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 "btree_cache.h"
   6 #include "btree_iter.h"
   7 #include "btree_key_cache.h"
   8 #include "btree_locking.h"
   9 #include "btree_update.h"
  10 #include "debug.h"
  11 #include "extents.h"
  12 #include "journal.h"
  13
  14 #include <linux/prefetch.h>
  15 #include <trace/events/bcachefs.h>
  16
  17 static inline bool is_btree_node(struct btree_iter *iter, unsigned l)
  18 {
  19         return l < BTREE_MAX_DEPTH &&
  20                 (unsigned long) iter->l[l].b >= 128;
  21 }
  22
  23 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
  24 {
  25         struct bpos pos = iter->pos;
  26
  27         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
  28             bkey_cmp(pos, POS_MAX))
  29                 pos = bkey_successor(pos);
  30         return pos;
  31 }
  32
  33 static inline bool btree_iter_pos_before_node(struct btree_iter *iter,
  34                                               struct btree *b)
  35 {
  36         return bkey_cmp(btree_iter_search_key(iter), b->data->min_key) < 0;
  37 }
  38
  39 static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
  40                                              struct btree *b)
  41 {
  42         return bkey_cmp(b->key.k.p, btree_iter_search_key(iter)) < 0;
  43 }
  44
  45 static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
  46                                           struct btree *b)
  47 {
  48         return iter->btree_id == b->c.btree_id &&
  49                 !btree_iter_pos_before_node(iter, b) &&
  50                 !btree_iter_pos_after_node(iter, b);
  51 }
  52
  53 /* Btree node locking: */
  54
  55 void bch2_btree_node_unlock_write(struct btree *b, struct btree_iter *iter)
  56 {
  57         bch2_btree_node_unlock_write_inlined(b, iter);
  58 }
  59
  60 void __bch2_btree_node_lock_write(struct btree *b, struct btree_iter *iter)
  61 {
  62         struct btree_iter *linked;
  63         unsigned readers = 0;
  64
  65         EBUG_ON(!btree_node_intent_locked(iter, b->c.level));
  66
  67         trans_for_each_iter(iter->trans, linked)
  68                 if (linked->l[b->c.level].b == b &&
  69                     btree_node_read_locked(linked, b->c.level))
  70                         readers++;
  71
  72         /*
  73          * Must drop our read locks before calling six_lock_write() -
  74          * six_unlock() won't do wakeups until the reader count
  75          * goes to 0, and it's safe because we have the node intent
  76          * locked:
  77          */
  78         atomic64_sub(__SIX_VAL(read_lock, readers),
  79                      &b->c.lock.state.counter);
  80         btree_node_lock_type(iter->trans->c, b, SIX_LOCK_write);
  81         atomic64_add(__SIX_VAL(read_lock, readers),
  82                      &b->c.lock.state.counter);
  83 }
  84
  85 bool __bch2_btree_node_relock(struct btree_iter *iter, unsigned level)
  86 {
  87         struct btree *b = btree_iter_node(iter, level);
  88         int want = __btree_lock_want(iter, level);
  89
  90         if (!is_btree_node(iter, level))
  91                 return false;
  92
  93         if (race_fault())
  94                 return false;
  95
  96         if (six_relock_type(&b->c.lock, want, iter->l[level].lock_seq) ||
  97             (btree_node_lock_seq_matches(iter, b, level) &&
  98              btree_node_lock_increment(iter->trans, b, level, want))) {
  99                 mark_btree_node_locked(iter, level, want);
 100                 return true;
 101         } else {
 102                 return false;
 103         }
 104 }
 105
 106 static bool bch2_btree_node_upgrade(struct btree_iter *iter, unsigned level)
 107 {
 108         struct btree *b = iter->l[level].b;
 109
 110         EBUG_ON(btree_lock_want(iter, level) != BTREE_NODE_INTENT_LOCKED);
 111
 112         if (!is_btree_node(iter, level))
 113                 return false;
 114
 115         if (btree_node_intent_locked(iter, level))
 116                 return true;
 117
 118         if (race_fault())
 119                 return false;
 120
 121         if (btree_node_locked(iter, level)
 122             ? six_lock_tryupgrade(&b->c.lock)
 123             : six_relock_type(&b->c.lock, SIX_LOCK_intent, iter->l[level].lock_seq))
 124                 goto success;
 125
 126         if (btree_node_lock_seq_matches(iter, b, level) &&
 127             btree_node_lock_increment(iter->trans, b, level, BTREE_NODE_INTENT_LOCKED)) {
 128                 btree_node_unlock(iter, level);
 129                 goto success;
 130         }
 131
 132         return false;
 133 success:
 134         mark_btree_node_intent_locked(iter, level);
 135         return true;
 136 }
 137
 138 static inline bool btree_iter_get_locks(struct btree_iter *iter,
 139                                         bool upgrade, bool trace)
 140 {
 141         unsigned l = iter->level;
 142         int fail_idx = -1;
 143
 144         do {
 145                 if (!btree_iter_node(iter, l))
 146                         break;
 147
 148                 if (!(upgrade
 149                       ? bch2_btree_node_upgrade(iter, l)
 150                       : bch2_btree_node_relock(iter, l))) {
 151                         if (trace)
 152                                 (upgrade
 153                                  ? trace_node_upgrade_fail
 154                                  : trace_node_relock_fail)(l, iter->l[l].lock_seq,
 155                                                 is_btree_node(iter, l)
 156                                                 ? 0
 157                                                 : (unsigned long) iter->l[l].b,
 158                                                 is_btree_node(iter, l)
 159                                                 ? iter->l[l].b->c.lock.state.seq
 160                                                 : 0);
 161
 162                         fail_idx = l;
 163                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
 164                 }
 165
 166                 l++;
 167         } while (l < iter->locks_want);
 168
 169         /*
 170          * When we fail to get a lock, we have to ensure that any child nodes
 171          * can't be relocked so bch2_btree_iter_traverse has to walk back up to
 172          * the node that we failed to relock:
 173          */
 174         while (fail_idx >= 0) {
 175                 btree_node_unlock(iter, fail_idx);
 176                 iter->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS;
 177                 --fail_idx;
 178         }
 179
 180         if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
 181                 iter->uptodate = BTREE_ITER_NEED_PEEK;
 182
 183         bch2_btree_trans_verify_locks(iter->trans);
 184
 185         return iter->uptodate < BTREE_ITER_NEED_RELOCK;
 186 }
 187
 188 static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b,
 189                                   enum btree_iter_type type)
 190 {
 191         return  type != BTREE_ITER_CACHED
 192                 ? container_of(_b, struct btree, c)->key.k.p
 193                 : container_of(_b, struct bkey_cached, c)->key.pos;
 194 }
 195
 196 /* Slowpath: */
 197 bool __bch2_btree_node_lock(struct btree *b, struct bpos pos,
 198                             unsigned level, struct btree_iter *iter,
 199                             enum six_lock_type type,
 200                             six_lock_should_sleep_fn should_sleep_fn, void *p,
 201                             unsigned long ip)
 202 {
 203         struct btree_trans *trans = iter->trans;
 204         struct btree_iter *linked, *deadlock_iter = NULL;
 205         u64 start_time = local_clock();
 206         unsigned reason = 9;
 207
 208         /* Check if it's safe to block: */
 209         trans_for_each_iter(trans, linked) {
 210                 if (!linked->nodes_locked)
 211                         continue;
 212
 213                 /*
 214                  * Can't block taking an intent lock if we have _any_ nodes read
 215                  * locked:
 216                  *
 217                  * - Our read lock blocks another thread with an intent lock on
 218                  *   the same node from getting a write lock, and thus from
 219                  *   dropping its intent lock
 220                  *
 221                  * - And the other thread may have multiple nodes intent locked:
 222                  *   both the node we want to intent lock, and the node we
 223                  *   already have read locked - deadlock:
 224                  */
 225                 if (type == SIX_LOCK_intent &&
 226                     linked->nodes_locked != linked->nodes_intent_locked) {
 227                         if (!(trans->nounlock)) {
 228                                 linked->locks_want = max_t(unsigned,
 229                                                 linked->locks_want,
 230                                                 __fls(linked->nodes_locked) + 1);
 231                                 if (!btree_iter_get_locks(linked, true, false)) {
 232                                         deadlock_iter = linked;
 233                                         reason = 1;
 234                                 }
 235                         } else {
 236                                 deadlock_iter = linked;
 237                                 reason = 2;
 238                         }
 239                 }
 240
 241                 if (linked->btree_id != iter->btree_id) {
 242                         if (linked->btree_id > iter->btree_id) {
 243                                 deadlock_iter = linked;
 244                                 reason = 3;
 245                         }
 246                         continue;
 247                 }
 248
 249                 /*
 250                  * Within the same btree, cached iterators come before non
 251                  * cached iterators:
 252                  */
 253                 if (btree_iter_is_cached(linked) != btree_iter_is_cached(iter)) {
 254                         if (btree_iter_is_cached(iter)) {
 255                                 deadlock_iter = linked;
 256                                 reason = 4;
 257                         }
 258                         continue;
 259                 }
 260
 261                 /*
 262                  * Interior nodes must be locked before their descendants: if
 263                  * another iterator has possible descendants locked of the node
 264                  * we're about to lock, it must have the ancestors locked too:
 265                  */
 266                 if (level > __fls(linked->nodes_locked)) {
 267                         if (!(trans->nounlock)) {
 268                                 linked->locks_want =
 269                                         max(level + 1, max_t(unsigned,
 270                                             linked->locks_want,
 271                                             iter->locks_want));
 272                                 if (!btree_iter_get_locks(linked, true, false)) {
 273                                         deadlock_iter = linked;
 274                                         reason = 5;
 275                                 }
 276                         } else {
 277                                 deadlock_iter = linked;
 278                                 reason = 6;
 279                         }
 280                 }
 281
 282                 /* Must lock btree nodes in key order: */
 283                 if (btree_node_locked(linked, level) &&
 284                     bkey_cmp(pos, btree_node_pos((void *) linked->l[level].b,
 285                                                  btree_iter_type(linked))) <= 0) {
 286                         deadlock_iter = linked;
 287                         reason = 7;
 288                 }
 289
 290                 /*
 291                  * Recheck if this is a node we already have locked - since one
 292                  * of the get_locks() calls might've successfully
 293                  * upgraded/relocked it:
 294                  */
 295                 if (linked->l[level].b == b &&
 296                     btree_node_locked_type(linked, level) >= type) {
 297                         six_lock_increment(&b->c.lock, type);
 298                         return true;
 299                 }
 300         }
 301
 302         if (unlikely(deadlock_iter)) {
 303                 trace_trans_restart_would_deadlock(iter->trans->ip, ip,
 304                                 reason,
 305                                 deadlock_iter->btree_id,
 306                                 btree_iter_type(deadlock_iter),
 307                                 iter->btree_id,
 308                                 btree_iter_type(iter));
 309                 return false;
 310         }
 311
 312         if (six_trylock_type(&b->c.lock, type))
 313                 return true;
 314
 315         if (six_lock_type(&b->c.lock, type, should_sleep_fn, p))
 316                 return false;
 317
 318         bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)],
 319                                start_time);
 320         return true;
 321 }
 322
 323 /* Btree iterator locking: */
 324
 325 #ifdef CONFIG_BCACHEFS_DEBUG
 326 static void bch2_btree_iter_verify_locks(struct btree_iter *iter)
 327 {
 328         unsigned l;
 329
 330         if (!(iter->trans->iters_linked & (1ULL << iter->idx))) {
 331                 BUG_ON(iter->nodes_locked);
 332                 return;
 333         }
 334
 335         for (l = 0; is_btree_node(iter, l); l++) {
 336                 if (iter->uptodate >= BTREE_ITER_NEED_RELOCK &&
 337                     !btree_node_locked(iter, l))
 338                         continue;
 339
 340                 BUG_ON(btree_lock_want(iter, l) !=
 341                        btree_node_locked_type(iter, l));
 342         }
 343 }
 344
 345 void bch2_btree_trans_verify_locks(struct btree_trans *trans)
 346 {
 347         struct btree_iter *iter;
 348
 349         trans_for_each_iter(trans, iter)
 350                 bch2_btree_iter_verify_locks(iter);
 351 }
 352 #else
 353 static inline void bch2_btree_iter_verify_locks(struct btree_iter *iter) {}
 354 #endif
 355
 356 __flatten
 357 bool bch2_btree_iter_relock(struct btree_iter *iter, bool trace)
 358 {
 359         return btree_iter_get_locks(iter, false, trace);
 360 }
 361
 362 bool __bch2_btree_iter_upgrade(struct btree_iter *iter,
 363                                unsigned new_locks_want)
 364 {
 365         struct btree_iter *linked;
 366
 367         EBUG_ON(iter->locks_want >= new_locks_want);
 368
 369         iter->locks_want = new_locks_want;
 370
 371         if (btree_iter_get_locks(iter, true, true))
 372                 return true;
 373
 374         /*
 375          * Ancestor nodes must be locked before child nodes, so set locks_want
 376          * on iterators that might lock ancestors before us to avoid getting
 377          * -EINTR later:
 378          */
 379         trans_for_each_iter(iter->trans, linked)
 380                 if (linked != iter &&
 381                     linked->btree_id == iter->btree_id &&
 382                     linked->locks_want < new_locks_want) {
 383                         linked->locks_want = new_locks_want;
 384                         btree_iter_get_locks(linked, true, false);
 385                 }
 386
 387         return false;
 388 }
 389
 390 bool __bch2_btree_iter_upgrade_nounlock(struct btree_iter *iter,
 391                                         unsigned new_locks_want)
 392 {
 393         unsigned l = iter->level;
 394
 395         EBUG_ON(iter->locks_want >= new_locks_want);
 396
 397         iter->locks_want = new_locks_want;
 398
 399         do {
 400                 if (!btree_iter_node(iter, l))
 401                         break;
 402
 403                 if (!bch2_btree_node_upgrade(iter, l)) {
 404                         iter->locks_want = l;
 405                         return false;
 406                 }
 407
 408                 l++;
 409         } while (l < iter->locks_want);
 410
 411         return true;
 412 }
 413
 414 void __bch2_btree_iter_downgrade(struct btree_iter *iter,
 415                                  unsigned downgrade_to)
 416 {
 417         unsigned l, new_locks_want = downgrade_to ?:
 418                 (iter->flags & BTREE_ITER_INTENT ? 1 : 0);
 419
 420         if (iter->locks_want < downgrade_to) {
 421                 iter->locks_want = new_locks_want;
 422
 423                 while (iter->nodes_locked &&
 424                        (l = __fls(iter->nodes_locked)) >= iter->locks_want) {
 425                         if (l > iter->level) {
 426                                 btree_node_unlock(iter, l);
 427                         } else {
 428                                 if (btree_node_intent_locked(iter, l)) {
 429                                         six_lock_downgrade(&iter->l[l].b->c.lock);
 430                                         iter->nodes_intent_locked ^= 1 << l;
 431                                 }
 432                                 break;
 433                         }
 434                 }
 435         }
 436
 437         bch2_btree_trans_verify_locks(iter->trans);
 438 }
 439
 440 void bch2_trans_downgrade(struct btree_trans *trans)
 441 {
 442         struct btree_iter *iter;
 443
 444         trans_for_each_iter(trans, iter)
 445                 bch2_btree_iter_downgrade(iter);
 446 }
 447
 448 /* Btree transaction locking: */
 449
 450 bool bch2_trans_relock(struct btree_trans *trans)
 451 {
 452         struct btree_iter *iter;
 453         bool ret = true;
 454
 455         trans_for_each_iter(trans, iter)
 456                 if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
 457                         ret &= bch2_btree_iter_relock(iter, true);
 458
 459         return ret;
 460 }
 461
 462 void bch2_trans_unlock(struct btree_trans *trans)
 463 {
 464         struct btree_iter *iter;
 465
 466         trans_for_each_iter(trans, iter)
 467                 __bch2_btree_iter_unlock(iter);
 468 }
 469
 470 /* Btree iterator: */
 471
 472 #ifdef CONFIG_BCACHEFS_DEBUG
 473
 474 static void bch2_btree_iter_verify_cached(struct btree_iter *iter)
 475 {
 476         struct bkey_cached *ck;
 477         bool locked = btree_node_locked(iter, 0);
 478
 479         if (!bch2_btree_node_relock(iter, 0))
 480                 return;
 481
 482         ck = (void *) iter->l[0].b;
 483         BUG_ON(ck->key.btree_id != iter->btree_id ||
 484                bkey_cmp(ck->key.pos, iter->pos));
 485
 486         if (!locked)
 487                 btree_node_unlock(iter, 0);
 488 }
 489
 490 static void bch2_btree_iter_verify_level(struct btree_iter *iter,
 491                                          unsigned level)
 492 {
 493         struct bpos pos = btree_iter_search_key(iter);
 494         struct btree_iter_level *l = &iter->l[level];
 495         struct btree_node_iter tmp = l->iter;
 496         bool locked = btree_node_locked(iter, level);
 497         struct bkey_packed *p, *k;
 498         char buf1[100], buf2[100];
 499         const char *msg;
 500
 501         if (!bch2_debug_check_iterators)
 502                 return;
 503
 504         if (btree_iter_type(iter) == BTREE_ITER_CACHED) {
 505                 if (!level)
 506                         bch2_btree_iter_verify_cached(iter);
 507                 return;
 508         }
 509
 510         BUG_ON(iter->level < iter->min_depth);
 511
 512         if (!btree_iter_node(iter, level))
 513                 return;
 514
 515         if (!bch2_btree_node_relock(iter, level))
 516                 return;
 517
 518         /*
 519          * Ideally this invariant would always be true, and hopefully in the
 520          * future it will be, but for now set_pos_same_leaf() breaks it:
 521          */
 522         BUG_ON(iter->uptodate < BTREE_ITER_NEED_TRAVERSE &&
 523                !btree_iter_pos_in_node(iter, l->b));
 524
 525         /*
 526          * node iterators don't use leaf node iterator:
 527          */
 528         if (btree_iter_type(iter) == BTREE_ITER_NODES &&
 529             level <= iter->min_depth)
 530                 goto unlock;
 531
 532         bch2_btree_node_iter_verify(&l->iter, l->b);
 533
 534         /*
 535          * For interior nodes, the iterator will have skipped past
 536          * deleted keys:
 537          *
 538          * For extents, the iterator may have skipped past deleted keys (but not
 539          * whiteouts)
 540          */
 541         p = level || btree_node_type_is_extents(iter->btree_id)
 542                 ? bch2_btree_node_iter_prev_filter(&tmp, l->b, KEY_TYPE_discard)
 543                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 544         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 545
 546         if (p && bkey_iter_pos_cmp(l->b, p, &pos) >= 0) {
 547                 msg = "before";
 548                 goto err;
 549         }
 550
 551         if (k && bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
 552                 msg = "after";
 553                 goto err;
 554         }
 555 unlock:
 556         if (!locked)
 557                 btree_node_unlock(iter, level);
 558         return;
 559 err:
 560         strcpy(buf1, "(none)");
 561         strcpy(buf2, "(none)");
 562
 563         if (p) {
 564                 struct bkey uk = bkey_unpack_key(l->b, p);
 565                 bch2_bkey_to_text(&PBUF(buf1), &uk);
 566         }
 567
 568         if (k) {
 569                 struct bkey uk = bkey_unpack_key(l->b, k);
 570                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 571         }
 572
 573         panic("iterator should be %s key at level %u:\n"
 574               "iter pos %s %llu:%llu\n"
 575               "prev key %s\n"
 576               "cur  key %s\n",
 577               msg, level,
 578               iter->flags & BTREE_ITER_IS_EXTENTS ? ">" : "=>",
 579               iter->pos.inode, iter->pos.offset,
 580               buf1, buf2);
 581 }
 582
 583 static void bch2_btree_iter_verify(struct btree_iter *iter)
 584 {
 585         unsigned i;
 586
 587         bch2_btree_trans_verify_locks(iter->trans);
 588
 589         for (i = 0; i < BTREE_MAX_DEPTH; i++)
 590                 bch2_btree_iter_verify_level(iter, i);
 591 }
 592
 593 void bch2_btree_trans_verify_iters(struct btree_trans *trans, struct btree *b)
 594 {
 595         struct btree_iter *iter;
 596
 597         if (!bch2_debug_check_iterators)
 598                 return;
 599
 600         trans_for_each_iter_with_node(trans, b, iter)
 601                 bch2_btree_iter_verify_level(iter, b->c.level);
 602 }
 603
 604 #else
 605
 606 static inline void bch2_btree_iter_verify_level(struct btree_iter *iter, unsigned l) {}
 607 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 608
 609 #endif
 610
 611 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 612                                         struct btree *b,
 613                                         struct bset_tree *t,
 614                                         struct bkey_packed *k)
 615 {
 616         struct btree_node_iter_set *set;
 617
 618         btree_node_iter_for_each(iter, set)
 619                 if (set->end == t->end_offset) {
 620                         set->k = __btree_node_key_to_offset(b, k);
 621                         bch2_btree_node_iter_sort(iter, b);
 622                         return;
 623                 }
 624
 625         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 626 }
 627
 628 static void __bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 629                                                struct btree *b,
 630                                                struct bkey_packed *where)
 631 {
 632         struct btree_iter_level *l = &iter->l[b->c.level];
 633         struct bpos pos = btree_iter_search_key(iter);
 634
 635         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 636                 return;
 637
 638         if (bkey_iter_pos_cmp(l->b, where, &pos) < 0)
 639                 bch2_btree_node_iter_advance(&l->iter, l->b);
 640
 641         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 642 }
 643
 644 void bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 645                                       struct btree *b,
 646                                       struct bkey_packed *where)
 647 {
 648         struct btree_iter *linked;
 649
 650         trans_for_each_iter_with_node(iter->trans, b, linked) {
 651                 __bch2_btree_iter_fix_key_modified(linked, b, where);
 652                 bch2_btree_iter_verify_level(linked, b->c.level);
 653         }
 654 }
 655
 656 static void __bch2_btree_node_iter_fix(struct btree_iter *iter,
 657                                       struct btree *b,
 658                                       struct btree_node_iter *node_iter,
 659                                       struct bset_tree *t,
 660                                       struct bkey_packed *where,
 661                                       unsigned clobber_u64s,
 662                                       unsigned new_u64s)
 663 {
 664         const struct bkey_packed *end = btree_bkey_last(b, t);
 665         struct btree_node_iter_set *set;
 666         unsigned offset = __btree_node_key_to_offset(b, where);
 667         int shift = new_u64s - clobber_u64s;
 668         unsigned old_end = t->end_offset - shift;
 669         unsigned orig_iter_pos = node_iter->data[0].k;
 670         bool iter_current_key_modified =
 671                 orig_iter_pos >= offset &&
 672                 orig_iter_pos <= offset + clobber_u64s;
 673         struct bpos iter_pos = btree_iter_search_key(iter);
 674
 675         btree_node_iter_for_each(node_iter, set)
 676                 if (set->end == old_end)
 677                         goto found;
 678
 679         /* didn't find the bset in the iterator - might have to readd it: */
 680         if (new_u64s &&
 681             bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
 682                 bch2_btree_node_iter_push(node_iter, b, where, end);
 683                 goto fixup_done;
 684         } else {
 685                 /* Iterator is after key that changed */
 686                 return;
 687         }
 688 found:
 689         set->end = t->end_offset;
 690
 691         /* Iterator hasn't gotten to the key that changed yet: */
 692         if (set->k < offset)
 693                 return;
 694
 695         if (new_u64s &&
 696             bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
 697                 set->k = offset;
 698         } else if (set->k < offset + clobber_u64s) {
 699                 set->k = offset + new_u64s;
 700                 if (set->k == set->end)
 701                         bch2_btree_node_iter_set_drop(node_iter, set);
 702         } else {
 703                 /* Iterator is after key that changed */
 704                 set->k = (int) set->k + shift;
 705                 return;
 706         }
 707
 708         bch2_btree_node_iter_sort(node_iter, b);
 709 fixup_done:
 710         if (node_iter->data[0].k != orig_iter_pos)
 711                 iter_current_key_modified = true;
 712
 713         /*
 714          * When a new key is added, and the node iterator now points to that
 715          * key, the iterator might have skipped past deleted keys that should
 716          * come after the key the iterator now points to. We have to rewind to
 717          * before those deleted keys - otherwise
 718          * bch2_btree_node_iter_prev_all() breaks:
 719          */
 720         if (!bch2_btree_node_iter_end(node_iter) &&
 721             iter_current_key_modified &&
 722             (b->c.level ||
 723              btree_node_type_is_extents(iter->btree_id))) {
 724                 struct bset_tree *t;
 725                 struct bkey_packed *k, *k2, *p;
 726
 727                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 728
 729                 for_each_bset(b, t) {
 730                         bool set_pos = false;
 731
 732                         if (node_iter->data[0].end == t->end_offset)
 733                                 continue;
 734
 735                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 736
 737                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 738                                bkey_iter_cmp(b, k, p) < 0) {
 739                                 k2 = p;
 740                                 set_pos = true;
 741                         }
 742
 743                         if (set_pos)
 744                                 btree_node_iter_set_set_pos(node_iter,
 745                                                             b, t, k2);
 746                 }
 747         }
 748
 749         if (!b->c.level &&
 750             node_iter == &iter->l[0].iter &&
 751             iter_current_key_modified)
 752                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 753 }
 754
 755 void bch2_btree_node_iter_fix(struct btree_iter *iter,
 756                               struct btree *b,
 757                               struct btree_node_iter *node_iter,
 758                               struct bkey_packed *where,
 759                               unsigned clobber_u64s,
 760                               unsigned new_u64s)
 761 {
 762         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 763         struct btree_iter *linked;
 764
 765         if (node_iter != &iter->l[b->c.level].iter) {
 766                 __bch2_btree_node_iter_fix(iter, b, node_iter, t,
 767                                            where, clobber_u64s, new_u64s);
 768
 769                 if (bch2_debug_check_iterators)
 770                         bch2_btree_node_iter_verify(node_iter, b);
 771         }
 772
 773         trans_for_each_iter_with_node(iter->trans, b, linked) {
 774                 __bch2_btree_node_iter_fix(linked, b,
 775                                            &linked->l[b->c.level].iter, t,
 776                                            where, clobber_u64s, new_u64s);
 777                 bch2_btree_iter_verify_level(linked, b->c.level);
 778         }
 779 }
 780
 781 static inline struct bkey_s_c __btree_iter_unpack(struct btree_iter *iter,
 782                                                   struct btree_iter_level *l,
 783                                                   struct bkey *u,
 784                                                   struct bkey_packed *k)
 785 {
 786         struct bkey_s_c ret;
 787
 788         if (unlikely(!k)) {
 789                 /*
 790                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 791                  * a hole
 792                  */
 793                 u->type = KEY_TYPE_deleted;
 794                 return bkey_s_c_null;
 795         }
 796
 797         ret = bkey_disassemble(l->b, k, u);
 798
 799         if (bch2_debug_check_bkeys)
 800                 bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
 801
 802         return ret;
 803 }
 804
 805 /* peek_all() doesn't skip deleted keys */
 806 static inline struct bkey_s_c __btree_iter_peek_all(struct btree_iter *iter,
 807                                                     struct btree_iter_level *l,
 808                                                     struct bkey *u)
 809 {
 810         return __btree_iter_unpack(iter, l, u,
 811                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 812 }
 813
 814 static inline struct bkey_s_c __btree_iter_peek(struct btree_iter *iter,
 815                                                 struct btree_iter_level *l)
 816 {
 817         return __btree_iter_unpack(iter, l, &iter->k,
 818                         bch2_btree_node_iter_peek(&l->iter, l->b));
 819 }
 820
 821 static inline struct bkey_s_c __btree_iter_prev(struct btree_iter *iter,
 822                                                 struct btree_iter_level *l)
 823 {
 824         return __btree_iter_unpack(iter, l, &iter->k,
 825                         bch2_btree_node_iter_prev(&l->iter, l->b));
 826 }
 827
 828 static inline bool btree_iter_advance_to_pos(struct btree_iter *iter,
 829                                              struct btree_iter_level *l,
 830                                              int max_advance)
 831 {
 832         struct bpos pos = btree_iter_search_key(iter);
 833         struct bkey_packed *k;
 834         int nr_advanced = 0;
 835
 836         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 837                bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
 838                 if (max_advance > 0 && nr_advanced >= max_advance)
 839                         return false;
 840
 841                 bch2_btree_node_iter_advance(&l->iter, l->b);
 842                 nr_advanced++;
 843         }
 844
 845         return true;
 846 }
 847
 848 /*
 849  * Verify that iterator for parent node points to child node:
 850  */
 851 static void btree_iter_verify_new_node(struct btree_iter *iter, struct btree *b)
 852 {
 853         struct btree_iter_level *l;
 854         unsigned plevel;
 855         bool parent_locked;
 856         struct bkey_packed *k;
 857
 858         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 859                 return;
 860
 861         plevel = b->c.level + 1;
 862         if (!btree_iter_node(iter, plevel))
 863                 return;
 864
 865         parent_locked = btree_node_locked(iter, plevel);
 866
 867         if (!bch2_btree_node_relock(iter, plevel))
 868                 return;
 869
 870         l = &iter->l[plevel];
 871         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 872         if (!k ||
 873             bkey_deleted(k) ||
 874             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
 875                 char buf[100];
 876                 struct bkey uk = bkey_unpack_key(b, k);
 877
 878                 bch2_bkey_to_text(&PBUF(buf), &uk);
 879                 panic("parent iter doesn't point to new node:\n%s\n%llu:%llu\n",
 880                       buf, b->key.k.p.inode, b->key.k.p.offset);
 881         }
 882
 883         if (!parent_locked)
 884                 btree_node_unlock(iter, b->c.level + 1);
 885 }
 886
 887 static inline void __btree_iter_init(struct btree_iter *iter,
 888                                      unsigned level)
 889 {
 890         struct bpos pos = btree_iter_search_key(iter);
 891         struct btree_iter_level *l = &iter->l[level];
 892
 893         bch2_btree_node_iter_init(&l->iter, l->b, &pos);
 894
 895         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 896 }
 897
 898 static inline void btree_iter_node_set(struct btree_iter *iter,
 899                                        struct btree *b)
 900 {
 901         BUG_ON(btree_iter_type(iter) == BTREE_ITER_CACHED);
 902
 903         btree_iter_verify_new_node(iter, b);
 904
 905         EBUG_ON(!btree_iter_pos_in_node(iter, b));
 906         EBUG_ON(b->c.lock.state.seq & 1);
 907
 908         iter->l[b->c.level].lock_seq = b->c.lock.state.seq;
 909         iter->l[b->c.level].b = b;
 910         __btree_iter_init(iter, b->c.level);
 911 }
 912
 913 /*
 914  * A btree node is being replaced - update the iterator to point to the new
 915  * node:
 916  */
 917 void bch2_btree_iter_node_replace(struct btree_iter *iter, struct btree *b)
 918 {
 919         enum btree_node_locked_type t;
 920         struct btree_iter *linked;
 921
 922         trans_for_each_iter(iter->trans, linked)
 923                 if (btree_iter_type(linked) != BTREE_ITER_CACHED &&
 924                     btree_iter_pos_in_node(linked, b)) {
 925                         /*
 926                          * bch2_btree_iter_node_drop() has already been called -
 927                          * the old node we're replacing has already been
 928                          * unlocked and the pointer invalidated
 929                          */
 930                         BUG_ON(btree_node_locked(linked, b->c.level));
 931
 932                         t = btree_lock_want(linked, b->c.level);
 933                         if (t != BTREE_NODE_UNLOCKED) {
 934                                 six_lock_increment(&b->c.lock, t);
 935                                 mark_btree_node_locked(linked, b->c.level, t);
 936                         }
 937
 938                         btree_iter_node_set(linked, b);
 939                 }
 940 }
 941
 942 void bch2_btree_iter_node_drop(struct btree_iter *iter, struct btree *b)
 943 {
 944         struct btree_iter *linked;
 945         unsigned level = b->c.level;
 946
 947         trans_for_each_iter(iter->trans, linked)
 948                 if (linked->l[level].b == b) {
 949                         __btree_node_unlock(linked, level);
 950                         linked->l[level].b = BTREE_ITER_NO_NODE_DROP;
 951                 }
 952 }
 953
 954 /*
 955  * A btree node has been modified in such a way as to invalidate iterators - fix
 956  * them:
 957  */
 958 void bch2_btree_iter_reinit_node(struct btree_iter *iter, struct btree *b)
 959 {
 960         struct btree_iter *linked;
 961
 962         trans_for_each_iter_with_node(iter->trans, b, linked)
 963                 __btree_iter_init(linked, b->c.level);
 964 }
 965
 966 static int lock_root_check_fn(struct six_lock *lock, void *p)
 967 {
 968         struct btree *b = container_of(lock, struct btree, c.lock);
 969         struct btree **rootp = p;
 970
 971         return b == *rootp ? 0 : -1;
 972 }
 973
 974 static inline int btree_iter_lock_root(struct btree_iter *iter,
 975                                        unsigned depth_want,
 976                                        unsigned long trace_ip)
 977 {
 978         struct bch_fs *c = iter->trans->c;
 979         struct btree *b, **rootp = &c->btree_roots[iter->btree_id].b;
 980         enum six_lock_type lock_type;
 981         unsigned i;
 982
 983         EBUG_ON(iter->nodes_locked);
 984
 985         while (1) {
 986                 b = READ_ONCE(*rootp);
 987                 iter->level = READ_ONCE(b->c.level);
 988
 989                 if (unlikely(iter->level < depth_want)) {
 990                         /*
 991                          * the root is at a lower depth than the depth we want:
 992                          * got to the end of the btree, or we're walking nodes
 993                          * greater than some depth and there are no nodes >=
 994                          * that depth
 995                          */
 996                         iter->level = depth_want;
 997                         for (i = iter->level; i < BTREE_MAX_DEPTH; i++)
 998                                 iter->l[i].b = NULL;
 999                         return 1;
1000                 }
1001
1002                 lock_type = __btree_lock_want(iter, iter->level);
1003                 if (unlikely(!btree_node_lock(b, POS_MAX, iter->level,
1004                                               iter, lock_type,
1005                                               lock_root_check_fn, rootp,
1006                                               trace_ip)))
1007                         return -EINTR;
1008
1009                 if (likely(b == READ_ONCE(*rootp) &&
1010                            b->c.level == iter->level &&
1011                            !race_fault())) {
1012                         for (i = 0; i < iter->level; i++)
1013                                 iter->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
1014                         iter->l[iter->level].b = b;
1015                         for (i = iter->level + 1; i < BTREE_MAX_DEPTH; i++)
1016                                 iter->l[i].b = NULL;
1017
1018                         mark_btree_node_locked(iter, iter->level, lock_type);
1019                         btree_iter_node_set(iter, b);
1020                         return 0;
1021                 }
1022
1023                 six_unlock_type(&b->c.lock, lock_type);
1024         }
1025 }
1026
1027 noinline
1028 static void btree_iter_prefetch(struct btree_iter *iter)
1029 {
1030         struct bch_fs *c = iter->trans->c;
1031         struct btree_iter_level *l = &iter->l[iter->level];
1032         struct btree_node_iter node_iter = l->iter;
1033         struct bkey_packed *k;
1034         BKEY_PADDED(k) tmp;
1035         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1036                 ? (iter->level > 1 ? 0 :  2)
1037                 : (iter->level > 1 ? 1 : 16);
1038         bool was_locked = btree_node_locked(iter, iter->level);
1039
1040         while (nr) {
1041                 if (!bch2_btree_node_relock(iter, iter->level))
1042                         return;
1043
1044                 bch2_btree_node_iter_advance(&node_iter, l->b);
1045                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
1046                 if (!k)
1047                         break;
1048
1049                 bch2_bkey_unpack(l->b, &tmp.k, k);
1050                 bch2_btree_node_prefetch(c, iter, &tmp.k, iter->level - 1);
1051         }
1052
1053         if (!was_locked)
1054                 btree_node_unlock(iter, iter->level);
1055 }
1056
1057 static noinline void btree_node_mem_ptr_set(struct btree_iter *iter,
1058                                             unsigned plevel, struct btree *b)
1059 {
1060         struct btree_iter_level *l = &iter->l[plevel];
1061         bool locked = btree_node_locked(iter, plevel);
1062         struct bkey_packed *k;
1063         struct bch_btree_ptr_v2 *bp;
1064
1065         if (!bch2_btree_node_relock(iter, plevel))
1066                 return;
1067
1068         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1069         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
1070
1071         bp = (void *) bkeyp_val(&l->b->format, k);
1072         bp->mem_ptr = (unsigned long)b;
1073
1074         if (!locked)
1075                 btree_node_unlock(iter, plevel);
1076 }
1077
1078 static __always_inline int btree_iter_down(struct btree_iter *iter,
1079                                            unsigned long trace_ip)
1080 {
1081         struct bch_fs *c = iter->trans->c;
1082         struct btree_iter_level *l = &iter->l[iter->level];
1083         struct btree *b;
1084         unsigned level = iter->level - 1;
1085         enum six_lock_type lock_type = __btree_lock_want(iter, level);
1086         BKEY_PADDED(k) tmp;
1087
1088         EBUG_ON(!btree_node_locked(iter, iter->level));
1089
1090         bch2_bkey_unpack(l->b, &tmp.k,
1091                          bch2_btree_node_iter_peek(&l->iter, l->b));
1092
1093         b = bch2_btree_node_get(c, iter, &tmp.k, level, lock_type, trace_ip);
1094         if (unlikely(IS_ERR(b)))
1095                 return PTR_ERR(b);
1096
1097         mark_btree_node_locked(iter, level, lock_type);
1098         btree_iter_node_set(iter, b);
1099
1100         if (tmp.k.k.type == KEY_TYPE_btree_ptr_v2 &&
1101             unlikely(b != btree_node_mem_ptr(&tmp.k)))
1102                 btree_node_mem_ptr_set(iter, level + 1, b);
1103
1104         if (iter->flags & BTREE_ITER_PREFETCH)
1105                 btree_iter_prefetch(iter);
1106
1107         iter->level = level;
1108
1109         return 0;
1110 }
1111
1112 static void btree_iter_up(struct btree_iter *iter)
1113 {
1114         btree_node_unlock(iter, iter->level++);
1115 }
1116
1117 static int btree_iter_traverse_one(struct btree_iter *, unsigned long);
1118
1119 static int __btree_iter_traverse_all(struct btree_trans *trans, int ret)
1120 {
1121         struct bch_fs *c = trans->c;
1122         struct btree_iter *iter;
1123         u8 sorted[BTREE_ITER_MAX];
1124         unsigned i, nr_sorted = 0;
1125
1126         if (trans->in_traverse_all)
1127                 return -EINTR;
1128
1129         trans->in_traverse_all = true;
1130 retry_all:
1131         nr_sorted = 0;
1132
1133         trans_for_each_iter(trans, iter)
1134                 sorted[nr_sorted++] = iter->idx;
1135
1136 #define btree_iter_cmp_by_idx(_l, _r)                           \
1137                 btree_iter_lock_cmp(&trans->iters[_l], &trans->iters[_r])
1138
1139         bubble_sort(sorted, nr_sorted, btree_iter_cmp_by_idx);
1140 #undef btree_iter_cmp_by_idx
1141         bch2_trans_unlock(trans);
1142         cond_resched();
1143
1144         if (unlikely(ret == -ENOMEM)) {
1145                 struct closure cl;
1146
1147                 closure_init_stack(&cl);
1148
1149                 do {
1150                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1151                         closure_sync(&cl);
1152                 } while (ret);
1153         }
1154
1155         if (unlikely(ret == -EIO)) {
1156                 trans->error = true;
1157                 goto out;
1158         }
1159
1160         BUG_ON(ret && ret != -EINTR);
1161
1162         /* Now, redo traversals in correct order: */
1163         for (i = 0; i < nr_sorted; i++) {
1164                 unsigned idx = sorted[i];
1165
1166                 /*
1167                  * sucessfully traversing one iterator can cause another to be
1168                  * unlinked, in btree_key_cache_fill()
1169                  */
1170                 if (!(trans->iters_linked & (1ULL << idx)))
1171                         continue;
1172
1173                 ret = btree_iter_traverse_one(&trans->iters[idx], _THIS_IP_);
1174                 if (ret)
1175                         goto retry_all;
1176         }
1177
1178         if (hweight64(trans->iters_live) > 1)
1179                 ret = -EINTR;
1180         else
1181                 trans_for_each_iter(trans, iter)
1182                         if (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) {
1183                                 ret = -EINTR;
1184                                 break;
1185                         }
1186 out:
1187         bch2_btree_cache_cannibalize_unlock(c);
1188
1189         trans->in_traverse_all = false;
1190         return ret;
1191 }
1192
1193 int bch2_btree_iter_traverse_all(struct btree_trans *trans)
1194 {
1195         return __btree_iter_traverse_all(trans, 0);
1196 }
1197
1198 static inline bool btree_iter_good_node(struct btree_iter *iter,
1199                                         unsigned l, int check_pos)
1200 {
1201         if (!is_btree_node(iter, l) ||
1202             !bch2_btree_node_relock(iter, l))
1203                 return false;
1204
1205         if (check_pos <= 0 && btree_iter_pos_before_node(iter, iter->l[l].b))
1206                 return false;
1207         if (check_pos >= 0 && btree_iter_pos_after_node(iter, iter->l[l].b))
1208                 return false;
1209         return true;
1210 }
1211
1212 static inline unsigned btree_iter_up_until_good_node(struct btree_iter *iter,
1213                                                      int check_pos)
1214 {
1215         unsigned l = iter->level;
1216
1217         while (btree_iter_node(iter, l) &&
1218                !btree_iter_good_node(iter, l, check_pos)) {
1219                 btree_node_unlock(iter, l);
1220                 iter->l[l].b = BTREE_ITER_NO_NODE_UP;
1221                 l++;
1222         }
1223
1224         return l;
1225 }
1226
1227 /*
1228  * This is the main state machine for walking down the btree - walks down to a
1229  * specified depth
1230  *
1231  * Returns 0 on success, -EIO on error (error reading in a btree node).
1232  *
1233  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1234  * stashed in the iterator and returned from bch2_trans_exit().
1235  */
1236 static int btree_iter_traverse_one(struct btree_iter *iter,
1237                                    unsigned long trace_ip)
1238 {
1239         unsigned depth_want = iter->level;
1240
1241         /*
1242          * if we need interior nodes locked, call btree_iter_relock() to make
1243          * sure we walk back up enough that we lock them:
1244          */
1245         if (iter->uptodate == BTREE_ITER_NEED_RELOCK ||
1246             iter->locks_want > 1)
1247                 bch2_btree_iter_relock(iter, false);
1248
1249         if (btree_iter_type(iter) == BTREE_ITER_CACHED)
1250                 return bch2_btree_iter_traverse_cached(iter);
1251
1252         if (iter->uptodate < BTREE_ITER_NEED_RELOCK)
1253                 return 0;
1254
1255         if (unlikely(iter->level >= BTREE_MAX_DEPTH))
1256                 return 0;
1257
1258         /*
1259          * XXX: correctly using BTREE_ITER_UPTODATE should make using check_pos
1260          * here unnecessary
1261          */
1262         iter->level = btree_iter_up_until_good_node(iter, 0);
1263
1264         /*
1265          * If we've got a btree node locked (i.e. we aren't about to relock the
1266          * root) - advance its node iterator if necessary:
1267          *
1268          * XXX correctly using BTREE_ITER_UPTODATE should make this unnecessary
1269          */
1270         if (is_btree_node(iter, iter->level)) {
1271                 BUG_ON(!btree_iter_pos_in_node(iter, iter->l[iter->level].b));
1272
1273                 btree_iter_advance_to_pos(iter, &iter->l[iter->level], -1);
1274         }
1275
1276         /*
1277          * Note: iter->nodes[iter->level] may be temporarily NULL here - that
1278          * would indicate to other code that we got to the end of the btree,
1279          * here it indicates that relocking the root failed - it's critical that
1280          * btree_iter_lock_root() comes next and that it can't fail
1281          */
1282         while (iter->level > depth_want) {
1283                 int ret = btree_iter_node(iter, iter->level)
1284                         ? btree_iter_down(iter, trace_ip)
1285                         : btree_iter_lock_root(iter, depth_want, trace_ip);
1286                 if (unlikely(ret)) {
1287                         if (ret == 1)
1288                                 return 0;
1289
1290                         iter->level = depth_want;
1291
1292                         if (ret == -EIO) {
1293                                 iter->flags |= BTREE_ITER_ERROR;
1294                                 iter->l[iter->level].b =
1295                                         BTREE_ITER_NO_NODE_ERROR;
1296                         } else {
1297                                 iter->l[iter->level].b =
1298                                         BTREE_ITER_NO_NODE_DOWN;
1299                         }
1300                         return ret;
1301                 }
1302         }
1303
1304         iter->uptodate = BTREE_ITER_NEED_PEEK;
1305
1306         bch2_btree_iter_verify(iter);
1307         return 0;
1308 }
1309
1310 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
1311 {
1312         struct btree_trans *trans = iter->trans;
1313         int ret;
1314
1315         ret =   bch2_trans_cond_resched(trans) ?:
1316                 btree_iter_traverse_one(iter, _RET_IP_);
1317         if (unlikely(ret))
1318                 ret = __btree_iter_traverse_all(trans, ret);
1319
1320         return ret;
1321 }
1322
1323 static inline void bch2_btree_iter_checks(struct btree_iter *iter)
1324 {
1325         enum btree_iter_type type = btree_iter_type(iter);
1326
1327         EBUG_ON(iter->btree_id >= BTREE_ID_NR);
1328
1329         BUG_ON((type == BTREE_ITER_KEYS ||
1330                 type == BTREE_ITER_CACHED) &&
1331                (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
1332                 bkey_cmp(iter->pos, iter->k.p) > 0));
1333
1334         bch2_btree_iter_verify_locks(iter);
1335         bch2_btree_iter_verify_level(iter, iter->level);
1336 }
1337
1338 /* Iterate across nodes (leaf and interior nodes) */
1339
1340 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1341 {
1342         struct btree *b;
1343         int ret;
1344
1345         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
1346         bch2_btree_iter_checks(iter);
1347
1348         if (iter->uptodate == BTREE_ITER_UPTODATE)
1349                 return iter->l[iter->level].b;
1350
1351         ret = bch2_btree_iter_traverse(iter);
1352         if (ret)
1353                 return NULL;
1354
1355         b = btree_iter_node(iter, iter->level);
1356         if (!b)
1357                 return NULL;
1358
1359         BUG_ON(bkey_cmp(b->key.k.p, iter->pos) < 0);
1360
1361         iter->pos = b->key.k.p;
1362         iter->uptodate = BTREE_ITER_UPTODATE;
1363
1364         bch2_btree_iter_verify(iter);
1365
1366         return b;
1367 }
1368
1369 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1370 {
1371         struct btree *b;
1372         int ret;
1373
1374         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
1375         bch2_btree_iter_checks(iter);
1376
1377         /* already got to end? */
1378         if (!btree_iter_node(iter, iter->level))
1379                 return NULL;
1380
1381         bch2_trans_cond_resched(iter->trans);
1382
1383         btree_iter_up(iter);
1384
1385         if (!bch2_btree_node_relock(iter, iter->level))
1386                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
1387
1388         ret = bch2_btree_iter_traverse(iter);
1389         if (ret)
1390                 return NULL;
1391
1392         /* got to end? */
1393         b = btree_iter_node(iter, iter->level);
1394         if (!b)
1395                 return NULL;
1396
1397         if (bkey_cmp(iter->pos, b->key.k.p) < 0) {
1398                 /*
1399                  * Haven't gotten to the end of the parent node: go back down to
1400                  * the next child node
1401                  */
1402
1403                 /*
1404                  * We don't really want to be unlocking here except we can't
1405                  * directly tell btree_iter_traverse() "traverse to this level"
1406                  * except by setting iter->level, so we have to unlock so we
1407                  * don't screw up our lock invariants:
1408                  */
1409                 if (btree_node_read_locked(iter, iter->level))
1410                         btree_node_unlock(iter, iter->level);
1411
1412                 iter->pos       = bkey_successor(iter->pos);
1413                 iter->level     = iter->min_depth;
1414
1415                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1416                 ret = bch2_btree_iter_traverse(iter);
1417                 if (ret)
1418                         return NULL;
1419
1420                 b = iter->l[iter->level].b;
1421         }
1422
1423         iter->pos = b->key.k.p;
1424         iter->uptodate = BTREE_ITER_UPTODATE;
1425
1426         bch2_btree_iter_verify(iter);
1427
1428         return b;
1429 }
1430
1431 /* Iterate across keys (in leaf nodes only) */
1432
1433 void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_pos)
1434 {
1435         struct btree_iter_level *l = &iter->l[0];
1436
1437         EBUG_ON(iter->level != 0);
1438         EBUG_ON(bkey_cmp(new_pos, iter->pos) < 0);
1439         EBUG_ON(!btree_node_locked(iter, 0));
1440         EBUG_ON(bkey_cmp(new_pos, l->b->key.k.p) > 0);
1441
1442         bkey_init(&iter->k);
1443         iter->k.p = iter->pos = new_pos;
1444         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
1445
1446         btree_iter_advance_to_pos(iter, l, -1);
1447
1448         /*
1449          * XXX:
1450          * keeping a node locked that's outside (even just outside) iter->pos
1451          * breaks __bch2_btree_node_lock(). This seems to only affect
1452          * bch2_btree_node_get_sibling so for now it's fixed there, but we
1453          * should try to get rid of this corner case.
1454          *
1455          * (this behaviour is currently needed for BTREE_INSERT_NOUNLOCK)
1456          */
1457
1458         if (bch2_btree_node_iter_end(&l->iter) &&
1459             btree_iter_pos_after_node(iter, l->b))
1460                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1461 }
1462
1463 static void btree_iter_pos_changed(struct btree_iter *iter, int cmp)
1464 {
1465         unsigned l = iter->level;
1466
1467         if (!cmp)
1468                 goto out;
1469
1470         if (unlikely(btree_iter_type(iter) == BTREE_ITER_CACHED)) {
1471                 btree_node_unlock(iter, 0);
1472                 iter->l[0].b = BTREE_ITER_NO_NODE_UP;
1473                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1474                 return;
1475         }
1476
1477         l = btree_iter_up_until_good_node(iter, cmp);
1478
1479         if (btree_iter_node(iter, l)) {
1480                 /*
1481                  * We might have to skip over many keys, or just a few: try
1482                  * advancing the node iterator, and if we have to skip over too
1483                  * many keys just reinit it (or if we're rewinding, since that
1484                  * is expensive).
1485                  */
1486                 if (cmp < 0 ||
1487                     !btree_iter_advance_to_pos(iter, &iter->l[l], 8))
1488                         __btree_iter_init(iter, l);
1489
1490                 /* Don't leave it locked if we're not supposed to: */
1491                 if (btree_lock_want(iter, l) == BTREE_NODE_UNLOCKED)
1492                         btree_node_unlock(iter, l);
1493         }
1494 out:
1495         if (l != iter->level)
1496                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1497         else
1498                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
1499 }
1500
1501 void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos,
1502                                bool strictly_greater)
1503 {
1504         struct bpos old = btree_iter_search_key(iter);
1505         int cmp;
1506
1507         iter->flags &= ~BTREE_ITER_IS_EXTENTS;
1508         iter->flags |= strictly_greater ? BTREE_ITER_IS_EXTENTS : 0;
1509
1510         bkey_init(&iter->k);
1511         iter->k.p = iter->pos = new_pos;
1512
1513         cmp = bkey_cmp(btree_iter_search_key(iter), old);
1514
1515         btree_iter_pos_changed(iter, cmp);
1516 }
1517
1518 void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
1519 {
1520         int cmp = bkey_cmp(new_pos, iter->pos);
1521
1522         bkey_init(&iter->k);
1523         iter->k.p = iter->pos = new_pos;
1524
1525         btree_iter_pos_changed(iter, cmp);
1526 }
1527
1528 static inline bool btree_iter_set_pos_to_next_leaf(struct btree_iter *iter)
1529 {
1530         struct btree_iter_level *l = &iter->l[0];
1531         bool ret;
1532
1533         bkey_init(&iter->k);
1534         iter->k.p = iter->pos = l->b->key.k.p;
1535
1536         ret = bkey_cmp(iter->pos, POS_MAX) != 0;
1537         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1538                 iter->k.p = iter->pos = bkey_successor(iter->pos);
1539
1540         btree_iter_pos_changed(iter, 1);
1541         return ret;
1542 }
1543
1544 static inline bool btree_iter_set_pos_to_prev_leaf(struct btree_iter *iter)
1545 {
1546         struct btree_iter_level *l = &iter->l[0];
1547         bool ret;
1548
1549         bkey_init(&iter->k);
1550         iter->k.p = iter->pos = l->b->data->min_key;
1551         iter->uptodate  = BTREE_ITER_NEED_TRAVERSE;
1552
1553         ret = bkey_cmp(iter->pos, POS_MIN) != 0;
1554         if (ret) {
1555                 iter->k.p = iter->pos = bkey_predecessor(iter->pos);
1556
1557                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
1558                         iter->k.p = iter->pos = bkey_predecessor(iter->pos);
1559         }
1560
1561         btree_iter_pos_changed(iter, -1);
1562         return ret;
1563 }
1564
1565 /**
1566  * btree_iter_peek_uptodate - given an iterator that is uptodate, return the key
1567  * it currently points to
1568  */
1569 static inline struct bkey_s_c btree_iter_peek_uptodate(struct btree_iter *iter)
1570 {
1571         struct btree_iter_level *l = &iter->l[0];
1572         struct bkey_s_c ret = { .k = &iter->k };
1573
1574         if (!bkey_deleted(&iter->k)) {
1575                 struct bkey_packed *_k =
1576                         __bch2_btree_node_iter_peek_all(&l->iter, l->b);
1577
1578                 ret.v = bkeyp_val(&l->b->format, _k);
1579
1580                 if (bch2_debug_check_iterators) {
1581                         struct bkey k = bkey_unpack_key(l->b, _k);
1582
1583                         BUG_ON(memcmp(&k, &iter->k, sizeof(k)));
1584                 }
1585
1586                 if (bch2_debug_check_bkeys)
1587                         bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
1588         }
1589
1590         return ret;
1591 }
1592
1593 /**
1594  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
1595  * current position
1596  */
1597 struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
1598 {
1599         struct btree_iter_level *l = &iter->l[0];
1600         struct bkey_s_c k;
1601         int ret;
1602
1603         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1604         bch2_btree_iter_checks(iter);
1605
1606         if (iter->uptodate == BTREE_ITER_UPTODATE &&
1607             !bkey_deleted(&iter->k))
1608                 return btree_iter_peek_uptodate(iter);
1609
1610         while (1) {
1611                 ret = bch2_btree_iter_traverse(iter);
1612                 if (unlikely(ret))
1613                         return bkey_s_c_err(ret);
1614
1615                 k = __btree_iter_peek(iter, l);
1616                 if (likely(k.k))
1617                         break;
1618
1619                 if (!btree_iter_set_pos_to_next_leaf(iter))
1620                         return bkey_s_c_null;
1621         }
1622
1623         /*
1624          * iter->pos should always be equal to the key we just
1625          * returned - except extents can straddle iter->pos:
1626          */
1627         if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
1628             bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
1629                 iter->pos = bkey_start_pos(k.k);
1630
1631         iter->uptodate = BTREE_ITER_UPTODATE;
1632
1633         bch2_btree_iter_verify_level(iter, 0);
1634         return k;
1635 }
1636
1637 /**
1638  * bch2_btree_iter_next: returns first key greater than iterator's current
1639  * position
1640  */
1641 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
1642 {
1643         if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
1644                 return bkey_s_c_null;
1645
1646         bch2_btree_iter_set_pos(iter,
1647                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1648                 ? iter->k.p
1649                 : bkey_successor(iter->k.p));
1650
1651         return bch2_btree_iter_peek(iter);
1652 }
1653
1654 static struct bkey_s_c __btree_trans_updates_peek(struct btree_iter *iter)
1655 {
1656         struct bpos pos = btree_iter_search_key(iter);
1657         struct btree_trans *trans = iter->trans;
1658         struct btree_insert_entry *i;
1659
1660         trans_for_each_update2(trans, i)
1661                 if ((cmp_int(iter->btree_id,    i->iter->btree_id) ?:
1662                      bkey_cmp(pos,              i->k->k.p)) <= 0)
1663                         break;
1664
1665         return i < trans->updates2 + trans->nr_updates2 &&
1666                 iter->btree_id == i->iter->btree_id
1667                 ? bkey_i_to_s_c(i->k)
1668                 : bkey_s_c_null;
1669 }
1670
1671 static struct bkey_s_c __bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
1672 {
1673         struct btree_iter_level *l = &iter->l[0];
1674         struct bkey_s_c k = __btree_iter_peek(iter, l);
1675         struct bkey_s_c u = __btree_trans_updates_peek(iter);
1676
1677         if (k.k && (!u.k || bkey_cmp(k.k->p, u.k->p) < 0))
1678                 return k;
1679         if (u.k && bkey_cmp(u.k->p, l->b->key.k.p) <= 0) {
1680                 iter->k = *u.k;
1681                 return u;
1682         }
1683         return bkey_s_c_null;
1684 }
1685
1686 struct bkey_s_c bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
1687 {
1688         struct bkey_s_c k;
1689         int ret;
1690
1691         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1692         bch2_btree_iter_checks(iter);
1693
1694         while (1) {
1695                 ret = bch2_btree_iter_traverse(iter);
1696                 if (unlikely(ret))
1697                         return bkey_s_c_err(ret);
1698
1699                 k = __bch2_btree_iter_peek_with_updates(iter);
1700
1701                 if (k.k && bkey_deleted(k.k)) {
1702                         bch2_btree_iter_set_pos(iter,
1703                                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1704                                 ? iter->k.p
1705                                 : bkey_successor(iter->k.p));
1706                         continue;
1707                 }
1708
1709                 if (likely(k.k))
1710                         break;
1711
1712                 if (!btree_iter_set_pos_to_next_leaf(iter))
1713                         return bkey_s_c_null;
1714         }
1715
1716         /*
1717          * iter->pos should always be equal to the key we just
1718          * returned - except extents can straddle iter->pos:
1719          */
1720         if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
1721             bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
1722                 iter->pos = bkey_start_pos(k.k);
1723
1724         iter->uptodate = BTREE_ITER_UPTODATE;
1725         return k;
1726 }
1727
1728 struct bkey_s_c bch2_btree_iter_next_with_updates(struct btree_iter *iter)
1729 {
1730         if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
1731                 return bkey_s_c_null;
1732
1733         bch2_btree_iter_set_pos(iter,
1734                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1735                 ? iter->k.p
1736                 : bkey_successor(iter->k.p));
1737
1738         return bch2_btree_iter_peek_with_updates(iter);
1739 }
1740
1741 /**
1742  * bch2_btree_iter_peek_prev: returns first key less than or equal to
1743  * iterator's current position
1744  */
1745 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
1746 {
1747         struct bpos pos = iter->pos;
1748         struct btree_iter_level *l = &iter->l[0];
1749         struct bkey_s_c k;
1750         int ret;
1751
1752         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1753         bch2_btree_iter_checks(iter);
1754
1755         if (iter->uptodate == BTREE_ITER_UPTODATE &&
1756             !bkey_deleted(&iter->k))
1757                 return btree_iter_peek_uptodate(iter);
1758
1759         while (1) {
1760                 ret = bch2_btree_iter_traverse(iter);
1761                 if (unlikely(ret))
1762                         return bkey_s_c_err(ret);
1763
1764                 k = __btree_iter_peek(iter, l);
1765                 if (!k.k || bkey_cmp(bkey_start_pos(k.k), pos) > 0)
1766                         k = __btree_iter_prev(iter, l);
1767
1768                 if (likely(k.k))
1769                         break;
1770
1771                 if (!btree_iter_set_pos_to_prev_leaf(iter))
1772                         return bkey_s_c_null;
1773         }
1774
1775         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), pos) > 0);
1776         iter->pos       = bkey_start_pos(k.k);
1777         iter->uptodate  = BTREE_ITER_UPTODATE;
1778         return k;
1779 }
1780
1781 /**
1782  * bch2_btree_iter_prev: returns first key less than iterator's current
1783  * position
1784  */
1785 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
1786 {
1787         struct bpos pos = bkey_start_pos(&iter->k);
1788
1789         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1790         bch2_btree_iter_checks(iter);
1791
1792         if (unlikely(!bkey_cmp(pos, POS_MIN)))
1793                 return bkey_s_c_null;
1794
1795         bch2_btree_iter_set_pos(iter, bkey_predecessor(pos));
1796
1797         return bch2_btree_iter_peek_prev(iter);
1798 }
1799
1800 static inline struct bkey_s_c
1801 __bch2_btree_iter_peek_slot_extents(struct btree_iter *iter)
1802 {
1803         struct btree_iter_level *l = &iter->l[0];
1804         struct btree_node_iter node_iter;
1805         struct bkey_s_c k;
1806         struct bkey n;
1807         int ret;
1808
1809         /* keys & holes can't span inode numbers: */
1810         if (iter->pos.offset == KEY_OFFSET_MAX) {
1811                 if (iter->pos.inode == KEY_INODE_MAX)
1812                         return bkey_s_c_null;
1813
1814                 bch2_btree_iter_set_pos(iter, bkey_successor(iter->pos));
1815
1816                 ret = bch2_btree_iter_traverse(iter);
1817                 if (unlikely(ret))
1818                         return bkey_s_c_err(ret);
1819         }
1820
1821         /*
1822          * iterator is now at the correct position for inserting at iter->pos,
1823          * but we need to keep iterating until we find the first non whiteout so
1824          * we know how big a hole we have, if any:
1825          */
1826
1827         node_iter = l->iter;
1828         k = __btree_iter_unpack(iter, l, &iter->k,
1829                 bch2_btree_node_iter_peek(&node_iter, l->b));
1830
1831         if (k.k && bkey_cmp(bkey_start_pos(k.k), iter->pos) <= 0) {
1832                 /*
1833                  * We're not setting iter->uptodate because the node iterator
1834                  * doesn't necessarily point at the key we're returning:
1835                  */
1836
1837                 EBUG_ON(bkey_cmp(k.k->p, iter->pos) <= 0);
1838                 bch2_btree_iter_verify_level(iter, 0);
1839                 return k;
1840         }
1841
1842         /* hole */
1843
1844         if (!k.k)
1845                 k.k = &l->b->key.k;
1846
1847         bkey_init(&n);
1848         n.p = iter->pos;
1849         bch2_key_resize(&n,
1850                         min_t(u64, KEY_SIZE_MAX,
1851                               (k.k->p.inode == n.p.inode
1852                                ? bkey_start_offset(k.k)
1853                                : KEY_OFFSET_MAX) -
1854                               n.p.offset));
1855
1856         EBUG_ON(!n.size);
1857
1858         iter->k = n;
1859         iter->uptodate = BTREE_ITER_UPTODATE;
1860
1861         bch2_btree_iter_verify_level(iter, 0);
1862         return (struct bkey_s_c) { &iter->k, NULL };
1863 }
1864
1865 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
1866 {
1867         struct btree_iter_level *l = &iter->l[0];
1868         struct bkey_s_c k;
1869         int ret;
1870
1871         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1872         bch2_btree_iter_checks(iter);
1873
1874         if (iter->uptodate == BTREE_ITER_UPTODATE)
1875                 return btree_iter_peek_uptodate(iter);
1876
1877         ret = bch2_btree_iter_traverse(iter);
1878         if (unlikely(ret))
1879                 return bkey_s_c_err(ret);
1880
1881         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1882                 return __bch2_btree_iter_peek_slot_extents(iter);
1883
1884         k = __btree_iter_peek_all(iter, l, &iter->k);
1885
1886         EBUG_ON(k.k && bkey_deleted(k.k) && bkey_cmp(k.k->p, iter->pos) == 0);
1887
1888         if (!k.k || bkey_cmp(iter->pos, k.k->p)) {
1889                 /* hole */
1890                 bkey_init(&iter->k);
1891                 iter->k.p = iter->pos;
1892                 k = (struct bkey_s_c) { &iter->k, NULL };
1893         }
1894
1895         iter->uptodate = BTREE_ITER_UPTODATE;
1896         bch2_btree_iter_verify_level(iter, 0);
1897         return k;
1898 }
1899
1900 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
1901 {
1902         if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
1903                 return bkey_s_c_null;
1904
1905         bch2_btree_iter_set_pos(iter,
1906                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1907                 ? iter->k.p
1908                 : bkey_successor(iter->k.p));
1909
1910         return bch2_btree_iter_peek_slot(iter);
1911 }
1912
1913 struct bkey_s_c bch2_btree_iter_peek_cached(struct btree_iter *iter)
1914 {
1915         struct bkey_cached *ck;
1916         int ret;
1917
1918         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_CACHED);
1919         bch2_btree_iter_checks(iter);
1920
1921         ret = bch2_btree_iter_traverse(iter);
1922         if (unlikely(ret))
1923                 return bkey_s_c_err(ret);
1924
1925         ck = (void *) iter->l[0].b;
1926
1927         EBUG_ON(iter->btree_id != ck->key.btree_id ||
1928                 bkey_cmp(iter->pos, ck->key.pos));
1929         BUG_ON(!ck->valid);
1930
1931         return bkey_i_to_s_c(ck->k);
1932 }
1933
1934 static inline void bch2_btree_iter_init(struct btree_trans *trans,
1935                         struct btree_iter *iter, enum btree_id btree_id,
1936                         struct bpos pos, unsigned flags)
1937 {
1938         struct bch_fs *c = trans->c;
1939         unsigned i;
1940
1941         if (btree_node_type_is_extents(btree_id) &&
1942             !(flags & BTREE_ITER_NODES))
1943                 flags |= BTREE_ITER_IS_EXTENTS;
1944
1945         iter->trans                     = trans;
1946         iter->pos                       = pos;
1947         bkey_init(&iter->k);
1948         iter->k.p                       = pos;
1949         iter->flags                     = flags;
1950         iter->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1951         iter->btree_id                  = btree_id;
1952         iter->level                     = 0;
1953         iter->min_depth                 = 0;
1954         iter->locks_want                = flags & BTREE_ITER_INTENT ? 1 : 0;
1955         iter->nodes_locked              = 0;
1956         iter->nodes_intent_locked       = 0;
1957         for (i = 0; i < ARRAY_SIZE(iter->l); i++)
1958                 iter->l[i].b            = BTREE_ITER_NO_NODE_INIT;
1959
1960         prefetch(c->btree_roots[btree_id].b);
1961 }
1962
1963 /* new transactional stuff: */
1964
1965 static inline void __bch2_trans_iter_free(struct btree_trans *trans,
1966                                           unsigned idx)
1967 {
1968         __bch2_btree_iter_unlock(&trans->iters[idx]);
1969         trans->iters_linked             &= ~(1ULL << idx);
1970         trans->iters_live               &= ~(1ULL << idx);
1971         trans->iters_touched            &= ~(1ULL << idx);
1972 }
1973
1974 int bch2_trans_iter_put(struct btree_trans *trans,
1975                         struct btree_iter *iter)
1976 {
1977         int ret;
1978
1979         if (IS_ERR_OR_NULL(iter))
1980                 return 0;
1981
1982         BUG_ON(trans->iters + iter->idx != iter);
1983
1984         ret = btree_iter_err(iter);
1985
1986         if (!(trans->iters_touched & (1ULL << iter->idx)) &&
1987             !(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT))
1988                 __bch2_trans_iter_free(trans, iter->idx);
1989
1990         trans->iters_live       &= ~(1ULL << iter->idx);
1991         return ret;
1992 }
1993
1994 int bch2_trans_iter_free(struct btree_trans *trans,
1995                          struct btree_iter *iter)
1996 {
1997         if (IS_ERR_OR_NULL(iter))
1998                 return 0;
1999
2000         trans->iters_touched &= ~(1ULL << iter->idx);
2001
2002         return bch2_trans_iter_put(trans, iter);
2003 }
2004
2005 noinline __cold
2006 static void btree_trans_iter_alloc_fail(struct btree_trans *trans)
2007 {
2008
2009         struct btree_iter *iter;
2010
2011         trans_for_each_iter(trans, iter)
2012                 pr_err("iter: btree %s pos %llu:%llu%s%s%s %ps",
2013                        bch2_btree_ids[iter->btree_id],
2014                        iter->pos.inode,
2015                        iter->pos.offset,
2016                        (trans->iters_live & (1ULL << iter->idx)) ? " live" : "",
2017                        (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "",
2018                        iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "",
2019                        (void *) iter->ip_allocated);
2020         panic("trans iter oveflow\n");
2021 }
2022
2023 static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans)
2024 {
2025         unsigned idx;
2026
2027         if (unlikely(trans->iters_linked ==
2028                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
2029                 btree_trans_iter_alloc_fail(trans);
2030
2031         idx = __ffs64(~trans->iters_linked);
2032
2033         trans->iters_linked     |= 1ULL << idx;
2034         trans->iters[idx].idx    = idx;
2035         trans->iters[idx].flags  = 0;
2036         return &trans->iters[idx];
2037 }
2038
2039 static inline void btree_iter_copy(struct btree_iter *dst,
2040                                    struct btree_iter *src)
2041 {
2042         unsigned i, idx = dst->idx;
2043
2044         *dst = *src;
2045         dst->idx = idx;
2046         dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
2047
2048         for (i = 0; i < BTREE_MAX_DEPTH; i++)
2049                 if (btree_node_locked(dst, i))
2050                         six_lock_increment(&dst->l[i].b->c.lock,
2051                                            __btree_lock_want(dst, i));
2052
2053         dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
2054         dst->flags &= ~BTREE_ITER_SET_POS_AFTER_COMMIT;
2055 }
2056
2057 static inline struct bpos bpos_diff(struct bpos l, struct bpos r)
2058 {
2059         if (bkey_cmp(l, r) > 0)
2060                 swap(l, r);
2061
2062         return POS(r.inode - l.inode, r.offset - l.offset);
2063 }
2064
2065 static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans,
2066                                                  unsigned btree_id, struct bpos pos,
2067                                                  unsigned flags)
2068 {
2069         struct btree_iter *iter, *best = NULL;
2070
2071         trans_for_each_iter(trans, iter) {
2072                 if (btree_iter_type(iter) != (flags & BTREE_ITER_TYPE))
2073                         continue;
2074
2075                 if (iter->btree_id != btree_id)
2076                         continue;
2077
2078                 if (best &&
2079                     bkey_cmp(bpos_diff(best->pos, pos),
2080                              bpos_diff(iter->pos, pos)) < 0)
2081                         continue;
2082
2083                 best = iter;
2084         }
2085
2086         if (!best) {
2087                 iter = btree_trans_iter_alloc(trans);
2088                 bch2_btree_iter_init(trans, iter, btree_id, pos, flags);
2089         } else if ((trans->iters_live & (1ULL << best->idx)) ||
2090                    (best->flags & BTREE_ITER_KEEP_UNTIL_COMMIT)) {
2091                 iter = btree_trans_iter_alloc(trans);
2092                 btree_iter_copy(iter, best);
2093         } else {
2094                 iter = best;
2095         }
2096
2097         iter->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
2098         iter->flags &= ~BTREE_ITER_USER_FLAGS;
2099         iter->flags |= flags & BTREE_ITER_USER_FLAGS;
2100
2101         if (iter->flags & BTREE_ITER_INTENT)
2102                 bch2_btree_iter_upgrade(iter, 1);
2103         else
2104                 bch2_btree_iter_downgrade(iter);
2105
2106         BUG_ON(iter->btree_id != btree_id);
2107         BUG_ON((iter->flags ^ flags) & BTREE_ITER_TYPE);
2108         BUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
2109         BUG_ON(iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT);
2110         BUG_ON(trans->iters_live & (1ULL << iter->idx));
2111
2112         trans->iters_live       |= 1ULL << iter->idx;
2113         trans->iters_touched    |= 1ULL << iter->idx;
2114
2115         return iter;
2116 }
2117
2118 struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans,
2119                                          enum btree_id btree_id,
2120                                          struct bpos pos, unsigned flags)
2121 {
2122         struct btree_iter *iter =
2123                 __btree_trans_get_iter(trans, btree_id, pos, flags);
2124
2125         __bch2_btree_iter_set_pos(iter, pos,
2126                 btree_node_type_is_extents(btree_id));
2127         return iter;
2128 }
2129
2130 struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *trans,
2131                                             enum btree_id btree_id,
2132                                             struct bpos pos,
2133                                             unsigned locks_want,
2134                                             unsigned depth,
2135                                             unsigned flags)
2136 {
2137         struct btree_iter *iter =
2138                 __btree_trans_get_iter(trans, btree_id, pos,
2139                                        flags|BTREE_ITER_NODES);
2140         unsigned i;
2141
2142         BUG_ON(bkey_cmp(iter->pos, pos));
2143
2144         iter->locks_want = locks_want;
2145         iter->level     = depth;
2146         iter->min_depth = depth;
2147
2148         for (i = 0; i < ARRAY_SIZE(iter->l); i++)
2149                 iter->l[i].b            = NULL;
2150         iter->l[iter->level].b          = BTREE_ITER_NO_NODE_INIT;
2151
2152         return iter;
2153 }
2154
2155 struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *trans,
2156                                         struct btree_iter *src)
2157 {
2158         struct btree_iter *iter;
2159
2160         iter = btree_trans_iter_alloc(trans);
2161         btree_iter_copy(iter, src);
2162
2163         trans->iters_live |= 1ULL << iter->idx;
2164         /*
2165          * We don't need to preserve this iter since it's cheap to copy it
2166          * again - this will cause trans_iter_put() to free it right away:
2167          */
2168         trans->iters_touched &= ~(1ULL << iter->idx);
2169
2170         return iter;
2171 }
2172
2173 static int bch2_trans_preload_mem(struct btree_trans *trans, size_t size)
2174 {
2175         if (size > trans->mem_bytes) {
2176                 size_t old_bytes = trans->mem_bytes;
2177                 size_t new_bytes = roundup_pow_of_two(size);
2178                 void *new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2179
2180                 if (!new_mem)
2181                         return -ENOMEM;
2182
2183                 trans->mem = new_mem;
2184                 trans->mem_bytes = new_bytes;
2185
2186                 if (old_bytes) {
2187                         trace_trans_restart_mem_realloced(trans->ip, new_bytes);
2188                         return -EINTR;
2189                 }
2190         }
2191
2192         return 0;
2193 }
2194
2195 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2196 {
2197         void *p;
2198         int ret;
2199
2200         ret = bch2_trans_preload_mem(trans, trans->mem_top + size);
2201         if (ret)
2202                 return ERR_PTR(ret);
2203
2204         p = trans->mem + trans->mem_top;
2205         trans->mem_top += size;
2206         return p;
2207 }
2208
2209 inline void bch2_trans_unlink_iters(struct btree_trans *trans)
2210 {
2211         u64 iters = trans->iters_linked &
2212                 ~trans->iters_touched &
2213                 ~trans->iters_live;
2214
2215         while (iters) {
2216                 unsigned idx = __ffs64(iters);
2217
2218                 iters &= ~(1ULL << idx);
2219                 __bch2_trans_iter_free(trans, idx);
2220         }
2221 }
2222
2223 void bch2_trans_reset(struct btree_trans *trans, unsigned flags)
2224 {
2225         struct btree_iter *iter;
2226
2227         trans_for_each_iter(trans, iter)
2228                 iter->flags &= ~(BTREE_ITER_KEEP_UNTIL_COMMIT|
2229                                  BTREE_ITER_SET_POS_AFTER_COMMIT);
2230
2231         bch2_trans_unlink_iters(trans);
2232
2233         trans->iters_touched &= trans->iters_live;
2234
2235         trans->nr_updates               = 0;
2236         trans->nr_updates2              = 0;
2237         trans->mem_top                  = 0;
2238
2239         trans->extra_journal_entries    = NULL;
2240         trans->extra_journal_entry_u64s = 0;
2241
2242         if (trans->fs_usage_deltas) {
2243                 trans->fs_usage_deltas->used = 0;
2244                 memset(&trans->fs_usage_deltas->memset_start, 0,
2245                        (void *) &trans->fs_usage_deltas->memset_end -
2246                        (void *) &trans->fs_usage_deltas->memset_start);
2247         }
2248
2249         if (!(flags & TRANS_RESET_NOTRAVERSE))
2250                 bch2_btree_iter_traverse_all(trans);
2251 }
2252
2253 static void bch2_trans_alloc_iters(struct btree_trans *trans, struct bch_fs *c)
2254 {
2255         size_t iters_bytes      = sizeof(struct btree_iter) * BTREE_ITER_MAX;
2256         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2257         void *p = NULL;
2258
2259         BUG_ON(trans->used_mempool);
2260
2261 #ifdef __KERNEL__
2262         p = this_cpu_xchg(c->btree_iters_bufs->iter, NULL);
2263 #endif
2264         if (!p)
2265                 p = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
2266
2267         trans->iters            = p; p += iters_bytes;
2268         trans->updates          = p; p += updates_bytes;
2269         trans->updates2         = p; p += updates_bytes;
2270 }
2271
2272 void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
2273                      unsigned expected_nr_iters,
2274                      size_t expected_mem_bytes)
2275 {
2276         memset(trans, 0, sizeof(*trans));
2277         trans->c                = c;
2278         trans->ip               = _RET_IP_;
2279
2280         /*
2281          * reallocating iterators currently completely breaks
2282          * bch2_trans_iter_put(), we always allocate the max:
2283          */
2284         bch2_trans_alloc_iters(trans, c);
2285
2286         if (expected_mem_bytes) {
2287                 trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes);
2288                 trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL);
2289         }
2290
2291         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2292
2293 #ifdef CONFIG_BCACHEFS_DEBUG
2294         trans->pid = current->pid;
2295         mutex_lock(&c->btree_trans_lock);
2296         list_add(&trans->list, &c->btree_trans_list);
2297         mutex_unlock(&c->btree_trans_lock);
2298 #endif
2299 }
2300
2301 int bch2_trans_exit(struct btree_trans *trans)
2302 {
2303         struct bch_fs *c = trans->c;
2304
2305         bch2_trans_unlock(trans);
2306
2307 #ifdef CONFIG_BCACHEFS_DEBUG
2308         mutex_lock(&trans->c->btree_trans_lock);
2309         list_del(&trans->list);
2310         mutex_unlock(&trans->c->btree_trans_lock);
2311 #endif
2312
2313         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2314
2315         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
2316
2317         kfree(trans->fs_usage_deltas);
2318         kfree(trans->mem);
2319
2320 #ifdef __KERNEL__
2321         /*
2322          * Userspace doesn't have a real percpu implementation:
2323          */
2324         trans->iters = this_cpu_xchg(c->btree_iters_bufs->iter, trans->iters);
2325 #endif
2326         if (trans->iters)
2327                 mempool_free(trans->iters, &trans->c->btree_iters_pool);
2328
2329         trans->mem      = (void *) 0x1;
2330         trans->iters    = (void *) 0x1;
2331
2332         return trans->error ? -EIO : 0;
2333 }
2334
2335 static void __maybe_unused
2336 bch2_btree_iter_node_to_text(struct printbuf *out,
2337                              struct btree_bkey_cached_common *_b,
2338                              enum btree_iter_type type)
2339 {
2340         pr_buf(out, "    %px l=%u %s:",
2341                _b, _b->level, bch2_btree_ids[_b->btree_id]);
2342         bch2_bpos_to_text(out, btree_node_pos(_b, type));
2343 }
2344
2345 void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
2346 {
2347 #ifdef CONFIG_BCACHEFS_DEBUG
2348         struct btree_trans *trans;
2349         struct btree_iter *iter;
2350         struct btree *b;
2351         unsigned l;
2352
2353         mutex_lock(&c->btree_trans_lock);
2354         list_for_each_entry(trans, &c->btree_trans_list, list) {
2355                 pr_buf(out, "%i %px %ps\n", trans->pid, trans, (void *) trans->ip);
2356
2357                 trans_for_each_iter(trans, iter) {
2358                         if (!iter->nodes_locked)
2359                                 continue;
2360
2361                         pr_buf(out, "  iter %u %s:",
2362                                iter->idx,
2363                                bch2_btree_ids[iter->btree_id]);
2364                         bch2_bpos_to_text(out, iter->pos);
2365                         pr_buf(out, "\n");
2366
2367                         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
2368                                 if (btree_node_locked(iter, l)) {
2369                                         pr_buf(out, "    %s l=%u ",
2370                                                btree_node_intent_locked(iter, l) ? "i" : "r", l);
2371                                         bch2_btree_iter_node_to_text(out,
2372                                                         (void *) iter->l[l].b,
2373                                                         btree_iter_type(iter));
2374                                         pr_buf(out, "\n");
2375                                 }
2376                         }
2377                 }
2378
2379                 b = READ_ONCE(trans->locking);
2380                 if (b) {
2381                         pr_buf(out, "  locking iter %u l=%u %s:",
2382                                trans->locking_iter_idx,
2383                                trans->locking_level,
2384                                bch2_btree_ids[trans->locking_btree_id]);
2385                         bch2_bpos_to_text(out, trans->locking_pos);
2386
2387
2388                         pr_buf(out, " node ");
2389                         bch2_btree_iter_node_to_text(out,
2390                                         (void *) b,
2391                                         btree_iter_type(&trans->iters[trans->locking_iter_idx]));
2392                         pr_buf(out, "\n");
2393                 }
2394         }
2395         mutex_unlock(&c->btree_trans_lock);
2396 #endif
2397 }
2398
2399 void bch2_fs_btree_iter_exit(struct bch_fs *c)
2400 {
2401         mempool_exit(&c->btree_iters_pool);
2402         cleanup_srcu_struct(&c->btree_trans_barrier);
2403 }
2404
2405 int bch2_fs_btree_iter_init(struct bch_fs *c)
2406 {
2407         unsigned nr = BTREE_ITER_MAX;
2408
2409         INIT_LIST_HEAD(&c->btree_trans_list);
2410         mutex_init(&c->btree_trans_lock);
2411
2412         return  init_srcu_struct(&c->btree_trans_barrier) ?:
2413                 mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
2414                         sizeof(struct btree_iter) * nr +
2415                         sizeof(struct btree_insert_entry) * nr +
2416                         sizeof(struct btree_insert_entry) * nr);
2417 }