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 #include "subvolume.h"
  17
  18 #include <linux/prefetch.h>
  19 #include <trace/events/bcachefs.h>
  20
  21 static void btree_trans_verify_sorted(struct btree_trans *);
  22 static void btree_path_check_sort(struct btree_trans *, struct btree_path *, int);
  23
  24 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
  25 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
  26                                        struct btree_path *);
  27
  28 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
  29
  30 /*
  31  * Unlocks before scheduling
  32  * Note: does not revalidate iterator
  33  */
  34 static inline int bch2_trans_cond_resched(struct btree_trans *trans)
  35 {
  36         if (need_resched() || race_fault()) {
  37                 bch2_trans_unlock(trans);
  38                 schedule();
  39                 return bch2_trans_relock(trans) ? 0 : -EINTR;
  40         } else {
  41                 return 0;
  42         }
  43 }
  44
  45 static inline int __btree_path_cmp(const struct btree_path *l,
  46                                    enum btree_id        r_btree_id,
  47                                    bool                 r_cached,
  48                                    struct bpos          r_pos,
  49                                    unsigned             r_level)
  50 {
  51         return   cmp_int(l->btree_id,   r_btree_id) ?:
  52                  cmp_int((int) l->cached,       (int) r_cached) ?:
  53                  bpos_cmp(l->pos,       r_pos) ?:
  54                 -cmp_int(l->level,      r_level);
  55 }
  56
  57 static inline int btree_path_cmp(const struct btree_path *l,
  58                                  const struct btree_path *r)
  59 {
  60         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  61 }
  62
  63 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  64 {
  65         /* Are we iterating over keys in all snapshots? */
  66         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  67                 p = bpos_successor(p);
  68         } else {
  69                 p = bpos_nosnap_successor(p);
  70                 p.snapshot = iter->snapshot;
  71         }
  72
  73         return p;
  74 }
  75
  76 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  77 {
  78         /* Are we iterating over keys in all snapshots? */
  79         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  80                 p = bpos_predecessor(p);
  81         } else {
  82                 p = bpos_nosnap_predecessor(p);
  83                 p.snapshot = iter->snapshot;
  84         }
  85
  86         return p;
  87 }
  88
  89 static inline bool is_btree_node(struct btree_path *path, unsigned l)
  90 {
  91         return l < BTREE_MAX_DEPTH &&
  92                 (unsigned long) path->l[l].b >= 128;
  93 }
  94
  95 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
  96 {
  97         struct bpos pos = iter->pos;
  98
  99         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 100             bkey_cmp(pos, POS_MAX))
 101                 pos = bkey_successor(iter, pos);
 102         return pos;
 103 }
 104
 105 static inline bool btree_path_pos_before_node(struct btree_path *path,
 106                                               struct btree *b)
 107 {
 108         return bpos_cmp(path->pos, b->data->min_key) < 0;
 109 }
 110
 111 static inline bool btree_path_pos_after_node(struct btree_path *path,
 112                                              struct btree *b)
 113 {
 114         return bpos_cmp(b->key.k.p, path->pos) < 0;
 115 }
 116
 117 static inline bool btree_path_pos_in_node(struct btree_path *path,
 118                                           struct btree *b)
 119 {
 120         return path->btree_id == b->c.btree_id &&
 121                 !btree_path_pos_before_node(path, b) &&
 122                 !btree_path_pos_after_node(path, b);
 123 }
 124
 125 /* Btree node locking: */
 126
 127 void bch2_btree_node_unlock_write(struct btree_trans *trans,
 128                         struct btree_path *path, struct btree *b)
 129 {
 130         bch2_btree_node_unlock_write_inlined(trans, path, b);
 131 }
 132
 133 void __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree *b)
 134 {
 135         struct btree_path *linked;
 136         unsigned readers = 0;
 137
 138         trans_for_each_path(trans, linked)
 139                 if (linked->l[b->c.level].b == b &&
 140                     btree_node_read_locked(linked, b->c.level))
 141                         readers++;
 142
 143         /*
 144          * Must drop our read locks before calling six_lock_write() -
 145          * six_unlock() won't do wakeups until the reader count
 146          * goes to 0, and it's safe because we have the node intent
 147          * locked:
 148          */
 149         atomic64_sub(__SIX_VAL(read_lock, readers),
 150                      &b->c.lock.state.counter);
 151         btree_node_lock_type(trans->c, b, SIX_LOCK_write);
 152         atomic64_add(__SIX_VAL(read_lock, readers),
 153                      &b->c.lock.state.counter);
 154 }
 155
 156 bool __bch2_btree_node_relock(struct btree_trans *trans,
 157                               struct btree_path *path, unsigned level)
 158 {
 159         struct btree *b = btree_path_node(path, level);
 160         int want = __btree_lock_want(path, level);
 161
 162         if (!is_btree_node(path, level))
 163                 return false;
 164
 165         if (race_fault())
 166                 return false;
 167
 168         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 169             (btree_node_lock_seq_matches(path, b, level) &&
 170              btree_node_lock_increment(trans, b, level, want))) {
 171                 mark_btree_node_locked(path, level, want);
 172                 return true;
 173         } else {
 174                 return false;
 175         }
 176 }
 177
 178 bool bch2_btree_node_upgrade(struct btree_trans *trans,
 179                              struct btree_path *path, unsigned level)
 180 {
 181         struct btree *b = path->l[level].b;
 182
 183         if (!is_btree_node(path, level))
 184                 return false;
 185
 186         switch (btree_lock_want(path, level)) {
 187         case BTREE_NODE_UNLOCKED:
 188                 BUG_ON(btree_node_locked(path, level));
 189                 return true;
 190         case BTREE_NODE_READ_LOCKED:
 191                 BUG_ON(btree_node_intent_locked(path, level));
 192                 return bch2_btree_node_relock(trans, path, level);
 193         case BTREE_NODE_INTENT_LOCKED:
 194                 break;
 195         }
 196
 197         if (btree_node_intent_locked(path, level))
 198                 return true;
 199
 200         if (race_fault())
 201                 return false;
 202
 203         if (btree_node_locked(path, level)
 204             ? six_lock_tryupgrade(&b->c.lock)
 205             : six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 206                 goto success;
 207
 208         if (btree_node_lock_seq_matches(path, b, level) &&
 209             btree_node_lock_increment(trans, b, level, BTREE_NODE_INTENT_LOCKED)) {
 210                 btree_node_unlock(path, level);
 211                 goto success;
 212         }
 213
 214         return false;
 215 success:
 216         mark_btree_node_intent_locked(path, level);
 217         return true;
 218 }
 219
 220 static inline bool btree_path_get_locks(struct btree_trans *trans,
 221                                         struct btree_path *path,
 222                                         bool upgrade, unsigned long trace_ip)
 223 {
 224         unsigned l = path->level;
 225         int fail_idx = -1;
 226
 227         do {
 228                 if (!btree_path_node(path, l))
 229                         break;
 230
 231                 if (!(upgrade
 232                       ? bch2_btree_node_upgrade(trans, path, l)
 233                       : bch2_btree_node_relock(trans, path, l)))
 234                         fail_idx = l;
 235
 236                 l++;
 237         } while (l < path->locks_want);
 238
 239         /*
 240          * When we fail to get a lock, we have to ensure that any child nodes
 241          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 242          * the node that we failed to relock:
 243          */
 244         if (fail_idx >= 0) {
 245                 __bch2_btree_path_unlock(path);
 246                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 247
 248                 do {
 249                         path->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS;
 250                         --fail_idx;
 251                 } while (fail_idx >= 0);
 252         }
 253
 254         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 255                 path->uptodate = BTREE_ITER_UPTODATE;
 256
 257         bch2_trans_verify_locks(trans);
 258
 259         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 260 }
 261
 262 static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b,
 263                                   bool cached)
 264 {
 265         return !cached
 266                 ? container_of(_b, struct btree, c)->key.k.p
 267                 : container_of(_b, struct bkey_cached, c)->key.pos;
 268 }
 269
 270 /* Slowpath: */
 271 bool __bch2_btree_node_lock(struct btree_trans *trans,
 272                             struct btree_path *path,
 273                             struct btree *b,
 274                             struct bpos pos, unsigned level,
 275                             enum six_lock_type type,
 276                             six_lock_should_sleep_fn should_sleep_fn, void *p,
 277                             unsigned long ip)
 278 {
 279         struct btree_path *linked, *deadlock_path = NULL;
 280         u64 start_time = local_clock();
 281         unsigned reason = 9;
 282         bool ret;
 283
 284         /* Check if it's safe to block: */
 285         trans_for_each_path(trans, linked) {
 286                 if (!linked->nodes_locked)
 287                         continue;
 288
 289                 /*
 290                  * Can't block taking an intent lock if we have _any_ nodes read
 291                  * locked:
 292                  *
 293                  * - Our read lock blocks another thread with an intent lock on
 294                  *   the same node from getting a write lock, and thus from
 295                  *   dropping its intent lock
 296                  *
 297                  * - And the other thread may have multiple nodes intent locked:
 298                  *   both the node we want to intent lock, and the node we
 299                  *   already have read locked - deadlock:
 300                  */
 301                 if (type == SIX_LOCK_intent &&
 302                     linked->nodes_locked != linked->nodes_intent_locked) {
 303                         deadlock_path = linked;
 304                         reason = 1;
 305                 }
 306
 307                 if (linked->btree_id != path->btree_id) {
 308                         if (linked->btree_id > path->btree_id) {
 309                                 deadlock_path = linked;
 310                                 reason = 3;
 311                         }
 312                         continue;
 313                 }
 314
 315                 /*
 316                  * Within the same btree, cached paths come before non
 317                  * cached paths:
 318                  */
 319                 if (linked->cached != path->cached) {
 320                         if (path->cached) {
 321                                 deadlock_path = linked;
 322                                 reason = 4;
 323                         }
 324                         continue;
 325                 }
 326
 327                 /*
 328                  * Interior nodes must be locked before their descendants: if
 329                  * another path has possible descendants locked of the node
 330                  * we're about to lock, it must have the ancestors locked too:
 331                  */
 332                 if (level > __fls(linked->nodes_locked)) {
 333                         deadlock_path = linked;
 334                         reason = 5;
 335                 }
 336
 337                 /* Must lock btree nodes in key order: */
 338                 if (btree_node_locked(linked, level) &&
 339                     bpos_cmp(pos, btree_node_pos((void *) linked->l[level].b,
 340                                                  linked->cached)) <= 0) {
 341                         deadlock_path = linked;
 342                         reason = 7;
 343                         BUG_ON(trans->in_traverse_all);
 344                 }
 345         }
 346
 347         if (unlikely(deadlock_path)) {
 348                 trace_trans_restart_would_deadlock(trans->ip, ip,
 349                                 trans->in_traverse_all, reason,
 350                                 deadlock_path->btree_id,
 351                                 deadlock_path->cached,
 352                                 &deadlock_path->pos,
 353                                 path->btree_id,
 354                                 path->cached,
 355                                 &pos);
 356                 btree_trans_restart(trans);
 357                 return false;
 358         }
 359
 360         if (six_trylock_type(&b->c.lock, type))
 361                 return true;
 362
 363 #ifdef CONFIG_BCACHEFS_DEBUG
 364         trans->locking_path_idx = path->idx;
 365         trans->locking_pos      = pos;
 366         trans->locking_btree_id = path->btree_id;
 367         trans->locking_level    = level;
 368         trans->locking          = b;
 369 #endif
 370
 371         ret = six_lock_type(&b->c.lock, type, should_sleep_fn, p) == 0;
 372
 373 #ifdef CONFIG_BCACHEFS_DEBUG
 374         trans->locking = NULL;
 375 #endif
 376         if (ret)
 377                 bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)],
 378                                        start_time);
 379         return ret;
 380 }
 381
 382 /* Btree iterator locking: */
 383
 384 #ifdef CONFIG_BCACHEFS_DEBUG
 385
 386 static void bch2_btree_path_verify_locks(struct btree_path *path)
 387 {
 388         unsigned l;
 389
 390         if (!path->nodes_locked) {
 391                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
 392                        btree_path_node(path, path->level));
 393                 return;
 394         }
 395
 396         for (l = 0; btree_path_node(path, l); l++)
 397                 BUG_ON(btree_lock_want(path, l) !=
 398                        btree_node_locked_type(path, l));
 399 }
 400
 401 void bch2_trans_verify_locks(struct btree_trans *trans)
 402 {
 403         struct btree_path *path;
 404
 405         trans_for_each_path(trans, path)
 406                 bch2_btree_path_verify_locks(path);
 407 }
 408 #else
 409 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
 410 #endif
 411
 412 /* Btree path locking: */
 413
 414 /*
 415  * Only for btree_cache.c - only relocks intent locks
 416  */
 417 bool bch2_btree_path_relock_intent(struct btree_trans *trans,
 418                                    struct btree_path *path)
 419 {
 420         unsigned l;
 421
 422         for (l = path->level;
 423              l < path->locks_want && btree_path_node(path, l);
 424              l++) {
 425                 if (!bch2_btree_node_relock(trans, path, l)) {
 426                         __bch2_btree_path_unlock(path);
 427                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 428                         btree_trans_restart(trans);
 429                         return false;
 430                 }
 431         }
 432
 433         return true;
 434 }
 435
 436 __flatten
 437 static bool bch2_btree_path_relock(struct btree_trans *trans,
 438                         struct btree_path *path, unsigned long trace_ip)
 439 {
 440         bool ret = btree_path_get_locks(trans, path, false, trace_ip);
 441
 442         if (!ret)
 443                 btree_trans_restart(trans);
 444         return ret;
 445 }
 446
 447 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 448                                struct btree_path *path,
 449                                unsigned new_locks_want)
 450 {
 451         struct btree_path *linked;
 452
 453         EBUG_ON(path->locks_want >= new_locks_want);
 454
 455         path->locks_want = new_locks_want;
 456
 457         if (btree_path_get_locks(trans, path, true, _THIS_IP_))
 458                 return true;
 459
 460         /*
 461          * XXX: this is ugly - we'd prefer to not be mucking with other
 462          * iterators in the btree_trans here.
 463          *
 464          * On failure to upgrade the iterator, setting iter->locks_want and
 465          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 466          * get the locks we want on transaction restart.
 467          *
 468          * But if this iterator was a clone, on transaction restart what we did
 469          * to this iterator isn't going to be preserved.
 470          *
 471          * Possibly we could add an iterator field for the parent iterator when
 472          * an iterator is a copy - for now, we'll just upgrade any other
 473          * iterators with the same btree id.
 474          *
 475          * The code below used to be needed to ensure ancestor nodes get locked
 476          * before interior nodes - now that's handled by
 477          * bch2_btree_path_traverse_all().
 478          */
 479         trans_for_each_path(trans, linked)
 480                 if (linked != path &&
 481                     linked->cached == path->cached &&
 482                     linked->btree_id == path->btree_id &&
 483                     linked->locks_want < new_locks_want) {
 484                         linked->locks_want = new_locks_want;
 485                         btree_path_get_locks(trans, linked, true, _THIS_IP_);
 486                 }
 487
 488         return false;
 489 }
 490
 491 void __bch2_btree_path_downgrade(struct btree_path *path,
 492                                  unsigned new_locks_want)
 493 {
 494         unsigned l;
 495
 496         EBUG_ON(path->locks_want < new_locks_want);
 497
 498         path->locks_want = new_locks_want;
 499
 500         while (path->nodes_locked &&
 501                (l = __fls(path->nodes_locked)) >= path->locks_want) {
 502                 if (l > path->level) {
 503                         btree_node_unlock(path, l);
 504                 } else {
 505                         if (btree_node_intent_locked(path, l)) {
 506                                 six_lock_downgrade(&path->l[l].b->c.lock);
 507                                 path->nodes_intent_locked ^= 1 << l;
 508                         }
 509                         break;
 510                 }
 511         }
 512
 513         bch2_btree_path_verify_locks(path);
 514 }
 515
 516 void bch2_trans_downgrade(struct btree_trans *trans)
 517 {
 518         struct btree_path *path;
 519
 520         trans_for_each_path(trans, path)
 521                 bch2_btree_path_downgrade(path);
 522 }
 523
 524 /* Btree transaction locking: */
 525
 526 bool bch2_trans_relock(struct btree_trans *trans)
 527 {
 528         struct btree_path *path;
 529
 530         if (unlikely(trans->restarted))
 531                 return false;
 532
 533         trans_for_each_path(trans, path)
 534                 if (path->should_be_locked &&
 535                     !bch2_btree_path_relock(trans, path, _RET_IP_)) {
 536                         trace_trans_restart_relock(trans->ip, _RET_IP_,
 537                                         path->btree_id, &path->pos);
 538                         BUG_ON(!trans->restarted);
 539                         return false;
 540                 }
 541         return true;
 542 }
 543
 544 void bch2_trans_unlock(struct btree_trans *trans)
 545 {
 546         struct btree_path *path;
 547
 548         trans_for_each_path(trans, path)
 549                 __bch2_btree_path_unlock(path);
 550
 551         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
 552 }
 553
 554 /* Btree iterator: */
 555
 556 #ifdef CONFIG_BCACHEFS_DEBUG
 557
 558 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 559                                           struct btree_path *path)
 560 {
 561         struct bkey_cached *ck;
 562         bool locked = btree_node_locked(path, 0);
 563
 564         if (!bch2_btree_node_relock(trans, path, 0))
 565                 return;
 566
 567         ck = (void *) path->l[0].b;
 568         BUG_ON(ck->key.btree_id != path->btree_id ||
 569                bkey_cmp(ck->key.pos, path->pos));
 570
 571         if (!locked)
 572                 btree_node_unlock(path, 0);
 573 }
 574
 575 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 576                                 struct btree_path *path, unsigned level)
 577 {
 578         struct btree_path_level *l;
 579         struct btree_node_iter tmp;
 580         bool locked;
 581         struct bkey_packed *p, *k;
 582         char buf1[100], buf2[100], buf3[100];
 583         const char *msg;
 584
 585         if (!bch2_debug_check_iterators)
 586                 return;
 587
 588         l       = &path->l[level];
 589         tmp     = l->iter;
 590         locked  = btree_node_locked(path, level);
 591
 592         if (path->cached) {
 593                 if (!level)
 594                         bch2_btree_path_verify_cached(trans, path);
 595                 return;
 596         }
 597
 598         if (!btree_path_node(path, level))
 599                 return;
 600
 601         if (!bch2_btree_node_relock(trans, path, level))
 602                 return;
 603
 604         BUG_ON(!btree_path_pos_in_node(path, l->b));
 605
 606         bch2_btree_node_iter_verify(&l->iter, l->b);
 607
 608         /*
 609          * For interior nodes, the iterator will have skipped past deleted keys:
 610          */
 611         p = level
 612                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 613                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 614         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 615
 616         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 617                 msg = "before";
 618                 goto err;
 619         }
 620
 621         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 622                 msg = "after";
 623                 goto err;
 624         }
 625
 626         if (!locked)
 627                 btree_node_unlock(path, level);
 628         return;
 629 err:
 630         strcpy(buf2, "(none)");
 631         strcpy(buf3, "(none)");
 632
 633         bch2_bpos_to_text(&PBUF(buf1), path->pos);
 634
 635         if (p) {
 636                 struct bkey uk = bkey_unpack_key(l->b, p);
 637                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 638         }
 639
 640         if (k) {
 641                 struct bkey uk = bkey_unpack_key(l->b, k);
 642                 bch2_bkey_to_text(&PBUF(buf3), &uk);
 643         }
 644
 645         panic("path should be %s key at level %u:\n"
 646               "path pos %s\n"
 647               "prev key %s\n"
 648               "cur  key %s\n",
 649               msg, level, buf1, buf2, buf3);
 650 }
 651
 652 static void bch2_btree_path_verify(struct btree_trans *trans,
 653                                    struct btree_path *path)
 654 {
 655         struct bch_fs *c = trans->c;
 656         unsigned i;
 657
 658         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 659
 660         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 661                 if (!path->l[i].b) {
 662                         BUG_ON(!path->cached &&
 663                                c->btree_roots[path->btree_id].b->c.level > i);
 664                         break;
 665                 }
 666
 667                 bch2_btree_path_verify_level(trans, path, i);
 668         }
 669
 670         bch2_btree_path_verify_locks(path);
 671 }
 672
 673 void bch2_trans_verify_paths(struct btree_trans *trans)
 674 {
 675         struct btree_path *path;
 676
 677         trans_for_each_path(trans, path)
 678                 bch2_btree_path_verify(trans, path);
 679 }
 680
 681 static void bch2_btree_iter_verify(struct btree_iter *iter)
 682 {
 683         struct btree_trans *trans = iter->trans;
 684
 685         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 686
 687         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
 688
 689         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 690                iter->pos.snapshot != iter->snapshot);
 691
 692         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 693                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 694
 695         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 696                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 697                !btree_type_has_snapshots(iter->btree_id));
 698
 699         bch2_btree_path_verify(trans, iter->path);
 700 }
 701
 702 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 703 {
 704         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 705                !iter->pos.snapshot);
 706
 707         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 708                iter->pos.snapshot != iter->snapshot);
 709
 710         BUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
 711                bkey_cmp(iter->pos, iter->k.p) > 0);
 712 }
 713
 714 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 715 {
 716         struct btree_trans *trans = iter->trans;
 717         struct btree_iter copy;
 718         struct bkey_s_c prev;
 719         int ret = 0;
 720
 721         if (!bch2_debug_check_iterators)
 722                 return 0;
 723
 724         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 725                 return 0;
 726
 727         if (bkey_err(k) || !k.k)
 728                 return 0;
 729
 730         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 731                                           iter->snapshot,
 732                                           k.k->p.snapshot));
 733
 734         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 735                              BTREE_ITER_ALL_SNAPSHOTS);
 736         prev = bch2_btree_iter_prev(&copy);
 737         if (!prev.k)
 738                 goto out;
 739
 740         ret = bkey_err(prev);
 741         if (ret)
 742                 goto out;
 743
 744         if (!bkey_cmp(prev.k->p, k.k->p) &&
 745             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 746                                       prev.k->p.snapshot) > 0) {
 747                 char buf1[100], buf2[200];
 748
 749                 bch2_bkey_to_text(&PBUF(buf1), k.k);
 750                 bch2_bkey_to_text(&PBUF(buf2), prev.k);
 751
 752                 panic("iter snap %u\n"
 753                       "k    %s\n"
 754                       "prev %s\n",
 755                       iter->snapshot,
 756                       buf1, buf2);
 757         }
 758 out:
 759         bch2_trans_iter_exit(trans, &copy);
 760         return ret;
 761 }
 762
 763 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 764                             struct bpos pos, bool key_cache)
 765 {
 766         struct btree_path *path;
 767         unsigned idx;
 768         char buf[100];
 769
 770         trans_for_each_path_inorder(trans, path, idx) {
 771                 int cmp = cmp_int(path->btree_id, id) ?:
 772                         cmp_int(path->cached, key_cache);
 773
 774                 if (cmp > 0)
 775                         break;
 776                 if (cmp < 0)
 777                         continue;
 778
 779                 if (!(path->nodes_locked & 1) ||
 780                     !path->should_be_locked)
 781                         continue;
 782
 783                 if (!key_cache) {
 784                         if (bkey_cmp(pos, path->l[0].b->data->min_key) >= 0 &&
 785                             bkey_cmp(pos, path->l[0].b->key.k.p) <= 0)
 786                                 return;
 787                 } else {
 788                         if (!bkey_cmp(pos, path->pos))
 789                                 return;
 790                 }
 791         }
 792
 793         bch2_dump_trans_paths_updates(trans);
 794         panic("not locked: %s %s%s\n",
 795               bch2_btree_ids[id],
 796               (bch2_bpos_to_text(&PBUF(buf), pos), buf),
 797               key_cache ? " cached" : "");
 798 }
 799
 800 #else
 801
 802 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 803                                                 struct btree_path *path, unsigned l) {}
 804 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 805                                           struct btree_path *path) {}
 806 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 807 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 808 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 809
 810 #endif
 811
 812 /* Btree path: fixups after btree updates */
 813
 814 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 815                                         struct btree *b,
 816                                         struct bset_tree *t,
 817                                         struct bkey_packed *k)
 818 {
 819         struct btree_node_iter_set *set;
 820
 821         btree_node_iter_for_each(iter, set)
 822                 if (set->end == t->end_offset) {
 823                         set->k = __btree_node_key_to_offset(b, k);
 824                         bch2_btree_node_iter_sort(iter, b);
 825                         return;
 826                 }
 827
 828         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 829 }
 830
 831 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 832                                                struct btree *b,
 833                                                struct bkey_packed *where)
 834 {
 835         struct btree_path_level *l = &path->l[b->c.level];
 836
 837         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 838                 return;
 839
 840         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 841                 bch2_btree_node_iter_advance(&l->iter, l->b);
 842 }
 843
 844 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 845                                       struct btree *b,
 846                                       struct bkey_packed *where)
 847 {
 848         struct btree_path *path;
 849
 850         trans_for_each_path_with_node(trans, b, path) {
 851                 __bch2_btree_path_fix_key_modified(path, b, where);
 852                 bch2_btree_path_verify_level(trans, path, b->c.level);
 853         }
 854 }
 855
 856 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 857                                        struct btree *b,
 858                                        struct btree_node_iter *node_iter,
 859                                        struct bset_tree *t,
 860                                        struct bkey_packed *where,
 861                                        unsigned clobber_u64s,
 862                                        unsigned new_u64s)
 863 {
 864         const struct bkey_packed *end = btree_bkey_last(b, t);
 865         struct btree_node_iter_set *set;
 866         unsigned offset = __btree_node_key_to_offset(b, where);
 867         int shift = new_u64s - clobber_u64s;
 868         unsigned old_end = t->end_offset - shift;
 869         unsigned orig_iter_pos = node_iter->data[0].k;
 870         bool iter_current_key_modified =
 871                 orig_iter_pos >= offset &&
 872                 orig_iter_pos <= offset + clobber_u64s;
 873
 874         btree_node_iter_for_each(node_iter, set)
 875                 if (set->end == old_end)
 876                         goto found;
 877
 878         /* didn't find the bset in the iterator - might have to readd it: */
 879         if (new_u64s &&
 880             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 881                 bch2_btree_node_iter_push(node_iter, b, where, end);
 882                 goto fixup_done;
 883         } else {
 884                 /* Iterator is after key that changed */
 885                 return;
 886         }
 887 found:
 888         set->end = t->end_offset;
 889
 890         /* Iterator hasn't gotten to the key that changed yet: */
 891         if (set->k < offset)
 892                 return;
 893
 894         if (new_u64s &&
 895             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 896                 set->k = offset;
 897         } else if (set->k < offset + clobber_u64s) {
 898                 set->k = offset + new_u64s;
 899                 if (set->k == set->end)
 900                         bch2_btree_node_iter_set_drop(node_iter, set);
 901         } else {
 902                 /* Iterator is after key that changed */
 903                 set->k = (int) set->k + shift;
 904                 return;
 905         }
 906
 907         bch2_btree_node_iter_sort(node_iter, b);
 908 fixup_done:
 909         if (node_iter->data[0].k != orig_iter_pos)
 910                 iter_current_key_modified = true;
 911
 912         /*
 913          * When a new key is added, and the node iterator now points to that
 914          * key, the iterator might have skipped past deleted keys that should
 915          * come after the key the iterator now points to. We have to rewind to
 916          * before those deleted keys - otherwise
 917          * bch2_btree_node_iter_prev_all() breaks:
 918          */
 919         if (!bch2_btree_node_iter_end(node_iter) &&
 920             iter_current_key_modified &&
 921             b->c.level) {
 922                 struct bset_tree *t;
 923                 struct bkey_packed *k, *k2, *p;
 924
 925                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 926
 927                 for_each_bset(b, t) {
 928                         bool set_pos = false;
 929
 930                         if (node_iter->data[0].end == t->end_offset)
 931                                 continue;
 932
 933                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 934
 935                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 936                                bkey_iter_cmp(b, k, p) < 0) {
 937                                 k2 = p;
 938                                 set_pos = true;
 939                         }
 940
 941                         if (set_pos)
 942                                 btree_node_iter_set_set_pos(node_iter,
 943                                                             b, t, k2);
 944                 }
 945         }
 946 }
 947
 948 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 949                               struct btree_path *path,
 950                               struct btree *b,
 951                               struct btree_node_iter *node_iter,
 952                               struct bkey_packed *where,
 953                               unsigned clobber_u64s,
 954                               unsigned new_u64s)
 955 {
 956         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 957         struct btree_path *linked;
 958
 959         if (node_iter != &path->l[b->c.level].iter) {
 960                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 961                                            where, clobber_u64s, new_u64s);
 962
 963                 if (bch2_debug_check_iterators)
 964                         bch2_btree_node_iter_verify(node_iter, b);
 965         }
 966
 967         trans_for_each_path_with_node(trans, b, linked) {
 968                 __bch2_btree_node_iter_fix(linked, b,
 969                                            &linked->l[b->c.level].iter, t,
 970                                            where, clobber_u64s, new_u64s);
 971                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 972         }
 973 }
 974
 975 /* Btree path level: pointer to a particular btree node and node iter */
 976
 977 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
 978                                                   struct btree_path_level *l,
 979                                                   struct bkey *u,
 980                                                   struct bkey_packed *k)
 981 {
 982         struct bkey_s_c ret;
 983
 984         if (unlikely(!k)) {
 985                 /*
 986                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 987                  * a hole
 988                  */
 989                 u->type = KEY_TYPE_deleted;
 990                 return bkey_s_c_null;
 991         }
 992
 993         ret = bkey_disassemble(l->b, k, u);
 994
 995         /*
 996          * XXX: bch2_btree_bset_insert_key() generates invalid keys when we
 997          * overwrite extents - it sets k->type = KEY_TYPE_deleted on the key
 998          * being overwritten but doesn't change k->size. But this is ok, because
 999          * those keys are never written out, we just have to avoid a spurious
1000          * assertion here:
1001          */
1002         if (bch2_debug_check_bkeys && !bkey_deleted(ret.k))
1003                 bch2_bkey_debugcheck(c, l->b, ret);
1004
1005         return ret;
1006 }
1007
1008 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
1009                                                         struct btree_path_level *l,
1010                                                         struct bkey *u)
1011 {
1012         return __btree_iter_unpack(c, l, u,
1013                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
1014 }
1015
1016 static inline struct bkey_s_c btree_path_level_peek(struct bch_fs *c,
1017                                                     struct btree_path *path,
1018                                                     struct btree_path_level *l,
1019                                                     struct bkey *u)
1020 {
1021         struct bkey_s_c k = __btree_iter_unpack(c, l, u,
1022                         bch2_btree_node_iter_peek(&l->iter, l->b));
1023
1024         path->pos = k.k ? k.k->p : l->b->key.k.p;
1025         return k;
1026 }
1027
1028 static inline struct bkey_s_c btree_path_level_prev(struct bch_fs *c,
1029                                                     struct btree_path *path,
1030                                                     struct btree_path_level *l,
1031                                                     struct bkey *u)
1032 {
1033         struct bkey_s_c k = __btree_iter_unpack(c, l, u,
1034                         bch2_btree_node_iter_prev(&l->iter, l->b));
1035
1036         path->pos = k.k ? k.k->p : l->b->data->min_key;
1037         return k;
1038 }
1039
1040 static inline bool btree_path_advance_to_pos(struct btree_path *path,
1041                                              struct btree_path_level *l,
1042                                              int max_advance)
1043 {
1044         struct bkey_packed *k;
1045         int nr_advanced = 0;
1046
1047         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
1048                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
1049                 if (max_advance > 0 && nr_advanced >= max_advance)
1050                         return false;
1051
1052                 bch2_btree_node_iter_advance(&l->iter, l->b);
1053                 nr_advanced++;
1054         }
1055
1056         return true;
1057 }
1058
1059 /*
1060  * Verify that iterator for parent node points to child node:
1061  */
1062 static void btree_path_verify_new_node(struct btree_trans *trans,
1063                                        struct btree_path *path, struct btree *b)
1064 {
1065         struct btree_path_level *l;
1066         unsigned plevel;
1067         bool parent_locked;
1068         struct bkey_packed *k;
1069
1070         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
1071                 return;
1072
1073         plevel = b->c.level + 1;
1074         if (!btree_path_node(path, plevel))
1075                 return;
1076
1077         parent_locked = btree_node_locked(path, plevel);
1078
1079         if (!bch2_btree_node_relock(trans, path, plevel))
1080                 return;
1081
1082         l = &path->l[plevel];
1083         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1084         if (!k ||
1085             bkey_deleted(k) ||
1086             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
1087                 char buf1[100];
1088                 char buf2[100];
1089                 char buf3[100];
1090                 char buf4[100];
1091                 struct bkey uk = bkey_unpack_key(b, k);
1092
1093                 bch2_dump_btree_node(trans->c, l->b);
1094                 bch2_bpos_to_text(&PBUF(buf1), path->pos);
1095                 bch2_bkey_to_text(&PBUF(buf2), &uk);
1096                 bch2_bpos_to_text(&PBUF(buf3), b->data->min_key);
1097                 bch2_bpos_to_text(&PBUF(buf3), b->data->max_key);
1098                 panic("parent iter doesn't point to new node:\n"
1099                       "iter pos %s %s\n"
1100                       "iter key %s\n"
1101                       "new node %s-%s\n",
1102                       bch2_btree_ids[path->btree_id], buf1,
1103                       buf2, buf3, buf4);
1104         }
1105
1106         if (!parent_locked)
1107                 btree_node_unlock(path, plevel);
1108 }
1109
1110 static inline void __btree_path_level_init(struct btree_path *path,
1111                                            unsigned level)
1112 {
1113         struct btree_path_level *l = &path->l[level];
1114
1115         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1116
1117         /*
1118          * Iterators to interior nodes should always be pointed at the first non
1119          * whiteout:
1120          */
1121         if (level)
1122                 bch2_btree_node_iter_peek(&l->iter, l->b);
1123 }
1124
1125 static inline void btree_path_level_init(struct btree_trans *trans,
1126                                          struct btree_path *path,
1127                                          struct btree *b)
1128 {
1129         BUG_ON(path->cached);
1130
1131         btree_path_verify_new_node(trans, path, b);
1132
1133         EBUG_ON(!btree_path_pos_in_node(path, b));
1134         EBUG_ON(b->c.lock.state.seq & 1);
1135
1136         path->l[b->c.level].lock_seq = b->c.lock.state.seq;
1137         path->l[b->c.level].b = b;
1138         __btree_path_level_init(path, b->c.level);
1139 }
1140
1141 /* Btree path: fixups after btree node updates: */
1142
1143 /*
1144  * A btree node is being replaced - update the iterator to point to the new
1145  * node:
1146  */
1147 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
1148 {
1149         struct btree_path *path;
1150
1151         trans_for_each_path(trans, path)
1152                 if (!path->cached &&
1153                     btree_path_pos_in_node(path, b)) {
1154                         enum btree_node_locked_type t =
1155                                 btree_lock_want(path, b->c.level);
1156
1157                         if (path->nodes_locked &&
1158                             t != BTREE_NODE_UNLOCKED) {
1159                                 btree_node_unlock(path, b->c.level);
1160                                 six_lock_increment(&b->c.lock, t);
1161                                 mark_btree_node_locked(path, b->c.level, t);
1162                         }
1163
1164                         btree_path_level_init(trans, path, b);
1165                 }
1166 }
1167
1168 /*
1169  * A btree node has been modified in such a way as to invalidate iterators - fix
1170  * them:
1171  */
1172 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
1173 {
1174         struct btree_path *path;
1175
1176         trans_for_each_path_with_node(trans, b, path)
1177                 __btree_path_level_init(path, b->c.level);
1178 }
1179
1180 /* Btree path: traverse, set_pos: */
1181
1182 static int lock_root_check_fn(struct six_lock *lock, void *p)
1183 {
1184         struct btree *b = container_of(lock, struct btree, c.lock);
1185         struct btree **rootp = p;
1186
1187         return b == *rootp ? 0 : -1;
1188 }
1189
1190 static inline int btree_path_lock_root(struct btree_trans *trans,
1191                                        struct btree_path *path,
1192                                        unsigned depth_want,
1193                                        unsigned long trace_ip)
1194 {
1195         struct bch_fs *c = trans->c;
1196         struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
1197         enum six_lock_type lock_type;
1198         unsigned i;
1199
1200         EBUG_ON(path->nodes_locked);
1201
1202         while (1) {
1203                 b = READ_ONCE(*rootp);
1204                 path->level = READ_ONCE(b->c.level);
1205
1206                 if (unlikely(path->level < depth_want)) {
1207                         /*
1208                          * the root is at a lower depth than the depth we want:
1209                          * got to the end of the btree, or we're walking nodes
1210                          * greater than some depth and there are no nodes >=
1211                          * that depth
1212                          */
1213                         path->level = depth_want;
1214                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
1215                                 path->l[i].b = NULL;
1216                         return 1;
1217                 }
1218
1219                 lock_type = __btree_lock_want(path, path->level);
1220                 if (unlikely(!btree_node_lock(trans, path, b, SPOS_MAX,
1221                                               path->level, lock_type,
1222                                               lock_root_check_fn, rootp,
1223                                               trace_ip))) {
1224                         if (trans->restarted)
1225                                 return -EINTR;
1226                         continue;
1227                 }
1228
1229                 if (likely(b == READ_ONCE(*rootp) &&
1230                            b->c.level == path->level &&
1231                            !race_fault())) {
1232                         for (i = 0; i < path->level; i++)
1233                                 path->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
1234                         path->l[path->level].b = b;
1235                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
1236                                 path->l[i].b = NULL;
1237
1238                         mark_btree_node_locked(path, path->level, lock_type);
1239                         btree_path_level_init(trans, path, b);
1240                         return 0;
1241                 }
1242
1243                 six_unlock_type(&b->c.lock, lock_type);
1244         }
1245 }
1246
1247 noinline
1248 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
1249 {
1250         struct bch_fs *c = trans->c;
1251         struct btree_path_level *l = path_l(path);
1252         struct btree_node_iter node_iter = l->iter;
1253         struct bkey_packed *k;
1254         struct bkey_buf tmp;
1255         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1256                 ? (path->level > 1 ? 0 :  2)
1257                 : (path->level > 1 ? 1 : 16);
1258         bool was_locked = btree_node_locked(path, path->level);
1259         int ret = 0;
1260
1261         bch2_bkey_buf_init(&tmp);
1262
1263         while (nr && !ret) {
1264                 if (!bch2_btree_node_relock(trans, path, path->level))
1265                         break;
1266
1267                 bch2_btree_node_iter_advance(&node_iter, l->b);
1268                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
1269                 if (!k)
1270                         break;
1271
1272                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
1273                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
1274                                                path->level - 1);
1275         }
1276
1277         if (!was_locked)
1278                 btree_node_unlock(path, path->level);
1279
1280         bch2_bkey_buf_exit(&tmp, c);
1281         return ret;
1282 }
1283
1284 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
1285                                             struct btree_path *path,
1286                                             unsigned plevel, struct btree *b)
1287 {
1288         struct btree_path_level *l = &path->l[plevel];
1289         bool locked = btree_node_locked(path, plevel);
1290         struct bkey_packed *k;
1291         struct bch_btree_ptr_v2 *bp;
1292
1293         if (!bch2_btree_node_relock(trans, path, plevel))
1294                 return;
1295
1296         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1297         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
1298
1299         bp = (void *) bkeyp_val(&l->b->format, k);
1300         bp->mem_ptr = (unsigned long)b;
1301
1302         if (!locked)
1303                 btree_node_unlock(path, plevel);
1304 }
1305
1306 static __always_inline int btree_path_down(struct btree_trans *trans,
1307                                            struct btree_path *path,
1308                                            unsigned flags,
1309                                            unsigned long trace_ip)
1310 {
1311         struct bch_fs *c = trans->c;
1312         struct btree_path_level *l = path_l(path);
1313         struct btree *b;
1314         unsigned level = path->level - 1;
1315         enum six_lock_type lock_type = __btree_lock_want(path, level);
1316         struct bkey_buf tmp;
1317         int ret;
1318
1319         EBUG_ON(!btree_node_locked(path, path->level));
1320
1321         bch2_bkey_buf_init(&tmp);
1322         bch2_bkey_buf_unpack(&tmp, c, l->b,
1323                          bch2_btree_node_iter_peek(&l->iter, l->b));
1324
1325         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
1326         ret = PTR_ERR_OR_ZERO(b);
1327         if (unlikely(ret))
1328                 goto err;
1329
1330         mark_btree_node_locked(path, level, lock_type);
1331         btree_path_level_init(trans, path, b);
1332
1333         if (tmp.k->k.type == KEY_TYPE_btree_ptr_v2 &&
1334             unlikely(b != btree_node_mem_ptr(tmp.k)))
1335                 btree_node_mem_ptr_set(trans, path, level + 1, b);
1336
1337         if (flags & BTREE_ITER_PREFETCH)
1338                 ret = btree_path_prefetch(trans, path);
1339
1340         if (btree_node_read_locked(path, level + 1))
1341                 btree_node_unlock(path, level + 1);
1342         path->level = level;
1343
1344         bch2_btree_path_verify_locks(path);
1345 err:
1346         bch2_bkey_buf_exit(&tmp, c);
1347         return ret;
1348 }
1349
1350 static int btree_path_traverse_one(struct btree_trans *, struct btree_path *,
1351                                    unsigned, unsigned long);
1352
1353 static int __btree_path_traverse_all(struct btree_trans *trans, int ret,
1354                                      unsigned long trace_ip)
1355 {
1356         struct bch_fs *c = trans->c;
1357         struct btree_path *path;
1358         int i;
1359
1360         if (trans->in_traverse_all)
1361                 return -EINTR;
1362
1363         trans->in_traverse_all = true;
1364 retry_all:
1365         trans->restarted = false;
1366
1367         trans_for_each_path(trans, path)
1368                 path->should_be_locked = false;
1369
1370         btree_trans_verify_sorted(trans);
1371
1372         for (i = trans->nr_sorted - 2; i >= 0; --i) {
1373                 struct btree_path *path1 = trans->paths + trans->sorted[i];
1374                 struct btree_path *path2 = trans->paths + trans->sorted[i + 1];
1375
1376                 if (path1->btree_id == path2->btree_id &&
1377                     path1->locks_want < path2->locks_want)
1378                         __bch2_btree_path_upgrade(trans, path1, path2->locks_want);
1379                 else if (!path1->locks_want && path2->locks_want)
1380                         __bch2_btree_path_upgrade(trans, path1, 1);
1381         }
1382
1383         bch2_trans_unlock(trans);
1384         cond_resched();
1385
1386         if (unlikely(ret == -ENOMEM)) {
1387                 struct closure cl;
1388
1389                 closure_init_stack(&cl);
1390
1391                 do {
1392                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1393                         closure_sync(&cl);
1394                 } while (ret);
1395         }
1396
1397         if (unlikely(ret == -EIO))
1398                 goto out;
1399
1400         BUG_ON(ret && ret != -EINTR);
1401
1402         /* Now, redo traversals in correct order: */
1403         i = 0;
1404         while (i < trans->nr_sorted) {
1405                 path = trans->paths + trans->sorted[i];
1406
1407                 EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
1408
1409                 ret = btree_path_traverse_one(trans, path, 0, _THIS_IP_);
1410                 if (ret)
1411                         goto retry_all;
1412
1413                 EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
1414
1415                 if (path->nodes_locked ||
1416                     !btree_path_node(path, path->level))
1417                         i++;
1418         }
1419
1420         /*
1421          * BTREE_ITER_NEED_RELOCK is ok here - if we called bch2_trans_unlock()
1422          * and relock(), relock() won't relock since path->should_be_locked
1423          * isn't set yet, which is all fine
1424          */
1425         trans_for_each_path(trans, path)
1426                 BUG_ON(path->uptodate >= BTREE_ITER_NEED_TRAVERSE);
1427 out:
1428         bch2_btree_cache_cannibalize_unlock(c);
1429
1430         trans->in_traverse_all = false;
1431
1432         trace_trans_traverse_all(trans->ip, trace_ip);
1433         return ret;
1434 }
1435
1436 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
1437 {
1438         return __btree_path_traverse_all(trans, 0, _RET_IP_);
1439 }
1440
1441 static inline bool btree_path_good_node(struct btree_trans *trans,
1442                                         struct btree_path *path,
1443                                         unsigned l, int check_pos)
1444 {
1445         if (!is_btree_node(path, l) ||
1446             !bch2_btree_node_relock(trans, path, l))
1447                 return false;
1448
1449         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1450                 return false;
1451         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1452                 return false;
1453         return true;
1454 }
1455
1456 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1457                                                      struct btree_path *path,
1458                                                      int check_pos)
1459 {
1460         unsigned i, l = path->level;
1461
1462         while (btree_path_node(path, l) &&
1463                !btree_path_good_node(trans, path, l, check_pos)) {
1464                 btree_node_unlock(path, l);
1465                 path->l[l].b = BTREE_ITER_NO_NODE_UP;
1466                 l++;
1467         }
1468
1469         /* If we need intent locks, take them too: */
1470         for (i = l + 1;
1471              i < path->locks_want && btree_path_node(path, i);
1472              i++)
1473                 if (!bch2_btree_node_relock(trans, path, i))
1474                         while (l <= i) {
1475                                 btree_node_unlock(path, l);
1476                                 path->l[l].b = BTREE_ITER_NO_NODE_UP;
1477                                 l++;
1478                         }
1479
1480         return l;
1481 }
1482
1483 /*
1484  * This is the main state machine for walking down the btree - walks down to a
1485  * specified depth
1486  *
1487  * Returns 0 on success, -EIO on error (error reading in a btree node).
1488  *
1489  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1490  * stashed in the iterator and returned from bch2_trans_exit().
1491  */
1492 static int btree_path_traverse_one(struct btree_trans *trans,
1493                                    struct btree_path *path,
1494                                    unsigned flags,
1495                                    unsigned long trace_ip)
1496 {
1497         unsigned depth_want = path->level;
1498         int ret = 0;
1499
1500         if (unlikely(trans->restarted)) {
1501                 ret = -EINTR;
1502                 goto out;
1503         }
1504
1505         /*
1506          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1507          * and re-traverse the path without a transaction restart:
1508          */
1509         if (path->should_be_locked) {
1510                 ret = bch2_btree_path_relock(trans, path, trace_ip) ? 0 : -EINTR;
1511                 goto out;
1512         }
1513
1514         if (path->cached) {
1515                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1516                 goto out;
1517         }
1518
1519         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1520                 goto out;
1521
1522         path->level = btree_path_up_until_good_node(trans, path, 0);
1523
1524         /*
1525          * Note: path->nodes[path->level] may be temporarily NULL here - that
1526          * would indicate to other code that we got to the end of the btree,
1527          * here it indicates that relocking the root failed - it's critical that
1528          * btree_path_lock_root() comes next and that it can't fail
1529          */
1530         while (path->level > depth_want) {
1531                 ret = btree_path_node(path, path->level)
1532                         ? btree_path_down(trans, path, flags, trace_ip)
1533                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1534                 if (unlikely(ret)) {
1535                         if (ret == 1) {
1536                                 /*
1537                                  * No nodes at this level - got to the end of
1538                                  * the btree:
1539                                  */
1540                                 ret = 0;
1541                                 goto out;
1542                         }
1543
1544                         __bch2_btree_path_unlock(path);
1545                         path->level = depth_want;
1546
1547                         if (ret == -EIO)
1548                                 path->l[path->level].b =
1549                                         BTREE_ITER_NO_NODE_ERROR;
1550                         else
1551                                 path->l[path->level].b =
1552                                         BTREE_ITER_NO_NODE_DOWN;
1553                         goto out;
1554                 }
1555         }
1556
1557         path->uptodate = BTREE_ITER_UPTODATE;
1558 out:
1559         BUG_ON((ret == -EINTR) != !!trans->restarted);
1560         bch2_btree_path_verify(trans, path);
1561         return ret;
1562 }
1563
1564 static int __btree_path_traverse_all(struct btree_trans *, int, unsigned long);
1565
1566 int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
1567                                           struct btree_path *path, unsigned flags)
1568 {
1569         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
1570                 return 0;
1571
1572         return  bch2_trans_cond_resched(trans) ?:
1573                 btree_path_traverse_one(trans, path, flags, _RET_IP_);
1574 }
1575
1576 static void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1577                             struct btree_path *src)
1578 {
1579         unsigned i;
1580
1581         memcpy(&dst->pos, &src->pos,
1582                sizeof(struct btree_path) - offsetof(struct btree_path, pos));
1583
1584         for (i = 0; i < BTREE_MAX_DEPTH; i++)
1585                 if (btree_node_locked(dst, i))
1586                         six_lock_increment(&dst->l[i].b->c.lock,
1587                                            __btree_lock_want(dst, i));
1588
1589         btree_path_check_sort(trans, dst, 0);
1590 }
1591
1592 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1593                                            bool intent)
1594 {
1595         struct btree_path *new = btree_path_alloc(trans, src);
1596
1597         btree_path_copy(trans, new, src);
1598         __btree_path_get(new, intent);
1599         return new;
1600 }
1601
1602 inline struct btree_path * __must_check
1603 bch2_btree_path_make_mut(struct btree_trans *trans,
1604                          struct btree_path *path, bool intent)
1605 {
1606         if (path->ref > 1 || path->preserve) {
1607                 __btree_path_put(path, intent);
1608                 path = btree_path_clone(trans, path, intent);
1609                 path->preserve = false;
1610 #ifdef CONFIG_BCACHEFS_DEBUG
1611                 path->ip_allocated = _RET_IP_;
1612 #endif
1613                 btree_trans_verify_sorted(trans);
1614         }
1615
1616         return path;
1617 }
1618
1619 static struct btree_path * __must_check
1620 btree_path_set_pos(struct btree_trans *trans,
1621                    struct btree_path *path, struct bpos new_pos,
1622                    bool intent)
1623 {
1624         int cmp = bpos_cmp(new_pos, path->pos);
1625         unsigned l = path->level;
1626
1627         EBUG_ON(trans->restarted);
1628         EBUG_ON(!path->ref);
1629
1630         if (!cmp)
1631                 return path;
1632
1633         path = bch2_btree_path_make_mut(trans, path, intent);
1634
1635         path->pos               = new_pos;
1636         path->should_be_locked  = false;
1637
1638         btree_path_check_sort(trans, path, cmp);
1639
1640         if (unlikely(path->cached)) {
1641                 btree_node_unlock(path, 0);
1642                 path->l[0].b = BTREE_ITER_NO_NODE_CACHED;
1643                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1644                 goto out;
1645         }
1646
1647         l = btree_path_up_until_good_node(trans, path, cmp);
1648
1649         if (btree_path_node(path, l)) {
1650                 /*
1651                  * We might have to skip over many keys, or just a few: try
1652                  * advancing the node iterator, and if we have to skip over too
1653                  * many keys just reinit it (or if we're rewinding, since that
1654                  * is expensive).
1655                  */
1656                 if (cmp < 0 ||
1657                     !btree_path_advance_to_pos(path, &path->l[l], 8))
1658                         __btree_path_level_init(path, l);
1659         }
1660
1661         if (l != path->level) {
1662                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1663                 __bch2_btree_path_unlock(path);
1664         }
1665 out:
1666         bch2_btree_path_verify(trans, path);
1667         return path;
1668 }
1669
1670 /* Btree path: main interface: */
1671
1672 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1673 {
1674         struct btree_path *next;
1675
1676         next = prev_btree_path(trans, path);
1677         if (next && !btree_path_cmp(next, path))
1678                 return next;
1679
1680         next = next_btree_path(trans, path);
1681         if (next && !btree_path_cmp(next, path))
1682                 return next;
1683
1684         return NULL;
1685 }
1686
1687 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1688 {
1689         struct btree_path *next;
1690
1691         next = prev_btree_path(trans, path);
1692         if (next && next->level == path->level && path_l(next)->b == path_l(path)->b)
1693                 return next;
1694
1695         next = next_btree_path(trans, path);
1696         if (next && next->level == path->level && path_l(next)->b == path_l(path)->b)
1697                 return next;
1698
1699         return NULL;
1700 }
1701
1702 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1703 {
1704         __bch2_btree_path_unlock(path);
1705         btree_path_list_remove(trans, path);
1706         trans->paths_allocated &= ~(1ULL << path->idx);
1707 }
1708
1709 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1710 {
1711         struct btree_path *dup;
1712
1713         EBUG_ON(trans->paths + path->idx != path);
1714         EBUG_ON(!path->ref);
1715
1716         if (!__btree_path_put(path, intent))
1717                 return;
1718
1719         /*
1720          * Perhaps instead we should check for duplicate paths in traverse_all:
1721          */
1722         if (path->preserve &&
1723             (dup = have_path_at_pos(trans, path))) {
1724                 dup->preserve = true;
1725                 path->preserve = false;
1726                 goto free;
1727         }
1728
1729         if (!path->preserve &&
1730             (dup = have_node_at_pos(trans, path)))
1731                 goto free;
1732         return;
1733 free:
1734         if (path->should_be_locked &&
1735             !btree_node_locked(dup, path->level))
1736                 return;
1737
1738         dup->should_be_locked |= path->should_be_locked;
1739         __bch2_path_free(trans, path);
1740 }
1741
1742 noinline __cold
1743 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1744 {
1745         struct btree_path *path;
1746         struct btree_insert_entry *i;
1747         unsigned idx;
1748         char buf1[300], buf2[300];
1749
1750         btree_trans_verify_sorted(trans);
1751
1752         trans_for_each_path_inorder(trans, path, idx)
1753                 printk(KERN_ERR "path: idx %u ref %u:%u%s%s btree %s pos %s locks %u %pS\n",
1754                        path->idx, path->ref, path->intent_ref,
1755                        path->should_be_locked ? " S" : "",
1756                        path->preserve ? " P" : "",
1757                        bch2_btree_ids[path->btree_id],
1758                        (bch2_bpos_to_text(&PBUF(buf1), path->pos), buf1),
1759                        path->nodes_locked,
1760 #ifdef CONFIG_BCACHEFS_DEBUG
1761                        (void *) path->ip_allocated
1762 #else
1763                        NULL
1764 #endif
1765                        );
1766
1767         trans_for_each_update(trans, i) {
1768                 struct bkey u;
1769                 struct bkey_s_c old = bch2_btree_path_peek_slot(i->path, &u);
1770
1771                 printk(KERN_ERR "update: btree %s %pS\n  old %s\n  new %s",
1772                        bch2_btree_ids[i->btree_id],
1773                        (void *) i->ip_allocated,
1774                        (bch2_bkey_val_to_text(&PBUF(buf1), trans->c, old), buf1),
1775                        (bch2_bkey_val_to_text(&PBUF(buf2), trans->c, bkey_i_to_s_c(i->k)), buf2));
1776         }
1777 }
1778
1779 static struct btree_path *btree_path_alloc(struct btree_trans *trans,
1780                                            struct btree_path *pos)
1781 {
1782         struct btree_path *path;
1783         unsigned idx;
1784
1785         if (unlikely(trans->paths_allocated ==
1786                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1)))) {
1787                 bch2_dump_trans_paths_updates(trans);
1788                 panic("trans path oveflow\n");
1789         }
1790
1791         idx = __ffs64(~trans->paths_allocated);
1792         trans->paths_allocated |= 1ULL << idx;
1793
1794         path = &trans->paths[idx];
1795
1796         path->idx               = idx;
1797         path->ref               = 0;
1798         path->intent_ref        = 0;
1799         path->nodes_locked      = 0;
1800         path->nodes_intent_locked = 0;
1801
1802         btree_path_list_add(trans, pos, path);
1803         return path;
1804 }
1805
1806 struct btree_path *bch2_path_get(struct btree_trans *trans, bool cached,
1807                                  enum btree_id btree_id, struct bpos pos,
1808                                  unsigned locks_want, unsigned level,
1809                                  bool intent)
1810 {
1811         struct btree_path *path, *path_pos = NULL;
1812         int i;
1813
1814         BUG_ON(trans->restarted);
1815
1816         trans_for_each_path_inorder(trans, path, i) {
1817                 if (__btree_path_cmp(path,
1818                                      btree_id,
1819                                      cached,
1820                                      pos,
1821                                      level) > 0)
1822                         break;
1823
1824                 path_pos = path;
1825         }
1826
1827         if (path_pos &&
1828             path_pos->cached    == cached &&
1829             path_pos->btree_id  == btree_id &&
1830             path_pos->level     == level) {
1831                 __btree_path_get(path_pos, intent);
1832                 path = btree_path_set_pos(trans, path_pos, pos, intent);
1833                 path->preserve = true;
1834         } else {
1835                 path = btree_path_alloc(trans, path_pos);
1836                 path_pos = NULL;
1837
1838                 __btree_path_get(path, intent);
1839                 path->pos                       = pos;
1840                 path->btree_id                  = btree_id;
1841                 path->cached                    = cached;
1842                 path->preserve                  = true;
1843                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1844                 path->should_be_locked          = false;
1845                 path->level                     = level;
1846                 path->locks_want                = locks_want;
1847                 path->nodes_locked              = 0;
1848                 path->nodes_intent_locked       = 0;
1849                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1850                         path->l[i].b            = BTREE_ITER_NO_NODE_INIT;
1851 #ifdef CONFIG_BCACHEFS_DEBUG
1852                 path->ip_allocated              = _RET_IP_;
1853 #endif
1854                 btree_trans_verify_sorted(trans);
1855         }
1856
1857         if (path->intent_ref)
1858                 locks_want = max(locks_want, level + 1);
1859
1860         /*
1861          * If the path has locks_want greater than requested, we don't downgrade
1862          * it here - on transaction restart because btree node split needs to
1863          * upgrade locks, we might be putting/getting the iterator again.
1864          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1865          * a successful transaction commit.
1866          */
1867
1868         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1869         if (locks_want > path->locks_want) {
1870                 path->locks_want = locks_want;
1871                 btree_path_get_locks(trans, path, true, _THIS_IP_);
1872         }
1873
1874         return path;
1875 }
1876
1877 inline struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1878 {
1879
1880         struct bkey_s_c k;
1881
1882         BUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1883
1884         if (!path->cached) {
1885                 struct btree_path_level *l = path_l(path);
1886                 struct bkey_packed *_k =
1887                         bch2_btree_node_iter_peek_all(&l->iter, l->b);
1888
1889                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1890
1891                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_cmp(k.k->p, path->pos) == 0);
1892
1893                 if (!k.k || bpos_cmp(path->pos, k.k->p))
1894                         goto hole;
1895         } else {
1896                 struct bkey_cached *ck = (void *) path->l[0].b;
1897
1898                 EBUG_ON(path->btree_id != ck->key.btree_id ||
1899                         bkey_cmp(path->pos, ck->key.pos));
1900
1901                 /* BTREE_ITER_CACHED_NOFILL? */
1902                 if (unlikely(!ck->valid))
1903                         goto hole;
1904
1905                 k = bkey_i_to_s_c(ck->k);
1906         }
1907
1908         return k;
1909 hole:
1910         bkey_init(u);
1911         u->p = path->pos;
1912         return (struct bkey_s_c) { u, NULL };
1913 }
1914
1915 /* Btree iterators: */
1916
1917 int __must_check
1918 __bch2_btree_iter_traverse(struct btree_iter *iter)
1919 {
1920         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1921 }
1922
1923 int __must_check
1924 bch2_btree_iter_traverse(struct btree_iter *iter)
1925 {
1926         int ret;
1927
1928         iter->path = btree_path_set_pos(iter->trans, iter->path,
1929                                         btree_iter_search_key(iter),
1930                                         iter->flags & BTREE_ITER_INTENT);
1931
1932         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1933         if (ret)
1934                 return ret;
1935
1936         iter->path->should_be_locked = true;
1937         return 0;
1938 }
1939
1940 /* Iterate across nodes (leaf and interior nodes) */
1941
1942 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1943 {
1944         struct btree_trans *trans = iter->trans;
1945         struct btree *b = NULL;
1946         int ret;
1947
1948         EBUG_ON(iter->path->cached);
1949         bch2_btree_iter_verify(iter);
1950
1951         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1952         if (ret)
1953                 goto err;
1954
1955         b = btree_path_node(iter->path, iter->path->level);
1956         if (!b)
1957                 goto out;
1958
1959         BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0);
1960
1961         bkey_init(&iter->k);
1962         iter->k.p = iter->pos = b->key.k.p;
1963
1964         iter->path = btree_path_set_pos(trans, iter->path, b->key.k.p,
1965                                         iter->flags & BTREE_ITER_INTENT);
1966         iter->path->should_be_locked = true;
1967         BUG_ON(iter->path->uptodate);
1968 out:
1969         bch2_btree_iter_verify_entry_exit(iter);
1970         bch2_btree_iter_verify(iter);
1971
1972         return b;
1973 err:
1974         b = ERR_PTR(ret);
1975         goto out;
1976 }
1977
1978 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1979 {
1980         struct btree_trans *trans = iter->trans;
1981         struct btree_path *path = iter->path;
1982         struct btree *b = NULL;
1983         unsigned l;
1984         int ret;
1985
1986         BUG_ON(trans->restarted);
1987         EBUG_ON(iter->path->cached);
1988         bch2_btree_iter_verify(iter);
1989
1990         /* already at end? */
1991         if (!btree_path_node(path, path->level))
1992                 return NULL;
1993
1994         /* got to end? */
1995         if (!btree_path_node(path, path->level + 1)) {
1996                 btree_node_unlock(path, path->level);
1997                 path->l[path->level].b = BTREE_ITER_NO_NODE_UP;
1998                 path->level++;
1999                 return NULL;
2000         }
2001
2002         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
2003                 __bch2_btree_path_unlock(path);
2004                 path->l[path->level].b = BTREE_ITER_NO_NODE_GET_LOCKS;
2005                 path->l[path->level + 1].b = BTREE_ITER_NO_NODE_GET_LOCKS;
2006                 btree_trans_restart(trans);
2007                 ret = -EINTR;
2008                 goto err;
2009         }
2010
2011         b = btree_path_node(path, path->level + 1);
2012
2013         if (!bpos_cmp(iter->pos, b->key.k.p)) {
2014                 btree_node_unlock(path, path->level);
2015                 path->l[path->level].b = BTREE_ITER_NO_NODE_UP;
2016                 path->level++;
2017         } else {
2018                 /*
2019                  * Haven't gotten to the end of the parent node: go back down to
2020                  * the next child node
2021                  */
2022                 path = iter->path =
2023                         btree_path_set_pos(trans, path, bpos_successor(iter->pos),
2024                                            iter->flags & BTREE_ITER_INTENT);
2025
2026                 path->level = iter->min_depth;
2027
2028                 for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
2029                         if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
2030                                 btree_node_unlock(path, l);
2031
2032                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
2033                 bch2_btree_iter_verify(iter);
2034
2035                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
2036                 if (ret)
2037                         goto err;
2038
2039                 b = path->l[path->level].b;
2040         }
2041
2042         bkey_init(&iter->k);
2043         iter->k.p = iter->pos = b->key.k.p;
2044
2045         iter->path = btree_path_set_pos(trans, iter->path, b->key.k.p,
2046                                         iter->flags & BTREE_ITER_INTENT);
2047         iter->path->should_be_locked = true;
2048         BUG_ON(iter->path->uptodate);
2049 out:
2050         bch2_btree_iter_verify_entry_exit(iter);
2051         bch2_btree_iter_verify(iter);
2052
2053         return b;
2054 err:
2055         b = ERR_PTR(ret);
2056         goto out;
2057 }
2058
2059 /* Iterate across keys (in leaf nodes only) */
2060
2061 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
2062 {
2063         struct bpos pos = iter->k.p;
2064         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
2065                     ? bpos_cmp(pos, SPOS_MAX)
2066                     : bkey_cmp(pos, SPOS_MAX)) != 0;
2067
2068         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
2069                 pos = bkey_successor(iter, pos);
2070         bch2_btree_iter_set_pos(iter, pos);
2071         return ret;
2072 }
2073
2074 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
2075 {
2076         struct bpos pos = bkey_start_pos(&iter->k);
2077         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
2078                     ? bpos_cmp(pos, POS_MIN)
2079                     : bkey_cmp(pos, POS_MIN)) != 0;
2080
2081         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
2082                 pos = bkey_predecessor(iter, pos);
2083         bch2_btree_iter_set_pos(iter, pos);
2084         return ret;
2085 }
2086
2087 /**
2088  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2089  * current position
2090  */
2091 struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
2092 {
2093         struct btree_trans *trans = iter->trans;
2094         struct bpos search_key = btree_iter_search_key(iter);
2095         struct bkey_i *next_update;
2096         struct bkey_s_c k;
2097         int ret, cmp;
2098
2099         EBUG_ON(iter->path->cached || iter->path->level);
2100         bch2_btree_iter_verify(iter);
2101         bch2_btree_iter_verify_entry_exit(iter);
2102
2103         while (1) {
2104                 iter->path = btree_path_set_pos(trans, iter->path, search_key,
2105                                    iter->flags & BTREE_ITER_INTENT);
2106
2107                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2108                 if (unlikely(ret)) {
2109                         /* ensure that iter->k is consistent with iter->pos: */
2110                         bch2_btree_iter_set_pos(iter, iter->pos);
2111                         k = bkey_s_c_err(ret);
2112                         goto out;
2113                 }
2114
2115                 next_update = iter->flags & BTREE_ITER_WITH_UPDATES
2116                         ? btree_trans_peek_updates(trans, iter->btree_id, search_key)
2117                         : NULL;
2118                 k = btree_path_level_peek_all(trans->c, &iter->path->l[0], &iter->k);
2119
2120                 /* * In the btree, deleted keys sort before non deleted: */
2121                 if (k.k && bkey_deleted(k.k) &&
2122                     (!next_update ||
2123                      bpos_cmp(k.k->p, next_update->k.p) <= 0)) {
2124                         search_key = k.k->p;
2125                         continue;
2126                 }
2127
2128                 if (next_update &&
2129                     bpos_cmp(next_update->k.p,
2130                              k.k ? k.k->p : iter->path->l[0].b->key.k.p) <= 0) {
2131                         iter->k = next_update->k;
2132                         k = bkey_i_to_s_c(next_update);
2133                 }
2134
2135                 if (likely(k.k)) {
2136                         /*
2137                          * We can never have a key in a leaf node at POS_MAX, so
2138                          * we don't have to check these successor() calls:
2139                          */
2140                         if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2141                             !bch2_snapshot_is_ancestor(trans->c,
2142                                                        iter->snapshot,
2143                                                        k.k->p.snapshot)) {
2144                                 search_key = bpos_successor(k.k->p);
2145                                 continue;
2146                         }
2147
2148                         if (bkey_whiteout(k.k) &&
2149                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2150                                 search_key = bkey_successor(iter, k.k->p);
2151                                 continue;
2152                         }
2153
2154                         break;
2155                 } else if (likely(bpos_cmp(iter->path->l[0].b->key.k.p, SPOS_MAX))) {
2156                         /* Advance to next leaf node: */
2157                         search_key = bpos_successor(iter->path->l[0].b->key.k.p);
2158                 } else {
2159                         /* End of btree: */
2160                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2161                         k = bkey_s_c_null;
2162                         goto out;
2163                 }
2164         }
2165
2166         /*
2167          * iter->pos should be mononotically increasing, and always be equal to
2168          * the key we just returned - except extents can straddle iter->pos:
2169          */
2170         if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2171                 iter->pos = k.k->p;
2172         else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
2173                 iter->pos = bkey_start_pos(k.k);
2174
2175         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2176                 iter->pos.snapshot = iter->snapshot;
2177
2178         cmp = bpos_cmp(k.k->p, iter->path->pos);
2179         if (cmp) {
2180                 iter->path = bch2_btree_path_make_mut(trans, iter->path,
2181                                         iter->flags & BTREE_ITER_INTENT);
2182                 iter->path->pos = k.k->p;
2183                 btree_path_check_sort(trans, iter->path, cmp);
2184         }
2185 out:
2186         iter->path->should_be_locked = true;
2187
2188         bch2_btree_iter_verify_entry_exit(iter);
2189         bch2_btree_iter_verify(iter);
2190         ret = bch2_btree_iter_verify_ret(iter, k);
2191         if (unlikely(ret))
2192                 return bkey_s_c_err(ret);
2193
2194         return k;
2195 }
2196
2197 /**
2198  * bch2_btree_iter_next: returns first key greater than iterator's current
2199  * position
2200  */
2201 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2202 {
2203         if (!bch2_btree_iter_advance(iter))
2204                 return bkey_s_c_null;
2205
2206         return bch2_btree_iter_peek(iter);
2207 }
2208
2209 /**
2210  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2211  * iterator's current position
2212  */
2213 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2214 {
2215         struct btree_trans *trans = iter->trans;
2216         struct bpos search_key = iter->pos;
2217         struct btree_path *saved_path = NULL;
2218         struct bkey_s_c k;
2219         struct bkey saved_k;
2220         const struct bch_val *saved_v;
2221         int ret;
2222
2223         EBUG_ON(iter->path->cached || iter->path->level);
2224         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2225         bch2_btree_iter_verify(iter);
2226         bch2_btree_iter_verify_entry_exit(iter);
2227
2228         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2229                 search_key.snapshot = U32_MAX;
2230
2231         while (1) {
2232                 iter->path = btree_path_set_pos(trans, iter->path, search_key,
2233                                                 iter->flags & BTREE_ITER_INTENT);
2234
2235                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2236                 if (unlikely(ret)) {
2237                         /* ensure that iter->k is consistent with iter->pos: */
2238                         bch2_btree_iter_set_pos(iter, iter->pos);
2239                         k = bkey_s_c_err(ret);
2240                         goto out;
2241                 }
2242
2243                 k = btree_path_level_peek(trans->c, iter->path,
2244                                           &iter->path->l[0], &iter->k);
2245                 if (!k.k ||
2246                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2247                      ? bpos_cmp(bkey_start_pos(k.k), search_key) >= 0
2248                      : bpos_cmp(k.k->p, search_key) > 0))
2249                         k = btree_path_level_prev(trans->c, iter->path,
2250                                                   &iter->path->l[0], &iter->k);
2251
2252                 btree_path_check_sort(trans, iter->path, 0);
2253
2254                 if (likely(k.k)) {
2255                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2256                                 if (k.k->p.snapshot == iter->snapshot)
2257                                         goto got_key;
2258
2259                                 /*
2260                                  * If we have a saved candidate, and we're no
2261                                  * longer at the same _key_ (not pos), return
2262                                  * that candidate
2263                                  */
2264                                 if (saved_path && bkey_cmp(k.k->p, saved_k.p)) {
2265                                         bch2_path_put(trans, iter->path,
2266                                                       iter->flags & BTREE_ITER_INTENT);
2267                                         iter->path = saved_path;
2268                                         saved_path = NULL;
2269                                         iter->k = saved_k;
2270                                         k.v     = saved_v;
2271                                         goto got_key;
2272                                 }
2273
2274                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2275                                                               iter->snapshot,
2276                                                               k.k->p.snapshot)) {
2277                                         if (saved_path)
2278                                                 bch2_path_put(trans, saved_path,
2279                                                       iter->flags & BTREE_ITER_INTENT);
2280                                         saved_path = btree_path_clone(trans, iter->path,
2281                                                                 iter->flags & BTREE_ITER_INTENT);
2282                                         saved_k = *k.k;
2283                                         saved_v = k.v;
2284                                 }
2285
2286                                 search_key = bpos_predecessor(k.k->p);
2287                                 continue;
2288                         }
2289 got_key:
2290                         if (bkey_whiteout(k.k) &&
2291                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2292                                 search_key = bkey_predecessor(iter, k.k->p);
2293                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2294                                         search_key.snapshot = U32_MAX;
2295                                 continue;
2296                         }
2297
2298                         break;
2299                 } else if (likely(bpos_cmp(iter->path->l[0].b->data->min_key, POS_MIN))) {
2300                         /* Advance to previous leaf node: */
2301                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2302                 } else {
2303                         /* Start of btree: */
2304                         bch2_btree_iter_set_pos(iter, POS_MIN);
2305                         k = bkey_s_c_null;
2306                         goto out;
2307                 }
2308         }
2309
2310         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
2311
2312         /* Extents can straddle iter->pos: */
2313         if (bkey_cmp(k.k->p, iter->pos) < 0)
2314                 iter->pos = k.k->p;
2315
2316         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2317                 iter->pos.snapshot = iter->snapshot;
2318 out:
2319         if (saved_path)
2320                 bch2_path_put(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2321         iter->path->should_be_locked = true;
2322
2323         bch2_btree_iter_verify_entry_exit(iter);
2324         bch2_btree_iter_verify(iter);
2325
2326         return k;
2327 }
2328
2329 /**
2330  * bch2_btree_iter_prev: returns first key less than iterator's current
2331  * position
2332  */
2333 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2334 {
2335         if (!bch2_btree_iter_rewind(iter))
2336                 return bkey_s_c_null;
2337
2338         return bch2_btree_iter_peek_prev(iter);
2339 }
2340
2341 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2342 {
2343         struct btree_trans *trans = iter->trans;
2344         struct bpos search_key;
2345         struct bkey_s_c k;
2346         int ret;
2347
2348         EBUG_ON(iter->path->level);
2349         bch2_btree_iter_verify(iter);
2350         bch2_btree_iter_verify_entry_exit(iter);
2351
2352         /* extents can't span inode numbers: */
2353         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2354             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2355                 if (iter->pos.inode == KEY_INODE_MAX)
2356                         return bkey_s_c_null;
2357
2358                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2359         }
2360
2361         search_key = btree_iter_search_key(iter);
2362         iter->path = btree_path_set_pos(trans, iter->path, search_key,
2363                                         iter->flags & BTREE_ITER_INTENT);
2364
2365         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2366         if (unlikely(ret))
2367                 return bkey_s_c_err(ret);
2368
2369         if ((iter->flags & BTREE_ITER_CACHED) ||
2370             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2371                 struct bkey_i *next_update;
2372
2373                 next_update = iter->flags & BTREE_ITER_WITH_UPDATES
2374                         ? btree_trans_peek_updates(trans, iter->btree_id, search_key)
2375                         : NULL;
2376
2377                 if (next_update &&
2378                     !bpos_cmp(next_update->k.p, iter->pos)) {
2379                         iter->k = next_update->k;
2380                         k = bkey_i_to_s_c(next_update);
2381                 } else {
2382                         k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2383                 }
2384         } else {
2385                 struct bpos next;
2386
2387                 if (iter->flags & BTREE_ITER_INTENT) {
2388                         struct btree_iter iter2;
2389
2390                         bch2_trans_copy_iter(&iter2, iter);
2391                         k = bch2_btree_iter_peek(&iter2);
2392
2393                         if (k.k && !bkey_err(k)) {
2394                                 iter->k = iter2.k;
2395                                 k.k = &iter->k;
2396                         }
2397                         bch2_trans_iter_exit(trans, &iter2);
2398                 } else {
2399                         struct bpos pos = iter->pos;
2400
2401                         k = bch2_btree_iter_peek(iter);
2402                         iter->pos = pos;
2403                 }
2404
2405                 if (unlikely(bkey_err(k)))
2406                         return k;
2407
2408                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2409
2410                 if (bkey_cmp(iter->pos, next) < 0) {
2411                         bkey_init(&iter->k);
2412                         iter->k.p = iter->pos;
2413                         bch2_key_resize(&iter->k,
2414                                         min_t(u64, KEY_SIZE_MAX,
2415                                               (next.inode == iter->pos.inode
2416                                                ? next.offset
2417                                                : KEY_OFFSET_MAX) -
2418                                               iter->pos.offset));
2419
2420                         k = (struct bkey_s_c) { &iter->k, NULL };
2421                         EBUG_ON(!k.k->size);
2422                 }
2423         }
2424
2425         iter->path->should_be_locked = true;
2426
2427         bch2_btree_iter_verify_entry_exit(iter);
2428         bch2_btree_iter_verify(iter);
2429         ret = bch2_btree_iter_verify_ret(iter, k);
2430         if (unlikely(ret))
2431                 return bkey_s_c_err(ret);
2432
2433         return k;
2434 }
2435
2436 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2437 {
2438         if (!bch2_btree_iter_advance(iter))
2439                 return bkey_s_c_null;
2440
2441         return bch2_btree_iter_peek_slot(iter);
2442 }
2443
2444 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2445 {
2446         if (!bch2_btree_iter_rewind(iter))
2447                 return bkey_s_c_null;
2448
2449         return bch2_btree_iter_peek_slot(iter);
2450 }
2451
2452 /* new transactional stuff: */
2453
2454 static inline void btree_path_verify_sorted_ref(struct btree_trans *trans,
2455                                                 struct btree_path *path)
2456 {
2457         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2458         EBUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2459         EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
2460 }
2461
2462 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2463 {
2464 #ifdef CONFIG_BCACHEFS_DEBUG
2465         unsigned i;
2466
2467         for (i = 0; i < trans->nr_sorted; i++)
2468                 btree_path_verify_sorted_ref(trans, trans->paths + trans->sorted[i]);
2469 #endif
2470 }
2471
2472 static void btree_trans_verify_sorted(struct btree_trans *trans)
2473 {
2474 #ifdef CONFIG_BCACHEFS_DEBUG
2475         struct btree_path *path, *prev = NULL;
2476         unsigned i;
2477
2478         trans_for_each_path_inorder(trans, path, i) {
2479                 BUG_ON(prev && btree_path_cmp(prev, path) > 0);
2480                 prev = path;
2481         }
2482 #endif
2483 }
2484
2485 static inline void btree_path_swap(struct btree_trans *trans,
2486                                    struct btree_path *l, struct btree_path *r)
2487 {
2488         swap(l->sorted_idx, r->sorted_idx);
2489         swap(trans->sorted[l->sorted_idx],
2490              trans->sorted[r->sorted_idx]);
2491
2492         btree_path_verify_sorted_ref(trans, l);
2493         btree_path_verify_sorted_ref(trans, r);
2494 }
2495
2496 static void btree_path_check_sort(struct btree_trans *trans, struct btree_path *path,
2497                                   int cmp)
2498 {
2499         struct btree_path *n;
2500
2501         if (cmp <= 0) {
2502                 n = prev_btree_path(trans, path);
2503                 if (n && btree_path_cmp(n, path) > 0) {
2504                         do {
2505                                 btree_path_swap(trans, n, path);
2506                                 n = prev_btree_path(trans, path);
2507                         } while (n && btree_path_cmp(n, path) > 0);
2508
2509                         goto out;
2510                 }
2511         }
2512
2513         if (cmp >= 0) {
2514                 n = next_btree_path(trans, path);
2515                 if (n && btree_path_cmp(path, n) > 0) {
2516                         do {
2517                                 btree_path_swap(trans, path, n);
2518                                 n = next_btree_path(trans, path);
2519                         } while (n && btree_path_cmp(path, n) > 0);
2520                 }
2521         }
2522 out:
2523         btree_trans_verify_sorted(trans);
2524 }
2525
2526 static inline void btree_path_list_remove(struct btree_trans *trans,
2527                                           struct btree_path *path)
2528 {
2529         unsigned i;
2530
2531         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2532
2533         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2534
2535         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2536                 trans->paths[trans->sorted[i]].sorted_idx = i;
2537
2538         path->sorted_idx = U8_MAX;
2539
2540         btree_trans_verify_sorted_refs(trans);
2541 }
2542
2543 static inline void btree_path_list_add(struct btree_trans *trans,
2544                                        struct btree_path *pos,
2545                                        struct btree_path *path)
2546 {
2547         unsigned i;
2548
2549         btree_trans_verify_sorted_refs(trans);
2550
2551         path->sorted_idx = pos ? pos->sorted_idx + 1 : 0;
2552
2553         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2554
2555         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2556                 trans->paths[trans->sorted[i]].sorted_idx = i;
2557
2558         btree_trans_verify_sorted_refs(trans);
2559 }
2560
2561 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2562 {
2563         if (iter->path)
2564                 bch2_path_put(trans, iter->path,
2565                               iter->flags & BTREE_ITER_INTENT);
2566         iter->path = NULL;
2567 }
2568
2569 static void __bch2_trans_iter_init(struct btree_trans *trans,
2570                                    struct btree_iter *iter,
2571                                    unsigned btree_id, struct bpos pos,
2572                                    unsigned locks_want,
2573                                    unsigned depth,
2574                                    unsigned flags)
2575 {
2576         EBUG_ON(trans->restarted);
2577
2578         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
2579             btree_node_type_is_extents(btree_id))
2580                 flags |= BTREE_ITER_IS_EXTENTS;
2581
2582         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
2583             !btree_type_has_snapshots(btree_id))
2584                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
2585
2586         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
2587             btree_type_has_snapshots(btree_id))
2588                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
2589
2590         iter->trans     = trans;
2591         iter->path      = NULL;
2592         iter->btree_id  = btree_id;
2593         iter->min_depth = depth;
2594         iter->flags     = flags;
2595         iter->snapshot  = pos.snapshot;
2596         iter->pos       = pos;
2597         iter->k.type    = KEY_TYPE_deleted;
2598         iter->k.p       = pos;
2599         iter->k.size    = 0;
2600
2601         iter->path = bch2_path_get(trans,
2602                                    flags & BTREE_ITER_CACHED,
2603                                    btree_id,
2604                                    iter->pos,
2605                                    locks_want,
2606                                    depth,
2607                                    flags & BTREE_ITER_INTENT);
2608 }
2609
2610 void bch2_trans_iter_init(struct btree_trans *trans,
2611                           struct btree_iter *iter,
2612                           unsigned btree_id, struct bpos pos,
2613                           unsigned flags)
2614 {
2615         __bch2_trans_iter_init(trans, iter, btree_id, pos,
2616                                0, 0, flags);
2617 }
2618
2619 void bch2_trans_node_iter_init(struct btree_trans *trans,
2620                                struct btree_iter *iter,
2621                                enum btree_id btree_id,
2622                                struct bpos pos,
2623                                unsigned locks_want,
2624                                unsigned depth,
2625                                unsigned flags)
2626 {
2627         __bch2_trans_iter_init(trans, iter, btree_id, pos, locks_want, depth,
2628                                BTREE_ITER_NOT_EXTENTS|
2629                                __BTREE_ITER_ALL_SNAPSHOTS|
2630                                BTREE_ITER_ALL_SNAPSHOTS|
2631                                flags);
2632         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2633         BUG_ON(iter->path->level        != depth);
2634         BUG_ON(iter->min_depth          != depth);
2635 }
2636
2637 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2638 {
2639         *dst = *src;
2640         if (src->path)
2641                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2642 }
2643
2644 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2645 {
2646         size_t new_top = trans->mem_top + size;
2647         void *p;
2648
2649         if (new_top > trans->mem_bytes) {
2650                 size_t old_bytes = trans->mem_bytes;
2651                 size_t new_bytes = roundup_pow_of_two(new_top);
2652                 void *new_mem;
2653
2654                 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2655
2656                 new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2657                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2658                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2659                         new_bytes = BTREE_TRANS_MEM_MAX;
2660                         kfree(trans->mem);
2661                 }
2662
2663                 if (!new_mem)
2664                         return ERR_PTR(-ENOMEM);
2665
2666                 trans->mem = new_mem;
2667                 trans->mem_bytes = new_bytes;
2668
2669                 if (old_bytes) {
2670                         trace_trans_restart_mem_realloced(trans->ip, _RET_IP_, new_bytes);
2671                         btree_trans_restart(trans);
2672                         return ERR_PTR(-EINTR);
2673                 }
2674         }
2675
2676         p = trans->mem + trans->mem_top;
2677         trans->mem_top += size;
2678         memset(p, 0, size);
2679         return p;
2680 }
2681
2682 /**
2683  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2684  * @trans: transaction to reset
2685  *
2686  * While iterating over nodes or updating nodes a attempt to lock a btree
2687  * node may return EINTR when the trylock fails. When this occurs
2688  * bch2_trans_begin() should be called and the transaction retried.
2689  */
2690 void bch2_trans_begin(struct btree_trans *trans)
2691 {
2692         struct btree_insert_entry *i;
2693         struct btree_path *path;
2694
2695         trans_for_each_update(trans, i)
2696                 __btree_path_put(i->path, true);
2697
2698         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
2699         trans->extra_journal_res        = 0;
2700         trans->nr_updates               = 0;
2701         trans->mem_top                  = 0;
2702
2703         trans->hooks                    = NULL;
2704         trans->extra_journal_entries    = NULL;
2705         trans->extra_journal_entry_u64s = 0;
2706
2707         if (trans->fs_usage_deltas) {
2708                 trans->fs_usage_deltas->used = 0;
2709                 memset(&trans->fs_usage_deltas->memset_start, 0,
2710                        (void *) &trans->fs_usage_deltas->memset_end -
2711                        (void *) &trans->fs_usage_deltas->memset_start);
2712         }
2713
2714         trans_for_each_path(trans, path) {
2715                 path->should_be_locked = false;
2716
2717                 /*
2718                  * XXX: we probably shouldn't be doing this if the transaction
2719                  * was restarted, but currently we still overflow transaction
2720                  * iterators if we do that
2721                  */
2722                 if (!path->ref && !path->preserve)
2723                         __bch2_path_free(trans, path);
2724                 else if (!path->ref)
2725                         path->preserve = false;
2726         }
2727
2728         bch2_trans_cond_resched(trans);
2729
2730         if (trans->restarted)
2731                 bch2_btree_path_traverse_all(trans);
2732
2733         trans->restarted = false;
2734 }
2735
2736 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
2737 {
2738         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
2739         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2740         void *p = NULL;
2741
2742         BUG_ON(trans->used_mempool);
2743
2744 #ifdef __KERNEL__
2745         p = this_cpu_xchg(c->btree_paths_bufs->path , NULL);
2746 #endif
2747         if (!p)
2748                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
2749
2750         trans->paths            = p; p += paths_bytes;
2751         trans->updates          = p; p += updates_bytes;
2752 }
2753
2754 void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
2755                      unsigned expected_nr_iters,
2756                      size_t expected_mem_bytes)
2757         __acquires(&c->btree_trans_barrier)
2758 {
2759         memset(trans, 0, sizeof(*trans));
2760         trans->c                = c;
2761         trans->ip               = _RET_IP_;
2762
2763         bch2_trans_alloc_paths(trans, c);
2764
2765         if (expected_mem_bytes) {
2766                 trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes);
2767                 trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL);
2768
2769                 if (!unlikely(trans->mem)) {
2770                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2771                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2772                 }
2773         }
2774
2775         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2776
2777 #ifdef CONFIG_BCACHEFS_DEBUG
2778         trans->pid = current->pid;
2779         mutex_lock(&c->btree_trans_lock);
2780         list_add(&trans->list, &c->btree_trans_list);
2781         mutex_unlock(&c->btree_trans_lock);
2782 #endif
2783 }
2784
2785 static void check_btree_paths_leaked(struct btree_trans *trans)
2786 {
2787 #ifdef CONFIG_BCACHEFS_DEBUG
2788         struct bch_fs *c = trans->c;
2789         struct btree_path *path;
2790
2791         trans_for_each_path(trans, path)
2792                 if (path->ref)
2793                         goto leaked;
2794         return;
2795 leaked:
2796         bch_err(c, "btree paths leaked from %pS!", (void *) trans->ip);
2797         trans_for_each_path(trans, path)
2798                 if (path->ref)
2799                         printk(KERN_ERR "  btree %s %pS\n",
2800                                bch2_btree_ids[path->btree_id],
2801                                (void *) path->ip_allocated);
2802         /* Be noisy about this: */
2803         bch2_fatal_error(c);
2804 #endif
2805 }
2806
2807 void bch2_trans_exit(struct btree_trans *trans)
2808         __releases(&c->btree_trans_barrier)
2809 {
2810         struct btree_insert_entry *i;
2811         struct bch_fs *c = trans->c;
2812
2813         bch2_trans_unlock(trans);
2814
2815         trans_for_each_update(trans, i)
2816                 __btree_path_put(i->path, true);
2817         trans->nr_updates               = 0;
2818
2819         check_btree_paths_leaked(trans);
2820
2821 #ifdef CONFIG_BCACHEFS_DEBUG
2822         mutex_lock(&c->btree_trans_lock);
2823         list_del(&trans->list);
2824         mutex_unlock(&c->btree_trans_lock);
2825 #endif
2826
2827         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2828
2829         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
2830
2831         if (trans->fs_usage_deltas) {
2832                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
2833                     REPLICAS_DELTA_LIST_MAX)
2834                         mempool_free(trans->fs_usage_deltas,
2835                                      &c->replicas_delta_pool);
2836                 else
2837                         kfree(trans->fs_usage_deltas);
2838         }
2839
2840         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
2841                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
2842         else
2843                 kfree(trans->mem);
2844
2845 #ifdef __KERNEL__
2846         /*
2847          * Userspace doesn't have a real percpu implementation:
2848          */
2849         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
2850 #endif
2851
2852         if (trans->paths)
2853                 mempool_free(trans->paths, &c->btree_paths_pool);
2854
2855         trans->mem      = (void *) 0x1;
2856         trans->paths    = (void *) 0x1;
2857 }
2858
2859 static void __maybe_unused
2860 bch2_btree_path_node_to_text(struct printbuf *out,
2861                              struct btree_bkey_cached_common *_b,
2862                              bool cached)
2863 {
2864         pr_buf(out, "    l=%u %s:",
2865                _b->level, bch2_btree_ids[_b->btree_id]);
2866         bch2_bpos_to_text(out, btree_node_pos(_b, cached));
2867 }
2868
2869 #ifdef CONFIG_BCACHEFS_DEBUG
2870 static bool trans_has_locks(struct btree_trans *trans)
2871 {
2872         struct btree_path *path;
2873
2874         trans_for_each_path(trans, path)
2875                 if (path->nodes_locked)
2876                         return true;
2877         return false;
2878 }
2879 #endif
2880
2881 void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
2882 {
2883 #ifdef CONFIG_BCACHEFS_DEBUG
2884         struct btree_trans *trans;
2885         struct btree_path *path;
2886         struct btree *b;
2887         unsigned l;
2888
2889         mutex_lock(&c->btree_trans_lock);
2890         list_for_each_entry(trans, &c->btree_trans_list, list) {
2891                 if (!trans_has_locks(trans))
2892                         continue;
2893
2894                 pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip);
2895
2896                 trans_for_each_path(trans, path) {
2897                         if (!path->nodes_locked)
2898                                 continue;
2899
2900                         pr_buf(out, "  path %u %c l=%u %s:",
2901                                path->idx,
2902                                path->cached ? 'c' : 'b',
2903                                path->level,
2904                                bch2_btree_ids[path->btree_id]);
2905                         bch2_bpos_to_text(out, path->pos);
2906                         pr_buf(out, "\n");
2907
2908                         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
2909                                 if (btree_node_locked(path, l)) {
2910                                         pr_buf(out, "    %s l=%u ",
2911                                                btree_node_intent_locked(path, l) ? "i" : "r", l);
2912                                         bch2_btree_path_node_to_text(out,
2913                                                         (void *) path->l[l].b,
2914                                                         path->cached);
2915                                         pr_buf(out, "\n");
2916                                 }
2917                         }
2918                 }
2919
2920                 b = READ_ONCE(trans->locking);
2921                 if (b) {
2922                         path = &trans->paths[trans->locking_path_idx];
2923                         pr_buf(out, "  locking path %u %c l=%u %s:",
2924                                trans->locking_path_idx,
2925                                path->cached ? 'c' : 'b',
2926                                trans->locking_level,
2927                                bch2_btree_ids[trans->locking_btree_id]);
2928                         bch2_bpos_to_text(out, trans->locking_pos);
2929
2930                         pr_buf(out, " node ");
2931                         bch2_btree_path_node_to_text(out,
2932                                         (void *) b, path->cached);
2933                         pr_buf(out, "\n");
2934                 }
2935         }
2936         mutex_unlock(&c->btree_trans_lock);
2937 #endif
2938 }
2939
2940 void bch2_fs_btree_iter_exit(struct bch_fs *c)
2941 {
2942         mempool_exit(&c->btree_trans_mem_pool);
2943         mempool_exit(&c->btree_paths_pool);
2944         cleanup_srcu_struct(&c->btree_trans_barrier);
2945 }
2946
2947 int bch2_fs_btree_iter_init(struct bch_fs *c)
2948 {
2949         unsigned nr = BTREE_ITER_MAX;
2950
2951         INIT_LIST_HEAD(&c->btree_trans_list);
2952         mutex_init(&c->btree_trans_lock);
2953
2954         return  init_srcu_struct(&c->btree_trans_barrier) ?:
2955                 mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
2956                         sizeof(struct btree_path) * nr +
2957                         sizeof(struct btree_insert_entry) * nr) ?:
2958                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
2959                                           BTREE_TRANS_MEM_MAX);
2960 }