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 static inline int __btree_path_cmp(const struct btree_path *l,
  31                                    enum btree_id        r_btree_id,
  32                                    bool                 r_cached,
  33                                    struct bpos          r_pos,
  34                                    unsigned             r_level)
  35 {
  36         return   cmp_int(l->btree_id,   r_btree_id) ?:
  37                  cmp_int(l->cached,     r_cached) ?:
  38                  bpos_cmp(l->pos,       r_pos) ?:
  39                 -cmp_int(l->level,      r_level);
  40 }
  41
  42 static inline int btree_path_cmp(const struct btree_path *l,
  43                                  const struct btree_path *r)
  44 {
  45         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  46 }
  47
  48 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  49 {
  50         /* Are we iterating over keys in all snapshots? */
  51         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  52                 p = bpos_successor(p);
  53         } else {
  54                 p = bpos_nosnap_successor(p);
  55                 p.snapshot = iter->snapshot;
  56         }
  57
  58         return p;
  59 }
  60
  61 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  62 {
  63         /* Are we iterating over keys in all snapshots? */
  64         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  65                 p = bpos_predecessor(p);
  66         } else {
  67                 p = bpos_nosnap_predecessor(p);
  68                 p.snapshot = iter->snapshot;
  69         }
  70
  71         return p;
  72 }
  73
  74 static inline bool is_btree_node(struct btree_path *path, unsigned l)
  75 {
  76         return l < BTREE_MAX_DEPTH &&
  77                 (unsigned long) path->l[l].b >= 128;
  78 }
  79
  80 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
  81 {
  82         struct bpos pos = iter->pos;
  83
  84         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
  85             bkey_cmp(pos, POS_MAX))
  86                 pos = bkey_successor(iter, pos);
  87         return pos;
  88 }
  89
  90 static inline bool btree_path_pos_before_node(struct btree_path *path,
  91                                               struct btree *b)
  92 {
  93         return bpos_cmp(path->pos, b->data->min_key) < 0;
  94 }
  95
  96 static inline bool btree_path_pos_after_node(struct btree_path *path,
  97                                              struct btree *b)
  98 {
  99         return bpos_cmp(b->key.k.p, path->pos) < 0;
 100 }
 101
 102 static inline bool btree_path_pos_in_node(struct btree_path *path,
 103                                           struct btree *b)
 104 {
 105         return path->btree_id == b->c.btree_id &&
 106                 !btree_path_pos_before_node(path, b) &&
 107                 !btree_path_pos_after_node(path, b);
 108 }
 109
 110 /* Btree node locking: */
 111
 112 void bch2_btree_node_unlock_write(struct btree_trans *trans,
 113                         struct btree_path *path, struct btree *b)
 114 {
 115         bch2_btree_node_unlock_write_inlined(trans, path, b);
 116 }
 117
 118 void __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree *b)
 119 {
 120         struct btree_path *linked;
 121         unsigned readers = 0;
 122
 123         trans_for_each_path(trans, linked)
 124                 if (linked->l[b->c.level].b == b &&
 125                     btree_node_read_locked(linked, b->c.level))
 126                         readers++;
 127
 128         /*
 129          * Must drop our read locks before calling six_lock_write() -
 130          * six_unlock() won't do wakeups until the reader count
 131          * goes to 0, and it's safe because we have the node intent
 132          * locked:
 133          */
 134         atomic64_sub(__SIX_VAL(read_lock, readers),
 135                      &b->c.lock.state.counter);
 136         btree_node_lock_type(trans->c, b, SIX_LOCK_write);
 137         atomic64_add(__SIX_VAL(read_lock, readers),
 138                      &b->c.lock.state.counter);
 139 }
 140
 141 bool __bch2_btree_node_relock(struct btree_trans *trans,
 142                               struct btree_path *path, unsigned level)
 143 {
 144         struct btree *b = btree_path_node(path, level);
 145         int want = __btree_lock_want(path, level);
 146
 147         if (!is_btree_node(path, level))
 148                 return false;
 149
 150         if (race_fault())
 151                 return false;
 152
 153         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 154             (btree_node_lock_seq_matches(path, b, level) &&
 155              btree_node_lock_increment(trans, b, level, want))) {
 156                 mark_btree_node_locked(path, level, want);
 157                 return true;
 158         } else {
 159                 return false;
 160         }
 161 }
 162
 163 static bool bch2_btree_node_upgrade(struct btree_trans *trans,
 164                                     struct btree_path *path, unsigned level)
 165 {
 166         struct btree *b = path->l[level].b;
 167
 168         EBUG_ON(btree_lock_want(path, level) != BTREE_NODE_INTENT_LOCKED);
 169
 170         if (!is_btree_node(path, level))
 171                 return false;
 172
 173         if (btree_node_intent_locked(path, level))
 174                 return true;
 175
 176         if (race_fault())
 177                 return false;
 178
 179         if (btree_node_locked(path, level)
 180             ? six_lock_tryupgrade(&b->c.lock)
 181             : six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 182                 goto success;
 183
 184         if (btree_node_lock_seq_matches(path, b, level) &&
 185             btree_node_lock_increment(trans, b, level, BTREE_NODE_INTENT_LOCKED)) {
 186                 btree_node_unlock(path, level);
 187                 goto success;
 188         }
 189
 190         return false;
 191 success:
 192         mark_btree_node_intent_locked(path, level);
 193         return true;
 194 }
 195
 196 static inline bool btree_path_get_locks(struct btree_trans *trans,
 197                                         struct btree_path *path,
 198                                         bool upgrade, unsigned long trace_ip)
 199 {
 200         unsigned l = path->level;
 201         int fail_idx = -1;
 202
 203         do {
 204                 if (!btree_path_node(path, l))
 205                         break;
 206
 207                 if (!(upgrade
 208                       ? bch2_btree_node_upgrade(trans, path, l)
 209                       : bch2_btree_node_relock(trans, path, l)))
 210                         fail_idx = l;
 211
 212                 l++;
 213         } while (l < path->locks_want);
 214
 215         /*
 216          * When we fail to get a lock, we have to ensure that any child nodes
 217          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 218          * the node that we failed to relock:
 219          */
 220         if (fail_idx >= 0) {
 221                 __bch2_btree_path_unlock(path);
 222                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 223
 224                 do {
 225                         path->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS;
 226                         --fail_idx;
 227                 } while (fail_idx >= 0);
 228         }
 229
 230         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 231                 path->uptodate = BTREE_ITER_UPTODATE;
 232
 233         bch2_trans_verify_locks(trans);
 234
 235         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 236 }
 237
 238 static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b,
 239                                   bool cached)
 240 {
 241         return !cached
 242                 ? container_of(_b, struct btree, c)->key.k.p
 243                 : container_of(_b, struct bkey_cached, c)->key.pos;
 244 }
 245
 246 /* Slowpath: */
 247 bool __bch2_btree_node_lock(struct btree_trans *trans,
 248                             struct btree_path *path,
 249                             struct btree *b,
 250                             struct bpos pos, unsigned level,
 251                             enum six_lock_type type,
 252                             six_lock_should_sleep_fn should_sleep_fn, void *p,
 253                             unsigned long ip)
 254 {
 255         struct btree_path *linked, *deadlock_path = NULL;
 256         u64 start_time = local_clock();
 257         unsigned reason = 9;
 258         bool ret;
 259
 260         /* Check if it's safe to block: */
 261         trans_for_each_path(trans, linked) {
 262                 if (!linked->nodes_locked)
 263                         continue;
 264
 265                 /*
 266                  * Can't block taking an intent lock if we have _any_ nodes read
 267                  * locked:
 268                  *
 269                  * - Our read lock blocks another thread with an intent lock on
 270                  *   the same node from getting a write lock, and thus from
 271                  *   dropping its intent lock
 272                  *
 273                  * - And the other thread may have multiple nodes intent locked:
 274                  *   both the node we want to intent lock, and the node we
 275                  *   already have read locked - deadlock:
 276                  */
 277                 if (type == SIX_LOCK_intent &&
 278                     linked->nodes_locked != linked->nodes_intent_locked) {
 279                         deadlock_path = linked;
 280                         reason = 1;
 281                 }
 282
 283                 if (linked->btree_id != path->btree_id) {
 284                         if (linked->btree_id > path->btree_id) {
 285                                 deadlock_path = linked;
 286                                 reason = 3;
 287                         }
 288                         continue;
 289                 }
 290
 291                 /*
 292                  * Within the same btree, cached paths come before non
 293                  * cached paths:
 294                  */
 295                 if (linked->cached != path->cached) {
 296                         if (path->cached) {
 297                                 deadlock_path = linked;
 298                                 reason = 4;
 299                         }
 300                         continue;
 301                 }
 302
 303                 /*
 304                  * Interior nodes must be locked before their descendants: if
 305                  * another path has possible descendants locked of the node
 306                  * we're about to lock, it must have the ancestors locked too:
 307                  */
 308                 if (level > __fls(linked->nodes_locked)) {
 309                         deadlock_path = linked;
 310                         reason = 5;
 311                 }
 312
 313                 /* Must lock btree nodes in key order: */
 314                 if (btree_node_locked(linked, level) &&
 315                     bpos_cmp(pos, btree_node_pos((void *) linked->l[level].b,
 316                                                  linked->cached)) <= 0) {
 317                         deadlock_path = linked;
 318                         reason = 7;
 319                         BUG_ON(trans->in_traverse_all);
 320                 }
 321         }
 322
 323         if (unlikely(deadlock_path)) {
 324                 trace_trans_restart_would_deadlock(trans->ip, ip,
 325                                 trans->in_traverse_all, reason,
 326                                 deadlock_path->btree_id,
 327                                 deadlock_path->cached,
 328                                 &deadlock_path->pos,
 329                                 path->btree_id,
 330                                 path->cached,
 331                                 &pos);
 332                 btree_trans_restart(trans);
 333                 return false;
 334         }
 335
 336         if (six_trylock_type(&b->c.lock, type))
 337                 return true;
 338
 339 #ifdef CONFIG_BCACHEFS_DEBUG
 340         trans->locking_path_idx = path->idx;
 341         trans->locking_pos      = pos;
 342         trans->locking_btree_id = path->btree_id;
 343         trans->locking_level    = level;
 344         trans->locking          = b;
 345 #endif
 346
 347         ret = six_lock_type(&b->c.lock, type, should_sleep_fn, p) == 0;
 348
 349 #ifdef CONFIG_BCACHEFS_DEBUG
 350         trans->locking = NULL;
 351 #endif
 352         if (ret)
 353                 bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)],
 354                                        start_time);
 355         return ret;
 356 }
 357
 358 /* Btree iterator locking: */
 359
 360 #ifdef CONFIG_BCACHEFS_DEBUG
 361
 362 static void bch2_btree_path_verify_locks(struct btree_path *path)
 363 {
 364         unsigned l;
 365
 366         if (!path->nodes_locked) {
 367                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE);
 368                 return;
 369         }
 370
 371         for (l = 0; btree_path_node(path, l); l++)
 372                 BUG_ON(btree_lock_want(path, l) !=
 373                        btree_node_locked_type(path, l));
 374 }
 375
 376 void bch2_trans_verify_locks(struct btree_trans *trans)
 377 {
 378         struct btree_path *path;
 379
 380         trans_for_each_path(trans, path)
 381                 bch2_btree_path_verify_locks(path);
 382 }
 383 #else
 384 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
 385 #endif
 386
 387 /* Btree path locking: */
 388
 389 /*
 390  * Only for btree_cache.c - only relocks intent locks
 391  */
 392 bool bch2_btree_path_relock_intent(struct btree_trans *trans,
 393                                    struct btree_path *path)
 394 {
 395         unsigned l;
 396
 397         for (l = path->level;
 398              l < path->locks_want && btree_path_node(path, l);
 399              l++) {
 400                 if (!bch2_btree_node_relock(trans, path, l)) {
 401                         __bch2_btree_path_unlock(path);
 402                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 403                         btree_trans_restart(trans);
 404                         return false;
 405                 }
 406         }
 407
 408         return true;
 409 }
 410
 411 __flatten
 412 static bool bch2_btree_path_relock(struct btree_trans *trans,
 413                         struct btree_path *path, unsigned long trace_ip)
 414 {
 415         bool ret = btree_path_get_locks(trans, path, false, trace_ip);
 416
 417         if (!ret)
 418                 btree_trans_restart(trans);
 419         return ret;
 420 }
 421
 422 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 423                                struct btree_path *path,
 424                                unsigned new_locks_want)
 425 {
 426         struct btree_path *linked;
 427
 428         EBUG_ON(path->locks_want >= new_locks_want);
 429
 430         path->locks_want = new_locks_want;
 431
 432         if (btree_path_get_locks(trans, path, true, _THIS_IP_))
 433                 return true;
 434
 435         /*
 436          * XXX: this is ugly - we'd prefer to not be mucking with other
 437          * iterators in the btree_trans here.
 438          *
 439          * On failure to upgrade the iterator, setting iter->locks_want and
 440          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 441          * get the locks we want on transaction restart.
 442          *
 443          * But if this iterator was a clone, on transaction restart what we did
 444          * to this iterator isn't going to be preserved.
 445          *
 446          * Possibly we could add an iterator field for the parent iterator when
 447          * an iterator is a copy - for now, we'll just upgrade any other
 448          * iterators with the same btree id.
 449          *
 450          * The code below used to be needed to ensure ancestor nodes get locked
 451          * before interior nodes - now that's handled by
 452          * bch2_btree_path_traverse_all().
 453          */
 454         trans_for_each_path(trans, linked)
 455                 if (linked != path &&
 456                     linked->cached == path->cached &&
 457                     linked->btree_id == path->btree_id &&
 458                     linked->locks_want < new_locks_want) {
 459                         linked->locks_want = new_locks_want;
 460                         btree_path_get_locks(trans, linked, true, _THIS_IP_);
 461                 }
 462
 463         return false;
 464 }
 465
 466 void __bch2_btree_path_downgrade(struct btree_path *path,
 467                                  unsigned new_locks_want)
 468 {
 469         unsigned l;
 470
 471         EBUG_ON(path->locks_want < new_locks_want);
 472
 473         path->locks_want = new_locks_want;
 474
 475         while (path->nodes_locked &&
 476                (l = __fls(path->nodes_locked)) >= path->locks_want) {
 477                 if (l > path->level) {
 478                         btree_node_unlock(path, l);
 479                 } else {
 480                         if (btree_node_intent_locked(path, l)) {
 481                                 six_lock_downgrade(&path->l[l].b->c.lock);
 482                                 path->nodes_intent_locked ^= 1 << l;
 483                         }
 484                         break;
 485                 }
 486         }
 487
 488         bch2_btree_path_verify_locks(path);
 489 }
 490
 491 void bch2_trans_downgrade(struct btree_trans *trans)
 492 {
 493         struct btree_path *path;
 494
 495         trans_for_each_path(trans, path)
 496                 bch2_btree_path_downgrade(path);
 497 }
 498
 499 /* Btree transaction locking: */
 500
 501 bool bch2_trans_relock(struct btree_trans *trans)
 502 {
 503         struct btree_path *path;
 504
 505         if (unlikely(trans->restarted))
 506                 return false;
 507
 508         trans_for_each_path(trans, path)
 509                 if (path->should_be_locked &&
 510                     !bch2_btree_path_relock(trans, path, _RET_IP_)) {
 511                         trace_trans_restart_relock(trans->ip, _RET_IP_,
 512                                         path->btree_id, &path->pos);
 513                         BUG_ON(!trans->restarted);
 514                         return false;
 515                 }
 516         return true;
 517 }
 518
 519 void bch2_trans_unlock(struct btree_trans *trans)
 520 {
 521         struct btree_path *path;
 522
 523         trans_for_each_path(trans, path)
 524                 __bch2_btree_path_unlock(path);
 525
 526         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
 527 }
 528
 529 /* Btree iterator: */
 530
 531 #ifdef CONFIG_BCACHEFS_DEBUG
 532
 533 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 534                                           struct btree_path *path)
 535 {
 536         struct bkey_cached *ck;
 537         bool locked = btree_node_locked(path, 0);
 538
 539         if (!bch2_btree_node_relock(trans, path, 0))
 540                 return;
 541
 542         ck = (void *) path->l[0].b;
 543         BUG_ON(ck->key.btree_id != path->btree_id ||
 544                bkey_cmp(ck->key.pos, path->pos));
 545
 546         if (!locked)
 547                 btree_node_unlock(path, 0);
 548 }
 549
 550 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 551                                 struct btree_path *path, unsigned level)
 552 {
 553         struct btree_path_level *l;
 554         struct btree_node_iter tmp;
 555         bool locked;
 556         struct bkey_packed *p, *k;
 557         char buf1[100], buf2[100], buf3[100];
 558         const char *msg;
 559
 560         if (!bch2_debug_check_iterators)
 561                 return;
 562
 563         l       = &path->l[level];
 564         tmp     = l->iter;
 565         locked  = btree_node_locked(path, level);
 566
 567         if (path->cached) {
 568                 if (!level)
 569                         bch2_btree_path_verify_cached(trans, path);
 570                 return;
 571         }
 572
 573         if (!btree_path_node(path, level))
 574                 return;
 575
 576         if (!bch2_btree_node_relock(trans, path, level))
 577                 return;
 578
 579         BUG_ON(!btree_path_pos_in_node(path, l->b));
 580
 581         bch2_btree_node_iter_verify(&l->iter, l->b);
 582
 583         /*
 584          * For interior nodes, the iterator will have skipped past deleted keys:
 585          */
 586         p = level
 587                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 588                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 589         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 590
 591         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 592                 msg = "before";
 593                 goto err;
 594         }
 595
 596         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 597                 msg = "after";
 598                 goto err;
 599         }
 600
 601         if (!locked)
 602                 btree_node_unlock(path, level);
 603         return;
 604 err:
 605         strcpy(buf2, "(none)");
 606         strcpy(buf3, "(none)");
 607
 608         bch2_bpos_to_text(&PBUF(buf1), path->pos);
 609
 610         if (p) {
 611                 struct bkey uk = bkey_unpack_key(l->b, p);
 612                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 613         }
 614
 615         if (k) {
 616                 struct bkey uk = bkey_unpack_key(l->b, k);
 617                 bch2_bkey_to_text(&PBUF(buf3), &uk);
 618         }
 619
 620         panic("path should be %s key at level %u:\n"
 621               "path pos %s\n"
 622               "prev key %s\n"
 623               "cur  key %s\n",
 624               msg, level, buf1, buf2, buf3);
 625 }
 626
 627 static void bch2_btree_path_verify(struct btree_trans *trans,
 628                                    struct btree_path *path)
 629 {
 630         struct bch_fs *c = trans->c;
 631         unsigned i;
 632
 633         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 634
 635         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 636                 if (!path->l[i].b) {
 637                         BUG_ON(c->btree_roots[path->btree_id].b->c.level > i);
 638                         break;
 639                 }
 640
 641                 bch2_btree_path_verify_level(trans, path, i);
 642         }
 643
 644         bch2_btree_path_verify_locks(path);
 645 }
 646
 647 void bch2_trans_verify_paths(struct btree_trans *trans)
 648 {
 649         struct btree_path *path;
 650
 651         trans_for_each_path(trans, path)
 652                 bch2_btree_path_verify(trans, path);
 653 }
 654
 655 static void bch2_btree_iter_verify(struct btree_iter *iter)
 656 {
 657         struct btree_trans *trans = iter->trans;
 658
 659         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 660
 661         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
 662
 663         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 664                iter->pos.snapshot != iter->snapshot);
 665
 666         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 667                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 668
 669         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 670                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 671                !btree_type_has_snapshots(iter->btree_id));
 672
 673         bch2_btree_path_verify(trans, iter->path);
 674 }
 675
 676 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 677 {
 678         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 679                !iter->pos.snapshot);
 680
 681         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 682                iter->pos.snapshot != iter->snapshot);
 683
 684         BUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
 685                bkey_cmp(iter->pos, iter->k.p) > 0);
 686 }
 687
 688 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 689 {
 690         struct btree_trans *trans = iter->trans;
 691         struct btree_iter copy;
 692         struct bkey_s_c prev;
 693         int ret = 0;
 694
 695         if (!bch2_debug_check_iterators)
 696                 return 0;
 697
 698         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 699                 return 0;
 700
 701         if (bkey_err(k) || !k.k)
 702                 return 0;
 703
 704         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 705                                           iter->snapshot,
 706                                           k.k->p.snapshot));
 707
 708         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 709                              BTREE_ITER_ALL_SNAPSHOTS);
 710         prev = bch2_btree_iter_prev(&copy);
 711         if (!prev.k)
 712                 goto out;
 713
 714         ret = bkey_err(prev);
 715         if (ret)
 716                 goto out;
 717
 718         if (!bkey_cmp(prev.k->p, k.k->p) &&
 719             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 720                                       prev.k->p.snapshot) > 0) {
 721                 char buf1[100], buf2[200];
 722
 723                 bch2_bkey_to_text(&PBUF(buf1), k.k);
 724                 bch2_bkey_to_text(&PBUF(buf2), prev.k);
 725
 726                 panic("iter snap %u\n"
 727                       "k    %s\n"
 728                       "prev %s\n",
 729                       iter->snapshot,
 730                       buf1, buf2);
 731         }
 732 out:
 733         bch2_trans_iter_exit(trans, &copy);
 734         return ret;
 735 }
 736
 737 #else
 738
 739 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 740                                                 struct btree_path *path, unsigned l) {}
 741 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 742                                           struct btree_path *path) {}
 743 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 744 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 745 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 746
 747 #endif
 748
 749 /* Btree path: fixups after btree updates */
 750
 751 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 752                                         struct btree *b,
 753                                         struct bset_tree *t,
 754                                         struct bkey_packed *k)
 755 {
 756         struct btree_node_iter_set *set;
 757
 758         btree_node_iter_for_each(iter, set)
 759                 if (set->end == t->end_offset) {
 760                         set->k = __btree_node_key_to_offset(b, k);
 761                         bch2_btree_node_iter_sort(iter, b);
 762                         return;
 763                 }
 764
 765         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 766 }
 767
 768 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 769                                                struct btree *b,
 770                                                struct bkey_packed *where)
 771 {
 772         struct btree_path_level *l = &path->l[b->c.level];
 773
 774         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 775                 return;
 776
 777         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 778                 bch2_btree_node_iter_advance(&l->iter, l->b);
 779 }
 780
 781 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 782                                       struct btree *b,
 783                                       struct bkey_packed *where)
 784 {
 785         struct btree_path *path;
 786
 787         trans_for_each_path_with_node(trans, b, path) {
 788                 __bch2_btree_path_fix_key_modified(path, b, where);
 789                 bch2_btree_path_verify_level(trans, path, b->c.level);
 790         }
 791 }
 792
 793 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 794                                        struct btree *b,
 795                                        struct btree_node_iter *node_iter,
 796                                        struct bset_tree *t,
 797                                        struct bkey_packed *where,
 798                                        unsigned clobber_u64s,
 799                                        unsigned new_u64s)
 800 {
 801         const struct bkey_packed *end = btree_bkey_last(b, t);
 802         struct btree_node_iter_set *set;
 803         unsigned offset = __btree_node_key_to_offset(b, where);
 804         int shift = new_u64s - clobber_u64s;
 805         unsigned old_end = t->end_offset - shift;
 806         unsigned orig_iter_pos = node_iter->data[0].k;
 807         bool iter_current_key_modified =
 808                 orig_iter_pos >= offset &&
 809                 orig_iter_pos <= offset + clobber_u64s;
 810
 811         btree_node_iter_for_each(node_iter, set)
 812                 if (set->end == old_end)
 813                         goto found;
 814
 815         /* didn't find the bset in the iterator - might have to readd it: */
 816         if (new_u64s &&
 817             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 818                 bch2_btree_node_iter_push(node_iter, b, where, end);
 819                 goto fixup_done;
 820         } else {
 821                 /* Iterator is after key that changed */
 822                 return;
 823         }
 824 found:
 825         set->end = t->end_offset;
 826
 827         /* Iterator hasn't gotten to the key that changed yet: */
 828         if (set->k < offset)
 829                 return;
 830
 831         if (new_u64s &&
 832             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 833                 set->k = offset;
 834         } else if (set->k < offset + clobber_u64s) {
 835                 set->k = offset + new_u64s;
 836                 if (set->k == set->end)
 837                         bch2_btree_node_iter_set_drop(node_iter, set);
 838         } else {
 839                 /* Iterator is after key that changed */
 840                 set->k = (int) set->k + shift;
 841                 return;
 842         }
 843
 844         bch2_btree_node_iter_sort(node_iter, b);
 845 fixup_done:
 846         if (node_iter->data[0].k != orig_iter_pos)
 847                 iter_current_key_modified = true;
 848
 849         /*
 850          * When a new key is added, and the node iterator now points to that
 851          * key, the iterator might have skipped past deleted keys that should
 852          * come after the key the iterator now points to. We have to rewind to
 853          * before those deleted keys - otherwise
 854          * bch2_btree_node_iter_prev_all() breaks:
 855          */
 856         if (!bch2_btree_node_iter_end(node_iter) &&
 857             iter_current_key_modified &&
 858             b->c.level) {
 859                 struct bset_tree *t;
 860                 struct bkey_packed *k, *k2, *p;
 861
 862                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 863
 864                 for_each_bset(b, t) {
 865                         bool set_pos = false;
 866
 867                         if (node_iter->data[0].end == t->end_offset)
 868                                 continue;
 869
 870                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 871
 872                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 873                                bkey_iter_cmp(b, k, p) < 0) {
 874                                 k2 = p;
 875                                 set_pos = true;
 876                         }
 877
 878                         if (set_pos)
 879                                 btree_node_iter_set_set_pos(node_iter,
 880                                                             b, t, k2);
 881                 }
 882         }
 883 }
 884
 885 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 886                               struct btree_path *path,
 887                               struct btree *b,
 888                               struct btree_node_iter *node_iter,
 889                               struct bkey_packed *where,
 890                               unsigned clobber_u64s,
 891                               unsigned new_u64s)
 892 {
 893         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 894         struct btree_path *linked;
 895
 896         if (node_iter != &path->l[b->c.level].iter) {
 897                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 898                                            where, clobber_u64s, new_u64s);
 899
 900                 if (bch2_debug_check_iterators)
 901                         bch2_btree_node_iter_verify(node_iter, b);
 902         }
 903
 904         trans_for_each_path_with_node(trans, b, linked) {
 905                 __bch2_btree_node_iter_fix(linked, b,
 906                                            &linked->l[b->c.level].iter, t,
 907                                            where, clobber_u64s, new_u64s);
 908                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 909         }
 910 }
 911
 912 /* Btree path level: pointer to a particular btree node and node iter */
 913
 914 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
 915                                                   struct btree_path_level *l,
 916                                                   struct bkey *u,
 917                                                   struct bkey_packed *k)
 918 {
 919         struct bkey_s_c ret;
 920
 921         if (unlikely(!k)) {
 922                 /*
 923                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 924                  * a hole
 925                  */
 926                 u->type = KEY_TYPE_deleted;
 927                 return bkey_s_c_null;
 928         }
 929
 930         ret = bkey_disassemble(l->b, k, u);
 931
 932         /*
 933          * XXX: bch2_btree_bset_insert_key() generates invalid keys when we
 934          * overwrite extents - it sets k->type = KEY_TYPE_deleted on the key
 935          * being overwritten but doesn't change k->size. But this is ok, because
 936          * those keys are never written out, we just have to avoid a spurious
 937          * assertion here:
 938          */
 939         if (bch2_debug_check_bkeys && !bkey_deleted(ret.k))
 940                 bch2_bkey_debugcheck(c, l->b, ret);
 941
 942         return ret;
 943 }
 944
 945 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
 946                                                         struct btree_path_level *l,
 947                                                         struct bkey *u)
 948 {
 949         return __btree_iter_unpack(c, l, u,
 950                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 951 }
 952
 953 static inline struct bkey_s_c btree_path_level_peek(struct bch_fs *c,
 954                                                     struct btree_path *path,
 955                                                     struct btree_path_level *l,
 956                                                     struct bkey *u)
 957 {
 958         struct bkey_s_c k = __btree_iter_unpack(c, l, u,
 959                         bch2_btree_node_iter_peek(&l->iter, l->b));
 960
 961         path->pos = k.k ? k.k->p : l->b->key.k.p;
 962         return k;
 963 }
 964
 965 static inline struct bkey_s_c btree_path_level_prev(struct bch_fs *c,
 966                                                     struct btree_path *path,
 967                                                     struct btree_path_level *l,
 968                                                     struct bkey *u)
 969 {
 970         struct bkey_s_c k = __btree_iter_unpack(c, l, u,
 971                         bch2_btree_node_iter_prev(&l->iter, l->b));
 972
 973         path->pos = k.k ? k.k->p : l->b->data->min_key;
 974         return k;
 975 }
 976
 977 static inline bool btree_path_advance_to_pos(struct btree_path *path,
 978                                              struct btree_path_level *l,
 979                                              int max_advance)
 980 {
 981         struct bkey_packed *k;
 982         int nr_advanced = 0;
 983
 984         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 985                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 986                 if (max_advance > 0 && nr_advanced >= max_advance)
 987                         return false;
 988
 989                 bch2_btree_node_iter_advance(&l->iter, l->b);
 990                 nr_advanced++;
 991         }
 992
 993         return true;
 994 }
 995
 996 /*
 997  * Verify that iterator for parent node points to child node:
 998  */
 999 static void btree_path_verify_new_node(struct btree_trans *trans,
1000                                        struct btree_path *path, struct btree *b)
1001 {
1002         struct btree_path_level *l;
1003         unsigned plevel;
1004         bool parent_locked;
1005         struct bkey_packed *k;
1006
1007         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
1008                 return;
1009
1010         plevel = b->c.level + 1;
1011         if (!btree_path_node(path, plevel))
1012                 return;
1013
1014         parent_locked = btree_node_locked(path, plevel);
1015
1016         if (!bch2_btree_node_relock(trans, path, plevel))
1017                 return;
1018
1019         l = &path->l[plevel];
1020         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1021         if (!k ||
1022             bkey_deleted(k) ||
1023             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
1024                 char buf1[100];
1025                 char buf2[100];
1026                 char buf3[100];
1027                 char buf4[100];
1028                 struct bkey uk = bkey_unpack_key(b, k);
1029
1030                 bch2_dump_btree_node(trans->c, l->b);
1031                 bch2_bpos_to_text(&PBUF(buf1), path->pos);
1032                 bch2_bkey_to_text(&PBUF(buf2), &uk);
1033                 bch2_bpos_to_text(&PBUF(buf3), b->data->min_key);
1034                 bch2_bpos_to_text(&PBUF(buf3), b->data->max_key);
1035                 panic("parent iter doesn't point to new node:\n"
1036                       "iter pos %s %s\n"
1037                       "iter key %s\n"
1038                       "new node %s-%s\n",
1039                       bch2_btree_ids[path->btree_id], buf1,
1040                       buf2, buf3, buf4);
1041         }
1042
1043         if (!parent_locked)
1044                 btree_node_unlock(path, plevel);
1045 }
1046
1047 static inline void __btree_path_level_init(struct btree_path *path,
1048                                            unsigned level)
1049 {
1050         struct btree_path_level *l = &path->l[level];
1051
1052         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1053
1054         /*
1055          * Iterators to interior nodes should always be pointed at the first non
1056          * whiteout:
1057          */
1058         if (level)
1059                 bch2_btree_node_iter_peek(&l->iter, l->b);
1060 }
1061
1062 static inline void btree_path_level_init(struct btree_trans *trans,
1063                                          struct btree_path *path,
1064                                          struct btree *b)
1065 {
1066         BUG_ON(path->cached);
1067
1068         btree_path_verify_new_node(trans, path, b);
1069
1070         EBUG_ON(!btree_path_pos_in_node(path, b));
1071         EBUG_ON(b->c.lock.state.seq & 1);
1072
1073         path->l[b->c.level].lock_seq = b->c.lock.state.seq;
1074         path->l[b->c.level].b = b;
1075         __btree_path_level_init(path, b->c.level);
1076 }
1077
1078 /* Btree path: fixups after btree node updates: */
1079
1080 /*
1081  * A btree node is being replaced - update the iterator to point to the new
1082  * node:
1083  */
1084 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
1085 {
1086         struct btree_path *path;
1087
1088         trans_for_each_path(trans, path)
1089                 if (!path->cached &&
1090                     btree_path_pos_in_node(path, b)) {
1091                         enum btree_node_locked_type t =
1092                                 btree_lock_want(path, b->c.level);
1093
1094                         if (path->nodes_locked &&
1095                             t != BTREE_NODE_UNLOCKED) {
1096                                 btree_node_unlock(path, b->c.level);
1097                                 six_lock_increment(&b->c.lock, t);
1098                                 mark_btree_node_locked(path, b->c.level, t);
1099                         }
1100
1101                         btree_path_level_init(trans, path, b);
1102                 }
1103 }
1104
1105 /*
1106  * A btree node has been modified in such a way as to invalidate iterators - fix
1107  * them:
1108  */
1109 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
1110 {
1111         struct btree_path *path;
1112
1113         trans_for_each_path_with_node(trans, b, path)
1114                 __btree_path_level_init(path, b->c.level);
1115 }
1116
1117 /* Btree path: traverse, set_pos: */
1118
1119 static int lock_root_check_fn(struct six_lock *lock, void *p)
1120 {
1121         struct btree *b = container_of(lock, struct btree, c.lock);
1122         struct btree **rootp = p;
1123
1124         return b == *rootp ? 0 : -1;
1125 }
1126
1127 static inline int btree_path_lock_root(struct btree_trans *trans,
1128                                        struct btree_path *path,
1129                                        unsigned depth_want,
1130                                        unsigned long trace_ip)
1131 {
1132         struct bch_fs *c = trans->c;
1133         struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
1134         enum six_lock_type lock_type;
1135         unsigned i;
1136
1137         EBUG_ON(path->nodes_locked);
1138
1139         while (1) {
1140                 b = READ_ONCE(*rootp);
1141                 path->level = READ_ONCE(b->c.level);
1142
1143                 if (unlikely(path->level < depth_want)) {
1144                         /*
1145                          * the root is at a lower depth than the depth we want:
1146                          * got to the end of the btree, or we're walking nodes
1147                          * greater than some depth and there are no nodes >=
1148                          * that depth
1149                          */
1150                         path->level = depth_want;
1151                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
1152                                 path->l[i].b = NULL;
1153                         return 1;
1154                 }
1155
1156                 lock_type = __btree_lock_want(path, path->level);
1157                 if (unlikely(!btree_node_lock(trans, path, b, SPOS_MAX,
1158                                               path->level, lock_type,
1159                                               lock_root_check_fn, rootp,
1160                                               trace_ip))) {
1161                         if (trans->restarted)
1162                                 return -EINTR;
1163                         continue;
1164                 }
1165
1166                 if (likely(b == READ_ONCE(*rootp) &&
1167                            b->c.level == path->level &&
1168                            !race_fault())) {
1169                         for (i = 0; i < path->level; i++)
1170                                 path->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
1171                         path->l[path->level].b = b;
1172                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
1173                                 path->l[i].b = NULL;
1174
1175                         mark_btree_node_locked(path, path->level, lock_type);
1176                         btree_path_level_init(trans, path, b);
1177                         return 0;
1178                 }
1179
1180                 six_unlock_type(&b->c.lock, lock_type);
1181         }
1182 }
1183
1184 noinline
1185 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
1186 {
1187         struct bch_fs *c = trans->c;
1188         struct btree_path_level *l = path_l(path);
1189         struct btree_node_iter node_iter = l->iter;
1190         struct bkey_packed *k;
1191         struct bkey_buf tmp;
1192         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1193                 ? (path->level > 1 ? 0 :  2)
1194                 : (path->level > 1 ? 1 : 16);
1195         bool was_locked = btree_node_locked(path, path->level);
1196         int ret = 0;
1197
1198         bch2_bkey_buf_init(&tmp);
1199
1200         while (nr && !ret) {
1201                 if (!bch2_btree_node_relock(trans, path, path->level))
1202                         break;
1203
1204                 bch2_btree_node_iter_advance(&node_iter, l->b);
1205                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
1206                 if (!k)
1207                         break;
1208
1209                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
1210                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
1211                                                path->level - 1);
1212         }
1213
1214         if (!was_locked)
1215                 btree_node_unlock(path, path->level);
1216
1217         bch2_bkey_buf_exit(&tmp, c);
1218         return ret;
1219 }
1220
1221 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
1222                                             struct btree_path *path,
1223                                             unsigned plevel, struct btree *b)
1224 {
1225         struct btree_path_level *l = &path->l[plevel];
1226         bool locked = btree_node_locked(path, plevel);
1227         struct bkey_packed *k;
1228         struct bch_btree_ptr_v2 *bp;
1229
1230         if (!bch2_btree_node_relock(trans, path, plevel))
1231                 return;
1232
1233         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1234         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
1235
1236         bp = (void *) bkeyp_val(&l->b->format, k);
1237         bp->mem_ptr = (unsigned long)b;
1238
1239         if (!locked)
1240                 btree_node_unlock(path, plevel);
1241 }
1242
1243 static __always_inline int btree_path_down(struct btree_trans *trans,
1244                                            struct btree_path *path,
1245                                            unsigned flags,
1246                                            unsigned long trace_ip)
1247 {
1248         struct bch_fs *c = trans->c;
1249         struct btree_path_level *l = path_l(path);
1250         struct btree *b;
1251         unsigned level = path->level - 1;
1252         enum six_lock_type lock_type = __btree_lock_want(path, level);
1253         struct bkey_buf tmp;
1254         int ret;
1255
1256         EBUG_ON(!btree_node_locked(path, path->level));
1257
1258         bch2_bkey_buf_init(&tmp);
1259         bch2_bkey_buf_unpack(&tmp, c, l->b,
1260                          bch2_btree_node_iter_peek(&l->iter, l->b));
1261
1262         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
1263         ret = PTR_ERR_OR_ZERO(b);
1264         if (unlikely(ret))
1265                 goto err;
1266
1267         mark_btree_node_locked(path, level, lock_type);
1268         btree_path_level_init(trans, path, b);
1269
1270         if (tmp.k->k.type == KEY_TYPE_btree_ptr_v2 &&
1271             unlikely(b != btree_node_mem_ptr(tmp.k)))
1272                 btree_node_mem_ptr_set(trans, path, level + 1, b);
1273
1274         if (flags & BTREE_ITER_PREFETCH)
1275                 ret = btree_path_prefetch(trans, path);
1276
1277         if (btree_node_read_locked(path, level + 1))
1278                 btree_node_unlock(path, level + 1);
1279         path->level = level;
1280
1281         bch2_btree_path_verify_locks(path);
1282 err:
1283         bch2_bkey_buf_exit(&tmp, c);
1284         return ret;
1285 }
1286
1287 static int btree_path_traverse_one(struct btree_trans *, struct btree_path *,
1288                                    unsigned, unsigned long);
1289
1290 static int __btree_path_traverse_all(struct btree_trans *trans, int ret,
1291                                      unsigned long trace_ip)
1292 {
1293         struct bch_fs *c = trans->c;
1294         struct btree_path *path;
1295         int i;
1296
1297         if (trans->in_traverse_all)
1298                 return -EINTR;
1299
1300         trans->in_traverse_all = true;
1301 retry_all:
1302         trans->restarted = false;
1303
1304         trans_for_each_path(trans, path)
1305                 path->should_be_locked = false;
1306
1307         btree_trans_verify_sorted(trans);
1308
1309         for (i = trans->nr_sorted - 2; i >= 0; --i) {
1310                 struct btree_path *path1 = trans->paths + trans->sorted[i];
1311                 struct btree_path *path2 = trans->paths + trans->sorted[i + 1];
1312
1313                 if (path1->btree_id == path2->btree_id &&
1314                     path1->locks_want < path2->locks_want)
1315                         __bch2_btree_path_upgrade(trans, path1, path2->locks_want);
1316                 else if (!path1->locks_want && path2->locks_want)
1317                         __bch2_btree_path_upgrade(trans, path1, 1);
1318         }
1319
1320         bch2_trans_unlock(trans);
1321         cond_resched();
1322
1323         if (unlikely(ret == -ENOMEM)) {
1324                 struct closure cl;
1325
1326                 closure_init_stack(&cl);
1327
1328                 do {
1329                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1330                         closure_sync(&cl);
1331                 } while (ret);
1332         }
1333
1334         if (unlikely(ret == -EIO)) {
1335                 trans->error = true;
1336                 goto out;
1337         }
1338
1339         BUG_ON(ret && ret != -EINTR);
1340
1341         /* Now, redo traversals in correct order: */
1342         i = 0;
1343         while (i < trans->nr_sorted) {
1344                 path = trans->paths + trans->sorted[i];
1345
1346                 EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
1347
1348                 ret = btree_path_traverse_one(trans, path, 0, _THIS_IP_);
1349                 if (ret)
1350                         goto retry_all;
1351
1352                 EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
1353
1354                 if (path->nodes_locked)
1355                         i++;
1356         }
1357
1358         /*
1359          * BTREE_ITER_NEED_RELOCK is ok here - if we called bch2_trans_unlock()
1360          * and relock(), relock() won't relock since path->should_be_locked
1361          * isn't set yet, which is all fine
1362          */
1363         trans_for_each_path(trans, path)
1364                 BUG_ON(path->uptodate >= BTREE_ITER_NEED_TRAVERSE);
1365 out:
1366         bch2_btree_cache_cannibalize_unlock(c);
1367
1368         trans->in_traverse_all = false;
1369
1370         trace_trans_traverse_all(trans->ip, trace_ip);
1371         return ret;
1372 }
1373
1374 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
1375 {
1376         return __btree_path_traverse_all(trans, 0, _RET_IP_);
1377 }
1378
1379 static inline bool btree_path_good_node(struct btree_trans *trans,
1380                                         struct btree_path *path,
1381                                         unsigned l, int check_pos)
1382 {
1383         if (!is_btree_node(path, l) ||
1384             !bch2_btree_node_relock(trans, path, l))
1385                 return false;
1386
1387         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1388                 return false;
1389         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1390                 return false;
1391         return true;
1392 }
1393
1394 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1395                                                      struct btree_path *path,
1396                                                      int check_pos)
1397 {
1398         unsigned i, l = path->level;
1399
1400         while (btree_path_node(path, l) &&
1401                !btree_path_good_node(trans, path, l, check_pos)) {
1402                 btree_node_unlock(path, l);
1403                 path->l[l].b = BTREE_ITER_NO_NODE_UP;
1404                 l++;
1405         }
1406
1407         /* If we need intent locks, take them too: */
1408         for (i = l + 1;
1409              i < path->locks_want && btree_path_node(path, i);
1410              i++)
1411                 if (!bch2_btree_node_relock(trans, path, i))
1412                         while (l <= i) {
1413                                 btree_node_unlock(path, l);
1414                                 path->l[l].b = BTREE_ITER_NO_NODE_UP;
1415                                 l++;
1416                         }
1417
1418         return l;
1419 }
1420
1421 /*
1422  * This is the main state machine for walking down the btree - walks down to a
1423  * specified depth
1424  *
1425  * Returns 0 on success, -EIO on error (error reading in a btree node).
1426  *
1427  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1428  * stashed in the iterator and returned from bch2_trans_exit().
1429  */
1430 static int btree_path_traverse_one(struct btree_trans *trans,
1431                                    struct btree_path *path,
1432                                    unsigned flags,
1433                                    unsigned long trace_ip)
1434 {
1435         unsigned depth_want = path->level;
1436         int ret = 0;
1437
1438         /*
1439          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1440          * and re-traverse the path without a transaction restart:
1441          */
1442         if (path->should_be_locked) {
1443                 ret = bch2_btree_path_relock(trans, path, trace_ip) ? 0 : -EINTR;
1444                 goto out;
1445         }
1446
1447         if (path->cached) {
1448                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1449                 goto out;
1450         }
1451
1452         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1453                 goto out;
1454
1455         path->level = btree_path_up_until_good_node(trans, path, 0);
1456
1457         /*
1458          * Note: path->nodes[path->level] may be temporarily NULL here - that
1459          * would indicate to other code that we got to the end of the btree,
1460          * here it indicates that relocking the root failed - it's critical that
1461          * btree_path_lock_root() comes next and that it can't fail
1462          */
1463         while (path->level > depth_want) {
1464                 ret = btree_path_node(path, path->level)
1465                         ? btree_path_down(trans, path, flags, trace_ip)
1466                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1467                 if (unlikely(ret)) {
1468                         if (ret == 1) {
1469                                 /*
1470                                  * No nodes at this level - got to the end of
1471                                  * the btree:
1472                                  */
1473                                 ret = 0;
1474                                 goto out;
1475                         }
1476
1477                         __bch2_btree_path_unlock(path);
1478                         path->level = depth_want;
1479
1480                         if (ret == -EIO)
1481                                 path->l[path->level].b =
1482                                         BTREE_ITER_NO_NODE_ERROR;
1483                         else
1484                                 path->l[path->level].b =
1485                                         BTREE_ITER_NO_NODE_DOWN;
1486                         goto out;
1487                 }
1488         }
1489
1490         path->uptodate = BTREE_ITER_UPTODATE;
1491 out:
1492         BUG_ON((ret == -EINTR) != !!trans->restarted);
1493         bch2_btree_path_verify(trans, path);
1494         return ret;
1495 }
1496
1497 static int __btree_path_traverse_all(struct btree_trans *, int, unsigned long);
1498
1499 int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
1500                                           struct btree_path *path, unsigned flags)
1501 {
1502         int ret;
1503
1504         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
1505                 return 0;
1506
1507         ret =   bch2_trans_cond_resched(trans) ?:
1508                 btree_path_traverse_one(trans, path, flags, _RET_IP_);
1509         if (unlikely(ret) && hweight64(trans->paths_allocated) == 1) {
1510                 ret = __btree_path_traverse_all(trans, ret, _RET_IP_);
1511                 BUG_ON(ret == -EINTR);
1512         }
1513
1514         return ret;
1515 }
1516
1517 static void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1518                             struct btree_path *src)
1519 {
1520         unsigned i;
1521
1522         memcpy(&dst->pos, &src->pos,
1523                sizeof(struct btree_path) - offsetof(struct btree_path, pos));
1524
1525         for (i = 0; i < BTREE_MAX_DEPTH; i++)
1526                 if (btree_node_locked(dst, i))
1527                         six_lock_increment(&dst->l[i].b->c.lock,
1528                                            __btree_lock_want(dst, i));
1529
1530         btree_path_check_sort(trans, dst, 0);
1531 }
1532
1533 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1534                                            bool intent)
1535 {
1536         struct btree_path *new = btree_path_alloc(trans, src);
1537
1538         btree_path_copy(trans, new, src);
1539         __btree_path_get(new, intent);
1540         return new;
1541 }
1542
1543 inline struct btree_path * __must_check
1544 bch2_btree_path_make_mut(struct btree_trans *trans,
1545                          struct btree_path *path, bool intent)
1546 {
1547         if (path->ref > 1 || path->preserve) {
1548                 __btree_path_put(path, intent);
1549                 path = btree_path_clone(trans, path, intent);
1550                 path->preserve = false;
1551 #ifdef CONFIG_BCACHEFS_DEBUG
1552                 path->ip_allocated = _RET_IP_;
1553 #endif
1554                 btree_trans_verify_sorted(trans);
1555         }
1556
1557         return path;
1558 }
1559
1560 static struct btree_path * __must_check
1561 btree_path_set_pos(struct btree_trans *trans,
1562                    struct btree_path *path, struct bpos new_pos,
1563                    bool intent)
1564 {
1565         int cmp = bpos_cmp(new_pos, path->pos);
1566         unsigned l = path->level;
1567
1568         EBUG_ON(trans->restarted);
1569         EBUG_ON(!path->ref);
1570
1571         if (!cmp)
1572                 return path;
1573
1574         path = bch2_btree_path_make_mut(trans, path, intent);
1575
1576         path->pos               = new_pos;
1577         path->should_be_locked  = false;
1578
1579         btree_path_check_sort(trans, path, cmp);
1580
1581         if (unlikely(path->cached)) {
1582                 btree_node_unlock(path, 0);
1583                 path->l[0].b = BTREE_ITER_NO_NODE_CACHED;
1584                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1585                 goto out;
1586         }
1587
1588         l = btree_path_up_until_good_node(trans, path, cmp);
1589
1590         if (btree_path_node(path, l)) {
1591                 /*
1592                  * We might have to skip over many keys, or just a few: try
1593                  * advancing the node iterator, and if we have to skip over too
1594                  * many keys just reinit it (or if we're rewinding, since that
1595                  * is expensive).
1596                  */
1597                 if (cmp < 0 ||
1598                     !btree_path_advance_to_pos(path, &path->l[l], 8))
1599                         __btree_path_level_init(path, l);
1600         }
1601
1602         if (l != path->level) {
1603                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1604                 __bch2_btree_path_unlock(path);
1605         }
1606 out:
1607         bch2_btree_path_verify(trans, path);
1608         return path;
1609 }
1610
1611 /* Btree path: main interface: */
1612
1613 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1614 {
1615         struct btree_path *next;
1616
1617         next = prev_btree_path(trans, path);
1618         if (next && !btree_path_cmp(next, path))
1619                 return next;
1620
1621         next = next_btree_path(trans, path);
1622         if (next && !btree_path_cmp(next, path))
1623                 return next;
1624
1625         return NULL;
1626 }
1627
1628 static bool have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1629 {
1630         struct btree_path *next;
1631
1632         next = prev_btree_path(trans, path);
1633         if (next && path_l(next)->b == path_l(path)->b)
1634                 return true;
1635
1636         next = next_btree_path(trans, path);
1637         if (next && path_l(next)->b == path_l(path)->b)
1638                 return true;
1639
1640         return false;
1641 }
1642
1643 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1644 {
1645         __bch2_btree_path_unlock(path);
1646         btree_path_list_remove(trans, path);
1647         trans->paths_allocated &= ~(1ULL << path->idx);
1648 }
1649
1650 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1651 {
1652         struct btree_path *dup;
1653
1654         EBUG_ON(trans->paths + path->idx != path);
1655         EBUG_ON(!path->ref);
1656
1657         if (!__btree_path_put(path, intent))
1658                 return;
1659
1660         /*
1661          * Perhaps instead we should check for duplicate paths in traverse_all:
1662          */
1663         if (path->preserve &&
1664             (dup = have_path_at_pos(trans, path))) {
1665                 dup->preserve = true;
1666                 path->preserve = false;
1667         }
1668
1669         if (!path->preserve &&
1670             have_node_at_pos(trans, path))
1671                 __bch2_path_free(trans, path);
1672 }
1673
1674 noinline __cold
1675 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1676 {
1677         struct btree_path *path;
1678         struct btree_insert_entry *i;
1679         unsigned idx;
1680         char buf[300];
1681
1682         btree_trans_verify_sorted(trans);
1683
1684         trans_for_each_path_inorder(trans, path, idx)
1685                 printk(KERN_ERR "path: idx %u ref %u:%u%s btree %s pos %s %pS\n",
1686                        path->idx, path->ref, path->intent_ref,
1687                        path->preserve ? " preserve" : "",
1688                        bch2_btree_ids[path->btree_id],
1689                        (bch2_bpos_to_text(&PBUF(buf), path->pos), buf),
1690 #ifdef CONFIG_BCACHEFS_DEBUG
1691                        (void *) path->ip_allocated
1692 #else
1693                        NULL
1694 #endif
1695                        );
1696
1697         trans_for_each_update(trans, i)
1698                 printk(KERN_ERR "update: btree %s %s %pS\n",
1699                        bch2_btree_ids[i->btree_id],
1700                        (bch2_bkey_val_to_text(&PBUF(buf), trans->c, bkey_i_to_s_c(i->k)), buf),
1701                        (void *) i->ip_allocated);
1702 }
1703
1704 static struct btree_path *btree_path_alloc(struct btree_trans *trans,
1705                                            struct btree_path *pos)
1706 {
1707         struct btree_path *path;
1708         unsigned idx;
1709
1710         if (unlikely(trans->paths_allocated ==
1711                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1)))) {
1712                 bch2_dump_trans_paths_updates(trans);
1713                 panic("trans path oveflow\n");
1714         }
1715
1716         idx = __ffs64(~trans->paths_allocated);
1717         trans->paths_allocated |= 1ULL << idx;
1718
1719         path = &trans->paths[idx];
1720
1721         path->idx               = idx;
1722         path->ref               = 0;
1723         path->intent_ref        = 0;
1724         path->nodes_locked      = 0;
1725         path->nodes_intent_locked = 0;
1726
1727         btree_path_list_add(trans, pos, path);
1728         return path;
1729 }
1730
1731 struct btree_path *bch2_path_get(struct btree_trans *trans, bool cached,
1732                                  enum btree_id btree_id, struct bpos pos,
1733                                  unsigned locks_want, unsigned level,
1734                                  bool intent)
1735 {
1736         struct btree_path *path, *path_pos = NULL;
1737         int i;
1738
1739         BUG_ON(trans->restarted);
1740
1741         trans_for_each_path_inorder(trans, path, i) {
1742                 if (__btree_path_cmp(path,
1743                                      btree_id,
1744                                      cached,
1745                                      pos,
1746                                      level) > 0)
1747                         break;
1748
1749                 path_pos = path;
1750         }
1751
1752         if (path_pos &&
1753             path_pos->cached    == cached &&
1754             path_pos->btree_id  == btree_id &&
1755             path_pos->level     == level) {
1756                 __btree_path_get(path_pos, intent);
1757                 path = btree_path_set_pos(trans, path_pos, pos, intent);
1758                 path->preserve = true;
1759         } else {
1760                 path = btree_path_alloc(trans, path_pos);
1761                 path_pos = NULL;
1762
1763                 __btree_path_get(path, intent);
1764                 path->pos                       = pos;
1765                 path->btree_id                  = btree_id;
1766                 path->cached                    = cached;
1767                 path->preserve                  = true;
1768                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1769                 path->should_be_locked          = false;
1770                 path->level                     = level;
1771                 path->locks_want                = locks_want;
1772                 path->nodes_locked              = 0;
1773                 path->nodes_intent_locked       = 0;
1774                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1775                         path->l[i].b            = BTREE_ITER_NO_NODE_INIT;
1776 #ifdef CONFIG_BCACHEFS_DEBUG
1777                 path->ip_allocated              = _RET_IP_;
1778 #endif
1779                 btree_trans_verify_sorted(trans);
1780         }
1781
1782         if (path->intent_ref)
1783                 locks_want = max(locks_want, level + 1);
1784
1785         /*
1786          * If the path has locks_want greater than requested, we don't downgrade
1787          * it here - on transaction restart because btree node split needs to
1788          * upgrade locks, we might be putting/getting the iterator again.
1789          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1790          * a successful transaction commit.
1791          */
1792
1793         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1794         if (locks_want > path->locks_want) {
1795                 path->locks_want = locks_want;
1796                 btree_path_get_locks(trans, path, true, _THIS_IP_);
1797         }
1798
1799         return path;
1800 }
1801
1802 inline struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1803 {
1804
1805         struct bkey_s_c k;
1806
1807         BUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1808
1809         if (!path->cached) {
1810                 struct btree_path_level *l = path_l(path);
1811                 struct bkey_packed *_k =
1812                         bch2_btree_node_iter_peek_all(&l->iter, l->b);
1813
1814                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1815
1816                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_cmp(k.k->p, path->pos) == 0);
1817
1818                 if (!k.k || bpos_cmp(path->pos, k.k->p))
1819                         goto hole;
1820         } else {
1821                 struct bkey_cached *ck = (void *) path->l[0].b;
1822
1823                 EBUG_ON(path->btree_id != ck->key.btree_id ||
1824                         bkey_cmp(path->pos, ck->key.pos));
1825
1826                 /* BTREE_ITER_CACHED_NOFILL? */
1827                 if (unlikely(!ck->valid))
1828                         goto hole;
1829
1830                 k = bkey_i_to_s_c(ck->k);
1831         }
1832
1833         return k;
1834 hole:
1835         bkey_init(u);
1836         u->p = path->pos;
1837         return (struct bkey_s_c) { u, NULL };
1838 }
1839
1840 /* Btree iterators: */
1841
1842 int __must_check
1843 __bch2_btree_iter_traverse(struct btree_iter *iter)
1844 {
1845         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1846 }
1847
1848 int __must_check
1849 bch2_btree_iter_traverse(struct btree_iter *iter)
1850 {
1851         int ret;
1852
1853         iter->path = btree_path_set_pos(iter->trans, iter->path,
1854                                         btree_iter_search_key(iter),
1855                                         iter->flags & BTREE_ITER_INTENT);
1856
1857         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1858         if (ret)
1859                 return ret;
1860
1861         iter->path->should_be_locked = true;
1862         return 0;
1863 }
1864
1865 /* Iterate across nodes (leaf and interior nodes) */
1866
1867 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1868 {
1869         struct btree *b = NULL;
1870         int ret;
1871
1872         EBUG_ON(iter->path->cached);
1873         bch2_btree_iter_verify(iter);
1874
1875         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1876         if (ret)
1877                 goto out;
1878
1879         b = btree_path_node(iter->path, iter->path->level);
1880         if (!b)
1881                 goto out;
1882
1883         BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0);
1884
1885         bkey_init(&iter->k);
1886         iter->k.p = iter->pos = b->key.k.p;
1887         iter->path->should_be_locked = true;
1888 out:
1889         bch2_btree_iter_verify_entry_exit(iter);
1890         bch2_btree_iter_verify(iter);
1891
1892         return b;
1893 }
1894
1895 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1896 {
1897         struct btree_trans *trans = iter->trans;
1898         struct btree_path *path = iter->path;
1899         struct btree *b = NULL;
1900         int ret;
1901
1902         EBUG_ON(iter->path->cached);
1903         bch2_btree_iter_verify(iter);
1904
1905         /* already got to end? */
1906         if (!btree_path_node(path, path->level))
1907                 goto out;
1908
1909         bch2_trans_cond_resched(trans);
1910
1911         btree_node_unlock(path, path->level);
1912         path->l[path->level].b = BTREE_ITER_NO_NODE_UP;
1913         path->level++;
1914
1915         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1916         ret = bch2_btree_path_traverse(trans, path, iter->flags);
1917         if (ret)
1918                 goto out;
1919
1920         /* got to end? */
1921         b = btree_path_node(path, path->level);
1922         if (!b)
1923                 goto out;
1924
1925         if (bpos_cmp(iter->pos, b->key.k.p) < 0) {
1926                 /*
1927                  * Haven't gotten to the end of the parent node: go back down to
1928                  * the next child node
1929                  */
1930                 path = iter->path =
1931                         btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1932                                            iter->flags & BTREE_ITER_INTENT);
1933
1934                 /* Unlock to avoid screwing up our lock invariants: */
1935                 btree_node_unlock(path, path->level);
1936
1937                 path->level = iter->min_depth;
1938                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1939                 bch2_btree_iter_verify(iter);
1940
1941                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1942                 if (ret) {
1943                         b = NULL;
1944                         goto out;
1945                 }
1946
1947                 b = path->l[path->level].b;
1948         }
1949
1950         bkey_init(&iter->k);
1951         iter->k.p = iter->pos = b->key.k.p;
1952         iter->path->should_be_locked = true;
1953 out:
1954         bch2_btree_iter_verify_entry_exit(iter);
1955         bch2_btree_iter_verify(iter);
1956
1957         return b;
1958 }
1959
1960 /* Iterate across keys (in leaf nodes only) */
1961
1962 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1963 {
1964         struct bpos pos = iter->k.p;
1965         bool ret = bpos_cmp(pos, SPOS_MAX) != 0;
1966
1967         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1968                 pos = bkey_successor(iter, pos);
1969         bch2_btree_iter_set_pos(iter, pos);
1970         return ret;
1971 }
1972
1973 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1974 {
1975         struct bpos pos = bkey_start_pos(&iter->k);
1976         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1977                     ? bpos_cmp(pos, POS_MIN)
1978                     : bkey_cmp(pos, POS_MIN)) != 0;
1979
1980         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1981                 pos = bkey_predecessor(iter, pos);
1982         bch2_btree_iter_set_pos(iter, pos);
1983         return ret;
1984 }
1985
1986 /**
1987  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
1988  * current position
1989  */
1990 struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
1991 {
1992         struct btree_trans *trans = iter->trans;
1993         struct bpos search_key = btree_iter_search_key(iter);
1994         struct bkey_i *next_update;
1995         struct bkey_s_c k;
1996         int ret, cmp;
1997
1998         EBUG_ON(iter->path->cached || iter->path->level);
1999         bch2_btree_iter_verify(iter);
2000         bch2_btree_iter_verify_entry_exit(iter);
2001
2002         while (1) {
2003                 iter->path = btree_path_set_pos(trans, iter->path, search_key,
2004                                    iter->flags & BTREE_ITER_INTENT);
2005
2006                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2007                 if (unlikely(ret)) {
2008                         /* ensure that iter->k is consistent with iter->pos: */
2009                         bch2_btree_iter_set_pos(iter, iter->pos);
2010                         k = bkey_s_c_err(ret);
2011                         goto out;
2012                 }
2013
2014                 next_update = iter->flags & BTREE_ITER_WITH_UPDATES
2015                         ? btree_trans_peek_updates(trans, iter->btree_id, search_key)
2016                         : NULL;
2017                 k = btree_path_level_peek_all(trans->c, &iter->path->l[0], &iter->k);
2018
2019                 /* * In the btree, deleted keys sort before non deleted: */
2020                 if (k.k && bkey_deleted(k.k) &&
2021                     (!next_update ||
2022                      bpos_cmp(k.k->p, next_update->k.p) <= 0)) {
2023                         search_key = k.k->p;
2024                         continue;
2025                 }
2026
2027                 if (next_update &&
2028                     bpos_cmp(next_update->k.p,
2029                              k.k ? k.k->p : iter->path->l[0].b->key.k.p) <= 0) {
2030                         iter->k = next_update->k;
2031                         k = bkey_i_to_s_c(next_update);
2032                 }
2033
2034                 if (likely(k.k)) {
2035                         /*
2036                          * We can never have a key in a leaf node at POS_MAX, so
2037                          * we don't have to check these successor() calls:
2038                          */
2039                         if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2040                             !bch2_snapshot_is_ancestor(trans->c,
2041                                                        iter->snapshot,
2042                                                        k.k->p.snapshot)) {
2043                                 search_key = bpos_successor(k.k->p);
2044                                 continue;
2045                         }
2046
2047                         if (bkey_whiteout(k.k) &&
2048                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2049                                 search_key = bkey_successor(iter, k.k->p);
2050                                 continue;
2051                         }
2052
2053                         break;
2054                 } else if (likely(bpos_cmp(iter->path->l[0].b->key.k.p, SPOS_MAX))) {
2055                         /* Advance to next leaf node: */
2056                         search_key = bpos_successor(iter->path->l[0].b->key.k.p);
2057                 } else {
2058                         /* End of btree: */
2059                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2060                         k = bkey_s_c_null;
2061                         goto out;
2062                 }
2063         }
2064
2065         /*
2066          * iter->pos should be mononotically increasing, and always be equal to
2067          * the key we just returned - except extents can straddle iter->pos:
2068          */
2069         if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2070                 iter->pos = k.k->p;
2071         else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
2072                 iter->pos = bkey_start_pos(k.k);
2073
2074         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2075                 iter->pos.snapshot = iter->snapshot;
2076
2077         cmp = bpos_cmp(k.k->p, iter->path->pos);
2078         if (cmp) {
2079                 iter->path = bch2_btree_path_make_mut(trans, iter->path,
2080                                         iter->flags & BTREE_ITER_INTENT);
2081                 iter->path->pos = k.k->p;
2082                 btree_path_check_sort(trans, iter->path, cmp);
2083         }
2084 out:
2085         iter->path->should_be_locked = true;
2086
2087         bch2_btree_iter_verify_entry_exit(iter);
2088         bch2_btree_iter_verify(iter);
2089         ret = bch2_btree_iter_verify_ret(iter, k);
2090         if (unlikely(ret))
2091                 return bkey_s_c_err(ret);
2092
2093         return k;
2094 }
2095
2096 /**
2097  * bch2_btree_iter_next: returns first key greater than iterator's current
2098  * position
2099  */
2100 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2101 {
2102         if (!bch2_btree_iter_advance(iter))
2103                 return bkey_s_c_null;
2104
2105         return bch2_btree_iter_peek(iter);
2106 }
2107
2108 /**
2109  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2110  * iterator's current position
2111  */
2112 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2113 {
2114         struct btree_trans *trans = iter->trans;
2115         struct bpos search_key = iter->pos;
2116         struct btree_path *saved_path = NULL;
2117         struct bkey_s_c k;
2118         struct bkey saved_k;
2119         const struct bch_val *saved_v;
2120         int ret;
2121
2122         EBUG_ON(iter->path->cached || iter->path->level);
2123         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2124         bch2_btree_iter_verify(iter);
2125         bch2_btree_iter_verify_entry_exit(iter);
2126
2127         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2128                 search_key.snapshot = U32_MAX;
2129
2130         while (1) {
2131                 iter->path = btree_path_set_pos(trans, iter->path, search_key,
2132                                                 iter->flags & BTREE_ITER_INTENT);
2133
2134                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2135                 if (unlikely(ret)) {
2136                         /* ensure that iter->k is consistent with iter->pos: */
2137                         bch2_btree_iter_set_pos(iter, iter->pos);
2138                         k = bkey_s_c_err(ret);
2139                         goto out;
2140                 }
2141
2142                 k = btree_path_level_peek(trans->c, iter->path,
2143                                           &iter->path->l[0], &iter->k);
2144                 if (!k.k ||
2145                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2146                      ? bpos_cmp(bkey_start_pos(k.k), search_key) >= 0
2147                      : bpos_cmp(k.k->p, search_key) > 0))
2148                         k = btree_path_level_prev(trans->c, iter->path,
2149                                                   &iter->path->l[0], &iter->k);
2150
2151                 btree_path_check_sort(trans, iter->path, 0);
2152
2153                 if (likely(k.k)) {
2154                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2155                                 if (k.k->p.snapshot == iter->snapshot)
2156                                         goto got_key;
2157
2158                                 /*
2159                                  * If we have a saved candidate, and we're no
2160                                  * longer at the same _key_ (not pos), return
2161                                  * that candidate
2162                                  */
2163                                 if (saved_path && bkey_cmp(k.k->p, saved_k.p)) {
2164                                         bch2_path_put(trans, iter->path,
2165                                                       iter->flags & BTREE_ITER_INTENT);
2166                                         iter->path = saved_path;
2167                                         saved_path = NULL;
2168                                         iter->k = saved_k;
2169                                         k.v     = saved_v;
2170                                         goto got_key;
2171                                 }
2172
2173                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2174                                                               iter->snapshot,
2175                                                               k.k->p.snapshot)) {
2176                                         if (saved_path)
2177                                                 bch2_path_put(trans, saved_path,
2178                                                       iter->flags & BTREE_ITER_INTENT);
2179                                         saved_path = btree_path_clone(trans, iter->path,
2180                                                                 iter->flags & BTREE_ITER_INTENT);
2181                                         saved_k = *k.k;
2182                                         saved_v = k.v;
2183                                 }
2184
2185                                 search_key = bpos_predecessor(k.k->p);
2186                                 continue;
2187                         }
2188 got_key:
2189                         if (bkey_whiteout(k.k) &&
2190                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2191                                 search_key = bkey_predecessor(iter, k.k->p);
2192                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2193                                         search_key.snapshot = U32_MAX;
2194                                 continue;
2195                         }
2196
2197                         break;
2198                 } else if (likely(bpos_cmp(iter->path->l[0].b->data->min_key, POS_MIN))) {
2199                         /* Advance to previous leaf node: */
2200                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2201                 } else {
2202                         /* Start of btree: */
2203                         bch2_btree_iter_set_pos(iter, POS_MIN);
2204                         k = bkey_s_c_null;
2205                         goto out;
2206                 }
2207         }
2208
2209         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
2210
2211         /* Extents can straddle iter->pos: */
2212         if (bkey_cmp(k.k->p, iter->pos) < 0)
2213                 iter->pos = k.k->p;
2214
2215         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2216                 iter->pos.snapshot = iter->snapshot;
2217 out:
2218         if (saved_path)
2219                 bch2_path_put(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2220         iter->path->should_be_locked = true;
2221
2222         bch2_btree_iter_verify_entry_exit(iter);
2223         bch2_btree_iter_verify(iter);
2224
2225         return k;
2226 }
2227
2228 /**
2229  * bch2_btree_iter_prev: returns first key less than iterator's current
2230  * position
2231  */
2232 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2233 {
2234         if (!bch2_btree_iter_rewind(iter))
2235                 return bkey_s_c_null;
2236
2237         return bch2_btree_iter_peek_prev(iter);
2238 }
2239
2240 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2241 {
2242         struct btree_trans *trans = iter->trans;
2243         struct bpos search_key;
2244         struct bkey_s_c k;
2245         int ret;
2246
2247         EBUG_ON(iter->path->level);
2248         bch2_btree_iter_verify(iter);
2249         bch2_btree_iter_verify_entry_exit(iter);
2250
2251         /* extents can't span inode numbers: */
2252         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2253             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2254                 if (iter->pos.inode == KEY_INODE_MAX)
2255                         return bkey_s_c_null;
2256
2257                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2258         }
2259
2260         search_key = btree_iter_search_key(iter);
2261         iter->path = btree_path_set_pos(trans, iter->path, search_key,
2262                                         iter->flags & BTREE_ITER_INTENT);
2263
2264         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2265         if (unlikely(ret))
2266                 return bkey_s_c_err(ret);
2267
2268         if ((iter->flags & BTREE_ITER_CACHED) ||
2269             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2270                 struct bkey_i *next_update;
2271
2272                 next_update = iter->flags & BTREE_ITER_WITH_UPDATES
2273                         ? btree_trans_peek_updates(trans, iter->btree_id, search_key)
2274                         : NULL;
2275
2276                 if (next_update &&
2277                     !bpos_cmp(next_update->k.p, iter->pos)) {
2278                         iter->k = next_update->k;
2279                         k = bkey_i_to_s_c(next_update);
2280                 } else {
2281                         k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2282                 }
2283         } else {
2284                 struct bpos next;
2285
2286                 if (iter->flags & BTREE_ITER_INTENT) {
2287                         struct btree_iter iter2;
2288
2289                         bch2_trans_copy_iter(&iter2, iter);
2290                         k = bch2_btree_iter_peek(&iter2);
2291
2292                         if (k.k && !bkey_err(k)) {
2293                                 iter->k = iter2.k;
2294                                 k.k = &iter->k;
2295                         }
2296                         bch2_trans_iter_exit(trans, &iter2);
2297                 } else {
2298                         struct bpos pos = iter->pos;
2299
2300                         k = bch2_btree_iter_peek(iter);
2301                         iter->pos = pos;
2302                 }
2303
2304                 if (unlikely(bkey_err(k)))
2305                         return k;
2306
2307                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2308
2309                 if (bkey_cmp(iter->pos, next) < 0) {
2310                         bkey_init(&iter->k);
2311                         iter->k.p = iter->pos;
2312                         bch2_key_resize(&iter->k,
2313                                         min_t(u64, KEY_SIZE_MAX,
2314                                               (next.inode == iter->pos.inode
2315                                                ? next.offset
2316                                                : KEY_OFFSET_MAX) -
2317                                               iter->pos.offset));
2318
2319                         k = (struct bkey_s_c) { &iter->k, NULL };
2320                         EBUG_ON(!k.k->size);
2321                 }
2322         }
2323
2324         iter->path->should_be_locked = true;
2325
2326         bch2_btree_iter_verify_entry_exit(iter);
2327         bch2_btree_iter_verify(iter);
2328         ret = bch2_btree_iter_verify_ret(iter, k);
2329         if (unlikely(ret))
2330                 return bkey_s_c_err(ret);
2331
2332         return k;
2333 }
2334
2335 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2336 {
2337         if (!bch2_btree_iter_advance(iter))
2338                 return bkey_s_c_null;
2339
2340         return bch2_btree_iter_peek_slot(iter);
2341 }
2342
2343 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2344 {
2345         if (!bch2_btree_iter_rewind(iter))
2346                 return bkey_s_c_null;
2347
2348         return bch2_btree_iter_peek_slot(iter);
2349 }
2350
2351 /* new transactional stuff: */
2352
2353 static inline void btree_path_verify_sorted_ref(struct btree_trans *trans,
2354                                                 struct btree_path *path)
2355 {
2356         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2357         EBUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2358         EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
2359 }
2360
2361 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2362 {
2363 #ifdef CONFIG_BCACHEFS_DEBUG
2364         unsigned i;
2365
2366         for (i = 0; i < trans->nr_sorted; i++)
2367                 btree_path_verify_sorted_ref(trans, trans->paths + trans->sorted[i]);
2368 #endif
2369 }
2370
2371 static void btree_trans_verify_sorted(struct btree_trans *trans)
2372 {
2373 #ifdef CONFIG_BCACHEFS_DEBUG
2374         struct btree_path *path, *prev = NULL;
2375         unsigned i;
2376
2377         trans_for_each_path_inorder(trans, path, i) {
2378                 BUG_ON(prev && btree_path_cmp(prev, path) > 0);
2379                 prev = path;
2380         }
2381 #endif
2382 }
2383
2384 static inline void btree_path_swap(struct btree_trans *trans,
2385                                    struct btree_path *l, struct btree_path *r)
2386 {
2387         swap(l->sorted_idx, r->sorted_idx);
2388         swap(trans->sorted[l->sorted_idx],
2389              trans->sorted[r->sorted_idx]);
2390
2391         btree_path_verify_sorted_ref(trans, l);
2392         btree_path_verify_sorted_ref(trans, r);
2393 }
2394
2395 static void btree_path_check_sort(struct btree_trans *trans, struct btree_path *path,
2396                                   int cmp)
2397 {
2398         struct btree_path *n;
2399
2400         if (cmp <= 0) {
2401                 n = prev_btree_path(trans, path);
2402                 if (n && btree_path_cmp(n, path) > 0) {
2403                         do {
2404                                 btree_path_swap(trans, n, path);
2405                                 n = prev_btree_path(trans, path);
2406                         } while (n && btree_path_cmp(n, path) > 0);
2407
2408                         goto out;
2409                 }
2410         }
2411
2412         if (cmp >= 0) {
2413                 n = next_btree_path(trans, path);
2414                 if (n && btree_path_cmp(path, n) > 0) {
2415                         do {
2416                                 btree_path_swap(trans, path, n);
2417                                 n = next_btree_path(trans, path);
2418                         } while (n && btree_path_cmp(path, n) > 0);
2419                 }
2420         }
2421 out:
2422         btree_trans_verify_sorted(trans);
2423 }
2424
2425 static inline void btree_path_list_remove(struct btree_trans *trans,
2426                                           struct btree_path *path)
2427 {
2428         unsigned i;
2429
2430         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2431
2432         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2433
2434         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2435                 trans->paths[trans->sorted[i]].sorted_idx = i;
2436
2437         path->sorted_idx = U8_MAX;
2438
2439         btree_trans_verify_sorted_refs(trans);
2440 }
2441
2442 static inline void btree_path_list_add(struct btree_trans *trans,
2443                                        struct btree_path *pos,
2444                                        struct btree_path *path)
2445 {
2446         unsigned i;
2447
2448         btree_trans_verify_sorted_refs(trans);
2449
2450         path->sorted_idx = pos ? pos->sorted_idx + 1 : 0;
2451
2452         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2453
2454         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2455                 trans->paths[trans->sorted[i]].sorted_idx = i;
2456
2457         btree_trans_verify_sorted_refs(trans);
2458 }
2459
2460 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2461 {
2462         if (iter->path)
2463                 bch2_path_put(trans, iter->path,
2464                               iter->flags & BTREE_ITER_INTENT);
2465         iter->path = NULL;
2466 }
2467
2468 static void __bch2_trans_iter_init(struct btree_trans *trans,
2469                                    struct btree_iter *iter,
2470                                    unsigned btree_id, struct bpos pos,
2471                                    unsigned locks_want,
2472                                    unsigned depth,
2473                                    unsigned flags)
2474 {
2475         EBUG_ON(trans->restarted);
2476
2477         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
2478             btree_node_type_is_extents(btree_id))
2479                 flags |= BTREE_ITER_IS_EXTENTS;
2480
2481         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
2482             !btree_type_has_snapshots(btree_id))
2483                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
2484
2485         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
2486             btree_type_has_snapshots(btree_id))
2487                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
2488
2489         iter->trans     = trans;
2490         iter->path      = NULL;
2491         iter->btree_id  = btree_id;
2492         iter->min_depth = depth;
2493         iter->flags     = flags;
2494         iter->snapshot  = pos.snapshot;
2495         iter->pos       = pos;
2496         iter->k.type    = KEY_TYPE_deleted;
2497         iter->k.p       = pos;
2498         iter->k.size    = 0;
2499
2500         iter->path = bch2_path_get(trans,
2501                                    flags & BTREE_ITER_CACHED,
2502                                    btree_id,
2503                                    iter->pos,
2504                                    locks_want,
2505                                    depth,
2506                                    flags & BTREE_ITER_INTENT);
2507 }
2508
2509 void bch2_trans_iter_init(struct btree_trans *trans,
2510                           struct btree_iter *iter,
2511                           unsigned btree_id, struct bpos pos,
2512                           unsigned flags)
2513 {
2514         __bch2_trans_iter_init(trans, iter, btree_id, pos,
2515                                0, 0, flags);
2516 }
2517
2518 void bch2_trans_node_iter_init(struct btree_trans *trans,
2519                                struct btree_iter *iter,
2520                                enum btree_id btree_id,
2521                                struct bpos pos,
2522                                unsigned locks_want,
2523                                unsigned depth,
2524                                unsigned flags)
2525 {
2526         __bch2_trans_iter_init(trans, iter, btree_id, pos, locks_want, depth,
2527                                BTREE_ITER_NOT_EXTENTS|
2528                                __BTREE_ITER_ALL_SNAPSHOTS|
2529                                BTREE_ITER_ALL_SNAPSHOTS|
2530                                flags);
2531         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2532         BUG_ON(iter->path->level        != depth);
2533         BUG_ON(iter->min_depth          != depth);
2534 }
2535
2536 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2537 {
2538         *dst = *src;
2539         if (src->path)
2540                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2541 }
2542
2543 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2544 {
2545         size_t new_top = trans->mem_top + size;
2546         void *p;
2547
2548         if (new_top > trans->mem_bytes) {
2549                 size_t old_bytes = trans->mem_bytes;
2550                 size_t new_bytes = roundup_pow_of_two(new_top);
2551                 void *new_mem;
2552
2553                 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2554
2555                 new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2556                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2557                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2558                         new_bytes = BTREE_TRANS_MEM_MAX;
2559                         kfree(trans->mem);
2560                 }
2561
2562                 if (!new_mem)
2563                         return ERR_PTR(-ENOMEM);
2564
2565                 trans->mem = new_mem;
2566                 trans->mem_bytes = new_bytes;
2567
2568                 if (old_bytes) {
2569                         trace_trans_restart_mem_realloced(trans->ip, _RET_IP_, new_bytes);
2570                         btree_trans_restart(trans);
2571                         return ERR_PTR(-EINTR);
2572                 }
2573         }
2574
2575         p = trans->mem + trans->mem_top;
2576         trans->mem_top += size;
2577         memset(p, 0, size);
2578         return p;
2579 }
2580
2581 /**
2582  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2583  * @trans: transaction to reset
2584  *
2585  * While iterating over nodes or updating nodes a attempt to lock a btree
2586  * node may return EINTR when the trylock fails. When this occurs
2587  * bch2_trans_begin() should be called and the transaction retried.
2588  */
2589 void bch2_trans_begin(struct btree_trans *trans)
2590 {
2591         struct btree_insert_entry *i;
2592         struct btree_path *path;
2593
2594         trans_for_each_update(trans, i)
2595                 __btree_path_put(i->path, true);
2596
2597         trans->extra_journal_res        = 0;
2598         trans->nr_updates               = 0;
2599         trans->mem_top                  = 0;
2600
2601         trans->hooks                    = NULL;
2602         trans->extra_journal_entries    = NULL;
2603         trans->extra_journal_entry_u64s = 0;
2604
2605         if (trans->fs_usage_deltas) {
2606                 trans->fs_usage_deltas->used = 0;
2607                 memset(&trans->fs_usage_deltas->memset_start, 0,
2608                        (void *) &trans->fs_usage_deltas->memset_end -
2609                        (void *) &trans->fs_usage_deltas->memset_start);
2610         }
2611
2612         trans_for_each_path(trans, path) {
2613                 path->should_be_locked = false;
2614
2615                 /*
2616                  * XXX: we probably shouldn't be doing this if the transaction
2617                  * was restarted, but currently we still overflow transaction
2618                  * iterators if we do that
2619                  */
2620                 if (!path->ref && !path->preserve)
2621                         __bch2_path_free(trans, path);
2622                 else if (!path->ref)
2623                         path->preserve = false;
2624         }
2625
2626         bch2_trans_cond_resched(trans);
2627
2628         if (trans->restarted)
2629                 bch2_btree_path_traverse_all(trans);
2630
2631         trans->restarted = false;
2632 }
2633
2634 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
2635 {
2636         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
2637         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2638         void *p = NULL;
2639
2640         BUG_ON(trans->used_mempool);
2641
2642 #ifdef __KERNEL__
2643         p = this_cpu_xchg(c->btree_paths_bufs->path , NULL);
2644 #endif
2645         if (!p)
2646                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
2647
2648         trans->paths            = p; p += paths_bytes;
2649         trans->updates          = p; p += updates_bytes;
2650 }
2651
2652 void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
2653                      unsigned expected_nr_iters,
2654                      size_t expected_mem_bytes)
2655         __acquires(&c->btree_trans_barrier)
2656 {
2657         memset(trans, 0, sizeof(*trans));
2658         trans->c                = c;
2659         trans->ip               = _RET_IP_;
2660
2661         bch2_trans_alloc_paths(trans, c);
2662
2663         if (expected_mem_bytes) {
2664                 trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes);
2665                 trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL);
2666
2667                 if (!unlikely(trans->mem)) {
2668                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2669                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2670                 }
2671         }
2672
2673         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2674
2675 #ifdef CONFIG_BCACHEFS_DEBUG
2676         trans->pid = current->pid;
2677         mutex_lock(&c->btree_trans_lock);
2678         list_add(&trans->list, &c->btree_trans_list);
2679         mutex_unlock(&c->btree_trans_lock);
2680 #endif
2681 }
2682
2683 static void check_btree_paths_leaked(struct btree_trans *trans)
2684 {
2685 #ifdef CONFIG_BCACHEFS_DEBUG
2686         struct bch_fs *c = trans->c;
2687         struct btree_path *path;
2688
2689         trans_for_each_path(trans, path)
2690                 if (path->ref)
2691                         goto leaked;
2692         return;
2693 leaked:
2694         bch_err(c, "btree paths leaked from %pS!", (void *) trans->ip);
2695         trans_for_each_path(trans, path)
2696                 if (path->ref)
2697                         printk(KERN_ERR "  btree %s %pS\n",
2698                                bch2_btree_ids[path->btree_id],
2699                                (void *) path->ip_allocated);
2700         /* Be noisy about this: */
2701         bch2_fatal_error(c);
2702 #endif
2703 }
2704
2705 int bch2_trans_exit(struct btree_trans *trans)
2706         __releases(&c->btree_trans_barrier)
2707 {
2708         struct btree_insert_entry *i;
2709         struct bch_fs *c = trans->c;
2710
2711         bch2_trans_unlock(trans);
2712
2713         trans_for_each_update(trans, i)
2714                 __btree_path_put(i->path, true);
2715         trans->nr_updates               = 0;
2716
2717         check_btree_paths_leaked(trans);
2718
2719 #ifdef CONFIG_BCACHEFS_DEBUG
2720         mutex_lock(&c->btree_trans_lock);
2721         list_del(&trans->list);
2722         mutex_unlock(&c->btree_trans_lock);
2723 #endif
2724
2725         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2726
2727         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
2728
2729         if (trans->fs_usage_deltas) {
2730                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
2731                     REPLICAS_DELTA_LIST_MAX)
2732                         mempool_free(trans->fs_usage_deltas,
2733                                      &c->replicas_delta_pool);
2734                 else
2735                         kfree(trans->fs_usage_deltas);
2736         }
2737
2738         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
2739                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
2740         else
2741                 kfree(trans->mem);
2742
2743 #ifdef __KERNEL__
2744         /*
2745          * Userspace doesn't have a real percpu implementation:
2746          */
2747         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
2748 #endif
2749
2750         if (trans->paths)
2751                 mempool_free(trans->paths, &c->btree_paths_pool);
2752
2753         trans->mem      = (void *) 0x1;
2754         trans->paths    = (void *) 0x1;
2755
2756         return trans->error ? -EIO : 0;
2757 }
2758
2759 static void __maybe_unused
2760 bch2_btree_path_node_to_text(struct printbuf *out,
2761                              struct btree_bkey_cached_common *_b,
2762                              bool cached)
2763 {
2764         pr_buf(out, "    l=%u %s:",
2765                _b->level, bch2_btree_ids[_b->btree_id]);
2766         bch2_bpos_to_text(out, btree_node_pos(_b, cached));
2767 }
2768
2769 #ifdef CONFIG_BCACHEFS_DEBUG
2770 static bool trans_has_locks(struct btree_trans *trans)
2771 {
2772         struct btree_path *path;
2773
2774         trans_for_each_path(trans, path)
2775                 if (path->nodes_locked)
2776                         return true;
2777         return false;
2778 }
2779 #endif
2780
2781 void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
2782 {
2783 #ifdef CONFIG_BCACHEFS_DEBUG
2784         struct btree_trans *trans;
2785         struct btree_path *path;
2786         struct btree *b;
2787         unsigned l;
2788
2789         mutex_lock(&c->btree_trans_lock);
2790         list_for_each_entry(trans, &c->btree_trans_list, list) {
2791                 if (!trans_has_locks(trans))
2792                         continue;
2793
2794                 pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip);
2795
2796                 trans_for_each_path(trans, path) {
2797                         if (!path->nodes_locked)
2798                                 continue;
2799
2800                         pr_buf(out, "  path %u %c l=%u %s:",
2801                                path->idx,
2802                                path->cached ? 'c' : 'b',
2803                                path->level,
2804                                bch2_btree_ids[path->btree_id]);
2805                         bch2_bpos_to_text(out, path->pos);
2806                         pr_buf(out, "\n");
2807
2808                         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
2809                                 if (btree_node_locked(path, l)) {
2810                                         pr_buf(out, "    %s l=%u ",
2811                                                btree_node_intent_locked(path, l) ? "i" : "r", l);
2812                                         bch2_btree_path_node_to_text(out,
2813                                                         (void *) path->l[l].b,
2814                                                         path->cached);
2815                                         pr_buf(out, "\n");
2816                                 }
2817                         }
2818                 }
2819
2820                 b = READ_ONCE(trans->locking);
2821                 if (b) {
2822                         path = &trans->paths[trans->locking_path_idx];
2823                         pr_buf(out, "  locking path %u %c l=%u %s:",
2824                                trans->locking_path_idx,
2825                                path->cached ? 'c' : 'b',
2826                                trans->locking_level,
2827                                bch2_btree_ids[trans->locking_btree_id]);
2828                         bch2_bpos_to_text(out, trans->locking_pos);
2829
2830                         pr_buf(out, " node ");
2831                         bch2_btree_path_node_to_text(out,
2832                                         (void *) b, path->cached);
2833                         pr_buf(out, "\n");
2834                 }
2835         }
2836         mutex_unlock(&c->btree_trans_lock);
2837 #endif
2838 }
2839
2840 void bch2_fs_btree_iter_exit(struct bch_fs *c)
2841 {
2842         mempool_exit(&c->btree_trans_mem_pool);
2843         mempool_exit(&c->btree_paths_pool);
2844         cleanup_srcu_struct(&c->btree_trans_barrier);
2845 }
2846
2847 int bch2_fs_btree_iter_init(struct bch_fs *c)
2848 {
2849         unsigned nr = BTREE_ITER_MAX;
2850
2851         INIT_LIST_HEAD(&c->btree_trans_list);
2852         mutex_init(&c->btree_trans_lock);
2853
2854         return  init_srcu_struct(&c->btree_trans_barrier) ?:
2855                 mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
2856                         sizeof(struct btree_path) * nr +
2857                         sizeof(struct btree_insert_entry) * nr) ?:
2858                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
2859                                           BTREE_TRANS_MEM_MAX);
2860 }