);
1449         if (ret)
1450                 return ret;
1451
1452         iter->should_be_locked = true;
1453         return 0;
1454 }
1455
1456 /* Iterate across nodes (leaf and interior nodes) */
1457
1458 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1459 {
1460         struct btree *b;
1461         int ret;
1462
1463         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
1464         bch2_btree_iter_verify(iter);
1465
1466         ret = btree_iter_traverse(iter);
1467         if (ret)
1468                 return NULL;
1469
1470         b = btree_iter_node(iter, iter->level);
1471         if (!b)
1472                 return NULL;
1473
1474         BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0);
1475
1476         iter->pos = iter->real_pos = b->key.k.p;
1477
1478         bch2_btree_iter_verify(iter);
1479         iter->should_be_locked = true;
1480
1481         return b;
1482 }
1483
1484 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1485 {
1486         struct btree *b;
1487         int ret;
1488
1489         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
1490         bch2_btree_iter_verify(iter);
1491
1492         /* already got to end? */
1493         if (!btree_iter_node(iter, iter->level))
1494                 return NULL;
1495
1496         bch2_trans_cond_resched(iter->trans);
1497
1498         btree_node_unlock(iter, iter->level);
1499         iter->l[iter->level].b = BTREE_ITER_NO_NODE_UP;
1500         iter->level++;
1501
1502         btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1503         ret = btree_iter_traverse(iter);
1504         if (ret)
1505                 return NULL;
1506
1507         /* got to end? */
1508         b = btree_iter_node(iter, iter->level);
1509         if (!b)
1510                 return NULL;
1511
1512         if (bpos_cmp(iter->pos, b->key.k.p) < 0) {
1513                 /*
1514                  * Haven't gotten to the end of the parent node: go back down to
1515                  * the next child node
1516                  */
1517                 btree_iter_set_search_pos(iter, bpos_successor(iter->pos));
1518
1519                 /* Unlock to avoid screwing up our lock invariants: */
1520                 btree_node_unlock(iter, iter->level);
1521
1522                 iter->level = iter->min_depth;
1523                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1524                 bch2_btree_iter_verify(iter);
1525
1526                 ret = btree_iter_traverse(iter);
1527                 if (ret)
1528                         return NULL;
1529
1530                 b = iter->l[iter->level].b;
1531         }
1532
1533         iter->pos = iter->real_pos = b->key.k.p;
1534
1535         bch2_btree_iter_verify(iter);
1536         iter->should_be_locked = true;
1537
1538         return b;
1539 }
1540
1541 /* Iterate across keys (in leaf nodes only) */
1542
1543 static void btree_iter_set_search_pos(struct btree_iter *iter, struct bpos new_pos)
1544 {
1545 #ifdef CONFIG_BCACHEFS_DEBUG
1546         struct bpos old_pos = iter->real_pos;
1547 #endif
1548         int cmp = bpos_cmp(new_pos, iter->real_pos);
1549         unsigned l = iter->level;
1550
1551         if (!cmp)
1552                 goto out;
1553
1554         iter->real_pos = new_pos;
1555         iter->should_be_locked = false;
1556
1557         if (unlikely(btree_iter_type(iter) == BTREE_ITER_CACHED)) {
1558                 btree_node_unlock(iter, 0);
1559                 iter->l[0].b = BTREE_ITER_NO_NODE_CACHED;
1560                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1561                 return;
1562         }
1563
1564         l = btree_iter_up_until_good_node(iter, cmp);
1565
1566         if (btree_iter_node(iter, l)) {
1567                 /*
1568                  * We might have to skip over many keys, or just a few: try
1569                  * advancing the node iterator, and if we have to skip over too
1570                  * many keys just reinit it (or if we're rewinding, since that
1571                  * is expensive).
1572                  */
1573                 if (cmp < 0 ||
1574                     !btree_iter_advance_to_pos(iter, &iter->l[l], 8))
1575                         __btree_iter_init(iter, l);
1576
1577                 /* Don't leave it locked if we're not supposed to: */
1578                 if (btree_lock_want(iter, l) == BTREE_NODE_UNLOCKED)
1579                         btree_node_unlock(iter, l);
1580         }
1581 out:
1582         if (l != iter->level)
1583                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1584         else
1585                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
1586
1587         bch2_btree_iter_verify(iter);
1588 #ifdef CONFIG_BCACHEFS_DEBUG
1589         trace_iter_set_search_pos(iter->trans->ip, _RET_IP_,
1590                                   iter->btree_id,
1591                                   &old_pos, &new_pos, l);
1592 #endif
1593 }
1594
1595 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1596 {
1597         struct bpos pos = iter->k.p;
1598         bool ret = bpos_cmp(pos, POS_MAX) != 0;
1599
1600         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1601                 pos = bkey_successor(iter, pos);
1602         bch2_btree_iter_set_pos(iter, pos);
1603         return ret;
1604 }
1605
1606 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1607 {
1608         struct bpos pos = bkey_start_pos(&iter->k);
1609         bool ret = bpos_cmp(pos, POS_MIN) != 0;
1610
1611         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1612                 pos = bkey_predecessor(iter, pos);
1613         bch2_btree_iter_set_pos(iter, pos);
1614         return ret;
1615 }
1616
1617 static inline bool btree_iter_set_pos_to_next_leaf(struct btree_iter *iter)
1618 {
1619         struct bpos next_pos = iter->l[0].b->key.k.p;
1620         bool ret = bpos_cmp(next_pos, POS_MAX) != 0;
1621
1622         /*
1623          * Typically, we don't want to modify iter->pos here, since that
1624          * indicates where we searched from - unless we got to the end of the
1625          * btree, in that case we want iter->pos to reflect that:
1626          */
1627         if (ret)
1628                 btree_iter_set_search_pos(iter, bpos_successor(next_pos));
1629         else
1630                 bch2_btree_iter_set_pos(iter, POS_MAX);
1631
1632         return ret;
1633 }
1634
1635 static inline bool btree_iter_set_pos_to_prev_leaf(struct btree_iter *iter)
1636 {
1637         struct bpos next_pos = iter->l[0].b->data->min_key;
1638         bool ret = bpos_cmp(next_pos, POS_MIN) != 0;
1639
1640         if (ret)
1641                 btree_iter_set_search_pos(iter, bpos_predecessor(next_pos));
1642         else
1643                 bch2_btree_iter_set_pos(iter, POS_MIN);
1644
1645         return ret;
1646 }
1647
1648 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter,
1649                                                       struct bpos pos)
1650 {
1651         struct btree_insert_entry *i;
1652
1653         if (!(iter->flags & BTREE_ITER_WITH_UPDATES))
1654                 return NULL;
1655
1656         trans_for_each_update(iter->trans, i)
1657                 if ((cmp_int(iter->btree_id,    i->iter->btree_id) ?:
1658                      bkey_cmp(pos,              i->k->k.p)) <= 0) {
1659                         if (iter->btree_id ==   i->iter->btree_id)
1660                                 return i->k;
1661                         break;
1662                 }
1663
1664         return NULL;
1665 }
1666
1667 /**
1668  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
1669  * current position
1670  */
1671 struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
1672 {
1673         struct bpos search_key = btree_iter_search_key(iter);
1674         struct bkey_i *next_update;
1675         struct bkey_s_c k;
1676         int ret;
1677
1678         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1679         bch2_btree_iter_verify(iter);
1680         bch2_btree_iter_verify_entry_exit(iter);
1681 start:
1682         next_update = btree_trans_peek_updates(iter, search_key);
1683         btree_iter_set_search_pos(iter, search_key);
1684
1685         while (1) {
1686                 ret = btree_iter_traverse(iter);
1687                 if (unlikely(ret))
1688                         return bkey_s_c_err(ret);
1689
1690                 k = btree_iter_level_peek(iter, &iter->l[0]);
1691
1692                 if (next_update &&
1693                     bpos_cmp(next_update->k.p, iter->real_pos) <= 0) {
1694                         iter->k = next_update->k;
1695                         k = bkey_i_to_s_c(next_update);
1696                 }
1697
1698                 if (likely(k.k)) {
1699                         if (bkey_deleted(k.k)) {
1700                                 search_key = bkey_successor(iter, k.k->p);
1701                                 goto start;
1702                         }
1703
1704                         break;
1705                 }
1706
1707                 if (!btree_iter_set_pos_to_next_leaf(iter))
1708                         return bkey_s_c_null;
1709         }
1710
1711         /*
1712          * iter->pos should be mononotically increasing, and always be equal to
1713          * the key we just returned - except extents can straddle iter->pos:
1714          */
1715         if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
1716                 iter->pos = k.k->p;
1717         else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
1718                 iter->pos = bkey_start_pos(k.k);
1719
1720         bch2_btree_iter_verify_entry_exit(iter);
1721         bch2_btree_iter_verify(iter);
1722         iter->should_be_locked = true;
1723         return k;
1724 }
1725
1726 /**
1727  * bch2_btree_iter_next: returns first key greater than iterator's current
1728  * position
1729  */
1730 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
1731 {
1732         if (!bch2_btree_iter_advance(iter))
1733                 return bkey_s_c_null;
1734
1735         return bch2_btree_iter_peek(iter);
1736 }
1737
1738 /**
1739  * bch2_btree_iter_peek_prev: returns first key less than or equal to
1740  * iterator's current position
1741  */
1742 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
1743 {
1744         struct btree_iter_level *l = &iter->l[0];
1745         struct bkey_s_c k;
1746         int ret;
1747
1748         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1749         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
1750         bch2_btree_iter_verify(iter);
1751         bch2_btree_iter_verify_entry_exit(iter);
1752
1753         btree_iter_set_search_pos(iter, iter->pos);
1754
1755         while (1) {
1756                 ret = btree_iter_traverse(iter);
1757                 if (unlikely(ret)) {
1758                         k = bkey_s_c_err(ret);
1759                         goto no_key;
1760                 }
1761
1762                 k = btree_iter_level_peek(iter, l);
1763                 if (!k.k ||
1764                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
1765                      ? bkey_cmp(bkey_start_pos(k.k), iter->pos) >= 0
1766                      : bkey_cmp(k.k->p, iter->pos) > 0))
1767                         k = btree_iter_level_prev(iter, l);
1768
1769                 if (likely(k.k))
1770                         break;
1771
1772                 if (!btree_iter_set_pos_to_prev_leaf(iter)) {
1773                         k = bkey_s_c_null;
1774                         goto no_key;
1775                 }
1776         }
1777
1778         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
1779
1780         /* Extents can straddle iter->pos: */
1781         if (bkey_cmp(k.k->p, iter->pos) < 0)
1782                 iter->pos = k.k->p;
1783 out:
1784         bch2_btree_iter_verify_entry_exit(iter);
1785         bch2_btree_iter_verify(iter);
1786         iter->should_be_locked = true;
1787         return k;
1788 no_key:
1789         /*
1790          * btree_iter_level_peek() may have set iter->k to a key we didn't want, and
1791          * then we errored going to the previous leaf - make sure it's
1792          * consistent with iter->pos:
1793          */
1794         bkey_init(&iter->k);
1795         iter->k.p = iter->pos;
1796         goto out;
1797 }
1798
1799 /**
1800  * bch2_btree_iter_prev: returns first key less than iterator's current
1801  * position
1802  */
1803 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
1804 {
1805         if (!bch2_btree_iter_rewind(iter))
1806                 return bkey_s_c_null;
1807
1808         return bch2_btree_iter_peek_prev(iter);
1809 }
1810
1811 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
1812 {
1813         struct bpos search_key;
1814         struct bkey_s_c k;
1815         int ret;
1816
1817         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS &&
1818                 btree_iter_type(iter) != BTREE_ITER_CACHED);
1819         bch2_btree_iter_verify(iter);
1820         bch2_btree_iter_verify_entry_exit(iter);
1821
1822         /* extents can't span inode numbers: */
1823         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
1824             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
1825                 if (iter->pos.inode == KEY_INODE_MAX)
1826                         return bkey_s_c_null;
1827
1828                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
1829         }
1830
1831         search_key = btree_iter_search_key(iter);
1832         btree_iter_set_search_pos(iter, search_key);
1833
1834         ret = btree_iter_traverse(iter);
1835         if (unlikely(ret))
1836                 return bkey_s_c_err(ret);
1837
1838         if (btree_iter_type(iter) == BTREE_ITER_CACHED ||
1839             !(iter->flags & BTREE_ITER_IS_EXTENTS)) {
1840                 struct bkey_i *next_update;
1841                 struct bkey_cached *ck;
1842
1843                 switch (btree_iter_type(iter)) {
1844                 case BTREE_ITER_KEYS:
1845                         k = btree_iter_level_peek_all(iter, &iter->l[0]);
1846                         EBUG_ON(k.k && bkey_deleted(k.k) && bkey_cmp(k.k->p, iter->pos) == 0);
1847                         break;
1848                 case BTREE_ITER_CACHED:
1849                         ck = (void *) iter->l[0].b;
1850                         EBUG_ON(iter->btree_id != ck->key.btree_id ||
1851                                 bkey_cmp(iter->pos, ck->key.pos));
1852                         BUG_ON(!ck->valid);
1853
1854                         k = bkey_i_to_s_c(ck->k);
1855                         break;
1856                 case BTREE_ITER_NODES:
1857                         BUG();
1858                 }
1859
1860                 next_update = btree_trans_peek_updates(iter, search_key);
1861                 if (next_update &&
1862                     (!k.k || bpos_cmp(next_update->k.p, k.k->p) <= 0)) {
1863                         iter->k = next_update->k;
1864                         k = bkey_i_to_s_c(next_update);
1865                 }
1866         } else {
1867                 if ((iter->flags & BTREE_ITER_INTENT)) {
1868                         struct btree_iter *child =
1869                                 btree_iter_child_alloc(iter, _THIS_IP_);
1870
1871                         btree_iter_copy(child, iter);
1872                         k = bch2_btree_iter_peek(child);
1873
1874                         if (k.k && !bkey_err(k))
1875                                 iter->k = child->k;
1876                 } else {
1877                         struct bpos pos = iter->pos;
1878
1879                         k = bch2_btree_iter_peek(iter);
1880                         iter->pos = pos;
1881                 }
1882
1883                 if (unlikely(bkey_err(k)))
1884                         return k;
1885         }
1886
1887         if (!(iter->flags & BTREE_ITER_IS_EXTENTS)) {
1888                 if (!k.k ||
1889                     ((iter->flags & BTREE_ITER_ALL_SNAPSHOTS)
1890                      ? bpos_cmp(iter->pos, k.k->p)
1891                      : bkey_cmp(iter->pos, k.k->p))) {
1892                         bkey_init(&iter->k);
1893                         iter->k.p = iter->pos;
1894                         k = (struct bkey_s_c) { &iter->k, NULL };
1895                 }
1896         } else {
1897                 struct bpos next = k.k ? bkey_start_pos(k.k) : POS_MAX;
1898
1899                 if (bkey_cmp(iter->pos, next) < 0) {
1900                         bkey_init(&iter->k);
1901                         iter->k.p = iter->pos;
1902                         bch2_key_resize(&iter->k,
1903                                         min_t(u64, KEY_SIZE_MAX,
1904                                               (next.inode == iter->pos.inode
1905                                                ? next.offset
1906                                                : KEY_OFFSET_MAX) -
1907                                               iter->pos.offset));
1908
1909                         k = (struct bkey_s_c) { &iter->k, NULL };
1910                         EBUG_ON(!k.k->size);
1911                 }
1912         }
1913
1914         bch2_btree_iter_verify_entry_exit(iter);
1915         bch2_btree_iter_verify(iter);
1916         iter->should_be_locked = true;
1917
1918         return k;
1919 }
1920
1921 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
1922 {
1923         if (!bch2_btree_iter_advance(iter))
1924                 return bkey_s_c_null;
1925
1926         return bch2_btree_iter_peek_slot(iter);
1927 }
1928
1929 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
1930 {
1931         if (!bch2_btree_iter_rewind(iter))
1932                 return bkey_s_c_null;
1933
1934         return bch2_btree_iter_peek_slot(iter);
1935 }
1936
1937 static inline void bch2_btree_iter_init(struct btree_trans *trans,
1938                         struct btree_iter *iter, enum btree_id btree_id)
1939 {
1940         struct bch_fs *c = trans->c;
1941         unsigned i;
1942
1943         iter->trans                     = trans;
1944         iter->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1945         iter->btree_id                  = btree_id;
1946         iter->real_pos                  = POS_MIN;
1947         iter->level                     = 0;
1948         iter->min_depth                 = 0;
1949         iter->locks_want                = 0;
1950         iter->nodes_locked              = 0;
1951         iter->nodes_intent_locked       = 0;
1952         for (i = 0; i < ARRAY_SIZE(iter->l); i++)
1953                 iter->l[i].b            = BTREE_ITER_NO_NODE_INIT;
1954
1955         prefetch(c->btree_roots[btree_id].b);
1956 }
1957
1958 /* new transactional stuff: */
1959
1960 static void btree_iter_child_free(struct btree_iter *iter)
1961 {
1962         struct btree_iter *child = btree_iter_child(iter);
1963
1964         if (child) {
1965                 bch2_trans_iter_free(iter->trans, child);
1966                 iter->child_idx = U8_MAX;
1967         }
1968 }
1969
1970 static struct btree_iter *btree_iter_child_alloc(struct btree_iter *iter,
1971                                                  unsigned long ip)
1972 {
1973         struct btree_trans *trans = iter->trans;
1974         struct btree_iter *child = btree_iter_child(iter);
1975
1976         if (!child) {
1977                 child = btree_trans_iter_alloc(trans);
1978                 child->ip_allocated     = ip;
1979                 iter->child_idx         = child->idx;
1980
1981                 trans->iters_live       |= 1ULL << child->idx;
1982                 trans->iters_touched    |= 1ULL << child->idx;
1983         }
1984
1985         return child;
1986 }
1987
1988 static inline void __bch2_trans_iter_free(struct btree_trans *trans,
1989                                           unsigned idx)
1990 {
1991         btree_iter_child_free(&trans->iters[idx]);
1992
1993         __bch2_btree_iter_unlock(&trans->iters[idx]);
1994         trans->iters_linked             &= ~(1ULL << idx);
1995         trans->iters_live               &= ~(1ULL << idx);
1996         trans->iters_touched            &= ~(1ULL << idx);
1997 }
1998
1999 int bch2_trans_iter_put(struct btree_trans *trans,
2000                         struct btree_iter *iter)
2001 {
2002         int ret;
2003
2004         if (IS_ERR_OR_NULL(iter))
2005                 return 0;
2006
2007         BUG_ON(trans->iters + iter->idx != iter);
2008         BUG_ON(!btree_iter_live(trans, iter));
2009
2010         ret = btree_iter_err(iter);
2011
2012         if (!(trans->iters_touched & (1ULL << iter->idx)) &&
2013             !(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT))
2014                 __bch2_trans_iter_free(trans, iter->idx);
2015
2016         trans->iters_live       &= ~(1ULL << iter->idx);
2017         return ret;
2018 }
2019
2020 int bch2_trans_iter_free(struct btree_trans *trans,
2021                          struct btree_iter *iter)
2022 {
2023         if (IS_ERR_OR_NULL(iter))
2024                 return 0;
2025
2026         set_btree_iter_dontneed(trans, iter);
2027
2028         return bch2_trans_iter_put(trans, iter);
2029 }
2030
2031 noinline __cold
2032 static void btree_trans_iter_alloc_fail(struct btree_trans *trans)
2033 {
2034
2035         struct btree_iter *iter;
2036         struct btree_insert_entry *i;
2037         char buf[100];
2038
2039         trans_for_each_iter(trans, iter)
2040                 printk(KERN_ERR "iter: btree %s pos %s%s%s%s %pS\n",
2041                        bch2_btree_ids[iter->btree_id],
2042                        (bch2_bpos_to_text(&PBUF(buf), iter->pos), buf),
2043                        btree_iter_live(trans, iter) ? " live" : "",
2044                        (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "",
2045                        iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "",
2046                        (void *) iter->ip_allocated);
2047
2048         trans_for_each_update(trans, i) {
2049                 char buf[300];
2050
2051                 bch2_bkey_val_to_text(&PBUF(buf), trans->c, bkey_i_to_s_c(i->k));
2052                 printk(KERN_ERR "update: btree %s %s\n",
2053                        bch2_btree_ids[i->iter->btree_id], buf);
2054         }
2055         panic("trans iter oveflow\n");
2056 }
2057
2058 static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans)
2059 {
2060         struct btree_iter *iter;
2061         unsigned idx;
2062
2063         if (unlikely(trans->iters_linked ==
2064                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
2065                 btree_trans_iter_alloc_fail(trans);
2066
2067         idx = __ffs64(~trans->iters_linked);
2068         iter = &trans->iters[idx];
2069
2070         iter->trans             = trans;
2071         iter->idx               = idx;
2072         iter->child_idx         = U8_MAX;
2073         iter->flags             = 0;
2074         iter->nodes_locked      = 0;
2075         iter->nodes_intent_locked = 0;
2076         trans->iters_linked     |= 1ULL << idx;
2077         return iter;
2078 }
2079
2080 static void btree_iter_copy(struct btree_iter *dst, struct btree_iter *src)
2081 {
2082         unsigned i;
2083
2084         __bch2_btree_iter_unlock(dst);
2085         btree_iter_child_free(dst);
2086
2087         memcpy(&dst->flags, &src->flags,
2088                sizeof(struct btree_iter) - offsetof(struct btree_iter, flags));
2089
2090         for (i = 0; i < BTREE_MAX_DEPTH; i++)
2091                 if (btree_node_locked(dst, i))
2092                         six_lock_increment(&dst->l[i].b->c.lock,
2093                                            __btree_lock_want(dst, i));
2094
2095         dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
2096         dst->flags &= ~BTREE_ITER_SET_POS_AFTER_COMMIT;
2097 }
2098
2099 struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans,
2100                                          unsigned btree_id, struct bpos pos,
2101                                          unsigned locks_want,
2102                                          unsigned depth,
2103                                          unsigned flags)
2104 {
2105         struct btree_iter *iter, *best = NULL;
2106         struct bpos real_pos, pos_min = POS_MIN;
2107
2108         if ((flags & BTREE_ITER_TYPE) != BTREE_ITER_NODES &&
2109             btree_node_type_is_extents(btree_id) &&
2110             !(flags & BTREE_ITER_NOT_EXTENTS) &&
2111             !(flags & BTREE_ITER_ALL_SNAPSHOTS))
2112                 flags |= BTREE_ITER_IS_EXTENTS;
2113
2114         if ((flags & BTREE_ITER_TYPE) != BTREE_ITER_NODES &&
2115             !btree_type_has_snapshots(btree_id))
2116                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
2117
2118         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS))
2119                 pos.snapshot = btree_type_has_snapshots(btree_id)
2120                         ? U32_MAX : 0;
2121
2122         real_pos = pos;
2123
2124         if ((flags & BTREE_ITER_IS_EXTENTS) &&
2125             bkey_cmp(pos, POS_MAX))
2126                 real_pos = bpos_nosnap_successor(pos);
2127
2128         trans_for_each_iter(trans, iter) {
2129                 if (btree_iter_type(iter) != (flags & BTREE_ITER_TYPE))
2130                         continue;
2131
2132                 if (iter->btree_id != btree_id)
2133                         continue;
2134
2135                 if (best) {
2136                         int cmp = bkey_cmp(bpos_diff(best->real_pos, real_pos),
2137                                            bpos_diff(iter->real_pos, real_pos));
2138
2139                         if (cmp < 0 ||
2140                             ((cmp == 0 && btree_iter_keep(trans, iter))))
2141                                 continue;
2142                 }
2143
2144                 best = iter;
2145         }
2146
2147         if (!best) {
2148                 iter = btree_trans_iter_alloc(trans);
2149                 bch2_btree_iter_init(trans, iter, btree_id);
2150         } else if (btree_iter_keep(trans, best)) {
2151                 iter = btree_trans_iter_alloc(trans);
2152                 btree_iter_copy(iter, best);
2153         } else {
2154                 iter = best;
2155         }
2156
2157         trans->iters_live       |= 1ULL << iter->idx;
2158         trans->iters_touched    |= 1ULL << iter->idx;
2159
2160         iter->flags = flags;
2161
2162         iter->snapshot = pos.snapshot;
2163
2164         /*
2165          * If the iterator has locks_want greater than requested, we explicitly
2166          * do not downgrade it here - on transaction restart because btree node
2167          * split needs to upgrade locks, we might be putting/getting the
2168          * iterator again. Downgrading iterators only happens via an explicit
2169          * bch2_trans_downgrade().
2170          */
2171
2172         locks_want = min(locks_want, BTREE_MAX_DEPTH);
2173         if (locks_want > iter->locks_want) {
2174                 iter->locks_want = locks_want;
2175                 btree_iter_get_locks(iter, true, _THIS_IP_);
2176         }
2177
2178         while (iter->level != depth) {
2179                 btree_node_unlock(iter, iter->level);
2180                 iter->l[iter->level].b = BTREE_ITER_NO_NODE_INIT;
2181                 iter->uptodate = BTREE_ITER_NEED_TRAVERSE;
2182                 if (iter->level < depth)
2183                         iter->level++;
2184                 else
2185                         iter->level--;
2186         }
2187
2188         iter->min_depth = depth;
2189
2190         bch2_btree_iter_set_pos(iter, pos);
2191         btree_iter_set_search_pos(iter, real_pos);
2192
2193         trace_trans_get_iter(_RET_IP_, trans->ip,
2194                              btree_id,
2195                              &real_pos, locks_want, iter->uptodate,
2196                              best ? &best->real_pos     : &pos_min,
2197                              best ? best->locks_want    : U8_MAX,
2198                              best ? best->uptodate      : U8_MAX);
2199
2200         return iter;
2201 }
2202
2203 struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *trans,
2204                                             enum btree_id btree_id,
2205                                             struct bpos pos,
2206                                             unsigned locks_want,
2207                                             unsigned depth,
2208                                             unsigned flags)
2209 {
2210         struct btree_iter *iter =
2211                 __bch2_trans_get_iter(trans, btree_id, pos,
2212                                       locks_want, depth,
2213                                       BTREE_ITER_NODES|
2214                                       BTREE_ITER_NOT_EXTENTS|
2215                                       BTREE_ITER_ALL_SNAPSHOTS|
2216                                       flags);
2217
2218         BUG_ON(bkey_cmp(iter->pos, pos));
2219         BUG_ON(iter->locks_want != min(locks_want, BTREE_MAX_DEPTH));
2220         BUG_ON(iter->level      != depth);
2221         BUG_ON(iter->min_depth  != depth);
2222         iter->ip_allocated = _RET_IP_;
2223
2224         return iter;
2225 }
2226
2227 struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *trans,
2228                                         struct btree_iter *src)
2229 {
2230         struct btree_iter *iter;
2231
2232         iter = btree_trans_iter_alloc(trans);
2233         btree_iter_copy(iter, src);
2234
2235         trans->iters_live |= 1ULL << iter->idx;
2236         /*
2237          * We don't need to preserve this iter since it's cheap to copy it
2238          * again - this will cause trans_iter_put() to free it right away:
2239          */
2240         set_btree_iter_dontneed(trans, iter);
2241
2242         return iter;
2243 }
2244
2245 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2246 {
2247         size_t new_top = trans->mem_top + size;
2248         void *p;
2249
2250         if (new_top > trans->mem_bytes) {
2251                 size_t old_bytes = trans->mem_bytes;
2252                 size_t new_bytes = roundup_pow_of_two(new_top);
2253                 void *new_mem;
2254
2255                 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2256
2257                 new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2258                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2259                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2260                         new_bytes = BTREE_TRANS_MEM_MAX;
2261                         kfree(trans->mem);
2262                 }
2263
2264                 if (!new_mem)
2265                         return ERR_PTR(-ENOMEM);
2266
2267                 trans->mem = new_mem;
2268                 trans->mem_bytes = new_bytes;
2269
2270                 if (old_bytes) {
2271                         trace_trans_restart_mem_realloced(trans->ip, _RET_IP_, new_bytes);
2272                         return ERR_PTR(-EINTR);
2273                 }
2274         }
2275
2276         p = trans->mem + trans->mem_top;
2277         trans->mem_top += size;
2278         memset(p, 0, size);
2279         return p;
2280 }
2281
2282 inline void bch2_trans_unlink_iters(struct btree_trans *trans)
2283 {
2284         u64 iters = trans->iters_linked &
2285                 ~trans->iters_touched &
2286                 ~trans->iters_live;
2287
2288         while (iters) {
2289                 unsigned idx = __ffs64(iters);
2290
2291                 iters &= ~(1ULL << idx);
2292                 __bch2_trans_iter_free(trans, idx);
2293         }
2294 }
2295
2296 void bch2_trans_reset(struct btree_trans *trans, unsigned flags)
2297 {
2298         struct btree_iter *iter;
2299
2300         trans_for_each_iter(trans, iter) {
2301                 iter->flags &= ~(BTREE_ITER_KEEP_UNTIL_COMMIT|
2302                                  BTREE_ITER_SET_POS_AFTER_COMMIT);
2303                 iter->should_be_locked = false;
2304         }
2305
2306         bch2_trans_unlink_iters(trans);
2307
2308         trans->iters_touched &= trans->iters_live;
2309
2310         trans->extra_journal_res        = 0;
2311         trans->nr_updates               = 0;
2312         trans->mem_top                  = 0;
2313
2314         trans->hooks                    = NULL;
2315         trans->extra_journal_entries    = NULL;
2316         trans->extra_journal_entry_u64s = 0;
2317
2318         if (trans->fs_usage_deltas) {
2319                 trans->fs_usage_deltas->used = 0;
2320                 memset(&trans->fs_usage_deltas->memset_start, 0,
2321                        (void *) &trans->fs_usage_deltas->memset_end -
2322                        (void *) &trans->fs_usage_deltas->memset_start);
2323         }
2324
2325         if (!(flags & TRANS_RESET_NOUNLOCK))
2326                 bch2_trans_cond_resched(trans);
2327
2328         if (!(flags & TRANS_RESET_NOTRAVERSE) &&
2329             trans->iters_linked)
2330                 bch2_btree_iter_traverse_all(trans);
2331 }
2332
2333 static void bch2_trans_alloc_iters(struct btree_trans *trans, struct bch_fs *c)
2334 {
2335         size_t iters_bytes      = sizeof(struct btree_iter) * BTREE_ITER_MAX;
2336         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2337         void *p = NULL;
2338
2339         BUG_ON(trans->used_mempool);
2340
2341 #ifdef __KERNEL__
2342         p = this_cpu_xchg(c->btree_iters_bufs->iter, NULL);
2343 #endif
2344         if (!p)
2345                 p = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
2346
2347         trans->iters            = p; p += iters_bytes;
2348         trans->updates          = p; p += updates_bytes;
2349 }
2350
2351 void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
2352                      unsigned expected_nr_iters,
2353                      size_t expected_mem_bytes)
2354         __acquires(&c->btree_trans_barrier)
2355 {
2356         memset(trans, 0, sizeof(*trans));
2357         trans->c                = c;
2358         trans->ip               = _RET_IP_;
2359
2360         /*
2361          * reallocating iterators currently completely breaks
2362          * bch2_trans_iter_put(), we always allocate the max:
2363          */
2364         bch2_trans_alloc_iters(trans, c);
2365
2366         if (expected_mem_bytes) {
2367                 trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes);
2368                 trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL);
2369
2370                 if (!unlikely(trans->mem)) {
2371                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2372                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2373                 }
2374         }
2375
2376         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2377
2378 #ifdef CONFIG_BCACHEFS_DEBUG
2379         trans->pid = current->pid;
2380         mutex_lock(&c->btree_trans_lock);
2381         list_add(&trans->list, &c->btree_trans_list);
2382         mutex_unlock(&c->btree_trans_lock);
2383 #endif
2384 }
2385
2386 int bch2_trans_exit(struct btree_trans *trans)
2387         __releases(&c->btree_trans_barrier)
2388 {
2389         struct bch_fs *c = trans->c;
2390
2391         bch2_trans_unlock(trans);
2392
2393 #ifdef CONFIG_BCACHEFS_DEBUG
2394         if (trans->iters_live) {
2395                 struct btree_iter *iter;
2396
2397                 trans_for_each_iter(trans, iter)
2398                         btree_iter_child_free(iter);
2399         }
2400
2401         if (trans->iters_live) {
2402                 struct btree_iter *iter;
2403
2404                 bch_err(c, "btree iterators leaked!");
2405                 trans_for_each_iter(trans, iter)
2406                         if (btree_iter_live(trans, iter))
2407                                 printk(KERN_ERR "  btree %s allocated at %pS\n",
2408                                        bch2_btree_ids[iter->btree_id],
2409                                        (void *) iter->ip_allocated);
2410                 /* Be noisy about this: */
2411                 bch2_fatal_error(c);
2412         }
2413
2414         mutex_lock(&trans->c->btree_trans_lock);
2415         list_del(&trans->list);
2416         mutex_unlock(&trans->c->btree_trans_lock);
2417 #endif
2418
2419         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2420
2421         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
2422
2423         if (trans->fs_usage_deltas) {
2424                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
2425                     REPLICAS_DELTA_LIST_MAX)
2426                         mempool_free(trans->fs_usage_deltas,
2427                                      &trans->c->replicas_delta_pool);
2428                 else
2429                         kfree(trans->fs_usage_deltas);
2430         }
2431
2432         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
2433                 mempool_free(trans->mem, &trans->c->btree_trans_mem_pool);
2434         else
2435                 kfree(trans->mem);
2436
2437 #ifdef __KERNEL__
2438         /*
2439          * Userspace doesn't have a real percpu implementation:
2440          */
2441         trans->iters = this_cpu_xchg(c->btree_iters_bufs->iter, trans->iters);
2442 #endif
2443
2444         if (trans->iters)
2445                 mempool_free(trans->iters, &trans->c->btree_iters_pool);
2446
2447         trans->mem      = (void *) 0x1;
2448         trans->iters    = (void *) 0x1;
2449
2450         return trans->error ? -EIO : 0;
2451 }
2452
2453 static void __maybe_unused
2454 bch2_btree_iter_node_to_text(struct printbuf *out,
2455                              struct btree_bkey_cached_common *_b,
2456                              enum btree_iter_type type)
2457 {
2458         pr_buf(out, "    l=%u %s:",
2459                _b->level, bch2_btree_ids[_b->btree_id]);
2460         bch2_bpos_to_text(out, btree_node_pos(_b, type));
2461 }
2462
2463 #ifdef CONFIG_BCACHEFS_DEBUG
2464 static bool trans_has_btree_nodes_locked(struct btree_trans *trans)
2465 {
2466         struct btree_iter *iter;
2467
2468         trans_for_each_iter(trans, iter)
2469                 if (btree_iter_type(iter) != BTREE_ITER_CACHED &&
2470                     iter->nodes_locked)
2471                         return true;
2472         return false;
2473 }
2474 #endif
2475
2476 void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
2477 {
2478 #ifdef CONFIG_BCACHEFS_DEBUG
2479         struct btree_trans *trans;
2480         struct btree_iter *iter;
2481         struct btree *b;
2482         unsigned l;
2483
2484         mutex_lock(&c->btree_trans_lock);
2485         list_for_each_entry(trans, &c->btree_trans_list, list) {
2486                 if (!trans_has_btree_nodes_locked(trans))
2487                         continue;
2488
2489                 pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip);
2490
2491                 trans_for_each_iter(trans, iter) {
2492                         if (!iter->nodes_locked)
2493                                 continue;
2494
2495                         pr_buf(out, "  iter %u %c %s:",
2496                                iter->idx,
2497                                btree_iter_type(iter) == BTREE_ITER_CACHED ? 'c' : 'b',
2498                                bch2_btree_ids[iter->btree_id]);
2499                         bch2_bpos_to_text(out, iter->pos);
2500                         pr_buf(out, "\n");
2501
2502                         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
2503                                 if (btree_node_locked(iter, l)) {
2504                                         pr_buf(out, "    %s l=%u ",
2505                                                btree_node_intent_locked(iter, l) ? "i" : "r", l);
2506                                         bch2_btree_iter_node_to_text(out,
2507                                                         (void *) iter->l[l].b,
2508                                                         btree_iter_type(iter));
2509                                         pr_buf(out, "\n");
2510                                 }
2511                         }
2512                 }
2513
2514                 b = READ_ONCE(trans->locking);
2515                 if (b) {
2516                         iter = &trans->iters[trans->locking_iter_idx];
2517                         pr_buf(out, "  locking iter %u %c l=%u %s:",
2518                                trans->locking_iter_idx,
2519                                btree_iter_type(iter) == BTREE_ITER_CACHED ? 'c' : 'b',
2520                                trans->locking_level,
2521                                bch2_btree_ids[trans->locking_btree_id]);
2522                         bch2_bpos_to_text(out, trans->locking_pos);
2523
2524                         pr_buf(out, " node ");
2525                         bch2_btree_iter_node_to_text(out,
2526                                         (void *) b,
2527                                         btree_iter_type(iter));
2528                         pr_buf(out, "\n");
2529                 }
2530         }
2531         mutex_unlock(&c->btree_trans_lock);
2532 #endif
2533 }
2534
2535 void bch2_fs_btree_iter_exit(struct bch_fs *c)
2536 {
2537         mempool_exit(&c->btree_trans_mem_pool);
2538         mempool_exit(&c->btree_iters_pool);
2539         cleanup_srcu_struct(&c->btree_trans_barrier);
2540 }
2541
2542 int bch2_fs_btree_iter_init(struct bch_fs *c)
2543 {
2544         unsigned nr = BTREE_ITER_MAX;
2545
2546         INIT_LIST_HEAD(&c->btree_trans_list);
2547         mutex_init(&c->btree_trans_lock);
2548
2549         return  init_srcu_struct(&c->btree_trans_barrier) ?:
2550                 mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
2551                         sizeof(struct btree_iter) * nr +
2552                         sizeof(struct btree_insert_entry) * nr) ?:
2553                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
2554                                           BTREE_TRANS_MEM_MAX);
2555 }