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 "recovery.h"
  16 #include "replicas.h"
  17 #include "subvolume.h"
  18
  19 #include <linux/prandom.h>
  20 #include <linux/prefetch.h>
  21 #include <trace/events/bcachefs.h>
  22
  23 static void btree_trans_verify_sorted(struct btree_trans *);
  24 inline void bch2_btree_path_check_sort(struct btree_trans *, struct btree_path *, int);
  25
  26 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
  27 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
  28                                        struct btree_path *);
  29
  30 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
  31 {
  32 #ifdef CONFIG_BCACHEFS_DEBUG
  33         return iter->ip_allocated;
  34 #else
  35         return 0;
  36 #endif
  37 }
  38
  39 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
  40
  41 /*
  42  * Unlocks before scheduling
  43  * Note: does not revalidate iterator
  44  */
  45 static inline int bch2_trans_cond_resched(struct btree_trans *trans)
  46 {
  47         if (need_resched() || race_fault()) {
  48                 bch2_trans_unlock(trans);
  49                 schedule();
  50                 return bch2_trans_relock(trans);
  51         } else {
  52                 return 0;
  53         }
  54 }
  55
  56 static inline int __btree_path_cmp(const struct btree_path *l,
  57                                    enum btree_id        r_btree_id,
  58                                    bool                 r_cached,
  59                                    struct bpos          r_pos,
  60                                    unsigned             r_level)
  61 {
  62         /*
  63          * Must match lock ordering as defined by __bch2_btree_node_lock:
  64          */
  65         return   cmp_int(l->btree_id,   r_btree_id) ?:
  66                  cmp_int((int) l->cached,       (int) r_cached) ?:
  67                  bpos_cmp(l->pos,       r_pos) ?:
  68                 -cmp_int(l->level,      r_level);
  69 }
  70
  71 static inline int btree_path_cmp(const struct btree_path *l,
  72                                  const struct btree_path *r)
  73 {
  74         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  75 }
  76
  77 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  78 {
  79         /* Are we iterating over keys in all snapshots? */
  80         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  81                 p = bpos_successor(p);
  82         } else {
  83                 p = bpos_nosnap_successor(p);
  84                 p.snapshot = iter->snapshot;
  85         }
  86
  87         return p;
  88 }
  89
  90 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  91 {
  92         /* Are we iterating over keys in all snapshots? */
  93         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  94                 p = bpos_predecessor(p);
  95         } else {
  96                 p = bpos_nosnap_predecessor(p);
  97                 p.snapshot = iter->snapshot;
  98         }
  99
 100         return p;
 101 }
 102
 103 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
 104 {
 105         struct bpos pos = iter->pos;
 106
 107         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 108             bkey_cmp(pos, POS_MAX))
 109                 pos = bkey_successor(iter, pos);
 110         return pos;
 111 }
 112
 113 static inline bool btree_path_pos_before_node(struct btree_path *path,
 114                                               struct btree *b)
 115 {
 116         return bpos_cmp(path->pos, b->data->min_key) < 0;
 117 }
 118
 119 static inline bool btree_path_pos_after_node(struct btree_path *path,
 120                                              struct btree *b)
 121 {
 122         return bpos_cmp(b->key.k.p, path->pos) < 0;
 123 }
 124
 125 static inline bool btree_path_pos_in_node(struct btree_path *path,
 126                                           struct btree *b)
 127 {
 128         return path->btree_id == b->c.btree_id &&
 129                 !btree_path_pos_before_node(path, b) &&
 130                 !btree_path_pos_after_node(path, b);
 131 }
 132
 133 /* Btree node locking: */
 134
 135 void bch2_btree_node_unlock_write(struct btree_trans *trans,
 136                         struct btree_path *path, struct btree *b)
 137 {
 138         bch2_btree_node_unlock_write_inlined(trans, path, b);
 139 }
 140
 141 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *trans,
 142                                                   struct btree_path *skip,
 143                                                   struct btree *b,
 144                                                   unsigned level)
 145 {
 146         struct btree_path *path;
 147         struct six_lock_count ret = { 0, 0 };
 148
 149         if (IS_ERR_OR_NULL(b))
 150                 return ret;
 151
 152         trans_for_each_path(trans, path)
 153                 if (path != skip && path->l[level].b == b) {
 154                         ret.read += btree_node_read_locked(path, level);
 155                         ret.intent += btree_node_intent_locked(path, level);
 156                 }
 157
 158         return ret;
 159 }
 160
 161 static inline void six_lock_readers_add(struct six_lock *lock, int nr)
 162 {
 163         if (!lock->readers)
 164                 atomic64_add(__SIX_VAL(read_lock, nr), &lock->state.counter);
 165         else
 166                 this_cpu_add(*lock->readers, nr);
 167 }
 168
 169 void __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree *b)
 170 {
 171         int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->c.level).read;
 172
 173         /*
 174          * Must drop our read locks before calling six_lock_write() -
 175          * six_unlock() won't do wakeups until the reader count
 176          * goes to 0, and it's safe because we have the node intent
 177          * locked:
 178          */
 179         six_lock_readers_add(&b->c.lock, -readers);
 180         six_lock_write(&b->c.lock, NULL, NULL);
 181         six_lock_readers_add(&b->c.lock, readers);
 182 }
 183
 184 bool __bch2_btree_node_relock(struct btree_trans *trans,
 185                               struct btree_path *path, unsigned level)
 186 {
 187         struct btree *b = btree_path_node(path, level);
 188         int want = __btree_lock_want(path, level);
 189
 190         if (!is_btree_node(path, level))
 191                 goto fail;
 192
 193         if (race_fault())
 194                 goto fail;
 195
 196         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 197             (btree_node_lock_seq_matches(path, b, level) &&
 198              btree_node_lock_increment(trans, b, level, want))) {
 199                 mark_btree_node_locked(trans, path, level, want);
 200                 return true;
 201         }
 202 fail:
 203         if (b != ERR_PTR(-BCH_ERR_no_btree_node_cached) &&
 204             b != ERR_PTR(-BCH_ERR_no_btree_node_init))
 205                 trace_btree_node_relock_fail(trans, _RET_IP_, path, level);
 206         return false;
 207 }
 208
 209 bool bch2_btree_node_upgrade(struct btree_trans *trans,
 210                              struct btree_path *path, unsigned level)
 211 {
 212         struct btree *b = path->l[level].b;
 213
 214         if (!is_btree_node(path, level))
 215                 return false;
 216
 217         switch (btree_lock_want(path, level)) {
 218         case BTREE_NODE_UNLOCKED:
 219                 BUG_ON(btree_node_locked(path, level));
 220                 return true;
 221         case BTREE_NODE_READ_LOCKED:
 222                 BUG_ON(btree_node_intent_locked(path, level));
 223                 return bch2_btree_node_relock(trans, path, level);
 224         case BTREE_NODE_INTENT_LOCKED:
 225                 break;
 226         }
 227
 228         if (btree_node_intent_locked(path, level))
 229                 return true;
 230
 231         if (race_fault())
 232                 return false;
 233
 234         if (btree_node_locked(path, level)
 235             ? six_lock_tryupgrade(&b->c.lock)
 236             : six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 237                 goto success;
 238
 239         if (btree_node_lock_seq_matches(path, b, level) &&
 240             btree_node_lock_increment(trans, b, level, BTREE_NODE_INTENT_LOCKED)) {
 241                 btree_node_unlock(trans, path, level);
 242                 goto success;
 243         }
 244
 245         trace_btree_node_upgrade_fail(trans, _RET_IP_, path, level);
 246         return false;
 247 success:
 248         mark_btree_node_intent_locked(trans, path, level);
 249         return true;
 250 }
 251
 252 static inline bool btree_path_get_locks(struct btree_trans *trans,
 253                                         struct btree_path *path,
 254                                         bool upgrade)
 255 {
 256         unsigned l = path->level;
 257         int fail_idx = -1;
 258
 259         do {
 260                 if (!btree_path_node(path, l))
 261                         break;
 262
 263                 if (!(upgrade
 264                       ? bch2_btree_node_upgrade(trans, path, l)
 265                       : bch2_btree_node_relock(trans, path, l)))
 266                         fail_idx = l;
 267
 268                 l++;
 269         } while (l < path->locks_want);
 270
 271         /*
 272          * When we fail to get a lock, we have to ensure that any child nodes
 273          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 274          * the node that we failed to relock:
 275          */
 276         if (fail_idx >= 0) {
 277                 __bch2_btree_path_unlock(trans, path);
 278                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 279
 280                 do {
 281                         path->l[fail_idx].b = upgrade
 282                                 ? ERR_PTR(-BCH_ERR_no_btree_node_upgrade)
 283                                 : ERR_PTR(-BCH_ERR_no_btree_node_relock);
 284                         --fail_idx;
 285                 } while (fail_idx >= 0);
 286         }
 287
 288         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 289                 path->uptodate = BTREE_ITER_UPTODATE;
 290
 291         bch2_trans_verify_locks(trans);
 292
 293         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 294 }
 295
 296 static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b,
 297                                   bool cached)
 298 {
 299         return !cached
 300                 ? container_of(_b, struct btree, c)->key.k.p
 301                 : container_of(_b, struct bkey_cached, c)->key.pos;
 302 }
 303
 304 /* Slowpath: */
 305 int __bch2_btree_node_lock(struct btree_trans *trans,
 306                            struct btree_path *path,
 307                            struct btree *b,
 308                            struct bpos pos, unsigned level,
 309                            enum six_lock_type type,
 310                            six_lock_should_sleep_fn should_sleep_fn, void *p,
 311                            unsigned long ip)
 312 {
 313         struct btree_path *linked;
 314         unsigned reason;
 315
 316         /* Check if it's safe to block: */
 317         trans_for_each_path(trans, linked) {
 318                 if (!linked->nodes_locked)
 319                         continue;
 320
 321                 /*
 322                  * Can't block taking an intent lock if we have _any_ nodes read
 323                  * locked:
 324                  *
 325                  * - Our read lock blocks another thread with an intent lock on
 326                  *   the same node from getting a write lock, and thus from
 327                  *   dropping its intent lock
 328                  *
 329                  * - And the other thread may have multiple nodes intent locked:
 330                  *   both the node we want to intent lock, and the node we
 331                  *   already have read locked - deadlock:
 332                  */
 333                 if (type == SIX_LOCK_intent &&
 334                     linked->nodes_locked != linked->nodes_intent_locked) {
 335                         reason = 1;
 336                         goto deadlock;
 337                 }
 338
 339                 if (linked->btree_id != path->btree_id) {
 340                         if (linked->btree_id < path->btree_id)
 341                                 continue;
 342
 343                         reason = 3;
 344                         goto deadlock;
 345                 }
 346
 347                 /*
 348                  * Within the same btree, non-cached paths come before cached
 349                  * paths:
 350                  */
 351                 if (linked->cached != path->cached) {
 352                         if (!linked->cached)
 353                                 continue;
 354
 355                         reason = 4;
 356                         goto deadlock;
 357                 }
 358
 359                 /*
 360                  * Interior nodes must be locked before their descendants: if
 361                  * another path has possible descendants locked of the node
 362                  * we're about to lock, it must have the ancestors locked too:
 363                  */
 364                 if (level > __fls(linked->nodes_locked)) {
 365                         reason = 5;
 366                         goto deadlock;
 367                 }
 368
 369                 /* Must lock btree nodes in key order: */
 370                 if (btree_node_locked(linked, level) &&
 371                     bpos_cmp(pos, btree_node_pos((void *) linked->l[level].b,
 372                                                  linked->cached)) <= 0) {
 373                         reason = 7;
 374                         goto deadlock;
 375                 }
 376         }
 377
 378         return btree_node_lock_type(trans, path, b, pos, level,
 379                                     type, should_sleep_fn, p);
 380 deadlock:
 381         trace_trans_restart_would_deadlock(trans, ip, reason, linked, path, &pos);
 382         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
 383 }
 384
 385 /* Btree iterator locking: */
 386
 387 #ifdef CONFIG_BCACHEFS_DEBUG
 388
 389 static void bch2_btree_path_verify_locks(struct btree_path *path)
 390 {
 391         unsigned l;
 392
 393         if (!path->nodes_locked) {
 394                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
 395                        btree_path_node(path, path->level));
 396                 return;
 397         }
 398
 399         for (l = 0; btree_path_node(path, l); l++)
 400                 BUG_ON(btree_lock_want(path, l) !=
 401                        btree_node_locked_type(path, l));
 402 }
 403
 404 void bch2_trans_verify_locks(struct btree_trans *trans)
 405 {
 406         struct btree_path *path;
 407
 408         trans_for_each_path(trans, path)
 409                 bch2_btree_path_verify_locks(path);
 410 }
 411 #else
 412 static inline void bch2_btree_path_verify_locks(struct btree_path *path) {}
 413 #endif
 414
 415 /* Btree path locking: */
 416
 417 /*
 418  * Only for btree_cache.c - only relocks intent locks
 419  */
 420 int bch2_btree_path_relock_intent(struct btree_trans *trans,
 421                                   struct btree_path *path)
 422 {
 423         unsigned l;
 424
 425         for (l = path->level;
 426              l < path->locks_want && btree_path_node(path, l);
 427              l++) {
 428                 if (!bch2_btree_node_relock(trans, path, l)) {
 429                         __bch2_btree_path_unlock(trans, path);
 430                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 431                         trace_trans_restart_relock_path_intent(trans, _RET_IP_, path);
 432                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path_intent);
 433                 }
 434         }
 435
 436         return 0;
 437 }
 438
 439 __flatten
 440 static bool bch2_btree_path_relock_norestart(struct btree_trans *trans,
 441                         struct btree_path *path, unsigned long trace_ip)
 442 {
 443         return btree_path_get_locks(trans, path, false);
 444 }
 445
 446 static int bch2_btree_path_relock(struct btree_trans *trans,
 447                         struct btree_path *path, unsigned long trace_ip)
 448 {
 449         if (!bch2_btree_path_relock_norestart(trans, path, trace_ip)) {
 450                 trace_trans_restart_relock_path(trans, trace_ip, path);
 451                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
 452         }
 453
 454         return 0;
 455 }
 456
 457 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 458                                struct btree_path *path,
 459                                unsigned new_locks_want)
 460 {
 461         struct btree_path *linked;
 462
 463         EBUG_ON(path->locks_want >= new_locks_want);
 464
 465         path->locks_want = new_locks_want;
 466
 467         if (btree_path_get_locks(trans, path, true))
 468                 return true;
 469
 470         /*
 471          * XXX: this is ugly - we'd prefer to not be mucking with other
 472          * iterators in the btree_trans here.
 473          *
 474          * On failure to upgrade the iterator, setting iter->locks_want and
 475          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 476          * get the locks we want on transaction restart.
 477          *
 478          * But if this iterator was a clone, on transaction restart what we did
 479          * to this iterator isn't going to be preserved.
 480          *
 481          * Possibly we could add an iterator field for the parent iterator when
 482          * an iterator is a copy - for now, we'll just upgrade any other
 483          * iterators with the same btree id.
 484          *
 485          * The code below used to be needed to ensure ancestor nodes get locked
 486          * before interior nodes - now that's handled by
 487          * bch2_btree_path_traverse_all().
 488          */
 489         if (!path->cached && !trans->in_traverse_all)
 490                 trans_for_each_path(trans, linked)
 491                         if (linked != path &&
 492                             linked->cached == path->cached &&
 493                             linked->btree_id == path->btree_id &&
 494                             linked->locks_want < new_locks_want) {
 495                                 linked->locks_want = new_locks_want;
 496                                 btree_path_get_locks(trans, linked, true);
 497                         }
 498
 499         return false;
 500 }
 501
 502 void __bch2_btree_path_downgrade(struct btree_trans *trans,
 503                                  struct btree_path *path,
 504                                  unsigned new_locks_want)
 505 {
 506         unsigned l;
 507
 508         EBUG_ON(path->locks_want < new_locks_want);
 509
 510         path->locks_want = new_locks_want;
 511
 512         while (path->nodes_locked &&
 513                (l = __fls(path->nodes_locked)) >= path->locks_want) {
 514                 if (l > path->level) {
 515                         btree_node_unlock(trans, path, l);
 516                 } else {
 517                         if (btree_node_intent_locked(path, l)) {
 518                                 six_lock_downgrade(&path->l[l].b->c.lock);
 519                                 path->nodes_intent_locked ^= 1 << l;
 520                         }
 521                         break;
 522                 }
 523         }
 524
 525         bch2_btree_path_verify_locks(path);
 526 }
 527
 528 void bch2_trans_downgrade(struct btree_trans *trans)
 529 {
 530         struct btree_path *path;
 531
 532         trans_for_each_path(trans, path)
 533                 bch2_btree_path_downgrade(trans, path);
 534 }
 535
 536 /* Btree transaction locking: */
 537
 538 int bch2_trans_relock(struct btree_trans *trans)
 539 {
 540         struct btree_path *path;
 541
 542         if (unlikely(trans->restarted))
 543                 return -BCH_ERR_transaction_restart_relock;
 544
 545         trans_for_each_path(trans, path)
 546                 if (path->should_be_locked &&
 547                     bch2_btree_path_relock(trans, path, _RET_IP_)) {
 548                         trace_trans_restart_relock(trans, _RET_IP_, path);
 549                         BUG_ON(!trans->restarted);
 550                         return -BCH_ERR_transaction_restart_relock;
 551                 }
 552         return 0;
 553 }
 554
 555 void bch2_trans_unlock(struct btree_trans *trans)
 556 {
 557         struct btree_path *path;
 558
 559         trans_for_each_path(trans, path)
 560                 __bch2_btree_path_unlock(trans, path);
 561
 562         /*
 563          * bch2_gc_btree_init_recurse() doesn't use btree iterators for walking
 564          * btree nodes, it implements its own walking:
 565          */
 566         BUG_ON(!trans->is_initial_gc &&
 567                lock_class_is_held(&bch2_btree_node_lock_key));
 568 }
 569
 570 /* Btree iterator: */
 571
 572 #ifdef CONFIG_BCACHEFS_DEBUG
 573
 574 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 575                                           struct btree_path *path)
 576 {
 577         struct bkey_cached *ck;
 578         bool locked = btree_node_locked(path, 0);
 579
 580         if (!bch2_btree_node_relock(trans, path, 0))
 581                 return;
 582
 583         ck = (void *) path->l[0].b;
 584         BUG_ON(ck->key.btree_id != path->btree_id ||
 585                bkey_cmp(ck->key.pos, path->pos));
 586
 587         if (!locked)
 588                 btree_node_unlock(trans, path, 0);
 589 }
 590
 591 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 592                                 struct btree_path *path, unsigned level)
 593 {
 594         struct btree_path_level *l;
 595         struct btree_node_iter tmp;
 596         bool locked;
 597         struct bkey_packed *p, *k;
 598         struct printbuf buf1 = PRINTBUF;
 599         struct printbuf buf2 = PRINTBUF;
 600         struct printbuf buf3 = PRINTBUF;
 601         const char *msg;
 602
 603         if (!bch2_debug_check_iterators)
 604                 return;
 605
 606         l       = &path->l[level];
 607         tmp     = l->iter;
 608         locked  = btree_node_locked(path, level);
 609
 610         if (path->cached) {
 611                 if (!level)
 612                         bch2_btree_path_verify_cached(trans, path);
 613                 return;
 614         }
 615
 616         if (!btree_path_node(path, level))
 617                 return;
 618
 619         if (!bch2_btree_node_relock(trans, path, level))
 620                 return;
 621
 622         BUG_ON(!btree_path_pos_in_node(path, l->b));
 623
 624         bch2_btree_node_iter_verify(&l->iter, l->b);
 625
 626         /*
 627          * For interior nodes, the iterator will have skipped past deleted keys:
 628          */
 629         p = level
 630                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 631                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 632         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 633
 634         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 635                 msg = "before";
 636                 goto err;
 637         }
 638
 639         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 640                 msg = "after";
 641                 goto err;
 642         }
 643
 644         if (!locked)
 645                 btree_node_unlock(trans, path, level);
 646         return;
 647 err:
 648         bch2_bpos_to_text(&buf1, path->pos);
 649
 650         if (p) {
 651                 struct bkey uk = bkey_unpack_key(l->b, p);
 652                 bch2_bkey_to_text(&buf2, &uk);
 653         } else {
 654                 prt_printf(&buf2, "(none)");
 655         }
 656
 657         if (k) {
 658                 struct bkey uk = bkey_unpack_key(l->b, k);
 659                 bch2_bkey_to_text(&buf3, &uk);
 660         } else {
 661                 prt_printf(&buf3, "(none)");
 662         }
 663
 664         panic("path should be %s key at level %u:\n"
 665               "path pos %s\n"
 666               "prev key %s\n"
 667               "cur  key %s\n",
 668               msg, level, buf1.buf, buf2.buf, buf3.buf);
 669 }
 670
 671 static void bch2_btree_path_verify(struct btree_trans *trans,
 672                                    struct btree_path *path)
 673 {
 674         struct bch_fs *c = trans->c;
 675         unsigned i;
 676
 677         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 678
 679         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 680                 if (!path->l[i].b) {
 681                         BUG_ON(!path->cached &&
 682                                c->btree_roots[path->btree_id].b->c.level > i);
 683                         break;
 684                 }
 685
 686                 bch2_btree_path_verify_level(trans, path, i);
 687         }
 688
 689         bch2_btree_path_verify_locks(path);
 690 }
 691
 692 void bch2_trans_verify_paths(struct btree_trans *trans)
 693 {
 694         struct btree_path *path;
 695
 696         trans_for_each_path(trans, path)
 697                 bch2_btree_path_verify(trans, path);
 698 }
 699
 700 static void bch2_btree_iter_verify(struct btree_iter *iter)
 701 {
 702         struct btree_trans *trans = iter->trans;
 703
 704         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 705
 706         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
 707
 708         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 709                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 710
 711         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 712                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 713                !btree_type_has_snapshots(iter->btree_id));
 714
 715         if (iter->update_path)
 716                 bch2_btree_path_verify(trans, iter->update_path);
 717         bch2_btree_path_verify(trans, iter->path);
 718 }
 719
 720 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 721 {
 722         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 723                !iter->pos.snapshot);
 724
 725         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 726                iter->pos.snapshot != iter->snapshot);
 727
 728         BUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
 729                bkey_cmp(iter->pos, iter->k.p) > 0);
 730 }
 731
 732 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 733 {
 734         struct btree_trans *trans = iter->trans;
 735         struct btree_iter copy;
 736         struct bkey_s_c prev;
 737         int ret = 0;
 738
 739         if (!bch2_debug_check_iterators)
 740                 return 0;
 741
 742         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 743                 return 0;
 744
 745         if (bkey_err(k) || !k.k)
 746                 return 0;
 747
 748         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 749                                           iter->snapshot,
 750                                           k.k->p.snapshot));
 751
 752         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 753                              BTREE_ITER_NOPRESERVE|
 754                              BTREE_ITER_ALL_SNAPSHOTS);
 755         prev = bch2_btree_iter_prev(&copy);
 756         if (!prev.k)
 757                 goto out;
 758
 759         ret = bkey_err(prev);
 760         if (ret)
 761                 goto out;
 762
 763         if (!bkey_cmp(prev.k->p, k.k->p) &&
 764             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 765                                       prev.k->p.snapshot) > 0) {
 766                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 767
 768                 bch2_bkey_to_text(&buf1, k.k);
 769                 bch2_bkey_to_text(&buf2, prev.k);
 770
 771                 panic("iter snap %u\n"
 772                       "k    %s\n"
 773                       "prev %s\n",
 774                       iter->snapshot,
 775                       buf1.buf, buf2.buf);
 776         }
 777 out:
 778         bch2_trans_iter_exit(trans, &copy);
 779         return ret;
 780 }
 781
 782 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 783                             struct bpos pos, bool key_cache)
 784 {
 785         struct btree_path *path;
 786         unsigned idx;
 787         struct printbuf buf = PRINTBUF;
 788
 789         trans_for_each_path_inorder(trans, path, idx) {
 790                 int cmp = cmp_int(path->btree_id, id) ?:
 791                         cmp_int(path->cached, key_cache);
 792
 793                 if (cmp > 0)
 794                         break;
 795                 if (cmp < 0)
 796                         continue;
 797
 798                 if (!(path->nodes_locked & 1) ||
 799                     !path->should_be_locked)
 800                         continue;
 801
 802                 if (!key_cache) {
 803                         if (bkey_cmp(pos, path->l[0].b->data->min_key) >= 0 &&
 804                             bkey_cmp(pos, path->l[0].b->key.k.p) <= 0)
 805                                 return;
 806                 } else {
 807                         if (!bkey_cmp(pos, path->pos))
 808                                 return;
 809                 }
 810         }
 811
 812         bch2_dump_trans_paths_updates(trans);
 813         bch2_bpos_to_text(&buf, pos);
 814
 815         panic("not locked: %s %s%s\n",
 816               bch2_btree_ids[id], buf.buf,
 817               key_cache ? " cached" : "");
 818 }
 819
 820 #else
 821
 822 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 823                                                 struct btree_path *path, unsigned l) {}
 824 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 825                                           struct btree_path *path) {}
 826 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 827 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 828 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 829
 830 #endif
 831
 832 /* Btree path: fixups after btree updates */
 833
 834 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 835                                         struct btree *b,
 836                                         struct bset_tree *t,
 837                                         struct bkey_packed *k)
 838 {
 839         struct btree_node_iter_set *set;
 840
 841         btree_node_iter_for_each(iter, set)
 842                 if (set->end == t->end_offset) {
 843                         set->k = __btree_node_key_to_offset(b, k);
 844                         bch2_btree_node_iter_sort(iter, b);
 845                         return;
 846                 }
 847
 848         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 849 }
 850
 851 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 852                                                struct btree *b,
 853                                                struct bkey_packed *where)
 854 {
 855         struct btree_path_level *l = &path->l[b->c.level];
 856
 857         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 858                 return;
 859
 860         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 861                 bch2_btree_node_iter_advance(&l->iter, l->b);
 862 }
 863
 864 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 865                                       struct btree *b,
 866                                       struct bkey_packed *where)
 867 {
 868         struct btree_path *path;
 869
 870         trans_for_each_path_with_node(trans, b, path) {
 871                 __bch2_btree_path_fix_key_modified(path, b, where);
 872                 bch2_btree_path_verify_level(trans, path, b->c.level);
 873         }
 874 }
 875
 876 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 877                                        struct btree *b,
 878                                        struct btree_node_iter *node_iter,
 879                                        struct bset_tree *t,
 880                                        struct bkey_packed *where,
 881                                        unsigned clobber_u64s,
 882                                        unsigned new_u64s)
 883 {
 884         const struct bkey_packed *end = btree_bkey_last(b, t);
 885         struct btree_node_iter_set *set;
 886         unsigned offset = __btree_node_key_to_offset(b, where);
 887         int shift = new_u64s - clobber_u64s;
 888         unsigned old_end = t->end_offset - shift;
 889         unsigned orig_iter_pos = node_iter->data[0].k;
 890         bool iter_current_key_modified =
 891                 orig_iter_pos >= offset &&
 892                 orig_iter_pos <= offset + clobber_u64s;
 893
 894         btree_node_iter_for_each(node_iter, set)
 895                 if (set->end == old_end)
 896                         goto found;
 897
 898         /* didn't find the bset in the iterator - might have to readd it: */
 899         if (new_u64s &&
 900             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 901                 bch2_btree_node_iter_push(node_iter, b, where, end);
 902                 goto fixup_done;
 903         } else {
 904                 /* Iterator is after key that changed */
 905                 return;
 906         }
 907 found:
 908         set->end = t->end_offset;
 909
 910         /* Iterator hasn't gotten to the key that changed yet: */
 911         if (set->k < offset)
 912                 return;
 913
 914         if (new_u64s &&
 915             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 916                 set->k = offset;
 917         } else if (set->k < offset + clobber_u64s) {
 918                 set->k = offset + new_u64s;
 919                 if (set->k == set->end)
 920                         bch2_btree_node_iter_set_drop(node_iter, set);
 921         } else {
 922                 /* Iterator is after key that changed */
 923                 set->k = (int) set->k + shift;
 924                 return;
 925         }
 926
 927         bch2_btree_node_iter_sort(node_iter, b);
 928 fixup_done:
 929         if (node_iter->data[0].k != orig_iter_pos)
 930                 iter_current_key_modified = true;
 931
 932         /*
 933          * When a new key is added, and the node iterator now points to that
 934          * key, the iterator might have skipped past deleted keys that should
 935          * come after the key the iterator now points to. We have to rewind to
 936          * before those deleted keys - otherwise
 937          * bch2_btree_node_iter_prev_all() breaks:
 938          */
 939         if (!bch2_btree_node_iter_end(node_iter) &&
 940             iter_current_key_modified &&
 941             b->c.level) {
 942                 struct bset_tree *t;
 943                 struct bkey_packed *k, *k2, *p;
 944
 945                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 946
 947                 for_each_bset(b, t) {
 948                         bool set_pos = false;
 949
 950                         if (node_iter->data[0].end == t->end_offset)
 951                                 continue;
 952
 953                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 954
 955                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 956                                bkey_iter_cmp(b, k, p) < 0) {
 957                                 k2 = p;
 958                                 set_pos = true;
 959                         }
 960
 961                         if (set_pos)
 962                                 btree_node_iter_set_set_pos(node_iter,
 963                                                             b, t, k2);
 964                 }
 965         }
 966 }
 967
 968 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 969                               struct btree_path *path,
 970                               struct btree *b,
 971                               struct btree_node_iter *node_iter,
 972                               struct bkey_packed *where,
 973                               unsigned clobber_u64s,
 974                               unsigned new_u64s)
 975 {
 976         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 977         struct btree_path *linked;
 978
 979         if (node_iter != &path->l[b->c.level].iter) {
 980                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 981                                            where, clobber_u64s, new_u64s);
 982
 983                 if (bch2_debug_check_iterators)
 984                         bch2_btree_node_iter_verify(node_iter, b);
 985         }
 986
 987         trans_for_each_path_with_node(trans, b, linked) {
 988                 __bch2_btree_node_iter_fix(linked, b,
 989                                            &linked->l[b->c.level].iter, t,
 990                                            where, clobber_u64s, new_u64s);
 991                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 992         }
 993 }
 994
 995 /* Btree path level: pointer to a particular btree node and node iter */
 996
 997 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
 998                                                   struct btree_path_level *l,
 999                                                   struct bkey *u,
1000                                                   struct bkey_packed *k)
1001 {
1002         if (unlikely(!k)) {
1003                 /*
1004                  * signal to bch2_btree_iter_peek_slot() that we're currently at
1005                  * a hole
1006                  */
1007                 u->type = KEY_TYPE_deleted;
1008                 return bkey_s_c_null;
1009         }
1010
1011         return bkey_disassemble(l->b, k, u);
1012 }
1013
1014 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
1015                                                         struct btree_path_level *l,
1016                                                         struct bkey *u)
1017 {
1018         return __btree_iter_unpack(c, l, u,
1019                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
1020 }
1021
1022 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
1023                                                     struct btree_path *path,
1024                                                     struct btree_path_level *l,
1025                                                     struct bkey *u)
1026 {
1027         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
1028                         bch2_btree_node_iter_peek(&l->iter, l->b));
1029
1030         path->pos = k.k ? k.k->p : l->b->key.k.p;
1031         bch2_btree_path_verify_level(trans, path, l - path->l);
1032         return k;
1033 }
1034
1035 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
1036                                                     struct btree_path *path,
1037                                                     struct btree_path_level *l,
1038                                                     struct bkey *u)
1039 {
1040         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
1041                         bch2_btree_node_iter_prev(&l->iter, l->b));
1042
1043         path->pos = k.k ? k.k->p : l->b->data->min_key;
1044         bch2_btree_path_verify_level(trans, path, l - path->l);
1045         return k;
1046 }
1047
1048 static inline bool btree_path_advance_to_pos(struct btree_path *path,
1049                                              struct btree_path_level *l,
1050                                              int max_advance)
1051 {
1052         struct bkey_packed *k;
1053         int nr_advanced = 0;
1054
1055         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
1056                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
1057                 if (max_advance > 0 && nr_advanced >= max_advance)
1058                         return false;
1059
1060                 bch2_btree_node_iter_advance(&l->iter, l->b);
1061                 nr_advanced++;
1062         }
1063
1064         return true;
1065 }
1066
1067 /*
1068  * Verify that iterator for parent node points to child node:
1069  */
1070 static void btree_path_verify_new_node(struct btree_trans *trans,
1071                                        struct btree_path *path, struct btree *b)
1072 {
1073         struct bch_fs *c = trans->c;
1074         struct btree_path_level *l;
1075         unsigned plevel;
1076         bool parent_locked;
1077         struct bkey_packed *k;
1078
1079         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
1080                 return;
1081
1082         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
1083                 return;
1084
1085         plevel = b->c.level + 1;
1086         if (!btree_path_node(path, plevel))
1087                 return;
1088
1089         parent_locked = btree_node_locked(path, plevel);
1090
1091         if (!bch2_btree_node_relock(trans, path, plevel))
1092                 return;
1093
1094         l = &path->l[plevel];
1095         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1096         if (!k ||
1097             bkey_deleted(k) ||
1098             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
1099                 struct printbuf buf1 = PRINTBUF;
1100                 struct printbuf buf2 = PRINTBUF;
1101                 struct printbuf buf3 = PRINTBUF;
1102                 struct printbuf buf4 = PRINTBUF;
1103                 struct bkey uk = bkey_unpack_key(b, k);
1104
1105                 bch2_dump_btree_node(c, l->b);
1106                 bch2_bpos_to_text(&buf1, path->pos);
1107                 bch2_bkey_to_text(&buf2, &uk);
1108                 bch2_bpos_to_text(&buf3, b->data->min_key);
1109                 bch2_bpos_to_text(&buf3, b->data->max_key);
1110                 panic("parent iter doesn't point to new node:\n"
1111                       "iter pos %s %s\n"
1112                       "iter key %s\n"
1113                       "new node %s-%s\n",
1114                       bch2_btree_ids[path->btree_id],
1115                       buf1.buf, buf2.buf, buf3.buf, buf4.buf);
1116         }
1117
1118         if (!parent_locked)
1119                 btree_node_unlock(trans, path, plevel);
1120 }
1121
1122 static inline void __btree_path_level_init(struct btree_path *path,
1123                                            unsigned level)
1124 {
1125         struct btree_path_level *l = &path->l[level];
1126
1127         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1128
1129         /*
1130          * Iterators to interior nodes should always be pointed at the first non
1131          * whiteout:
1132          */
1133         if (level)
1134                 bch2_btree_node_iter_peek(&l->iter, l->b);
1135 }
1136
1137 static inline void btree_path_level_init(struct btree_trans *trans,
1138                                          struct btree_path *path,
1139                                          struct btree *b)
1140 {
1141         BUG_ON(path->cached);
1142
1143         btree_path_verify_new_node(trans, path, b);
1144
1145         EBUG_ON(!btree_path_pos_in_node(path, b));
1146         EBUG_ON(b->c.lock.state.seq & 1);
1147
1148         path->l[b->c.level].lock_seq = b->c.lock.state.seq;
1149         path->l[b->c.level].b = b;
1150         __btree_path_level_init(path, b->c.level);
1151 }
1152
1153 /* Btree path: fixups after btree node updates: */
1154
1155 /*
1156  * A btree node is being replaced - update the iterator to point to the new
1157  * node:
1158  */
1159 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
1160 {
1161         struct btree_path *path;
1162
1163         trans_for_each_path(trans, path)
1164                 if (!path->cached &&
1165                     btree_path_pos_in_node(path, b)) {
1166                         enum btree_node_locked_type t =
1167                                 btree_lock_want(path, b->c.level);
1168
1169                         if (path->nodes_locked &&
1170                             t != BTREE_NODE_UNLOCKED) {
1171                                 btree_node_unlock(trans, path, b->c.level);
1172                                 six_lock_increment(&b->c.lock, t);
1173                                 mark_btree_node_locked(trans, path, b->c.level, t);
1174                         }
1175
1176                         btree_path_level_init(trans, path, b);
1177                 }
1178 }
1179
1180 /*
1181  * A btree node has been modified in such a way as to invalidate iterators - fix
1182  * them:
1183  */
1184 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
1185 {
1186         struct btree_path *path;
1187
1188         trans_for_each_path_with_node(trans, b, path)
1189                 __btree_path_level_init(path, b->c.level);
1190 }
1191
1192 /* Btree path: traverse, set_pos: */
1193
1194 static int lock_root_check_fn(struct six_lock *lock, void *p)
1195 {
1196         struct btree *b = container_of(lock, struct btree, c.lock);
1197         struct btree **rootp = p;
1198
1199         if (b != *rootp)
1200                 return BCH_ERR_lock_fail_root_changed;
1201         return 0;
1202 }
1203
1204 static inline int btree_path_lock_root(struct btree_trans *trans,
1205                                        struct btree_path *path,
1206                                        unsigned depth_want,
1207                                        unsigned long trace_ip)
1208 {
1209         struct bch_fs *c = trans->c;
1210         struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
1211         enum six_lock_type lock_type;
1212         unsigned i;
1213         int ret;
1214
1215         EBUG_ON(path->nodes_locked);
1216
1217         while (1) {
1218                 b = READ_ONCE(*rootp);
1219                 path->level = READ_ONCE(b->c.level);
1220
1221                 if (unlikely(path->level < depth_want)) {
1222                         /*
1223                          * the root is at a lower depth than the depth we want:
1224                          * got to the end of the btree, or we're walking nodes
1225                          * greater than some depth and there are no nodes >=
1226                          * that depth
1227                          */
1228                         path->level = depth_want;
1229                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
1230                                 path->l[i].b = NULL;
1231                         return 1;
1232                 }
1233
1234                 lock_type = __btree_lock_want(path, path->level);
1235                 ret = btree_node_lock(trans, path, b, SPOS_MAX,
1236                                       path->level, lock_type,
1237                                       lock_root_check_fn, rootp,
1238                                       trace_ip);
1239                 if (unlikely(ret)) {
1240                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
1241                                 continue;
1242                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1243                                 return ret;
1244                         BUG();
1245                 }
1246
1247                 if (likely(b == READ_ONCE(*rootp) &&
1248                            b->c.level == path->level &&
1249                            !race_fault())) {
1250                         for (i = 0; i < path->level; i++)
1251                                 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
1252                         path->l[path->level].b = b;
1253                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
1254                                 path->l[i].b = NULL;
1255
1256                         mark_btree_node_locked(trans, path, path->level, lock_type);
1257                         btree_path_level_init(trans, path, b);
1258                         return 0;
1259                 }
1260
1261                 six_unlock_type(&b->c.lock, lock_type);
1262         }
1263 }
1264
1265 noinline
1266 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
1267 {
1268         struct bch_fs *c = trans->c;
1269         struct btree_path_level *l = path_l(path);
1270         struct btree_node_iter node_iter = l->iter;
1271         struct bkey_packed *k;
1272         struct bkey_buf tmp;
1273         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1274                 ? (path->level > 1 ? 0 :  2)
1275                 : (path->level > 1 ? 1 : 16);
1276         bool was_locked = btree_node_locked(path, path->level);
1277         int ret = 0;
1278
1279         bch2_bkey_buf_init(&tmp);
1280
1281         while (nr && !ret) {
1282                 if (!bch2_btree_node_relock(trans, path, path->level))
1283                         break;
1284
1285                 bch2_btree_node_iter_advance(&node_iter, l->b);
1286                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
1287                 if (!k)
1288                         break;
1289
1290                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
1291                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
1292                                                path->level - 1);
1293         }
1294
1295         if (!was_locked)
1296                 btree_node_unlock(trans, path, path->level);
1297
1298         bch2_bkey_buf_exit(&tmp, c);
1299         return ret;
1300 }
1301
1302 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
1303                                  struct btree_and_journal_iter *jiter)
1304 {
1305         struct bch_fs *c = trans->c;
1306         struct bkey_s_c k;
1307         struct bkey_buf tmp;
1308         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1309                 ? (path->level > 1 ? 0 :  2)
1310                 : (path->level > 1 ? 1 : 16);
1311         bool was_locked = btree_node_locked(path, path->level);
1312         int ret = 0;
1313
1314         bch2_bkey_buf_init(&tmp);
1315
1316         while (nr && !ret) {
1317                 if (!bch2_btree_node_relock(trans, path, path->level))
1318                         break;
1319
1320                 bch2_btree_and_journal_iter_advance(jiter);
1321                 k = bch2_btree_and_journal_iter_peek(jiter);
1322                 if (!k.k)
1323                         break;
1324
1325                 bch2_bkey_buf_reassemble(&tmp, c, k);
1326                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
1327                                                path->level - 1);
1328         }
1329
1330         if (!was_locked)
1331                 btree_node_unlock(trans, path, path->level);
1332
1333         bch2_bkey_buf_exit(&tmp, c);
1334         return ret;
1335 }
1336
1337 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
1338                                             struct btree_path *path,
1339                                             unsigned plevel, struct btree *b)
1340 {
1341         struct btree_path_level *l = &path->l[plevel];
1342         bool locked = btree_node_locked(path, plevel);
1343         struct bkey_packed *k;
1344         struct bch_btree_ptr_v2 *bp;
1345
1346         if (!bch2_btree_node_relock(trans, path, plevel))
1347                 return;
1348
1349         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1350         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
1351
1352         bp = (void *) bkeyp_val(&l->b->format, k);
1353         bp->mem_ptr = (unsigned long)b;
1354
1355         if (!locked)
1356                 btree_node_unlock(trans, path, plevel);
1357 }
1358
1359 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
1360                                                      struct btree_path *path,
1361                                                      unsigned flags,
1362                                                      struct bkey_buf *out)
1363 {
1364         struct bch_fs *c = trans->c;
1365         struct btree_path_level *l = path_l(path);
1366         struct btree_and_journal_iter jiter;
1367         struct bkey_s_c k;
1368         int ret = 0;
1369
1370         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
1371
1372         k = bch2_btree_and_journal_iter_peek(&jiter);
1373
1374         bch2_bkey_buf_reassemble(out, c, k);
1375
1376         if (flags & BTREE_ITER_PREFETCH)
1377                 ret = btree_path_prefetch_j(trans, path, &jiter);
1378
1379         bch2_btree_and_journal_iter_exit(&jiter);
1380         return ret;
1381 }
1382
1383 static __always_inline int btree_path_down(struct btree_trans *trans,
1384                                            struct btree_path *path,
1385                                            unsigned flags,
1386                                            unsigned long trace_ip)
1387 {
1388         struct bch_fs *c = trans->c;
1389         struct btree_path_level *l = path_l(path);
1390         struct btree *b;
1391         unsigned level = path->level - 1;
1392         enum six_lock_type lock_type = __btree_lock_want(path, level);
1393         bool replay_done = test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
1394         struct bkey_buf tmp;
1395         int ret;
1396
1397         EBUG_ON(!btree_node_locked(path, path->level));
1398
1399         bch2_bkey_buf_init(&tmp);
1400
1401         if (unlikely(!replay_done)) {
1402                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
1403                 if (ret)
1404                         goto err;
1405         } else {
1406                 bch2_bkey_buf_unpack(&tmp, c, l->b,
1407                                  bch2_btree_node_iter_peek(&l->iter, l->b));
1408
1409                 if (flags & BTREE_ITER_PREFETCH) {
1410                         ret = btree_path_prefetch(trans, path);
1411                         if (ret)
1412                                 goto err;
1413                 }
1414         }
1415
1416         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
1417         ret = PTR_ERR_OR_ZERO(b);
1418         if (unlikely(ret))
1419                 goto err;
1420
1421         if (likely(replay_done && tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
1422             unlikely(b != btree_node_mem_ptr(tmp.k)))
1423                 btree_node_mem_ptr_set(trans, path, level + 1, b);
1424
1425         if (btree_node_read_locked(path, level + 1))
1426                 btree_node_unlock(trans, path, level + 1);
1427
1428         mark_btree_node_locked(trans, path, level, lock_type);
1429         path->level = level;
1430         btree_path_level_init(trans, path, b);
1431
1432         bch2_btree_path_verify_locks(path);
1433 err:
1434         bch2_bkey_buf_exit(&tmp, c);
1435         return ret;
1436 }
1437
1438 static int btree_path_traverse_one(struct btree_trans *, struct btree_path *,
1439                                    unsigned, unsigned long);
1440
1441 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
1442 {
1443         struct bch_fs *c = trans->c;
1444         struct btree_path *path;
1445         unsigned long trace_ip = _RET_IP_;
1446         int i, ret = 0;
1447
1448         if (trans->in_traverse_all)
1449                 return -BCH_ERR_transaction_restart_in_traverse_all;
1450
1451         trans->in_traverse_all = true;
1452 retry_all:
1453         trans->restarted = 0;
1454         trans->traverse_all_idx = U8_MAX;
1455
1456         trans_for_each_path(trans, path)
1457                 path->should_be_locked = false;
1458
1459         btree_trans_verify_sorted(trans);
1460
1461         for (i = trans->nr_sorted - 2; i >= 0; --i) {
1462                 struct btree_path *path1 = trans->paths + trans->sorted[i];
1463                 struct btree_path *path2 = trans->paths + trans->sorted[i + 1];
1464
1465                 if (path1->btree_id == path2->btree_id &&
1466                     path1->locks_want < path2->locks_want)
1467                         __bch2_btree_path_upgrade(trans, path1, path2->locks_want);
1468                 else if (!path1->locks_want && path2->locks_want)
1469                         __bch2_btree_path_upgrade(trans, path1, 1);
1470         }
1471
1472         bch2_trans_unlock(trans);
1473         cond_resched();
1474
1475         if (unlikely(trans->memory_allocation_failure)) {
1476                 struct closure cl;
1477
1478                 closure_init_stack(&cl);
1479
1480                 do {
1481                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1482                         closure_sync(&cl);
1483                 } while (ret);
1484         }
1485
1486         /* Now, redo traversals in correct order: */
1487         trans->traverse_all_idx = 0;
1488         while (trans->traverse_all_idx < trans->nr_sorted) {
1489                 path = trans->paths + trans->sorted[trans->traverse_all_idx];
1490
1491                 /*
1492                  * Traversing a path can cause another path to be added at about
1493                  * the same position:
1494                  */
1495                 if (path->uptodate) {
1496                         ret = btree_path_traverse_one(trans, path, 0, _THIS_IP_);
1497                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1498                             ret == -ENOMEM)
1499                                 goto retry_all;
1500                         if (ret)
1501                                 goto err;
1502                         BUG_ON(path->uptodate);
1503                 } else {
1504                         trans->traverse_all_idx++;
1505                 }
1506         }
1507
1508         /*
1509          * BTREE_ITER_NEED_RELOCK is ok here - if we called bch2_trans_unlock()
1510          * and relock(), relock() won't relock since path->should_be_locked
1511          * isn't set yet, which is all fine
1512          */
1513         trans_for_each_path(trans, path)
1514                 BUG_ON(path->uptodate >= BTREE_ITER_NEED_TRAVERSE);
1515 err:
1516         bch2_btree_cache_cannibalize_unlock(c);
1517
1518         trans->in_traverse_all = false;
1519
1520         trace_trans_traverse_all(trans, trace_ip);
1521         return ret;
1522 }
1523
1524 static inline bool btree_path_good_node(struct btree_trans *trans,
1525                                         struct btree_path *path,
1526                                         unsigned l, int check_pos)
1527 {
1528         if (!is_btree_node(path, l) ||
1529             !bch2_btree_node_relock(trans, path, l))
1530                 return false;
1531
1532         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1533                 return false;
1534         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1535                 return false;
1536         return true;
1537 }
1538
1539 static void btree_path_set_level_down(struct btree_trans *trans,
1540                                       struct btree_path *path,
1541                                       unsigned new_level)
1542 {
1543         unsigned l;
1544
1545         path->level = new_level;
1546
1547         for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1548                 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1549                         btree_node_unlock(trans, path, l);
1550
1551         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1552         bch2_btree_path_verify(trans, path);
1553 }
1554
1555 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1556                                                      struct btree_path *path,
1557                                                      int check_pos)
1558 {
1559         unsigned i, l = path->level;
1560
1561         while (btree_path_node(path, l) &&
1562                !btree_path_good_node(trans, path, l, check_pos))
1563                 __btree_path_set_level_up(trans, path, l++);
1564
1565         /* If we need intent locks, take them too: */
1566         for (i = l + 1;
1567              i < path->locks_want && btree_path_node(path, i);
1568              i++)
1569                 if (!bch2_btree_node_relock(trans, path, i))
1570                         while (l <= i)
1571                                 __btree_path_set_level_up(trans, path, l++);
1572
1573         return l;
1574 }
1575
1576 /*
1577  * This is the main state machine for walking down the btree - walks down to a
1578  * specified depth
1579  *
1580  * Returns 0 on success, -EIO on error (error reading in a btree node).
1581  *
1582  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1583  * stashed in the iterator and returned from bch2_trans_exit().
1584  */
1585 static int btree_path_traverse_one(struct btree_trans *trans,
1586                                    struct btree_path *path,
1587                                    unsigned flags,
1588                                    unsigned long trace_ip)
1589 {
1590         unsigned depth_want = path->level;
1591         int ret = trans->restarted;
1592
1593         if (unlikely(ret))
1594                 goto out;
1595
1596         /*
1597          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1598          * and re-traverse the path without a transaction restart:
1599          */
1600         if (path->should_be_locked) {
1601                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1602                 goto out;
1603         }
1604
1605         if (path->cached) {
1606                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1607                 goto out;
1608         }
1609
1610         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1611                 goto out;
1612
1613         path->level = btree_path_up_until_good_node(trans, path, 0);
1614
1615         /*
1616          * Note: path->nodes[path->level] may be temporarily NULL here - that
1617          * would indicate to other code that we got to the end of the btree,
1618          * here it indicates that relocking the root failed - it's critical that
1619          * btree_path_lock_root() comes next and that it can't fail
1620          */
1621         while (path->level > depth_want) {
1622                 ret = btree_path_node(path, path->level)
1623                         ? btree_path_down(trans, path, flags, trace_ip)
1624                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1625                 if (unlikely(ret)) {
1626                         if (ret == 1) {
1627                                 /*
1628                                  * No nodes at this level - got to the end of
1629                                  * the btree:
1630                                  */
1631                                 ret = 0;
1632                                 goto out;
1633                         }
1634
1635                         __bch2_btree_path_unlock(trans, path);
1636                         path->level = depth_want;
1637                         path->l[path->level].b = ERR_PTR(ret);
1638                         goto out;
1639                 }
1640         }
1641
1642         path->uptodate = BTREE_ITER_UPTODATE;
1643 out:
1644         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1645         bch2_btree_path_verify(trans, path);
1646         return ret;
1647 }
1648
1649 int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
1650                                           struct btree_path *path, unsigned flags)
1651 {
1652         if (0 && IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
1653                 unsigned restart_probability_bits = 4 << min(trans->restart_count, 32U);
1654                 u64 mask = ~(~0ULL << restart_probability_bits);
1655
1656                 if ((prandom_u32() & mask) == mask) {
1657                         trace_transaction_restart_injected(trans, _RET_IP_);
1658                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_fault_inject);
1659                 }
1660         }
1661
1662         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
1663                 return 0;
1664
1665         return  bch2_trans_cond_resched(trans) ?:
1666                 btree_path_traverse_one(trans, path, flags, _RET_IP_);
1667 }
1668
1669 static void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1670                             struct btree_path *src)
1671 {
1672         unsigned i, offset = offsetof(struct btree_path, pos);
1673
1674         memcpy((void *) dst + offset,
1675                (void *) src + offset,
1676                sizeof(struct btree_path) - offset);
1677
1678         for (i = 0; i < BTREE_MAX_DEPTH; i++)
1679                 if (btree_node_locked(dst, i))
1680                         six_lock_increment(&dst->l[i].b->c.lock,
1681                                            __btree_lock_want(dst, i));
1682
1683         bch2_btree_path_check_sort(trans, dst, 0);
1684 }
1685
1686 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1687                                            bool intent)
1688 {
1689         struct btree_path *new = btree_path_alloc(trans, src);
1690
1691         btree_path_copy(trans, new, src);
1692         __btree_path_get(new, intent);
1693         return new;
1694 }
1695
1696 inline struct btree_path * __must_check
1697 bch2_btree_path_make_mut(struct btree_trans *trans,
1698                          struct btree_path *path, bool intent,
1699                          unsigned long ip)
1700 {
1701         if (path->ref > 1 || path->preserve) {
1702                 __btree_path_put(path, intent);
1703                 path = btree_path_clone(trans, path, intent);
1704                 path->preserve = false;
1705 #ifdef CONFIG_BCACHEFS_DEBUG
1706                 path->ip_allocated = ip;
1707 #endif
1708                 btree_trans_verify_sorted(trans);
1709         }
1710
1711         path->should_be_locked = false;
1712         return path;
1713 }
1714
1715 struct btree_path * __must_check
1716 bch2_btree_path_set_pos(struct btree_trans *trans,
1717                    struct btree_path *path, struct bpos new_pos,
1718                    bool intent, unsigned long ip)
1719 {
1720         int cmp = bpos_cmp(new_pos, path->pos);
1721         unsigned l = path->level;
1722
1723         EBUG_ON(trans->restarted);
1724         EBUG_ON(!path->ref);
1725
1726         if (!cmp)
1727                 return path;
1728
1729         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1730
1731         path->pos = new_pos;
1732
1733         bch2_btree_path_check_sort(trans, path, cmp);
1734
1735         if (unlikely(path->cached)) {
1736                 btree_node_unlock(trans, path, 0);
1737                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1738                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1739                 goto out;
1740         }
1741
1742         l = btree_path_up_until_good_node(trans, path, cmp);
1743
1744         if (btree_path_node(path, l)) {
1745                 BUG_ON(!btree_node_locked(path, l));
1746                 /*
1747                  * We might have to skip over many keys, or just a few: try
1748                  * advancing the node iterator, and if we have to skip over too
1749                  * many keys just reinit it (or if we're rewinding, since that
1750                  * is expensive).
1751                  */
1752                 if (cmp < 0 ||
1753                     !btree_path_advance_to_pos(path, &path->l[l], 8))
1754                         __btree_path_level_init(path, l);
1755         }
1756
1757         if (l != path->level) {
1758                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1759                 __bch2_btree_path_unlock(trans, path);
1760         }
1761 out:
1762         bch2_btree_path_verify(trans, path);
1763         return path;
1764 }
1765
1766 /* Btree path: main interface: */
1767
1768 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1769 {
1770         struct btree_path *sib;
1771
1772         sib = prev_btree_path(trans, path);
1773         if (sib && !btree_path_cmp(sib, path))
1774                 return sib;
1775
1776         sib = next_btree_path(trans, path);
1777         if (sib && !btree_path_cmp(sib, path))
1778                 return sib;
1779
1780         return NULL;
1781 }
1782
1783 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1784 {
1785         struct btree_path *sib;
1786
1787         sib = prev_btree_path(trans, path);
1788         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1789                 return sib;
1790
1791         sib = next_btree_path(trans, path);
1792         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1793                 return sib;
1794
1795         return NULL;
1796 }
1797
1798 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1799 {
1800         __bch2_btree_path_unlock(trans, path);
1801         btree_path_list_remove(trans, path);
1802         trans->paths_allocated &= ~(1ULL << path->idx);
1803 }
1804
1805 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1806 {
1807         struct btree_path *dup;
1808
1809         EBUG_ON(trans->paths + path->idx != path);
1810         EBUG_ON(!path->ref);
1811
1812         if (!__btree_path_put(path, intent))
1813                 return;
1814
1815         dup = path->preserve
1816                 ? have_path_at_pos(trans, path)
1817                 : have_node_at_pos(trans, path);
1818
1819         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1820                 return;
1821
1822         if (path->should_be_locked &&
1823             !trans->restarted &&
1824             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1825                 return;
1826
1827         if (dup) {
1828                 dup->preserve           |= path->preserve;
1829                 dup->should_be_locked   |= path->should_be_locked;
1830         }
1831
1832         __bch2_path_free(trans, path);
1833 }
1834
1835 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1836 {
1837         struct btree_insert_entry *i;
1838
1839         prt_printf(buf, "transaction updates for %s journal seq %llu",
1840                trans->fn, trans->journal_res.seq);
1841         prt_newline(buf);
1842         printbuf_indent_add(buf, 2);
1843
1844         trans_for_each_update(trans, i) {
1845                 struct bkey_s_c old = { &i->old_k, i->old_v };
1846
1847                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1848                        bch2_btree_ids[i->btree_id],
1849                        i->cached,
1850                        (void *) i->ip_allocated);
1851                 prt_newline(buf);
1852
1853                 prt_printf(buf, "  old ");
1854                 bch2_bkey_val_to_text(buf, trans->c, old);
1855                 prt_newline(buf);
1856
1857                 prt_printf(buf, "  new ");
1858                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1859                 prt_newline(buf);
1860         }
1861
1862         printbuf_indent_sub(buf, 2);
1863 }
1864
1865 noinline __cold
1866 void bch2_dump_trans_updates(struct btree_trans *trans)
1867 {
1868         struct printbuf buf = PRINTBUF;
1869
1870         bch2_trans_updates_to_text(&buf, trans);
1871         bch_err(trans->c, "%s", buf.buf);
1872         printbuf_exit(&buf);
1873 }
1874
1875 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1876 {
1877         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1878                    path->idx, path->ref, path->intent_ref,
1879                    path->preserve ? 'P' : ' ',
1880                    path->should_be_locked ? 'S' : ' ',
1881                    bch2_btree_ids[path->btree_id],
1882                    path->level);
1883         bch2_bpos_to_text(out, path->pos);
1884
1885         prt_printf(out, " locks %u", path->nodes_locked);
1886 #ifdef CONFIG_BCACHEFS_DEBUG
1887         prt_printf(out, " %pS", (void *) path->ip_allocated);
1888 #endif
1889         prt_newline(out);
1890 }
1891
1892 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1893 {
1894         struct btree_path *path;
1895         unsigned idx;
1896
1897         trans_for_each_path_inorder(trans, path, idx)
1898                 bch2_btree_path_to_text(out, path);
1899 }
1900
1901 noinline __cold
1902 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1903 {
1904         struct printbuf buf = PRINTBUF;
1905
1906         bch2_trans_paths_to_text(&buf, trans);
1907
1908         printk(KERN_ERR "%s", buf.buf);
1909         printbuf_exit(&buf);
1910
1911         bch2_dump_trans_updates(trans);
1912 }
1913
1914 noinline
1915 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1916 {
1917         struct btree_transaction_stats *s = btree_trans_stats(trans);
1918         struct printbuf buf = PRINTBUF;
1919
1920         bch2_trans_paths_to_text(&buf, trans);
1921
1922         if (!buf.allocation_failure) {
1923                 mutex_lock(&s->lock);
1924                 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1925                         s->nr_max_paths = hweight64(trans->paths_allocated);
1926                         swap(s->max_paths_text, buf.buf);
1927                 }
1928                 mutex_unlock(&s->lock);
1929         }
1930
1931         printbuf_exit(&buf);
1932 }
1933
1934 static struct btree_path *btree_path_alloc(struct btree_trans *trans,
1935                                            struct btree_path *pos)
1936 {
1937         struct btree_transaction_stats *s = btree_trans_stats(trans);
1938         struct btree_path *path;
1939         unsigned idx;
1940
1941         if (unlikely(trans->paths_allocated ==
1942                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1)))) {
1943                 bch2_dump_trans_paths_updates(trans);
1944                 panic("trans path oveflow\n");
1945         }
1946
1947         idx = __ffs64(~trans->paths_allocated);
1948         trans->paths_allocated |= 1ULL << idx;
1949
1950         if (s && unlikely(hweight64(trans->paths_allocated) > s->nr_max_paths))
1951                 bch2_trans_update_max_paths(trans);
1952
1953         path = &trans->paths[idx];
1954
1955         path->idx               = idx;
1956         path->ref               = 0;
1957         path->intent_ref        = 0;
1958         path->nodes_locked      = 0;
1959         path->nodes_intent_locked = 0;
1960
1961         btree_path_list_add(trans, pos, path);
1962         return path;
1963 }
1964
1965 struct btree_path *bch2_path_get(struct btree_trans *trans,
1966                                  enum btree_id btree_id, struct bpos pos,
1967                                  unsigned locks_want, unsigned level,
1968                                  unsigned flags, unsigned long ip)
1969 {
1970         struct btree_path *path, *path_pos = NULL;
1971         bool cached = flags & BTREE_ITER_CACHED;
1972         bool intent = flags & BTREE_ITER_INTENT;
1973         int i;
1974
1975         BUG_ON(trans->restarted);
1976         btree_trans_verify_sorted(trans);
1977         bch2_trans_verify_locks(trans);
1978
1979         trans_for_each_path_inorder(trans, path, i) {
1980                 if (__btree_path_cmp(path,
1981                                      btree_id,
1982                                      cached,
1983                                      pos,
1984                                      level) > 0)
1985                         break;
1986
1987                 path_pos = path;
1988         }
1989
1990         if (path_pos &&
1991             path_pos->cached    == cached &&
1992             path_pos->btree_id  == btree_id &&
1993             path_pos->level     == level) {
1994                 __btree_path_get(path_pos, intent);
1995                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1996         } else {
1997                 path = btree_path_alloc(trans, path_pos);
1998                 path_pos = NULL;
1999
2000                 __btree_path_get(path, intent);
2001                 path->pos                       = pos;
2002                 path->btree_id                  = btree_id;
2003                 path->cached                    = cached;
2004                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
2005                 path->should_be_locked          = false;
2006                 path->level                     = level;
2007                 path->locks_want                = locks_want;
2008                 path->nodes_locked              = 0;
2009                 path->nodes_intent_locked       = 0;
2010                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
2011                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
2012 #ifdef CONFIG_BCACHEFS_DEBUG
2013                 path->ip_allocated              = ip;
2014 #endif
2015                 btree_trans_verify_sorted(trans);
2016         }
2017
2018         if (!(flags & BTREE_ITER_NOPRESERVE))
2019                 path->preserve = true;
2020
2021         if (path->intent_ref)
2022                 locks_want = max(locks_want, level + 1);
2023
2024         /*
2025          * If the path has locks_want greater than requested, we don't downgrade
2026          * it here - on transaction restart because btree node split needs to
2027          * upgrade locks, we might be putting/getting the iterator again.
2028          * Downgrading iterators only happens via bch2_trans_downgrade(), after
2029          * a successful transaction commit.
2030          */
2031
2032         locks_want = min(locks_want, BTREE_MAX_DEPTH);
2033         if (locks_want > path->locks_want) {
2034                 path->locks_want = locks_want;
2035                 btree_path_get_locks(trans, path, true);
2036         }
2037
2038         return path;
2039 }
2040
2041 inline struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
2042 {
2043
2044         struct bkey_s_c k;
2045
2046         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
2047         EBUG_ON(!btree_node_locked(path, path->level));
2048
2049         if (!path->cached) {
2050                 struct btree_path_level *l = path_l(path);
2051                 struct bkey_packed *_k;
2052
2053                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
2054                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
2055
2056                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_cmp(k.k->p, path->pos) == 0);
2057
2058                 if (!k.k || bpos_cmp(path->pos, k.k->p))
2059                         goto hole;
2060         } else {
2061                 struct bkey_cached *ck = (void *) path->l[0].b;
2062
2063                 EBUG_ON(ck &&
2064                         (path->btree_id != ck->key.btree_id ||
2065                          bkey_cmp(path->pos, ck->key.pos)));
2066                 EBUG_ON(!ck || !ck->valid);
2067
2068                 *u = ck->k->k;
2069                 k = bkey_i_to_s_c(ck->k);
2070         }
2071
2072         return k;
2073 hole:
2074         bkey_init(u);
2075         u->p = path->pos;
2076         return (struct bkey_s_c) { u, NULL };
2077 }
2078
2079 /* Btree iterators: */
2080
2081 int __must_check
2082 __bch2_btree_iter_traverse(struct btree_iter *iter)
2083 {
2084         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
2085 }
2086
2087 int __must_check
2088 bch2_btree_iter_traverse(struct btree_iter *iter)
2089 {
2090         int ret;
2091
2092         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
2093                                         btree_iter_search_key(iter),
2094                                         iter->flags & BTREE_ITER_INTENT,
2095                                         btree_iter_ip_allocated(iter));
2096
2097         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
2098         if (ret)
2099                 return ret;
2100
2101         btree_path_set_should_be_locked(iter->path);
2102         return 0;
2103 }
2104
2105 /* Iterate across nodes (leaf and interior nodes) */
2106
2107 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
2108 {
2109         struct btree_trans *trans = iter->trans;
2110         struct btree *b = NULL;
2111         int ret;
2112
2113         EBUG_ON(iter->path->cached);
2114         bch2_btree_iter_verify(iter);
2115
2116         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2117         if (ret)
2118                 goto err;
2119
2120         b = btree_path_node(iter->path, iter->path->level);
2121         if (!b)
2122                 goto out;
2123
2124         BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0);
2125
2126         bkey_init(&iter->k);
2127         iter->k.p = iter->pos = b->key.k.p;
2128
2129         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
2130                                         iter->flags & BTREE_ITER_INTENT,
2131                                         btree_iter_ip_allocated(iter));
2132         btree_path_set_should_be_locked(iter->path);
2133 out:
2134         bch2_btree_iter_verify_entry_exit(iter);
2135         bch2_btree_iter_verify(iter);
2136
2137         return b;
2138 err:
2139         b = ERR_PTR(ret);
2140         goto out;
2141 }
2142
2143 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
2144 {
2145         struct btree_trans *trans = iter->trans;
2146         struct btree_path *path = iter->path;
2147         struct btree *b = NULL;
2148         int ret;
2149
2150         BUG_ON(trans->restarted);
2151         EBUG_ON(iter->path->cached);
2152         bch2_btree_iter_verify(iter);
2153
2154         /* already at end? */
2155         if (!btree_path_node(path, path->level))
2156                 return NULL;
2157
2158         /* got to end? */
2159         if (!btree_path_node(path, path->level + 1)) {
2160                 btree_path_set_level_up(trans, path);
2161                 return NULL;
2162         }
2163
2164         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
2165                 __bch2_btree_path_unlock(trans, path);
2166                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
2167                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
2168                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
2169                 trace_trans_restart_relock_next_node(trans, _THIS_IP_, path);
2170                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
2171                 goto err;
2172         }
2173
2174         b = btree_path_node(path, path->level + 1);
2175
2176         if (!bpos_cmp(iter->pos, b->key.k.p)) {
2177                 __btree_path_set_level_up(trans, path, path->level++);
2178         } else {
2179                 /*
2180                  * Haven't gotten to the end of the parent node: go back down to
2181                  * the next child node
2182                  */
2183                 path = iter->path =
2184                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
2185                                            iter->flags & BTREE_ITER_INTENT,
2186                                            btree_iter_ip_allocated(iter));
2187
2188                 btree_path_set_level_down(trans, path, iter->min_depth);
2189
2190                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
2191                 if (ret)
2192                         goto err;
2193
2194                 b = path->l[path->level].b;
2195         }
2196
2197         bkey_init(&iter->k);
2198         iter->k.p = iter->pos = b->key.k.p;
2199
2200         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
2201                                         iter->flags & BTREE_ITER_INTENT,
2202                                         btree_iter_ip_allocated(iter));
2203         btree_path_set_should_be_locked(iter->path);
2204         BUG_ON(iter->path->uptodate);
2205 out:
2206         bch2_btree_iter_verify_entry_exit(iter);
2207         bch2_btree_iter_verify(iter);
2208
2209         return b;
2210 err:
2211         b = ERR_PTR(ret);
2212         goto out;
2213 }
2214
2215 /* Iterate across keys (in leaf nodes only) */
2216
2217 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
2218 {
2219         if (likely(!(iter->flags & BTREE_ITER_ALL_LEVELS))) {
2220                 struct bpos pos = iter->k.p;
2221                 bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
2222                             ? bpos_cmp(pos, SPOS_MAX)
2223                             : bkey_cmp(pos, SPOS_MAX)) != 0;
2224
2225                 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
2226                         pos = bkey_successor(iter, pos);
2227                 bch2_btree_iter_set_pos(iter, pos);
2228                 return ret;
2229         } else {
2230                 if (!btree_path_node(iter->path, iter->path->level))
2231                         return true;
2232
2233                 iter->advanced = true;
2234                 return false;
2235         }
2236 }
2237
2238 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
2239 {
2240         struct bpos pos = bkey_start_pos(&iter->k);
2241         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
2242                     ? bpos_cmp(pos, POS_MIN)
2243                     : bkey_cmp(pos, POS_MIN)) != 0;
2244
2245         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
2246                 pos = bkey_predecessor(iter, pos);
2247         bch2_btree_iter_set_pos(iter, pos);
2248         return ret;
2249 }
2250
2251 static inline struct bkey_i *btree_trans_peek_updates(struct btree_trans *trans,
2252                                                       enum btree_id btree_id,
2253                                                       struct bpos pos)
2254 {
2255         struct btree_insert_entry *i;
2256         struct bkey_i *ret = NULL;
2257
2258         trans_for_each_update(trans, i) {
2259                 if (i->btree_id < btree_id)
2260                         continue;
2261                 if (i->btree_id > btree_id)
2262                         break;
2263                 if (bpos_cmp(i->k->k.p, pos) < 0)
2264                         continue;
2265                 if (i->key_cache_already_flushed)
2266                         continue;
2267                 if (!ret || bpos_cmp(i->k->k.p, ret->k.p) < 0)
2268                         ret = i->k;
2269         }
2270
2271         return ret;
2272 }
2273
2274 struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
2275                                        struct btree_iter *iter,
2276                                        struct bpos start_pos,
2277                                        struct bpos end_pos)
2278 {
2279         struct bkey_i *k;
2280
2281         if (bpos_cmp(start_pos, iter->journal_pos) < 0)
2282                 iter->journal_idx = 0;
2283
2284         k = bch2_journal_keys_peek_upto(trans->c, iter->btree_id, 0,
2285                                         start_pos, end_pos,
2286                                         &iter->journal_idx);
2287
2288         iter->journal_pos = k ? k->k.p : end_pos;
2289         return k;
2290 }
2291
2292 struct bkey_i *bch2_btree_journal_peek_slot(struct btree_trans *trans,
2293                                             struct btree_iter *iter,
2294                                             struct bpos pos)
2295 {
2296         return bch2_btree_journal_peek(trans, iter, pos, pos);
2297 }
2298
2299 static noinline
2300 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
2301                                          struct btree_iter *iter,
2302                                          struct bkey_s_c k)
2303 {
2304         struct bkey_i *next_journal =
2305                 bch2_btree_journal_peek(trans, iter, iter->path->pos,
2306                                 k.k ? k.k->p : iter->path->l[0].b->key.k.p);
2307
2308         if (next_journal) {
2309                 iter->k = next_journal->k;
2310                 k = bkey_i_to_s_c(next_journal);
2311         }
2312
2313         return k;
2314 }
2315
2316 /*
2317  * Checks btree key cache for key at iter->pos and returns it if present, or
2318  * bkey_s_c_null:
2319  */
2320 static noinline
2321 struct bkey_s_c __btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
2322 {
2323         struct btree_trans *trans = iter->trans;
2324         struct bch_fs *c = trans->c;
2325         struct bkey u;
2326         int ret;
2327
2328         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
2329                 return bkey_s_c_null;
2330
2331         if (!iter->key_cache_path)
2332                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
2333                                                      iter->flags & BTREE_ITER_INTENT, 0,
2334                                                      iter->flags|BTREE_ITER_CACHED,
2335                                                      _THIS_IP_);
2336
2337         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
2338                                         iter->flags & BTREE_ITER_INTENT,
2339                                         btree_iter_ip_allocated(iter));
2340
2341         ret = bch2_btree_path_traverse(trans, iter->key_cache_path, iter->flags|BTREE_ITER_CACHED);
2342         if (unlikely(ret))
2343                 return bkey_s_c_err(ret);
2344
2345         btree_path_set_should_be_locked(iter->key_cache_path);
2346
2347         return bch2_btree_path_peek_slot(iter->key_cache_path, &u);
2348 }
2349
2350 static noinline
2351 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
2352 {
2353         struct bkey_s_c ret = __btree_trans_peek_key_cache(iter, pos);
2354         int err = bkey_err(ret) ?: bch2_btree_path_relock(iter->trans, iter->path, _THIS_IP_);
2355
2356         return err ? bkey_s_c_err(err) : ret;
2357 }
2358
2359 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
2360 {
2361         struct btree_trans *trans = iter->trans;
2362         struct bkey_i *next_update;
2363         struct bkey_s_c k, k2;
2364         int ret;
2365
2366         EBUG_ON(iter->path->cached || iter->path->level);
2367         bch2_btree_iter_verify(iter);
2368
2369         while (1) {
2370                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2371                                         iter->flags & BTREE_ITER_INTENT,
2372                                         btree_iter_ip_allocated(iter));
2373
2374                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2375                 if (unlikely(ret)) {
2376                         /* ensure that iter->k is consistent with iter->pos: */
2377                         bch2_btree_iter_set_pos(iter, iter->pos);
2378                         k = bkey_s_c_err(ret);
2379                         goto out;
2380                 }
2381
2382                 btree_path_set_should_be_locked(iter->path);
2383
2384                 k = btree_path_level_peek_all(trans->c, &iter->path->l[0], &iter->k);
2385
2386                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2387                     k.k &&
2388                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2389                         k = k2;
2390                         ret = bkey_err(k);
2391                         if (ret) {
2392                                 bch2_btree_iter_set_pos(iter, iter->pos);
2393                                 goto out;
2394                         }
2395                 }
2396
2397                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2398                         k = btree_trans_peek_journal(trans, iter, k);
2399
2400                 next_update = iter->flags & BTREE_ITER_WITH_UPDATES
2401                         ? btree_trans_peek_updates(trans, iter->btree_id, search_key)
2402                         : NULL;
2403                 if (next_update &&
2404                     bpos_cmp(next_update->k.p,
2405                              k.k ? k.k->p : iter->path->l[0].b->key.k.p) <= 0) {
2406                         iter->k = next_update->k;
2407                         k = bkey_i_to_s_c(next_update);
2408                 }
2409
2410                 if (k.k && bkey_deleted(k.k)) {
2411                         /*
2412                          * If we've got a whiteout, and it's after the search
2413                          * key, advance the search key to the whiteout instead
2414                          * of just after the whiteout - it might be a btree
2415                          * whiteout, with a real key at the same position, since
2416                          * in the btree deleted keys sort before non deleted.
2417                          */
2418                         search_key = bpos_cmp(search_key, k.k->p)
2419                                 ? k.k->p
2420                                 : bpos_successor(k.k->p);
2421                         continue;
2422                 }
2423
2424                 if (likely(k.k)) {
2425                         break;
2426                 } else if (likely(bpos_cmp(iter->path->l[0].b->key.k.p, SPOS_MAX))) {
2427                         /* Advance to next leaf node: */
2428                         search_key = bpos_successor(iter->path->l[0].b->key.k.p);
2429                 } else {
2430                         /* End of btree: */
2431                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2432                         k = bkey_s_c_null;
2433                         goto out;
2434                 }
2435         }
2436 out:
2437         bch2_btree_iter_verify(iter);
2438
2439         return k;
2440 }
2441
2442 /**
2443  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2444  * current position
2445  */
2446 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2447 {
2448         struct btree_trans *trans = iter->trans;
2449         struct bpos search_key = btree_iter_search_key(iter);
2450         struct bkey_s_c k;
2451         struct bpos iter_pos;
2452         int ret;
2453
2454         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2455
2456         if (iter->update_path) {
2457                 bch2_path_put(trans, iter->update_path,
2458                               iter->flags & BTREE_ITER_INTENT);
2459                 iter->update_path = NULL;
2460         }
2461
2462         bch2_btree_iter_verify_entry_exit(iter);
2463
2464         while (1) {
2465                 k = __bch2_btree_iter_peek(iter, search_key);
2466                 if (!k.k || bkey_err(k))
2467                         goto out_no_locked;
2468
2469                 /*
2470                  * iter->pos should be mononotically increasing, and always be
2471                  * equal to the key we just returned - except extents can
2472                  * straddle iter->pos:
2473                  */
2474                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2475                         iter_pos = k.k->p;
2476                 else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
2477                         iter_pos = bkey_start_pos(k.k);
2478                 else
2479                         iter_pos = iter->pos;
2480
2481                 if (bkey_cmp(iter_pos, end) > 0) {
2482                         bch2_btree_iter_set_pos(iter, end);
2483                         k = bkey_s_c_null;
2484                         goto out_no_locked;
2485                 }
2486
2487                 if (iter->update_path &&
2488                     bkey_cmp(iter->update_path->pos, k.k->p)) {
2489                         bch2_path_put(trans, iter->update_path,
2490                                       iter->flags & BTREE_ITER_INTENT);
2491                         iter->update_path = NULL;
2492                 }
2493
2494                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2495                     (iter->flags & BTREE_ITER_INTENT) &&
2496                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2497                     !iter->update_path) {
2498                         struct bpos pos = k.k->p;
2499
2500                         if (pos.snapshot < iter->snapshot) {
2501                                 search_key = bpos_successor(k.k->p);
2502                                 continue;
2503                         }
2504
2505                         pos.snapshot = iter->snapshot;
2506
2507                         /*
2508                          * advance, same as on exit for iter->path, but only up
2509                          * to snapshot
2510                          */
2511                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2512                         iter->update_path = iter->path;
2513
2514                         iter->update_path = bch2_btree_path_set_pos(trans,
2515                                                 iter->update_path, pos,
2516                                                 iter->flags & BTREE_ITER_INTENT,
2517                                                 _THIS_IP_);
2518                 }
2519
2520                 /*
2521                  * We can never have a key in a leaf node at POS_MAX, so
2522                  * we don't have to check these successor() calls:
2523                  */
2524                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2525                     !bch2_snapshot_is_ancestor(trans->c,
2526                                                iter->snapshot,
2527                                                k.k->p.snapshot)) {
2528                         search_key = bpos_successor(k.k->p);
2529                         continue;
2530                 }
2531
2532                 if (bkey_whiteout(k.k) &&
2533                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2534                         search_key = bkey_successor(iter, k.k->p);
2535                         continue;
2536                 }
2537
2538                 break;
2539         }
2540
2541         iter->pos = iter_pos;
2542
2543         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2544                                 iter->flags & BTREE_ITER_INTENT,
2545                                 btree_iter_ip_allocated(iter));
2546
2547         btree_path_set_should_be_locked(iter->path);
2548 out_no_locked:
2549         if (iter->update_path) {
2550                 if (iter->update_path->uptodate &&
2551                     (ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_)))
2552                         k = bkey_s_c_err(ret);
2553                 else
2554                         btree_path_set_should_be_locked(iter->update_path);
2555         }
2556
2557         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2558                 iter->pos.snapshot = iter->snapshot;
2559
2560         ret = bch2_btree_iter_verify_ret(iter, k);
2561         if (unlikely(ret)) {
2562                 bch2_btree_iter_set_pos(iter, iter->pos);
2563                 k = bkey_s_c_err(ret);
2564         }
2565
2566         bch2_btree_iter_verify_entry_exit(iter);
2567
2568         return k;
2569 }
2570
2571 /**
2572  * bch2_btree_iter_peek_all_levels: returns the first key greater than or equal
2573  * to iterator's current position, returning keys from every level of the btree.
2574  * For keys at different levels of the btree that compare equal, the key from
2575  * the lower level (leaf) is returned first.
2576  */
2577 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *iter)
2578 {
2579         struct btree_trans *trans = iter->trans;
2580         struct bkey_s_c k;
2581         int ret;
2582
2583         EBUG_ON(iter->path->cached);
2584         bch2_btree_iter_verify(iter);
2585         BUG_ON(iter->path->level < iter->min_depth);
2586         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
2587         EBUG_ON(!(iter->flags & BTREE_ITER_ALL_LEVELS));
2588
2589         while (1) {
2590                 iter->path = bch2_btree_path_set_pos(trans, iter->path, iter->pos,
2591                                         iter->flags & BTREE_ITER_INTENT,
2592                                         btree_iter_ip_allocated(iter));
2593
2594                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2595                 if (unlikely(ret)) {
2596                         /* ensure that iter->k is consistent with iter->pos: */
2597                         bch2_btree_iter_set_pos(iter, iter->pos);
2598                         k = bkey_s_c_err(ret);
2599                         goto out_no_locked;
2600                 }
2601
2602                 /* Already at end? */
2603                 if (!btree_path_node(iter->path, iter->path->level)) {
2604                         k = bkey_s_c_null;
2605                         goto out_no_locked;
2606                 }
2607
2608                 k = btree_path_level_peek_all(trans->c,
2609                                 &iter->path->l[iter->path->level], &iter->k);
2610
2611                 /* Check if we should go up to the parent node: */
2612                 if (!k.k ||
2613                     (iter->advanced &&
2614                      !bpos_cmp(path_l(iter->path)->b->key.k.p, iter->pos))) {
2615                         iter->pos = path_l(iter->path)->b->key.k.p;
2616                         btree_path_set_level_up(trans, iter->path);
2617                         iter->advanced = false;
2618                         continue;
2619                 }
2620
2621                 /*
2622                  * Check if we should go back down to a leaf:
2623                  * If we're not in a leaf node, we only return the current key
2624                  * if it exactly matches iter->pos - otherwise we first have to
2625                  * go back to the leaf:
2626                  */
2627                 if (iter->path->level != iter->min_depth &&
2628                     (iter->advanced ||
2629                      !k.k ||
2630                      bpos_cmp(iter->pos, k.k->p))) {
2631                         btree_path_set_level_down(trans, iter->path, iter->min_depth);
2632                         iter->pos = bpos_successor(iter->pos);
2633                         iter->advanced = false;
2634                         continue;
2635                 }
2636
2637                 /* Check if we should go to the next key: */
2638                 if (iter->path->level == iter->min_depth &&
2639                     iter->advanced &&
2640                     k.k &&
2641                     !bpos_cmp(iter->pos, k.k->p)) {
2642                         iter->pos = bpos_successor(iter->pos);
2643                         iter->advanced = false;
2644                         continue;
2645                 }
2646
2647                 if (iter->advanced &&
2648                     iter->path->level == iter->min_depth &&
2649                     bpos_cmp(k.k->p, iter->pos))
2650                         iter->advanced = false;
2651
2652                 BUG_ON(iter->advanced);
2653                 BUG_ON(!k.k);
2654                 break;
2655         }
2656
2657         iter->pos = k.k->p;
2658         btree_path_set_should_be_locked(iter->path);
2659 out_no_locked:
2660         bch2_btree_iter_verify(iter);
2661
2662         return k;
2663 }
2664
2665 /**
2666  * bch2_btree_iter_next: returns first key greater than iterator's current
2667  * position
2668  */
2669 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2670 {
2671         if (!bch2_btree_iter_advance(iter))
2672                 return bkey_s_c_null;
2673
2674         return bch2_btree_iter_peek(iter);
2675 }
2676
2677 /**
2678  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2679  * iterator's current position
2680  */
2681 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2682 {
2683         struct btree_trans *trans = iter->trans;
2684         struct bpos search_key = iter->pos;
2685         struct btree_path *saved_path = NULL;
2686         struct bkey_s_c k;
2687         struct bkey saved_k;
2688         const struct bch_val *saved_v;
2689         int ret;
2690
2691         EBUG_ON(iter->path->cached || iter->path->level);
2692         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2693
2694         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2695                 return bkey_s_c_err(-EIO);
2696
2697         bch2_btree_iter_verify(iter);
2698         bch2_btree_iter_verify_entry_exit(iter);
2699
2700         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2701                 search_key.snapshot = U32_MAX;
2702
2703         while (1) {
2704                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2705                                                 iter->flags & BTREE_ITER_INTENT,
2706                                                 btree_iter_ip_allocated(iter));
2707
2708                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2709                 if (unlikely(ret)) {
2710                         /* ensure that iter->k is consistent with iter->pos: */
2711                         bch2_btree_iter_set_pos(iter, iter->pos);
2712                         k = bkey_s_c_err(ret);
2713                         goto out_no_locked;
2714                 }
2715
2716                 k = btree_path_level_peek(trans, iter->path,
2717                                           &iter->path->l[0], &iter->k);
2718                 if (!k.k ||
2719                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2720                      ? bpos_cmp(bkey_start_pos(k.k), search_key) >= 0
2721                      : bpos_cmp(k.k->p, search_key) > 0))
2722                         k = btree_path_level_prev(trans, iter->path,
2723                                                   &iter->path->l[0], &iter->k);
2724
2725                 bch2_btree_path_check_sort(trans, iter->path, 0);
2726
2727                 if (likely(k.k)) {
2728                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2729                                 if (k.k->p.snapshot == iter->snapshot)
2730                                         goto got_key;
2731
2732                                 /*
2733                                  * If we have a saved candidate, and we're no
2734                                  * longer at the same _key_ (not pos), return
2735                                  * that candidate
2736                                  */
2737                                 if (saved_path && bkey_cmp(k.k->p, saved_k.p)) {
2738                                         bch2_path_put(trans, iter->path,
2739                                                       iter->flags & BTREE_ITER_INTENT);
2740                                         iter->path = saved_path;
2741                                         saved_path = NULL;
2742                                         iter->k = saved_k;
2743                                         k.v     = saved_v;
2744                                         goto got_key;
2745                                 }
2746
2747                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2748                                                               iter->snapshot,
2749                                                               k.k->p.snapshot)) {
2750                                         if (saved_path)
2751                                                 bch2_path_put(trans, saved_path,
2752                                                       iter->flags & BTREE_ITER_INTENT);
2753                                         saved_path = btree_path_clone(trans, iter->path,
2754                                                                 iter->flags & BTREE_ITER_INTENT);
2755                                         saved_k = *k.k;
2756                                         saved_v = k.v;
2757                                 }
2758
2759                                 search_key = bpos_predecessor(k.k->p);
2760                                 continue;
2761                         }
2762 got_key:
2763                         if (bkey_whiteout(k.k) &&
2764                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2765                                 search_key = bkey_predecessor(iter, k.k->p);
2766                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2767                                         search_key.snapshot = U32_MAX;
2768                                 continue;
2769                         }
2770
2771                         break;
2772                 } else if (likely(bpos_cmp(iter->path->l[0].b->data->min_key, POS_MIN))) {
2773                         /* Advance to previous leaf node: */
2774                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2775                 } else {
2776                         /* Start of btree: */
2777                         bch2_btree_iter_set_pos(iter, POS_MIN);
2778                         k = bkey_s_c_null;
2779                         goto out_no_locked;
2780                 }
2781         }
2782
2783         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
2784
2785         /* Extents can straddle iter->pos: */
2786         if (bkey_cmp(k.k->p, iter->pos) < 0)
2787                 iter->pos = k.k->p;
2788
2789         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2790                 iter->pos.snapshot = iter->snapshot;
2791
2792         btree_path_set_should_be_locked(iter->path);
2793 out_no_locked:
2794         if (saved_path)
2795                 bch2_path_put(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2796
2797         bch2_btree_iter_verify_entry_exit(iter);
2798         bch2_btree_iter_verify(iter);
2799
2800         return k;
2801 }
2802
2803 /**
2804  * bch2_btree_iter_prev: returns first key less than iterator's current
2805  * position
2806  */
2807 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2808 {
2809         if (!bch2_btree_iter_rewind(iter))
2810                 return bkey_s_c_null;
2811
2812         return bch2_btree_iter_peek_prev(iter);
2813 }
2814
2815 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2816 {
2817         struct btree_trans *trans = iter->trans;
2818         struct bpos search_key;
2819         struct bkey_s_c k;
2820         int ret;
2821
2822         bch2_btree_iter_verify(iter);
2823         bch2_btree_iter_verify_entry_exit(iter);
2824         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2825         EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2826
2827         /* extents can't span inode numbers: */
2828         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2829             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2830                 if (iter->pos.inode == KEY_INODE_MAX)
2831                         return bkey_s_c_null;
2832
2833                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2834         }
2835
2836         search_key = btree_iter_search_key(iter);
2837         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2838                                         iter->flags & BTREE_ITER_INTENT,
2839                                         btree_iter_ip_allocated(iter));
2840
2841         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2842         if (unlikely(ret)) {
2843                 k = bkey_s_c_err(ret);
2844                 goto out_no_locked;
2845         }
2846
2847         if ((iter->flags & BTREE_ITER_CACHED) ||
2848             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2849                 struct bkey_i *next_update;
2850
2851                 if ((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2852                     (next_update = btree_trans_peek_updates(trans,
2853                                                 iter->btree_id, search_key)) &&
2854                     !bpos_cmp(next_update->k.p, iter->pos)) {
2855                         iter->k = next_update->k;
2856                         k = bkey_i_to_s_c(next_update);
2857                         goto out;
2858                 }
2859
2860                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2861                     (next_update = bch2_btree_journal_peek_slot(trans,
2862                                         iter, iter->pos))) {
2863                         iter->k = next_update->k;
2864                         k = bkey_i_to_s_c(next_update);
2865                         goto out;
2866                 }
2867
2868                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2869                     (k = __btree_trans_peek_key_cache(iter, iter->pos)).k) {
2870                         if (!bkey_err(k))
2871                                 iter->k = *k.k;
2872                         /* We're not returning a key from iter->path: */
2873                         goto out_no_locked;
2874                 }
2875
2876                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2877         } else {
2878                 struct bpos next;
2879
2880                 EBUG_ON(iter->path->level);
2881
2882                 if (iter->flags & BTREE_ITER_INTENT) {
2883                         struct btree_iter iter2;
2884                         struct bpos end = iter->pos;
2885
2886                         if (iter->flags & BTREE_ITER_IS_EXTENTS)
2887                                 end.offset = U64_MAX;
2888
2889                         bch2_trans_copy_iter(&iter2, iter);
2890                         k = bch2_btree_iter_peek_upto(&iter2, end);
2891
2892                         if (k.k && !bkey_err(k)) {
2893                                 iter->k = iter2.k;
2894                                 k.k = &iter->k;
2895                         }
2896                         bch2_trans_iter_exit(trans, &iter2);
2897                 } else {
2898                         struct bpos pos = iter->pos;
2899
2900                         k = bch2_btree_iter_peek(iter);
2901                         if (unlikely(bkey_err(k)))
2902                                 bch2_btree_iter_set_pos(iter, pos);
2903                         else
2904                                 iter->pos = pos;
2905                 }
2906
2907                 if (unlikely(bkey_err(k)))
2908                         goto out_no_locked;
2909
2910                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2911
2912                 if (bkey_cmp(iter->pos, next) < 0) {
2913                         bkey_init(&iter->k);
2914                         iter->k.p = iter->pos;
2915
2916                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2917                                 bch2_key_resize(&iter->k,
2918                                                 min_t(u64, KEY_SIZE_MAX,
2919                                                       (next.inode == iter->pos.inode
2920                                                        ? next.offset
2921                                                        : KEY_OFFSET_MAX) -
2922                                                       iter->pos.offset));
2923                                 EBUG_ON(!iter->k.size);
2924                         }
2925
2926                         k = (struct bkey_s_c) { &iter->k, NULL };
2927                 }
2928         }
2929 out:
2930         btree_path_set_should_be_locked(iter->path);
2931 out_no_locked:
2932         bch2_btree_iter_verify_entry_exit(iter);
2933         bch2_btree_iter_verify(iter);
2934         ret = bch2_btree_iter_verify_ret(iter, k);
2935         if (unlikely(ret))
2936                 return bkey_s_c_err(ret);
2937
2938         return k;
2939 }
2940
2941 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2942 {
2943         if (!bch2_btree_iter_advance(iter))
2944                 return bkey_s_c_null;
2945
2946         return bch2_btree_iter_peek_slot(iter);
2947 }
2948
2949 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2950 {
2951         if (!bch2_btree_iter_rewind(iter))
2952                 return bkey_s_c_null;
2953
2954         return bch2_btree_iter_peek_slot(iter);
2955 }
2956
2957 /* new transactional stuff: */
2958
2959 static inline void btree_path_verify_sorted_ref(struct btree_trans *trans,
2960                                                 struct btree_path *path)
2961 {
2962         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2963         EBUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2964         EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
2965 }
2966
2967 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2968 {
2969 #ifdef CONFIG_BCACHEFS_DEBUG
2970         unsigned i;
2971
2972         for (i = 0; i < trans->nr_sorted; i++)
2973                 btree_path_verify_sorted_ref(trans, trans->paths + trans->sorted[i]);
2974 #endif
2975 }
2976
2977 static void btree_trans_verify_sorted(struct btree_trans *trans)
2978 {
2979 #ifdef CONFIG_BCACHEFS_DEBUG
2980         struct btree_path *path, *prev = NULL;
2981         unsigned i;
2982
2983         if (!bch2_debug_check_iterators)
2984                 return;
2985
2986         trans_for_each_path_inorder(trans, path, i) {
2987                 if (prev && btree_path_cmp(prev, path) > 0) {
2988                         bch2_dump_trans_paths_updates(trans);
2989                         panic("trans paths out of order!\n");
2990                 }
2991                 prev = path;
2992         }
2993 #endif
2994 }
2995
2996 static inline void btree_path_swap(struct btree_trans *trans,
2997                                    struct btree_path *l, struct btree_path *r)
2998 {
2999         swap(l->sorted_idx, r->sorted_idx);
3000         swap(trans->sorted[l->sorted_idx],
3001              trans->sorted[r->sorted_idx]);
3002
3003         btree_path_verify_sorted_ref(trans, l);
3004         btree_path_verify_sorted_ref(trans, r);
3005 }
3006
3007 inline void bch2_btree_path_check_sort(struct btree_trans *trans, struct btree_path *path,
3008                                        int cmp)
3009 {
3010         struct btree_path *n;
3011
3012         if (cmp <= 0) {
3013                 n = prev_btree_path(trans, path);
3014                 if (n && btree_path_cmp(n, path) > 0) {
3015                         do {
3016                                 btree_path_swap(trans, n, path);
3017                                 n = prev_btree_path(trans, path);
3018                         } while (n && btree_path_cmp(n, path) > 0);
3019
3020                         goto out;
3021                 }
3022         }
3023
3024         if (cmp >= 0) {
3025                 n = next_btree_path(trans, path);
3026                 if (n && btree_path_cmp(path, n) > 0) {
3027                         do {
3028                                 btree_path_swap(trans, path, n);
3029                                 n = next_btree_path(trans, path);
3030                         } while (n && btree_path_cmp(path, n) > 0);
3031                 }
3032         }
3033 out:
3034         btree_trans_verify_sorted(trans);
3035 }
3036
3037 static inline void btree_path_list_remove(struct btree_trans *trans,
3038                                           struct btree_path *path)
3039 {
3040         unsigned i;
3041
3042         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
3043
3044         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
3045
3046         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
3047                 trans->paths[trans->sorted[i]].sorted_idx = i;
3048
3049         path->sorted_idx = U8_MAX;
3050
3051         btree_trans_verify_sorted_refs(trans);
3052 }
3053
3054 static inline void btree_path_list_add(struct btree_trans *trans,
3055                                        struct btree_path *pos,
3056                                        struct btree_path *path)
3057 {
3058         unsigned i;
3059
3060         btree_trans_verify_sorted_refs(trans);
3061
3062         path->sorted_idx = pos ? pos->sorted_idx + 1 : 0;
3063
3064         if (trans->in_traverse_all &&
3065             trans->traverse_all_idx != U8_MAX &&
3066             trans->traverse_all_idx >= path->sorted_idx)
3067                 trans->traverse_all_idx++;
3068
3069         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
3070
3071         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
3072                 trans->paths[trans->sorted[i]].sorted_idx = i;
3073
3074         btree_trans_verify_sorted_refs(trans);
3075 }
3076
3077 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
3078 {
3079         if (iter->path)
3080                 bch2_path_put(trans, iter->path,
3081                               iter->flags & BTREE_ITER_INTENT);
3082         if (iter->update_path)
3083                 bch2_path_put(trans, iter->update_path,
3084                               iter->flags & BTREE_ITER_INTENT);
3085         if (iter->key_cache_path)
3086                 bch2_path_put(trans, iter->key_cache_path,
3087                               iter->flags & BTREE_ITER_INTENT);
3088         iter->path = NULL;
3089         iter->update_path = NULL;
3090         iter->key_cache_path = NULL;
3091 }
3092
3093 static void __bch2_trans_iter_init(struct btree_trans *trans,
3094                                    struct btree_iter *iter,
3095                                    unsigned btree_id, struct bpos pos,
3096                                    unsigned locks_want,
3097                                    unsigned depth,
3098                                    unsigned flags,
3099                                    unsigned long ip)
3100 {
3101         EBUG_ON(trans->restarted);
3102
3103         if (flags & BTREE_ITER_ALL_LEVELS)
3104                 flags |= BTREE_ITER_ALL_SNAPSHOTS|__BTREE_ITER_ALL_SNAPSHOTS;
3105
3106         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
3107             btree_node_type_is_extents(btree_id))
3108                 flags |= BTREE_ITER_IS_EXTENTS;
3109
3110         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
3111             !btree_type_has_snapshots(btree_id))
3112                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
3113
3114         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
3115             btree_type_has_snapshots(btree_id))
3116                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
3117
3118         if (!test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags))
3119                 flags |= BTREE_ITER_WITH_JOURNAL;
3120
3121         iter->trans     = trans;
3122         iter->path      = NULL;
3123         iter->update_path = NULL;
3124         iter->key_cache_path = NULL;
3125         iter->btree_id  = btree_id;
3126         iter->min_depth = depth;
3127         iter->flags     = flags;
3128         iter->snapshot  = pos.snapshot;
3129         iter->pos       = pos;
3130         iter->k.type    = KEY_TYPE_deleted;
3131         iter->k.p       = pos;
3132         iter->k.size    = 0;
3133         iter->journal_idx = 0;
3134         iter->journal_pos = POS_MIN;
3135 #ifdef CONFIG_BCACHEFS_DEBUG
3136         iter->ip_allocated = ip;
3137 #endif
3138
3139         iter->path = bch2_path_get(trans, btree_id, iter->pos,
3140                                    locks_want, depth, flags, ip);
3141 }
3142
3143 void bch2_trans_iter_init(struct btree_trans *trans,
3144                           struct btree_iter *iter,
3145                           unsigned btree_id, struct bpos pos,
3146                           unsigned flags)
3147 {
3148         if (!btree_id_cached(trans->c, btree_id)) {
3149                 flags &= ~BTREE_ITER_CACHED;
3150                 flags &= ~BTREE_ITER_WITH_KEY_CACHE;
3151         } else if (!(flags & BTREE_ITER_CACHED))
3152                 flags |= BTREE_ITER_WITH_KEY_CACHE;
3153
3154         __bch2_trans_iter_init(trans, iter, btree_id, pos,
3155                                0, 0, flags, _RET_IP_);
3156 }
3157
3158 void bch2_trans_node_iter_init(struct btree_trans *trans,
3159                                struct btree_iter *iter,
3160                                enum btree_id btree_id,
3161                                struct bpos pos,
3162                                unsigned locks_want,
3163                                unsigned depth,
3164                                unsigned flags)
3165 {
3166         __bch2_trans_iter_init(trans, iter, btree_id, pos, locks_want, depth,
3167                                BTREE_ITER_NOT_EXTENTS|
3168                                __BTREE_ITER_ALL_SNAPSHOTS|
3169                                BTREE_ITER_ALL_SNAPSHOTS|
3170                                flags, _RET_IP_);
3171         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
3172         BUG_ON(iter->path->level        != depth);
3173         BUG_ON(iter->min_depth          != depth);
3174 }
3175
3176 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
3177 {
3178         *dst = *src;
3179         if (src->path)
3180                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
3181         if (src->update_path)
3182                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
3183         dst->key_cache_path = NULL;
3184 }
3185
3186 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
3187 {
3188         size_t new_top = trans->mem_top + size;
3189         void *p;
3190
3191         if (new_top > trans->mem_bytes) {
3192                 size_t old_bytes = trans->mem_bytes;
3193                 size_t new_bytes = roundup_pow_of_two(new_top);
3194                 void *new_mem;
3195
3196                 WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
3197
3198                 new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
3199                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
3200                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
3201                         new_bytes = BTREE_TRANS_MEM_MAX;
3202                         kfree(trans->mem);
3203                 }
3204
3205                 if (!new_mem)
3206                         return ERR_PTR(-ENOMEM);
3207
3208                 trans->mem = new_mem;
3209                 trans->mem_bytes = new_bytes;
3210
3211                 if (old_bytes) {
3212                         trace_trans_restart_mem_realloced(trans, _RET_IP_, new_bytes);
3213                         return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
3214                 }
3215         }
3216
3217         p = trans->mem + trans->mem_top;
3218         trans->mem_top += size;
3219         memset(p, 0, size);
3220         return p;
3221 }
3222
3223 /**
3224  * bch2_trans_begin() - reset a transaction after a interrupted attempt
3225  * @trans: transaction to reset
3226  *
3227  * While iterating over nodes or updating nodes a attempt to lock a btree node
3228  * may return BCH_ERR_transaction_restart when the trylock fails. When this
3229  * occurs bch2_trans_begin() should be called and the transaction retried.
3230  */
3231 u32 bch2_trans_begin(struct btree_trans *trans)
3232 {
3233         struct btree_path *path;
3234
3235         bch2_trans_reset_updates(trans);
3236
3237         trans->restart_count++;
3238         trans->mem_top                  = 0;
3239
3240         if (trans->fs_usage_deltas) {
3241                 trans->fs_usage_deltas->used = 0;
3242                 memset((void *) trans->fs_usage_deltas +
3243                        offsetof(struct replicas_delta_list, memset_start), 0,
3244                        (void *) &trans->fs_usage_deltas->memset_end -
3245                        (void *) &trans->fs_usage_deltas->memset_start);
3246         }
3247
3248         trans_for_each_path(trans, path) {
3249                 path->should_be_locked = false;
3250
3251                 /*
3252                  * If the transaction wasn't restarted, we're presuming to be
3253                  * doing something new: dont keep iterators excpt the ones that
3254                  * are in use - except for the subvolumes btree:
3255                  */
3256                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
3257                         path->preserve = false;
3258
3259                 /*
3260                  * XXX: we probably shouldn't be doing this if the transaction
3261                  * was restarted, but currently we still overflow transaction
3262                  * iterators if we do that
3263                  */
3264                 if (!path->ref && !path->preserve)
3265                         __bch2_path_free(trans, path);
3266                 else
3267                         path->preserve = false;
3268         }
3269
3270         if (!trans->restarted &&
3271             (need_resched() ||
3272              ktime_get_ns() - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
3273                 bch2_trans_unlock(trans);
3274                 cond_resched();
3275                 bch2_trans_relock(trans);
3276         }
3277
3278         trans->last_restarted_ip = _RET_IP_;
3279         if (trans->restarted)
3280                 bch2_btree_path_traverse_all(trans);
3281
3282         trans->last_begin_time = ktime_get_ns();
3283         return trans->restart_count;
3284 }
3285
3286 void bch2_trans_verify_not_restarted(struct btree_trans *trans, u32 restart_count)
3287 {
3288         if (trans_was_restarted(trans, restart_count))
3289                 panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
3290                       trans->restart_count, restart_count,
3291                       (void *) trans->last_restarted_ip);
3292 }
3293
3294 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
3295 {
3296         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
3297         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
3298         void *p = NULL;
3299
3300         BUG_ON(trans->used_mempool);
3301
3302 #ifdef __KERNEL__
3303         p = this_cpu_xchg(c->btree_paths_bufs->path , NULL);
3304 #endif
3305         if (!p)
3306                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
3307
3308         trans->paths            = p; p += paths_bytes;
3309         trans->updates          = p; p += updates_bytes;
3310 }
3311
3312 static inline unsigned bch2_trans_get_fn_idx(struct btree_trans *trans, struct bch_fs *c,
3313                                         const char *fn)
3314 {
3315         unsigned i;
3316
3317         for (i = 0; i < ARRAY_SIZE(c->btree_transaction_fns); i++)
3318                 if (!c->btree_transaction_fns[i] ||
3319                     c->btree_transaction_fns[i] == fn) {
3320                         c->btree_transaction_fns[i] = fn;
3321                         return i;
3322                 }
3323
3324         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
3325         return i;
3326 }
3327
3328 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
3329                        unsigned expected_nr_iters,
3330                        size_t expected_mem_bytes,
3331                        const char *fn)
3332         __acquires(&c->btree_trans_barrier)
3333 {
3334         struct btree_trans *pos;
3335
3336         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
3337
3338         memset(trans, 0, sizeof(*trans));
3339         trans->c                = c;
3340         trans->fn               = fn;
3341         trans->last_begin_time  = ktime_get_ns();
3342         trans->task             = current;
3343         trans->fn_idx           = bch2_trans_get_fn_idx(trans, c, fn);
3344
3345         bch2_trans_alloc_paths(trans, c);
3346
3347         if (expected_mem_bytes) {
3348                 trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes);
3349                 trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL);
3350
3351                 if (!unlikely(trans->mem)) {
3352                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
3353                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
3354                 }
3355         }
3356
3357         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
3358
3359         mutex_lock(&c->btree_trans_lock);
3360         list_for_each_entry(pos, &c->btree_trans_list, list) {
3361                 if (trans->task->pid < pos->task->pid) {
3362                         list_add_tail(&trans->list, &pos->list);
3363                         goto list_add_done;
3364                 }
3365         }
3366         list_add_tail(&trans->list, &c->btree_trans_list);
3367 list_add_done:
3368         mutex_unlock(&c->btree_trans_lock);
3369 }
3370
3371 static void check_btree_paths_leaked(struct btree_trans *trans)
3372 {
3373 #ifdef CONFIG_BCACHEFS_DEBUG
3374         struct bch_fs *c = trans->c;
3375         struct btree_path *path;
3376
3377         trans_for_each_path(trans, path)
3378                 if (path->ref)
3379                         goto leaked;
3380         return;
3381 leaked:
3382         bch_err(c, "btree paths leaked from %s!", trans->fn);
3383         trans_for_each_path(trans, path)
3384                 if (path->ref)
3385                         printk(KERN_ERR "  btree %s %pS\n",
3386                                bch2_btree_ids[path->btree_id],
3387                                (void *) path->ip_allocated);
3388         /* Be noisy about this: */
3389         bch2_fatal_error(c);
3390 #endif
3391 }
3392
3393 void bch2_trans_exit(struct btree_trans *trans)
3394         __releases(&c->btree_trans_barrier)
3395 {
3396         struct btree_insert_entry *i;
3397         struct bch_fs *c = trans->c;
3398
3399         bch2_trans_unlock(trans);
3400
3401         trans_for_each_update(trans, i)
3402                 __btree_path_put(i->path, true);
3403         trans->nr_updates               = 0;
3404
3405         check_btree_paths_leaked(trans);
3406
3407         mutex_lock(&c->btree_trans_lock);
3408         list_del(&trans->list);
3409         mutex_unlock(&c->btree_trans_lock);
3410
3411         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3412
3413         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3414
3415         kfree(trans->extra_journal_entries.data);
3416
3417         if (trans->fs_usage_deltas) {
3418                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3419                     REPLICAS_DELTA_LIST_MAX)
3420                         mempool_free(trans->fs_usage_deltas,
3421                                      &c->replicas_delta_pool);
3422                 else
3423                         kfree(trans->fs_usage_deltas);
3424         }
3425
3426         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3427                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3428         else
3429                 kfree(trans->mem);
3430
3431 #ifdef __KERNEL__
3432         /*
3433          * Userspace doesn't have a real percpu implementation:
3434          */
3435         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3436 #endif
3437
3438         if (trans->paths)
3439                 mempool_free(trans->paths, &c->btree_paths_pool);
3440
3441         trans->mem      = (void *) 0x1;
3442         trans->paths    = (void *) 0x1;
3443 }
3444
3445 static void __maybe_unused
3446 bch2_btree_path_node_to_text(struct printbuf *out,
3447                              struct btree_bkey_cached_common *b,
3448                              bool cached)
3449 {
3450         prt_printf(out, "    l=%u %s:",
3451                b->level, bch2_btree_ids[b->btree_id]);
3452         bch2_bpos_to_text(out, btree_node_pos(b, cached));
3453 }
3454
3455 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3456 {
3457         struct btree_path *path;
3458         struct btree_bkey_cached_common *b;
3459         static char lock_types[] = { 'r', 'i', 'w' };
3460         unsigned l;
3461
3462         prt_printf(out, "%i %s\n", trans->task->pid, trans->fn);
3463
3464         trans_for_each_path(trans, path) {
3465                 if (!path->nodes_locked)
3466                         continue;
3467
3468                 prt_printf(out, "  path %u %c l=%u %s:",
3469                        path->idx,
3470                        path->cached ? 'c' : 'b',
3471                        path->level,
3472                        bch2_btree_ids[path->btree_id]);
3473                 bch2_bpos_to_text(out, path->pos);
3474                 prt_printf(out, "\n");
3475
3476                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3477                         if (btree_node_locked(path, l) &&
3478                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3479                                 prt_printf(out, "    %s l=%u ",
3480                                        btree_node_intent_locked(path, l) ? "i" : "r", l);
3481                                 bch2_btree_path_node_to_text(out, b, path->cached);
3482                                 prt_printf(out, "\n");
3483                         }
3484                 }
3485         }
3486
3487         b = READ_ONCE(trans->locking);
3488         if (b) {
3489                 path = &trans->paths[trans->locking_path_idx];
3490                 prt_printf(out, "  locking path %u %c l=%u %c %s:",
3491                        trans->locking_path_idx,
3492                        path->cached ? 'c' : 'b',
3493                        trans->locking_level,
3494                        lock_types[trans->locking_lock_type],
3495                        bch2_btree_ids[trans->locking_btree_id]);
3496                 bch2_bpos_to_text(out, trans->locking_pos);
3497
3498                 prt_printf(out, " node ");
3499                 bch2_btree_path_node_to_text(out, b, path->cached);
3500                 prt_printf(out, "\n");
3501         }
3502 }
3503
3504 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3505 {
3506         if (c->btree_trans_barrier_initialized)
3507                 cleanup_srcu_struct(&c->btree_trans_barrier);
3508         mempool_exit(&c->btree_trans_mem_pool);
3509         mempool_exit(&c->btree_paths_pool);
3510 }
3511
3512 int bch2_fs_btree_iter_init(struct bch_fs *c)
3513 {
3514         unsigned i, nr = BTREE_ITER_MAX;
3515         int ret;
3516
3517         for (i = 0; i < ARRAY_SIZE(c->btree_transaction_stats); i++)
3518                 mutex_init(&c->btree_transaction_stats[i].lock);
3519
3520         INIT_LIST_HEAD(&c->btree_trans_list);
3521         mutex_init(&c->btree_trans_lock);
3522
3523         ret   = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3524                         sizeof(struct btree_path) * nr +
3525                         sizeof(struct btree_insert_entry) * nr) ?:
3526                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3527                                           BTREE_TRANS_MEM_MAX) ?:
3528                 init_srcu_struct(&c->btree_trans_barrier);
3529         if (!ret)
3530                 c->btree_trans_barrier_initialized = true;
3531         return ret;
3532 }