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,
 201                             void *p)
 202 {
 203         struct btree_trans *trans = iter->trans;
 204         struct btree_iter *linked;
 205         u64 start_time = local_clock();
 206         bool ret = true;
 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                                         ret = false;
 233                         } else {
 234                                 ret = false;
 235                         }
 236                 }
 237
 238                 /*
 239                  * Interior nodes must be locked before their descendants: if
 240                  * another iterator has possible descendants locked of the node
 241                  * we're about to lock, it must have the ancestors locked too:
 242                  */
 243                 if (linked->btree_id == iter->btree_id &&
 244                     level > __fls(linked->nodes_locked)) {
 245                         if (!(trans->nounlock)) {
 246                                 linked->locks_want =
 247                                         max(level + 1, max_t(unsigned,
 248                                             linked->locks_want,
 249                                             iter->locks_want));
 250                                 if (!btree_iter_get_locks(linked, true, false))
 251                                         ret = false;
 252                         } else {
 253                                 ret = false;
 254                         }
 255                 }
 256
 257                 /* Must lock btree nodes in key order: */
 258                 if ((cmp_int(iter->btree_id, linked->btree_id) ?:
 259                      -cmp_int(btree_iter_type(iter), btree_iter_type(linked))) < 0)
 260                         ret = false;
 261
 262                 if (iter->btree_id == linked->btree_id &&
 263                     btree_node_locked(linked, level) &&
 264                     bkey_cmp(pos, btree_node_pos((void *) linked->l[level].b,
 265                                                  btree_iter_type(linked))) <= 0)
 266                         ret = false;
 267
 268                 /*
 269                  * Recheck if this is a node we already have locked - since one
 270                  * of the get_locks() calls might've successfully
 271                  * upgraded/relocked it:
 272                  */
 273                 if (linked->l[level].b == b &&
 274                     btree_node_locked_type(linked, level) >= type) {
 275                         six_lock_increment(&b->c.lock, type);
 276                         return true;
 277                 }
 278         }
 279
 280         if (unlikely(!ret)) {
 281                 trace_trans_restart_would_deadlock(iter->trans->ip);
 282                 return false;
 283         }
 284
 285         if (six_trylock_type(&b->c.lock, type))
 286                 return true;
 287
 288         if (six_lock_type(&b->c.lock, type, should_sleep_fn, p))
 289                 return false;
 290
 291         bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)],
 292                                start_time);
 293         return true;
 294 }
 295
 296 /* Btree iterator locking: */
 297
 298 #ifdef CONFIG_BCACHEFS_DEBUG
 299 static void bch2_btree_iter_verify_locks(struct btree_iter *iter)
 300 {
 301         unsigned l;
 302
 303         if (!(iter->trans->iters_linked & (1ULL << iter->idx))) {
 304                 BUG_ON(iter->nodes_locked);
 305                 return;
 306         }
 307
 308         for (l = 0; is_btree_node(iter, l); l++) {
 309                 if (iter->uptodate >= BTREE_ITER_NEED_RELOCK &&
 310                     !btree_node_locked(iter, l))
 311                         continue;
 312
 313                 BUG_ON(btree_lock_want(iter, l) !=
 314                        btree_node_locked_type(iter, l));
 315         }
 316 }
 317
 318 void bch2_btree_trans_verify_locks(struct btree_trans *trans)
 319 {
 320         struct btree_iter *iter;
 321
 322         trans_for_each_iter_all(trans, iter)
 323                 bch2_btree_iter_verify_locks(iter);
 324 }
 325 #else
 326 static inline void bch2_btree_iter_verify_locks(struct btree_iter *iter) {}
 327 #endif
 328
 329 __flatten
 330 bool bch2_btree_iter_relock(struct btree_iter *iter, bool trace)
 331 {
 332         return btree_iter_get_locks(iter, false, trace);
 333 }
 334
 335 bool __bch2_btree_iter_upgrade(struct btree_iter *iter,
 336                                unsigned new_locks_want)
 337 {
 338         struct btree_iter *linked;
 339
 340         EBUG_ON(iter->locks_want >= new_locks_want);
 341
 342         iter->locks_want = new_locks_want;
 343
 344         if (btree_iter_get_locks(iter, true, true))
 345                 return true;
 346
 347         /*
 348          * Ancestor nodes must be locked before child nodes, so set locks_want
 349          * on iterators that might lock ancestors before us to avoid getting
 350          * -EINTR later:
 351          */
 352         trans_for_each_iter(iter->trans, linked)
 353                 if (linked != iter &&
 354                     linked->btree_id == iter->btree_id &&
 355                     linked->locks_want < new_locks_want) {
 356                         linked->locks_want = new_locks_want;
 357                         btree_iter_get_locks(linked, true, false);
 358                 }
 359
 360         return false;
 361 }
 362
 363 bool __bch2_btree_iter_upgrade_nounlock(struct btree_iter *iter,
 364                                         unsigned new_locks_want)
 365 {
 366         unsigned l = iter->level;
 367
 368         EBUG_ON(iter->locks_want >= new_locks_want);
 369
 370         iter->locks_want = new_locks_want;
 371
 372         do {
 373                 if (!btree_iter_node(iter, l))
 374                         break;
 375
 376                 if (!bch2_btree_node_upgrade(iter, l)) {
 377                         iter->locks_want = l;
 378                         return false;
 379                 }
 380
 381                 l++;
 382         } while (l < iter->locks_want);
 383
 384         return true;
 385 }
 386
 387 void __bch2_btree_iter_downgrade(struct btree_iter *iter,
 388                                  unsigned downgrade_to)
 389 {
 390         unsigned l, new_locks_want = downgrade_to ?:
 391                 (iter->flags & BTREE_ITER_INTENT ? 1 : 0);
 392
 393         if (iter->locks_want < downgrade_to) {
 394                 iter->locks_want = new_locks_want;
 395
 396                 while (iter->nodes_locked &&
 397                        (l = __fls(iter->nodes_locked)) >= iter->locks_want) {
 398                         if (l > iter->level) {
 399                                 btree_node_unlock(iter, l);
 400                         } else {
 401                                 if (btree_node_intent_locked(iter, l)) {
 402                                         six_lock_downgrade(&iter->l[l].b->c.lock);
 403                                         iter->nodes_intent_locked ^= 1 << l;
 404                                 }
 405                                 break;
 406                         }
 407                 }
 408         }
 409
 410         bch2_btree_trans_verify_locks(iter->trans);
 411 }
 412
 413 void bch2_trans_downgrade(struct btree_trans *trans)
 414 {
 415         struct btree_iter *iter;
 416
 417         trans_for_each_iter(trans, iter)
 418                 bch2_btree_iter_downgrade(iter);
 419 }
 420
 421 /* Btree transaction locking: */
 422
 423 bool bch2_trans_relock(struct btree_trans *trans)
 424 {
 425         struct btree_iter *iter;
 426         bool ret = true;
 427
 428         trans_for_each_iter(trans, iter)
 429                 if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
 430                         ret &= bch2_btree_iter_relock(iter, true);
 431
 432         return ret;
 433 }
 434
 435 void bch2_trans_unlock(struct btree_trans *trans)
 436 {
 437         struct btree_iter *iter;
 438
 439         trans_for_each_iter(trans, iter)
 440                 __bch2_btree_iter_unlock(iter);
 441 }
 442
 443 /* Btree iterator: */
 444
 445 #ifdef CONFIG_BCACHEFS_DEBUG
 446
 447 static void bch2_btree_iter_verify_cached(struct btree_iter *iter)
 448 {
 449         struct bkey_cached *ck;
 450         bool locked = btree_node_locked(iter, 0);
 451
 452         if (!bch2_btree_node_relock(iter, 0))
 453                 return;
 454
 455         ck = (void *) iter->l[0].b;
 456         BUG_ON(ck->key.btree_id != iter->btree_id ||
 457                bkey_cmp(ck->key.pos, iter->pos));
 458
 459         if (!locked)
 460                 btree_node_unlock(iter, 0);
 461 }
 462
 463 static void bch2_btree_iter_verify_level(struct btree_iter *iter,
 464                                          unsigned level)
 465 {
 466         struct bpos pos = btree_iter_search_key(iter);
 467         struct btree_iter_level *l = &iter->l[level];
 468         struct btree_node_iter tmp = l->iter;
 469         bool locked = btree_node_locked(iter, level);
 470         struct bkey_packed *p, *k;
 471         char buf1[100], buf2[100];
 472         const char *msg;
 473
 474         if (!debug_check_iterators(iter->trans->c))
 475                 return;
 476
 477         if (btree_iter_type(iter) == BTREE_ITER_CACHED) {
 478                 if (!level)
 479                         bch2_btree_iter_verify_cached(iter);
 480                 return;
 481         }
 482
 483         BUG_ON(iter->level < iter->min_depth);
 484
 485         if (!btree_iter_node(iter, level))
 486                 return;
 487
 488         if (!bch2_btree_node_relock(iter, level))
 489                 return;
 490
 491         /*
 492          * Ideally this invariant would always be true, and hopefully in the
 493          * future it will be, but for now set_pos_same_leaf() breaks it:
 494          */
 495         BUG_ON(iter->uptodate < BTREE_ITER_NEED_TRAVERSE &&
 496                !btree_iter_pos_in_node(iter, l->b));
 497
 498         /*
 499          * node iterators don't use leaf node iterator:
 500          */
 501         if (btree_iter_type(iter) == BTREE_ITER_NODES &&
 502             level <= iter->min_depth)
 503                 goto unlock;
 504
 505         bch2_btree_node_iter_verify(&l->iter, l->b);
 506
 507         /*
 508          * For interior nodes, the iterator will have skipped past
 509          * deleted keys:
 510          *
 511          * For extents, the iterator may have skipped past deleted keys (but not
 512          * whiteouts)
 513          */
 514         p = level || btree_node_type_is_extents(iter->btree_id)
 515                 ? bch2_btree_node_iter_prev_filter(&tmp, l->b, KEY_TYPE_discard)
 516                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 517         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 518
 519         if (p && bkey_iter_pos_cmp(l->b, p, &pos) >= 0) {
 520                 msg = "before";
 521                 goto err;
 522         }
 523
 524         if (k && bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
 525                 msg = "after";
 526                 goto err;
 527         }
 528 unlock:
 529         if (!locked)
 530                 btree_node_unlock(iter, level);
 531         return;
 532 err:
 533         strcpy(buf1, "(none)");
 534         strcpy(buf2, "(none)");
 535
 536         if (p) {
 537                 struct bkey uk = bkey_unpack_key(l->b, p);
 538                 bch2_bkey_to_text(&PBUF(buf1), &uk);
 539         }
 540
 541         if (k) {
 542                 struct bkey uk = bkey_unpack_key(l->b, k);
 543                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 544         }
 545
 546         panic("iterator should be %s key at level %u:\n"
 547               "iter pos %s %llu:%llu\n"
 548               "prev key %s\n"
 549               "cur  key %s\n",
 550               msg, level,
 551               iter->flags & BTREE_ITER_IS_EXTENTS ? ">" : "=>",
 552               iter->pos.inode, iter->pos.offset,
 553               buf1, buf2);
 554 }
 555
 556 static void bch2_btree_iter_verify(struct btree_iter *iter)
 557 {
 558         unsigned i;
 559
 560         bch2_btree_trans_verify_locks(iter->trans);
 561
 562         for (i = 0; i < BTREE_MAX_DEPTH; i++)
 563                 bch2_btree_iter_verify_level(iter, i);
 564 }
 565
 566 void bch2_btree_trans_verify_iters(struct btree_trans *trans, struct btree *b)
 567 {
 568         struct btree_iter *iter;
 569
 570         if (!debug_check_iterators(trans->c))
 571                 return;
 572
 573         trans_for_each_iter_with_node(trans, b, iter)
 574                 bch2_btree_iter_verify_level(iter, b->c.level);
 575 }
 576
 577 #else
 578
 579 static inline void bch2_btree_iter_verify_level(struct btree_iter *iter, unsigned l) {}
 580 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 581
 582 #endif
 583
 584 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 585                                         struct btree *b,
 586                                         struct bset_tree *t,
 587                                         struct bkey_packed *k)
 588 {
 589         struct btree_node_iter_set *set;
 590
 591         btree_node_iter_for_each(iter, set)
 592                 if (set->end == t->end_offset) {
 593                         set->k = __btree_node_key_to_offset(b, k);
 594                         bch2_btree_node_iter_sort(iter, b);
 595                         return;
 596                 }
 597
 598         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 599 }
 600
 601 static void __bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 602                                                struct btree *b,
 603                                                struct bkey_packed *where)
 604 {
 605         struct btree_iter_level *l = &iter->l[b->c.level];
 606         struct bpos pos = btree_iter_search_key(iter);
 607
 608         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 609                 return;
 610
 611         if (bkey_iter_pos_cmp(l->b, where, &pos) < 0)
 612                 bch2_btree_node_iter_advance(&l->iter, l->b);
 613
 614         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 615 }
 616
 617 void bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 618                                       struct btree *b,
 619                                       struct bkey_packed *where)
 620 {
 621         struct btree_iter *linked;
 622
 623         trans_for_each_iter_with_node(iter->trans, b, linked) {
 624                 __bch2_btree_iter_fix_key_modified(linked, b, where);
 625                 bch2_btree_iter_verify_level(linked, b->c.level);
 626         }
 627 }
 628
 629 static void __bch2_btree_node_iter_fix(struct btree_iter *iter,
 630                                       struct btree *b,
 631                                       struct btree_node_iter *node_iter,
 632                                       struct bset_tree *t,
 633                                       struct bkey_packed *where,
 634                                       unsigned clobber_u64s,
 635                                       unsigned new_u64s)
 636 {
 637         const struct bkey_packed *end = btree_bkey_last(b, t);
 638         struct btree_node_iter_set *set;
 639         unsigned offset = __btree_node_key_to_offset(b, where);
 640         int shift = new_u64s - clobber_u64s;
 641         unsigned old_end = t->end_offset - shift;
 642         unsigned orig_iter_pos = node_iter->data[0].k;
 643         bool iter_current_key_modified =
 644                 orig_iter_pos >= offset &&
 645                 orig_iter_pos <= offset + clobber_u64s;
 646         struct bpos iter_pos = btree_iter_search_key(iter);
 647
 648         btree_node_iter_for_each(node_iter, set)
 649                 if (set->end == old_end)
 650                         goto found;
 651
 652         /* didn't find the bset in the iterator - might have to readd it: */
 653         if (new_u64s &&
 654             bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
 655                 bch2_btree_node_iter_push(node_iter, b, where, end);
 656                 goto fixup_done;
 657         } else {
 658                 /* Iterator is after key that changed */
 659                 return;
 660         }
 661 found:
 662         set->end = t->end_offset;
 663
 664         /* Iterator hasn't gotten to the key that changed yet: */
 665         if (set->k < offset)
 666                 return;
 667
 668         if (new_u64s &&
 669             bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
 670                 set->k = offset;
 671         } else if (set->k < offset + clobber_u64s) {
 672                 set->k = offset + new_u64s;
 673                 if (set->k == set->end)
 674                         bch2_btree_node_iter_set_drop(node_iter, set);
 675         } else {
 676                 /* Iterator is after key that changed */
 677                 set->k = (int) set->k + shift;
 678                 return;
 679         }
 680
 681         bch2_btree_node_iter_sort(node_iter, b);
 682 fixup_done:
 683         if (node_iter->data[0].k != orig_iter_pos)
 684                 iter_current_key_modified = true;
 685
 686         /*
 687          * When a new key is added, and the node iterator now points to that
 688          * key, the iterator might have skipped past deleted keys that should
 689          * come after the key the iterator now points to. We have to rewind to
 690          * before those deleted keys - otherwise
 691          * bch2_btree_node_iter_prev_all() breaks:
 692          */
 693         if (!bch2_btree_node_iter_end(node_iter) &&
 694             iter_current_key_modified &&
 695             (b->c.level ||
 696              btree_node_type_is_extents(iter->btree_id))) {
 697                 struct bset_tree *t;
 698                 struct bkey_packed *k, *k2, *p;
 699
 700                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 701
 702                 for_each_bset(b, t) {
 703                         bool set_pos = false;
 704
 705                         if (node_iter->data[0].end == t->end_offset)
 706                                 continue;
 707
 708                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 709
 710                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 711                                bkey_iter_cmp(b, k, p) < 0) {
 712                                 k2 = p;
 713                                 set_pos = true;
 714                         }
 715
 716                         if (set_pos)
 717                                 btree_node_iter_set_set_pos(node_iter,
 718                                                             b, t, k2);
 719                 }
 720         }
 721
 722         if (!b->c.level &&
 723             node_iter == &iter->l[0].iter &&
 724             iter_current_key_modified)
 725                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 726 }
 727
 728 void bch2_btree_node_iter_fix(struct btree_iter *iter,
 729                               struct btree *b,
 730                               struct btree_node_iter *node_iter,
 731                               struct bkey_packed *where,
 732                               unsigned clobber_u64s,
 733                               unsigned new_u64s)
 734 {
 735         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 736         struct btree_iter *linked;
 737
 738         if (node_iter != &iter->l[b->c.level].iter) {
 739                 __bch2_btree_node_iter_fix(iter, b, node_iter, t,
 740                                            where, clobber_u64s, new_u64s);
 741
 742                 if (debug_check_iterators(iter->trans->c))
 743                         bch2_btree_node_iter_verify(node_iter, b);
 744         }
 745
 746         trans_for_each_iter_with_node(iter->trans, b, linked) {
 747                 __bch2_btree_node_iter_fix(linked, b,
 748                                            &linked->l[b->c.level].iter, t,
 749                                            where, clobber_u64s, new_u64s);
 750                 bch2_btree_iter_verify_level(linked, b->c.level);
 751         }
 752 }
 753
 754 static inline struct bkey_s_c __btree_iter_unpack(struct btree_iter *iter,
 755                                                   struct btree_iter_level *l,
 756                                                   struct bkey *u,
 757                                                   struct bkey_packed *k)
 758 {
 759         struct bkey_s_c ret;
 760
 761         if (unlikely(!k)) {
 762                 /*
 763                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 764                  * a hole
 765                  */
 766                 u->type = KEY_TYPE_deleted;
 767                 return bkey_s_c_null;
 768         }
 769
 770         ret = bkey_disassemble(l->b, k, u);
 771
 772         if (debug_check_bkeys(iter->trans->c))
 773                 bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
 774
 775         return ret;
 776 }
 777
 778 /* peek_all() doesn't skip deleted keys */
 779 static inline struct bkey_s_c __btree_iter_peek_all(struct btree_iter *iter,
 780                                                     struct btree_iter_level *l,
 781                                                     struct bkey *u)
 782 {
 783         return __btree_iter_unpack(iter, l, u,
 784                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 785 }
 786
 787 static inline struct bkey_s_c __btree_iter_peek(struct btree_iter *iter,
 788                                                 struct btree_iter_level *l)
 789 {
 790         return __btree_iter_unpack(iter, l, &iter->k,
 791                         bch2_btree_node_iter_peek(&l->iter, l->b));
 792 }
 793
 794 static inline struct bkey_s_c __btree_iter_prev(struct btree_iter *iter,
 795                                                 struct btree_iter_level *l)
 796 {
 797         return __btree_iter_unpack(iter, l, &iter->k,
 798                         bch2_btree_node_iter_prev(&l->iter, l->b));
 799 }
 800
 801 static inline bool btree_iter_advance_to_pos(struct btree_iter *iter,
 802                                              struct btree_iter_level *l,
 803                                              int max_advance)
 804 {
 805         struct bpos pos = btree_iter_search_key(iter);
 806         struct bkey_packed *k;
 807         int nr_advanced = 0;
 808
 809         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 810                bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
 811                 if (max_advance > 0 && nr_advanced >= max_advance)
 812                         return false;
 813
 814                 bch2_btree_node_iter_advance(&l->iter, l->b);
 815                 nr_advanced++;
 816         }
 817
 818         return true;
 819 }
 820
 821 /*
 822  * Verify that iterator for parent node points to child node:
 823  */
 824 static void btree_iter_verify_new_node(struct btree_iter *iter, struct btree *b)
 825 {
 826         struct btree_iter_level *l;
 827         unsigned plevel;
 828         bool parent_locked;
 829         struct bkey_packed *k;
 830
 831         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 832                 return;
 833
 834         plevel = b->c.level + 1;
 835         if (!btree_iter_node(iter, plevel))
 836                 return;
 837
 838         parent_locked = btree_node_locked(iter, plevel);
 839
 840         if (!bch2_btree_node_relock(iter, plevel))
 841                 return;
 842
 843         l = &iter->l[plevel];
 844         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 845         if (!k ||
 846             bkey_deleted(k) ||
 847             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
 848                 char buf[100];
 849                 struct bkey uk = bkey_unpack_key(b, k);
 850
 851                 bch2_bkey_to_text(&PBUF(buf), &uk);
 852                 panic("parent iter doesn't point to new node:\n%s\n%llu:%llu\n",
 853                       buf, b->key.k.p.inode, b->key.k.p.offset);
 854         }
 855
 856         if (!parent_locked)
 857                 btree_node_unlock(iter, b->c.level + 1);
 858 }
 859
 860 static inline void __btree_iter_init(struct btree_iter *iter,
 861                                      unsigned level)
 862 {
 863         struct bpos pos = btree_iter_search_key(iter);
 864         struct btree_iter_level *l = &iter->l[level];
 865
 866         bch2_btree_node_iter_init(&l->iter, l->b, &pos);
 867
 868         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 869 }
 870
 871 static inline void btree_iter_node_set(struct btree_iter *iter,
 872                                        struct btree *b)
 873 {
 874         BUG_ON(btree_iter_type(iter) == BTREE_ITER_CACHED);
 875
 876         btree_iter_verify_new_node(iter, b);
 877
 878         EBUG_ON(!btree_iter_pos_in_node(iter, b));
 879         EBUG_ON(b->c.lock.state.seq & 1);
 880
 881         iter->l[b->c.level].lock_seq = b->c.lock.state.seq;
 882         iter->l[b->c.level].b = b;
 883         __btree_iter_init(iter, b->c.level);
 884 }
 885
 886 /*
 887  * A btree node is being replaced - update the iterator to point to the new
 888  * node:
 889  */
 890 void bch2_btree_iter_node_replace(struct btree_iter *iter, struct btree *b)
 891 {
 892         enum btree_node_locked_type t;
 893         struct btree_iter *linked;
 894
 895         trans_for_each_iter(iter->trans, linked)
 896                 if (btree_iter_type(linked) != BTREE_ITER_CACHED &&
 897                     btree_iter_pos_in_node(linked, b)) {
 898                         /*
 899                          * bch2_btree_iter_node_drop() has already been called -
 900                          * the old node we're replacing has already been
 901                          * unlocked and the pointer invalidated
 902                          */
 903                         BUG_ON(btree_node_locked(linked, b->c.level));
 904
 905                         t = btree_lock_want(linked, b->c.level);
 906                         if (t != BTREE_NODE_UNLOCKED) {
 907                                 six_lock_increment(&b->c.lock, t);
 908                                 mark_btree_node_locked(linked, b->c.level, t);
 909                         }
 910
 911                         btree_iter_node_set(linked, b);
 912                 }
 913 }
 914
 915 void bch2_btree_iter_node_drop(struct btree_iter *iter, struct btree *b)
 916 {
 917         struct btree_iter *linked;
 918         unsigned level = b->c.level;
 919
 920         trans_for_each_iter(iter->trans, linked)
 921                 if (linked->l[level].b == b) {
 922                         __btree_node_unlock(linked, level);
 923                         linked->l[level].b = BTREE_ITER_NO_NODE_DROP;
 924                 }
 925 }
 926
 927 /*
 928  * A btree node has been modified in such a way as to invalidate iterators - fix
 929  * them:
 930  */
 931 void bch2_btree_iter_reinit_node(struct btree_iter *iter, struct btree *b)
 932 {
 933         struct btree_iter *linked;
 934
 935         trans_for_each_iter_with_node(iter->trans, b, linked)
 936                 __btree_iter_init(linked, b->c.level);
 937 }
 938
 939 static int lock_root_check_fn(struct six_lock *lock, void *p)
 940 {
 941         struct btree *b = container_of(lock, struct btree, c.lock);
 942         struct btree **rootp = p;
 943
 944         return b == *rootp ? 0 : -1;
 945 }
 946
 947 static inline int btree_iter_lock_root(struct btree_iter *iter,
 948                                        unsigned depth_want)
 949 {
 950         struct bch_fs *c = iter->trans->c;
 951         struct btree *b, **rootp = &c->btree_roots[iter->btree_id].b;
 952         enum six_lock_type lock_type;
 953         unsigned i;
 954
 955         EBUG_ON(iter->nodes_locked);
 956
 957         while (1) {
 958                 b = READ_ONCE(*rootp);
 959                 iter->level = READ_ONCE(b->c.level);
 960
 961                 if (unlikely(iter->level < depth_want)) {
 962                         /*
 963                          * the root is at a lower depth than the depth we want:
 964                          * got to the end of the btree, or we're walking nodes
 965                          * greater than some depth and there are no nodes >=
 966                          * that depth
 967                          */
 968                         iter->level = depth_want;
 969                         for (i = iter->level; i < BTREE_MAX_DEPTH; i++)
 970                                 iter->l[i].b = NULL;
 971                         return 1;
 972                 }
 973
 974                 lock_type = __btree_lock_want(iter, iter->level);
 975                 if (unlikely(!btree_node_lock(b, POS_MAX, iter->level,
 976                                               iter, lock_type,
 977                                               lock_root_check_fn, rootp)))
 978                         return -EINTR;
 979
 980                 if (likely(b == READ_ONCE(*rootp) &&
 981                            b->c.level == iter->level &&
 982                            !race_fault())) {
 983                         for (i = 0; i < iter->level; i++)
 984                                 iter->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
 985                         iter->l[iter->level].b = b;
 986                         for (i = iter->level + 1; i < BTREE_MAX_DEPTH; i++)
 987                                 iter->l[i].b = NULL;
 988
 989                         mark_btree_node_locked(iter, iter->level, lock_type);
 990                         btree_iter_node_set(iter, b);
 991                         return 0;
 992                 }
 993
 994                 six_unlock_type(&b->c.lock, lock_type);
 995         }
 996 }
 997
 998 noinline
 999 static void btree_iter_prefetch(struct btree_iter *iter)
1000 {
1001         struct bch_fs *c = iter->trans->c;
1002         struct btree_iter_level *l = &iter->l[iter->level];
1003         struct btree_node_iter node_iter = l->iter;
1004         struct bkey_packed *k;
1005         BKEY_PADDED(k) tmp;
1006         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
1007                 ? (iter->level > 1 ? 0 :  2)
1008                 : (iter->level > 1 ? 1 : 16);
1009         bool was_locked = btree_node_locked(iter, iter->level);
1010
1011         while (nr) {
1012                 if (!bch2_btree_node_relock(iter, iter->level))
1013                         return;
1014
1015                 bch2_btree_node_iter_advance(&node_iter, l->b);
1016                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
1017                 if (!k)
1018                         break;
1019
1020                 bch2_bkey_unpack(l->b, &tmp.k, k);
1021                 bch2_btree_node_prefetch(c, iter, &tmp.k, iter->level - 1);
1022         }
1023
1024         if (!was_locked)
1025                 btree_node_unlock(iter, iter->level);
1026 }
1027
1028 static noinline void btree_node_mem_ptr_set(struct btree_iter *iter,
1029                                             unsigned plevel, struct btree *b)
1030 {
1031         struct btree_iter_level *l = &iter->l[plevel];
1032         bool locked = btree_node_locked(iter, plevel);
1033         struct bkey_packed *k;
1034         struct bch_btree_ptr_v2 *bp;
1035
1036         if (!bch2_btree_node_relock(iter, plevel))
1037                 return;
1038
1039         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1040         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
1041
1042         bp = (void *) bkeyp_val(&l->b->format, k);
1043         bp->mem_ptr = (unsigned long)b;
1044
1045         if (!locked)
1046                 btree_node_unlock(iter, plevel);
1047 }
1048
1049 static __always_inline int btree_iter_down(struct btree_iter *iter)
1050 {
1051         struct bch_fs *c = iter->trans->c;
1052         struct btree_iter_level *l = &iter->l[iter->level];
1053         struct btree *b;
1054         unsigned level = iter->level - 1;
1055         enum six_lock_type lock_type = __btree_lock_want(iter, level);
1056         BKEY_PADDED(k) tmp;
1057
1058         EBUG_ON(!btree_node_locked(iter, iter->level));
1059
1060         bch2_bkey_unpack(l->b, &tmp.k,
1061                          bch2_btree_node_iter_peek(&l->iter, l->b));
1062
1063         b = bch2_btree_node_get(c, iter, &tmp.k, level, lock_type);
1064         if (unlikely(IS_ERR(b)))
1065                 return PTR_ERR(b);
1066
1067         mark_btree_node_locked(iter, level, lock_type);
1068         btree_iter_node_set(iter, b);
1069
1070         if (tmp.k.k.type == KEY_TYPE_btree_ptr_v2 &&
1071             unlikely(b != btree_node_mem_ptr(&tmp.k)))
1072                 btree_node_mem_ptr_set(iter, level + 1, b);
1073
1074         if (iter->flags & BTREE_ITER_PREFETCH)
1075                 btree_iter_prefetch(iter);
1076
1077         iter->level = level;
1078
1079         return 0;
1080 }
1081
1082 static void btree_iter_up(struct btree_iter *iter)
1083 {
1084         btree_node_unlock(iter, iter->level++);
1085 }
1086
1087 static int btree_iter_traverse_one(struct btree_iter *);
1088
1089 static int __btree_iter_traverse_all(struct btree_trans *trans, int ret)
1090 {
1091         struct bch_fs *c = trans->c;
1092         struct btree_iter *iter;
1093         u8 sorted[BTREE_ITER_MAX];
1094         unsigned i, nr_sorted = 0;
1095
1096         if (trans->in_traverse_all)
1097                 return -EINTR;
1098
1099         trans->in_traverse_all = true;
1100 retry_all:
1101         nr_sorted = 0;
1102
1103         trans_for_each_iter(trans, iter)
1104                 sorted[nr_sorted++] = iter->idx;
1105
1106 #define btree_iter_cmp_by_idx(_l, _r)                           \
1107                 btree_iter_cmp(&trans->iters[_l], &trans->iters[_r])
1108
1109         bubble_sort(sorted, nr_sorted, btree_iter_cmp_by_idx);
1110 #undef btree_iter_cmp_by_idx
1111         bch2_trans_unlock(trans);
1112
1113         if (unlikely(ret == -ENOMEM)) {
1114                 struct closure cl;
1115
1116                 closure_init_stack(&cl);
1117
1118                 do {
1119                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1120                         closure_sync(&cl);
1121                 } while (ret);
1122         }
1123
1124         if (unlikely(ret == -EIO)) {
1125                 trans->error = true;
1126                 goto out;
1127         }
1128
1129         BUG_ON(ret && ret != -EINTR);
1130
1131         /* Now, redo traversals in correct order: */
1132         for (i = 0; i < nr_sorted; i++) {
1133                 unsigned idx = sorted[i];
1134
1135                 /*
1136                  * sucessfully traversing one iterator can cause another to be
1137                  * unlinked, in btree_key_cache_fill()
1138                  */
1139                 if (!(trans->iters_linked & (1ULL << idx)))
1140                         continue;
1141
1142                 ret = btree_iter_traverse_one(&trans->iters[idx]);
1143                 if (ret)
1144                         goto retry_all;
1145         }
1146
1147         if (hweight64(trans->iters_live) > 1)
1148                 ret = -EINTR;
1149         else
1150                 trans_for_each_iter(trans, iter)
1151                         if (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) {
1152                                 ret = -EINTR;
1153                                 break;
1154                         }
1155 out:
1156         bch2_btree_cache_cannibalize_unlock(c);
1157
1158         trans->in_traverse_all = false;
1159         return ret;
1160 }
1161
1162 int bch2_btree_iter_traverse_all(struct btree_trans *trans)
1163 {
1164         return __btree_iter_traverse_all(trans, 0);
1165 }
1166
1167 static inline bool btree_iter_good_node(struct btree_iter *iter,
1168                                         unsigned l, int check_pos)
1169 {
1170         if (!is_btree_node(iter, l) ||
1171             !bch2_btree_node_relock(iter, l))
1172                 return false;
1173
1174         if (check_pos <= 0 && btree_iter_pos_before_node(iter, iter->l[l].b))
1175                 return false;
1176         if (check_pos >= 0 && btree_iter_pos_after_node(iter, iter->l[l].b))
1177                 return false;
1178         return true;
1179 }
1180
1181 static inline unsigned btree_iter_up_until_good_node(struct btree_iter *iter,
1182                                                      int check_pos)
1183 {
1184         unsigned l = iter->level;
1185
1186         while (btree_iter_node(iter, l) &&
1187                !btree_iter_good_node(iter, l, check_pos)) {
1188                 btree_node_unlock(iter, l);
1189                 iter->l[l].b = BTREE_ITER_NO_NODE_UP;
1190                 l++;
1191         }
1192
1193         return l;
1194 }
1195
1196 /*
1197  * This is the main state machine for walking down the btree - walks down to a
1198  * specified depth
1199  *
1200  * Returns 0 on success, -EIO on error (error reading in a btree node).
1201  *
1202  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1203  * stashed in the iterator and returned from bch2_trans_exit().
1204  */
1205 static int btree_iter_traverse_one(struct btree_iter *iter)
1206 {
1207         unsigned depth_want = iter->level;
1208
1209         /*
1210          * if we need interior nodes locked, call btree_iter_relock() to make
1211          * sure we walk back up enough that we lock them:
1212          */
1213         if (iter->uptodate == BTREE_ITER_NEED_RELOCK ||
1214             iter->locks_want > 1)
1215                 bch2_btree_iter_relock(iter, false);
1216
1217         if (btree_iter_type(iter) == BTREE_ITER_CACHED)
1218                 return bch2_btree_iter_traverse_cached(iter);
1219
1220         if (iter->uptodate < BTREE_ITER_NEED_RELOCK)
1221                 return 0;
1222
1223         if (unlikely(iter->level >= BTREE_MAX_DEPTH))
1224                 return 0;
1225
1226         /*
1227          * XXX: correctly using BTREE_ITER_UPTODATE should make using check_pos
1228          * here unnecessary
1229          */
1230         iter->level = btree_iter_up_until_good_node(iter, 0);
1231
1232         /*
1233          * If we've got a btree node locked (i.e. we aren't about to relock the
1234          * root) - advance its node iterator if necessary:
1235          *
1236          * XXX correctly using BTREE_ITER_UPTODATE should make this unnecessary
1237          */
1238         if (is_btree_node(iter, iter->level)) {
1239                 BUG_ON(!btree_iter_pos_in_node(iter, iter->l[iter->level].b));
1240
1241                 btree_iter_advance_to_pos(iter, &iter->l[iter->level], -1);
1242         }
1243
1244         /*
1245          * Note: iter->nodes[iter->level] may be temporarily NULL here - that
1246          * would indicate to other code that we got to the end of the btree,
1247          * here it indicates that relocking the root failed - it's critical that
1248          * btree_iter_lock_root() comes next and that it can't fail
1249          */
1250         while (iter->level > depth_want) {
1251                 int ret = btree_iter_node(iter, iter->level)
1252                         ? btree_iter_down(iter)
1253                         : btree_iter_lock_root(iter, depth_want);
1254                 if (unlikely(ret)) {
1255                         if (ret == 1)
1256                                 return 0;
1257
1258                         iter->level = depth_want;
1259
1260                         if (ret == -EIO) {
1261                                 iter->flags |= BTREE_ITER_ERROR;
1262                                 iter->l[iter->level].b =
1263                                         BTREE_ITER_NO_NODE_ERROR;
1264                         } else {
1265                                 iter->l[iter->level].b =
1266                                         BTREE_ITER_NO_NODE_DOWN;
1267                         }
1268                         return ret;
1269                 }
1270         }
1271
1272         iter->uptodate = BTREE_ITER_NEED_PEEK;
1273
1274         bch2_btree_iter_verify(iter);
1275         return 0;
1276 }
1277
1278 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
1279 {
1280         struct btree_trans *trans = iter->trans;
1281         int ret;
1282
1283         ret =   bch2_trans_cond_resched(trans) ?:
1284                 btree_iter_traverse_one(iter);
1285         if (unlikely(ret))
1286                 ret = __btree_iter_traverse_all(trans, ret);
1287
1288         return ret;
1289 }
1290
1291 static inline void bch2_btree_iter_checks(struct btree_iter *iter)
1292 {
1293         enum btree_iter_type type = btree_iter_type(iter);
1294
1295         EBUG_ON(iter->btree_id >= BTREE_ID_NR);
1296
1297         BUG_ON((type == BTREE_ITER_KEYS ||
1298                 type == BTREE_ITER_CACHED) &&
1299                (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
1300                 bkey_cmp(iter->pos, iter->k.p) > 0));
1301
1302         bch2_btree_iter_verify_locks(iter);
1303         bch2_btree_iter_verify_level(iter, iter->level);
1304 }
1305
1306 /* Iterate across nodes (leaf and interior nodes) */
1307
1308 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1309 {
1310         struct btree *b;
1311         int ret;
1312
1313         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
1314         bch2_btree_iter_checks(iter);
1315
1316         if (iter->uptodate == BTREE_ITER_UPTODATE)
1317                 return iter->l[iter->level].b;
1318
1319         ret = bch2_btree_iter_traverse(iter);
1320         if (ret)
1321                 return NULL;
1322
1323         b = btree_iter_node(iter, iter->level);
1324         if (!b)
1325                 return NULL;
1326
1327         BUG_ON(bkey_cmp(b->key.k.p, iter->pos) < 0);
1328
1329         iter->pos = b->key.k.p;
1330         iter->uptodate = BTREE_ITER_UPTODATE;
1331
1332         bch2_btree_iter_verify(iter);
1333
1334         return b;
1335 }
1336
1337 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1338 {
1339         struct btree *b;
1340         int ret;
1341
1342         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES);
1343         bch2_btree_iter_checks(iter);
1344
1345         /* already got to end? */
1346         if (!btree_iter_node(iter, iter->level))
1347                 return NULL;
1348
1349         bch2_trans_cond_resched(iter->trans);
1350
1351         btree_iter_up(iter);
1352
1353         if (!bch2_btree_node_relock(iter, iter->level))
1354                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
1355
1356         ret = bch2_btree_iter_traverse(iter);
1357         if (ret)
1358                 return NULL;
1359
1360         /* got to end? */
1361         b = btree_iter_node(iter, iter->level);
1362         if (!b)
1363                 return NULL;
1364
1365         if (bkey_cmp(iter->pos, b->key.k.p) < 0) {
1366                 /*
1367                  * Haven't gotten to the end of the parent node: go back down to
1368                  * the next child node
1369                  */
1370
1371                 /*
1372                  * We don't really want to be unlocking here except we can't
1373                  * directly tell btree_iter_traverse() "traverse to this level"
1374                  * except by setting iter->level, so we have to unlock so we
1375                  * don't screw up our lock invariants:
1376                  */
1377                 if (btree_node_read_locked(iter, iter->level))
1378                         btree_node_unlock(iter, iter->level);
1379
1380                 iter->pos       = bkey_successor(iter->pos);
1381                 iter->level     = iter->min_depth;
1382
1383                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1384                 ret = bch2_btree_iter_traverse(iter);
1385                 if (ret)
1386                         return NULL;
1387
1388                 b = iter->l[iter->level].b;
1389         }
1390
1391         iter->pos = b->key.k.p;
1392         iter->uptodate = BTREE_ITER_UPTODATE;
1393
1394         bch2_btree_iter_verify(iter);
1395
1396         return b;
1397 }
1398
1399 /* Iterate across keys (in leaf nodes only) */
1400
1401 void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_pos)
1402 {
1403         struct btree_iter_level *l = &iter->l[0];
1404
1405         EBUG_ON(iter->level != 0);
1406         EBUG_ON(bkey_cmp(new_pos, iter->pos) < 0);
1407         EBUG_ON(!btree_node_locked(iter, 0));
1408         EBUG_ON(bkey_cmp(new_pos, l->b->key.k.p) > 0);
1409
1410         bkey_init(&iter->k);
1411         iter->k.p = iter->pos = new_pos;
1412         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
1413
1414         btree_iter_advance_to_pos(iter, l, -1);
1415
1416         /*
1417          * XXX:
1418          * keeping a node locked that's outside (even just outside) iter->pos
1419          * breaks __bch2_btree_node_lock(). This seems to only affect
1420          * bch2_btree_node_get_sibling so for now it's fixed there, but we
1421          * should try to get rid of this corner case.
1422          *
1423          * (this behaviour is currently needed for BTREE_INSERT_NOUNLOCK)
1424          */
1425
1426         if (bch2_btree_node_iter_end(&l->iter) &&
1427             btree_iter_pos_after_node(iter, l->b))
1428                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1429 }
1430
1431 static void btree_iter_pos_changed(struct btree_iter *iter, int cmp)
1432 {
1433         unsigned l = iter->level;
1434
1435         if (!cmp)
1436                 goto out;
1437
1438         if (unlikely(btree_iter_type(iter) == BTREE_ITER_CACHED)) {
1439                 btree_node_unlock(iter, 0);
1440                 iter->l[0].b = BTREE_ITER_NO_NODE_UP;
1441                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1442                 return;
1443         }
1444
1445         l = btree_iter_up_until_good_node(iter, cmp);
1446
1447         if (btree_iter_node(iter, l)) {
1448                 /*
1449                  * We might have to skip over many keys, or just a few: try
1450                  * advancing the node iterator, and if we have to skip over too
1451                  * many keys just reinit it (or if we're rewinding, since that
1452                  * is expensive).
1453                  */
1454                 if (cmp < 0 ||
1455                     !btree_iter_advance_to_pos(iter, &iter->l[l], 8))
1456                         __btree_iter_init(iter, l);
1457
1458                 /* Don't leave it locked if we're not supposed to: */
1459                 if (btree_lock_want(iter, l) == BTREE_NODE_UNLOCKED)
1460                         btree_node_unlock(iter, l);
1461         }
1462 out:
1463         if (l != iter->level)
1464                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1465         else
1466                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
1467 }
1468
1469 void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos,
1470                                bool strictly_greater)
1471 {
1472         struct bpos old = btree_iter_search_key(iter);
1473         int cmp;
1474
1475         iter->flags &= ~BTREE_ITER_IS_EXTENTS;
1476         iter->flags |= strictly_greater ? BTREE_ITER_IS_EXTENTS : 0;
1477
1478         bkey_init(&iter->k);
1479         iter->k.p = iter->pos = new_pos;
1480
1481         cmp = bkey_cmp(btree_iter_search_key(iter), old);
1482
1483         btree_iter_pos_changed(iter, cmp);
1484 }
1485
1486 void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
1487 {
1488         int cmp = bkey_cmp(new_pos, iter->pos);
1489
1490         bkey_init(&iter->k);
1491         iter->k.p = iter->pos = new_pos;
1492
1493         btree_iter_pos_changed(iter, cmp);
1494 }
1495
1496 static inline bool btree_iter_set_pos_to_next_leaf(struct btree_iter *iter)
1497 {
1498         struct btree_iter_level *l = &iter->l[0];
1499         bool ret;
1500
1501         bkey_init(&iter->k);
1502         iter->k.p = iter->pos = l->b->key.k.p;
1503
1504         ret = bkey_cmp(iter->pos, POS_MAX) != 0;
1505         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1506                 iter->k.p = iter->pos = bkey_successor(iter->pos);
1507
1508         btree_iter_pos_changed(iter, 1);
1509         return ret;
1510 }
1511
1512 static inline bool btree_iter_set_pos_to_prev_leaf(struct btree_iter *iter)
1513 {
1514         struct btree_iter_level *l = &iter->l[0];
1515         bool ret;
1516
1517         bkey_init(&iter->k);
1518         iter->k.p = iter->pos = l->b->data->min_key;
1519         iter->uptodate  = BTREE_ITER_NEED_TRAVERSE;
1520
1521         ret = bkey_cmp(iter->pos, POS_MIN) != 0;
1522         if (ret) {
1523                 iter->k.p = iter->pos = bkey_predecessor(iter->pos);
1524
1525                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
1526                         iter->k.p = iter->pos = bkey_predecessor(iter->pos);
1527         }
1528
1529         btree_iter_pos_changed(iter, -1);
1530         return ret;
1531 }
1532
1533 /**
1534  * btree_iter_peek_uptodate - given an iterator that is uptodate, return the key
1535  * it currently points to
1536  */
1537 static inline struct bkey_s_c btree_iter_peek_uptodate(struct btree_iter *iter)
1538 {
1539         struct btree_iter_level *l = &iter->l[0];
1540         struct bkey_s_c ret = { .k = &iter->k };
1541
1542         if (!bkey_deleted(&iter->k)) {
1543                 struct bkey_packed *_k =
1544                         __bch2_btree_node_iter_peek_all(&l->iter, l->b);
1545
1546                 ret.v = bkeyp_val(&l->b->format, _k);
1547
1548                 if (debug_check_iterators(iter->trans->c)) {
1549                         struct bkey k = bkey_unpack_key(l->b, _k);
1550
1551                         BUG_ON(memcmp(&k, &iter->k, sizeof(k)));
1552                 }
1553
1554                 if (debug_check_bkeys(iter->trans->c))
1555                         bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
1556         }
1557
1558         return ret;
1559 }
1560
1561 /**
1562  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
1563  * current position
1564  */
1565 struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
1566 {
1567         struct btree_iter_level *l = &iter->l[0];
1568         struct bkey_s_c k;
1569         int ret;
1570
1571         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1572         bch2_btree_iter_checks(iter);
1573
1574         if (iter->uptodate == BTREE_ITER_UPTODATE &&
1575             !bkey_deleted(&iter->k))
1576                 return btree_iter_peek_uptodate(iter);
1577
1578         while (1) {
1579                 ret = bch2_btree_iter_traverse(iter);
1580                 if (unlikely(ret))
1581                         return bkey_s_c_err(ret);
1582
1583                 k = __btree_iter_peek(iter, l);
1584                 if (likely(k.k))
1585                         break;
1586
1587                 if (!btree_iter_set_pos_to_next_leaf(iter))
1588                         return bkey_s_c_null;
1589         }
1590
1591         /*
1592          * iter->pos should always be equal to the key we just
1593          * returned - except extents can straddle iter->pos:
1594          */
1595         if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
1596             bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
1597                 iter->pos = bkey_start_pos(k.k);
1598
1599         iter->uptodate = BTREE_ITER_UPTODATE;
1600
1601         bch2_btree_iter_verify_level(iter, 0);
1602         return k;
1603 }
1604
1605 /**
1606  * bch2_btree_iter_next: returns first key greater than iterator's current
1607  * position
1608  */
1609 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
1610 {
1611         if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
1612                 return bkey_s_c_null;
1613
1614         bch2_btree_iter_set_pos(iter,
1615                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1616                 ? iter->k.p
1617                 : bkey_successor(iter->k.p));
1618
1619         return bch2_btree_iter_peek(iter);
1620 }
1621
1622 static struct bkey_s_c __btree_trans_updates_peek(struct btree_iter *iter)
1623 {
1624         struct bpos pos = btree_iter_search_key(iter);
1625         struct btree_trans *trans = iter->trans;
1626         struct btree_insert_entry *i;
1627
1628         trans_for_each_update2(trans, i)
1629                 if ((cmp_int(iter->btree_id,    i->iter->btree_id) ?:
1630                      bkey_cmp(pos,              i->k->k.p)) <= 0)
1631                         break;
1632
1633         return i < trans->updates2 + trans->nr_updates2 &&
1634                 iter->btree_id == i->iter->btree_id
1635                 ? bkey_i_to_s_c(i->k)
1636                 : bkey_s_c_null;
1637 }
1638
1639 static struct bkey_s_c __bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
1640 {
1641         struct btree_iter_level *l = &iter->l[0];
1642         struct bkey_s_c k = __btree_iter_peek(iter, l);
1643         struct bkey_s_c u = __btree_trans_updates_peek(iter);
1644
1645         if (k.k && (!u.k || bkey_cmp(k.k->p, u.k->p) < 0))
1646                 return k;
1647         if (u.k && bkey_cmp(u.k->p, l->b->key.k.p) <= 0) {
1648                 iter->k = *u.k;
1649                 return u;
1650         }
1651         return bkey_s_c_null;
1652 }
1653
1654 struct bkey_s_c bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
1655 {
1656         struct bkey_s_c k;
1657         int ret;
1658
1659         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1660         bch2_btree_iter_checks(iter);
1661
1662         while (1) {
1663                 ret = bch2_btree_iter_traverse(iter);
1664                 if (unlikely(ret))
1665                         return bkey_s_c_err(ret);
1666
1667                 k = __bch2_btree_iter_peek_with_updates(iter);
1668
1669                 if (k.k && bkey_deleted(k.k)) {
1670                         bch2_btree_iter_set_pos(iter,
1671                                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1672                                 ? iter->k.p
1673                                 : bkey_successor(iter->k.p));
1674                         continue;
1675                 }
1676
1677                 if (likely(k.k))
1678                         break;
1679
1680                 if (!btree_iter_set_pos_to_next_leaf(iter))
1681                         return bkey_s_c_null;
1682         }
1683
1684         /*
1685          * iter->pos should always be equal to the key we just
1686          * returned - except extents can straddle iter->pos:
1687          */
1688         if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
1689             bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
1690                 iter->pos = bkey_start_pos(k.k);
1691
1692         iter->uptodate = BTREE_ITER_UPTODATE;
1693         return k;
1694 }
1695
1696 struct bkey_s_c bch2_btree_iter_next_with_updates(struct btree_iter *iter)
1697 {
1698         if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
1699                 return bkey_s_c_null;
1700
1701         bch2_btree_iter_set_pos(iter,
1702                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1703                 ? iter->k.p
1704                 : bkey_successor(iter->k.p));
1705
1706         return bch2_btree_iter_peek_with_updates(iter);
1707 }
1708
1709 /**
1710  * bch2_btree_iter_peek_prev: returns first key less than or equal to
1711  * iterator's current position
1712  */
1713 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
1714 {
1715         struct bpos pos = iter->pos;
1716         struct btree_iter_level *l = &iter->l[0];
1717         struct bkey_s_c k;
1718         int ret;
1719
1720         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1721         bch2_btree_iter_checks(iter);
1722
1723         if (iter->uptodate == BTREE_ITER_UPTODATE &&
1724             !bkey_deleted(&iter->k))
1725                 return btree_iter_peek_uptodate(iter);
1726
1727         while (1) {
1728                 ret = bch2_btree_iter_traverse(iter);
1729                 if (unlikely(ret))
1730                         return bkey_s_c_err(ret);
1731
1732                 k = __btree_iter_peek(iter, l);
1733                 if (!k.k || bkey_cmp(bkey_start_pos(k.k), pos) > 0)
1734                         k = __btree_iter_prev(iter, l);
1735
1736                 if (likely(k.k))
1737                         break;
1738
1739                 if (!btree_iter_set_pos_to_prev_leaf(iter))
1740                         return bkey_s_c_null;
1741         }
1742
1743         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), pos) > 0);
1744         iter->pos       = bkey_start_pos(k.k);
1745         iter->uptodate  = BTREE_ITER_UPTODATE;
1746         return k;
1747 }
1748
1749 /**
1750  * bch2_btree_iter_prev: returns first key less than iterator's current
1751  * position
1752  */
1753 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
1754 {
1755         struct bpos pos = bkey_start_pos(&iter->k);
1756
1757         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1758         bch2_btree_iter_checks(iter);
1759
1760         if (unlikely(!bkey_cmp(pos, POS_MIN)))
1761                 return bkey_s_c_null;
1762
1763         bch2_btree_iter_set_pos(iter, bkey_predecessor(pos));
1764
1765         return bch2_btree_iter_peek_prev(iter);
1766 }
1767
1768 static inline struct bkey_s_c
1769 __bch2_btree_iter_peek_slot_extents(struct btree_iter *iter)
1770 {
1771         struct btree_iter_level *l = &iter->l[0];
1772         struct btree_node_iter node_iter;
1773         struct bkey_s_c k;
1774         struct bkey n;
1775         int ret;
1776
1777         /* keys & holes can't span inode numbers: */
1778         if (iter->pos.offset == KEY_OFFSET_MAX) {
1779                 if (iter->pos.inode == KEY_INODE_MAX)
1780                         return bkey_s_c_null;
1781
1782                 bch2_btree_iter_set_pos(iter, bkey_successor(iter->pos));
1783
1784                 ret = bch2_btree_iter_traverse(iter);
1785                 if (unlikely(ret))
1786                         return bkey_s_c_err(ret);
1787         }
1788
1789         /*
1790          * iterator is now at the correct position for inserting at iter->pos,
1791          * but we need to keep iterating until we find the first non whiteout so
1792          * we know how big a hole we have, if any:
1793          */
1794
1795         node_iter = l->iter;
1796         k = __btree_iter_unpack(iter, l, &iter->k,
1797                 bch2_btree_node_iter_peek(&node_iter, l->b));
1798
1799         if (k.k && bkey_cmp(bkey_start_pos(k.k), iter->pos) <= 0) {
1800                 /*
1801                  * We're not setting iter->uptodate because the node iterator
1802                  * doesn't necessarily point at the key we're returning:
1803                  */
1804
1805                 EBUG_ON(bkey_cmp(k.k->p, iter->pos) <= 0);
1806                 bch2_btree_iter_verify_level(iter, 0);
1807                 return k;
1808         }
1809
1810         /* hole */
1811
1812         if (!k.k)
1813                 k.k = &l->b->key.k;
1814
1815         bkey_init(&n);
1816         n.p = iter->pos;
1817         bch2_key_resize(&n,
1818                         min_t(u64, KEY_SIZE_MAX,
1819                               (k.k->p.inode == n.p.inode
1820                                ? bkey_start_offset(k.k)
1821                                : KEY_OFFSET_MAX) -
1822                               n.p.offset));
1823
1824         EBUG_ON(!n.size);
1825
1826         iter->k = n;
1827         iter->uptodate = BTREE_ITER_UPTODATE;
1828
1829         bch2_btree_iter_verify_level(iter, 0);
1830         return (struct bkey_s_c) { &iter->k, NULL };
1831 }
1832
1833 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
1834 {
1835         struct btree_iter_level *l = &iter->l[0];
1836         struct bkey_s_c k;
1837         int ret;
1838
1839         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS);
1840         bch2_btree_iter_checks(iter);
1841
1842         if (iter->uptodate == BTREE_ITER_UPTODATE)
1843                 return btree_iter_peek_uptodate(iter);
1844
1845         ret = bch2_btree_iter_traverse(iter);
1846         if (unlikely(ret))
1847                 return bkey_s_c_err(ret);
1848
1849         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1850                 return __bch2_btree_iter_peek_slot_extents(iter);
1851
1852         k = __btree_iter_peek_all(iter, l, &iter->k);
1853
1854         EBUG_ON(k.k && bkey_deleted(k.k) && bkey_cmp(k.k->p, iter->pos) == 0);
1855
1856         if (!k.k || bkey_cmp(iter->pos, k.k->p)) {
1857                 /* hole */
1858                 bkey_init(&iter->k);
1859                 iter->k.p = iter->pos;
1860                 k = (struct bkey_s_c) { &iter->k, NULL };
1861         }
1862
1863         iter->uptodate = BTREE_ITER_UPTODATE;
1864         bch2_btree_iter_verify_level(iter, 0);
1865         return k;
1866 }
1867
1868 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
1869 {
1870         if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
1871                 return bkey_s_c_null;
1872
1873         bch2_btree_iter_set_pos(iter,
1874                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1875                 ? iter->k.p
1876                 : bkey_successor(iter->k.p));
1877
1878         return bch2_btree_iter_peek_slot(iter);
1879 }
1880
1881 struct bkey_s_c bch2_btree_iter_peek_cached(struct btree_iter *iter)
1882 {
1883         struct bkey_cached *ck;
1884         int ret;
1885
1886         EBUG_ON(btree_iter_type(iter) != BTREE_ITER_CACHED);
1887         bch2_btree_iter_checks(iter);
1888
1889         ret = bch2_btree_iter_traverse(iter);
1890         if (unlikely(ret))
1891                 return bkey_s_c_err(ret);
1892
1893         ck = (void *) iter->l[0].b;
1894
1895         EBUG_ON(iter->btree_id != ck->key.btree_id ||
1896                 bkey_cmp(iter->pos, ck->key.pos));
1897         BUG_ON(!ck->valid);
1898
1899         return bkey_i_to_s_c(ck->k);
1900 }
1901
1902 static inline void bch2_btree_iter_init(struct btree_trans *trans,
1903                         struct btree_iter *iter, enum btree_id btree_id,
1904                         struct bpos pos, unsigned flags)
1905 {
1906         struct bch_fs *c = trans->c;
1907         unsigned i;
1908
1909         if (btree_node_type_is_extents(btree_id) &&
1910             !(flags & BTREE_ITER_NODES))
1911                 flags |= BTREE_ITER_IS_EXTENTS;
1912
1913         iter->trans                     = trans;
1914         iter->pos                       = pos;
1915         bkey_init(&iter->k);
1916         iter->k.p                       = pos;
1917         iter->flags                     = flags;
1918         iter->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1919         iter->btree_id                  = btree_id;
1920         iter->level                     = 0;
1921         iter->min_depth                 = 0;
1922         iter->locks_want                = flags & BTREE_ITER_INTENT ? 1 : 0;
1923         iter->nodes_locked              = 0;
1924         iter->nodes_intent_locked       = 0;
1925         for (i = 0; i < ARRAY_SIZE(iter->l); i++)
1926                 iter->l[i].b            = BTREE_ITER_NO_NODE_INIT;
1927
1928         prefetch(c->btree_roots[btree_id].b);
1929 }
1930
1931 /* new transactional stuff: */
1932
1933 static inline void __bch2_trans_iter_free(struct btree_trans *trans,
1934                                           unsigned idx)
1935 {
1936         __bch2_btree_iter_unlock(&trans->iters[idx]);
1937         trans->iters_linked             &= ~(1ULL << idx);
1938         trans->iters_live               &= ~(1ULL << idx);
1939         trans->iters_touched            &= ~(1ULL << idx);
1940 }
1941
1942 int bch2_trans_iter_put(struct btree_trans *trans,
1943                         struct btree_iter *iter)
1944 {
1945         int ret;
1946
1947         if (IS_ERR_OR_NULL(iter))
1948                 return 0;
1949
1950         BUG_ON(trans->iters + iter->idx != iter);
1951
1952         ret = btree_iter_err(iter);
1953
1954         if (!(trans->iters_touched & (1ULL << iter->idx)) &&
1955             !(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT))
1956                 __bch2_trans_iter_free(trans, iter->idx);
1957
1958         trans->iters_live       &= ~(1ULL << iter->idx);
1959         return ret;
1960 }
1961
1962 int bch2_trans_iter_free(struct btree_trans *trans,
1963                          struct btree_iter *iter)
1964 {
1965         if (IS_ERR_OR_NULL(iter))
1966                 return 0;
1967
1968         trans->iters_touched &= ~(1ULL << iter->idx);
1969
1970         return bch2_trans_iter_put(trans, iter);
1971 }
1972
1973 static int bch2_trans_realloc_iters(struct btree_trans *trans,
1974                                     unsigned new_size)
1975 {
1976         void *p, *new_iters, *new_updates, *new_updates2;
1977         size_t iters_bytes;
1978         size_t updates_bytes;
1979
1980         new_size = roundup_pow_of_two(new_size);
1981
1982         BUG_ON(new_size > BTREE_ITER_MAX);
1983
1984         if (new_size <= trans->size)
1985                 return 0;
1986
1987         BUG_ON(trans->used_mempool);
1988
1989         bch2_trans_unlock(trans);
1990
1991         iters_bytes     = sizeof(struct btree_iter) * new_size;
1992         updates_bytes   = sizeof(struct btree_insert_entry) * new_size;
1993
1994         p = kmalloc(iters_bytes +
1995                     updates_bytes +
1996                     updates_bytes, GFP_NOFS);
1997         if (p)
1998                 goto success;
1999
2000         p = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
2001         new_size = BTREE_ITER_MAX;
2002
2003         trans->used_mempool = true;
2004 success:
2005         new_iters       = p; p += iters_bytes;
2006         new_updates     = p; p += updates_bytes;
2007         new_updates2    = p; p += updates_bytes;
2008
2009         memcpy(new_iters, trans->iters,
2010                sizeof(struct btree_iter) * trans->nr_iters);
2011         memcpy(new_updates, trans->updates,
2012                sizeof(struct btree_insert_entry) * trans->nr_updates);
2013         memcpy(new_updates2, trans->updates2,
2014                sizeof(struct btree_insert_entry) * trans->nr_updates2);
2015
2016         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
2017                 memset(trans->iters, POISON_FREE,
2018                        sizeof(struct btree_iter) * trans->nr_iters +
2019                        sizeof(struct btree_insert_entry) * trans->nr_iters);
2020
2021         if (trans->iters != trans->iters_onstack)
2022                 kfree(trans->iters);
2023
2024         trans->iters            = new_iters;
2025         trans->updates          = new_updates;
2026         trans->updates2         = new_updates2;
2027         trans->size             = new_size;
2028
2029         if (trans->iters_live) {
2030                 trace_trans_restart_iters_realloced(trans->ip, trans->size);
2031                 return -EINTR;
2032         }
2033
2034         return 0;
2035 }
2036
2037 static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans)
2038 {
2039         unsigned idx = __ffs64(~trans->iters_linked);
2040
2041         if (idx < trans->nr_iters)
2042                 goto got_slot;
2043
2044         if (trans->nr_iters == trans->size) {
2045                 int ret;
2046
2047                 if (trans->nr_iters >= BTREE_ITER_MAX) {
2048                         struct btree_iter *iter;
2049
2050                         trans_for_each_iter(trans, iter) {
2051                                 pr_err("iter: btree %s pos %llu:%llu%s%s%s %ps",
2052                                        bch2_btree_ids[iter->btree_id],
2053                                        iter->pos.inode,
2054                                        iter->pos.offset,
2055                                        (trans->iters_live & (1ULL << iter->idx)) ? " live" : "",
2056                                        (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "",
2057                                        iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "",
2058                                        (void *) iter->ip_allocated);
2059                         }
2060
2061                         panic("trans iter oveflow\n");
2062                 }
2063
2064                 ret = bch2_trans_realloc_iters(trans, trans->size * 2);
2065                 if (ret)
2066                         return ERR_PTR(ret);
2067         }
2068
2069         idx = trans->nr_iters++;
2070         BUG_ON(trans->nr_iters > trans->size);
2071
2072         trans->iters[idx].idx = idx;
2073 got_slot:
2074         BUG_ON(trans->iters_linked & (1ULL << idx));
2075         trans->iters_linked |= 1ULL << idx;
2076         trans->iters[idx].flags = 0;
2077         return &trans->iters[idx];
2078 }
2079
2080 static inline void btree_iter_copy(struct btree_iter *dst,
2081                                    struct btree_iter *src)
2082 {
2083         unsigned i, idx = dst->idx;
2084
2085         *dst = *src;
2086         dst->idx = idx;
2087         dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
2088
2089         for (i = 0; i < BTREE_MAX_DEPTH; i++)
2090                 if (btree_node_locked(dst, i))
2091                         six_lock_increment(&dst->l[i].b->c.lock,
2092                                            __btree_lock_want(dst, i));
2093
2094         dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
2095         dst->flags &= ~BTREE_ITER_SET_POS_AFTER_COMMIT;
2096 }
2097
2098 static inline struct bpos bpos_diff(struct bpos l, struct bpos r)
2099 {
2100         if (bkey_cmp(l, r) > 0)
2101                 swap(l, r);
2102
2103         return POS(r.inode - l.inode, r.offset - l.offset);
2104 }
2105
2106 static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans,
2107                                                  unsigned btree_id, struct bpos pos,
2108                                                  unsigned flags)
2109 {
2110         struct btree_iter *iter, *best = NULL;
2111
2112         BUG_ON(trans->nr_iters > BTREE_ITER_MAX);
2113
2114         trans_for_each_iter(trans, iter) {
2115                 if (btree_iter_type(iter) != (flags & BTREE_ITER_TYPE))
2116                         continue;
2117
2118                 if (iter->btree_id != btree_id)
2119                         continue;
2120
2121                 if (best &&
2122                     bkey_cmp(bpos_diff(best->pos, pos),
2123                              bpos_diff(iter->pos, pos)) < 0)
2124                         continue;
2125
2126                 best = iter;
2127         }
2128
2129         if (!best) {
2130                 iter = btree_trans_iter_alloc(trans);
2131                 if (IS_ERR(iter))
2132                         return iter;
2133
2134                 bch2_btree_iter_init(trans, iter, btree_id, pos, flags);
2135         } else if ((trans->iters_live & (1ULL << best->idx)) ||
2136                    (best->flags & BTREE_ITER_KEEP_UNTIL_COMMIT)) {
2137                 iter = btree_trans_iter_alloc(trans);
2138                 if (IS_ERR(iter))
2139                         return iter;
2140
2141                 btree_iter_copy(iter, best);
2142         } else {
2143                 iter = best;
2144         }
2145
2146         iter->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
2147         iter->flags &= ~BTREE_ITER_USER_FLAGS;
2148         iter->flags |= flags & BTREE_ITER_USER_FLAGS;
2149
2150         if (iter->flags & BTREE_ITER_INTENT)
2151                 bch2_btree_iter_upgrade(iter, 1);
2152         else
2153                 bch2_btree_iter_downgrade(iter);
2154
2155         BUG_ON(iter->btree_id != btree_id);
2156         BUG_ON((iter->flags ^ flags) & BTREE_ITER_TYPE);
2157         BUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
2158         BUG_ON(iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT);
2159         BUG_ON(trans->iters_live & (1ULL << iter->idx));
2160
2161         trans->iters_live       |= 1ULL << iter->idx;
2162         trans->iters_touched    |= 1ULL << iter->idx;
2163
2164         return iter;
2165 }
2166
2167 struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans,
2168                                          enum btree_id btree_id,
2169                                          struct bpos pos, unsigned flags)
2170 {
2171         struct btree_iter *iter =
2172                 __btree_trans_get_iter(trans, btree_id, pos, flags);
2173
2174         if (!IS_ERR(iter))
2175                 __bch2_btree_iter_set_pos(iter, pos,
2176                         btree_node_type_is_extents(btree_id));
2177         return iter;
2178 }
2179
2180 struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *trans,
2181                                             enum btree_id btree_id,
2182                                             struct bpos pos,
2183                                             unsigned locks_want,
2184                                             unsigned depth,
2185                                             unsigned flags)
2186 {
2187         struct btree_iter *iter =
2188                 __btree_trans_get_iter(trans, btree_id, pos,
2189                                        flags|BTREE_ITER_NODES);
2190         unsigned i;
2191
2192         BUG_ON(IS_ERR(iter));
2193         BUG_ON(bkey_cmp(iter->pos, pos));
2194
2195         iter->locks_want = locks_want;
2196         iter->level     = depth;
2197         iter->min_depth = depth;
2198
2199         for (i = 0; i < ARRAY_SIZE(iter->l); i++)
2200                 iter->l[i].b            = NULL;
2201         iter->l[iter->level].b          = BTREE_ITER_NO_NODE_INIT;
2202
2203         return iter;
2204 }
2205
2206 struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *trans,
2207                                         struct btree_iter *src)
2208 {
2209         struct btree_iter *iter;
2210
2211         iter = btree_trans_iter_alloc(trans);
2212         if (IS_ERR(iter))
2213                 return iter;
2214
2215         btree_iter_copy(iter, src);
2216
2217         trans->iters_live |= 1ULL << iter->idx;
2218         /*
2219          * We don't need to preserve this iter since it's cheap to copy it
2220          * again - this will cause trans_iter_put() to free it right away:
2221          */
2222         trans->iters_touched &= ~(1ULL << iter->idx);
2223
2224         return iter;
2225 }
2226
2227 static int bch2_trans_preload_mem(struct btree_trans *trans, size_t size)
2228 {
2229         if (size > trans->mem_bytes) {
2230                 size_t old_bytes = trans->mem_bytes;
2231                 size_t new_bytes = roundup_pow_of_two(size);
2232                 void *new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2233
2234                 if (!new_mem)
2235                         return -ENOMEM;
2236
2237                 trans->mem = new_mem;
2238                 trans->mem_bytes = new_bytes;
2239
2240                 if (old_bytes) {
2241                         trace_trans_restart_mem_realloced(trans->ip, new_bytes);
2242                         return -EINTR;
2243                 }
2244         }
2245
2246         return 0;
2247 }
2248
2249 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2250 {
2251         void *p;
2252         int ret;
2253
2254         ret = bch2_trans_preload_mem(trans, trans->mem_top + size);
2255         if (ret)
2256                 return ERR_PTR(ret);
2257
2258         p = trans->mem + trans->mem_top;
2259         trans->mem_top += size;
2260         return p;
2261 }
2262
2263 inline void bch2_trans_unlink_iters(struct btree_trans *trans)
2264 {
2265         u64 iters = trans->iters_linked &
2266                 ~trans->iters_touched &
2267                 ~trans->iters_live;
2268
2269         while (iters) {
2270                 unsigned idx = __ffs64(iters);
2271
2272                 iters &= ~(1ULL << idx);
2273                 __bch2_trans_iter_free(trans, idx);
2274         }
2275 }
2276
2277 void bch2_trans_reset(struct btree_trans *trans, unsigned flags)
2278 {
2279         struct btree_iter *iter;
2280
2281         trans_for_each_iter(trans, iter)
2282                 iter->flags &= ~(BTREE_ITER_KEEP_UNTIL_COMMIT|
2283                                  BTREE_ITER_SET_POS_AFTER_COMMIT);
2284
2285         bch2_trans_unlink_iters(trans);
2286
2287         trans->iters_touched &= trans->iters_live;
2288
2289         trans->need_reset               = 0;
2290         trans->nr_updates               = 0;
2291         trans->nr_updates2              = 0;
2292         trans->mem_top                  = 0;
2293
2294         trans->extra_journal_entries    = NULL;
2295         trans->extra_journal_entry_u64s = 0;
2296
2297         if (trans->fs_usage_deltas) {
2298                 trans->fs_usage_deltas->used = 0;
2299                 memset(&trans->fs_usage_deltas->memset_start, 0,
2300                        (void *) &trans->fs_usage_deltas->memset_end -
2301                        (void *) &trans->fs_usage_deltas->memset_start);
2302         }
2303
2304         if (!(flags & TRANS_RESET_NOTRAVERSE))
2305                 bch2_btree_iter_traverse_all(trans);
2306 }
2307
2308 void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
2309                      unsigned expected_nr_iters,
2310                      size_t expected_mem_bytes)
2311 {
2312         memset(trans, 0, offsetof(struct btree_trans, iters_onstack));
2313
2314         /*
2315          * reallocating iterators currently completely breaks
2316          * bch2_trans_iter_put():
2317          */
2318         expected_nr_iters = BTREE_ITER_MAX;
2319
2320         trans->c                = c;
2321         trans->ip               = _RET_IP_;
2322         trans->size             = ARRAY_SIZE(trans->iters_onstack);
2323         trans->iters            = trans->iters_onstack;
2324         trans->updates          = trans->updates_onstack;
2325         trans->updates2         = trans->updates2_onstack;
2326         trans->fs_usage_deltas  = NULL;
2327
2328         if (expected_nr_iters > trans->size)
2329                 bch2_trans_realloc_iters(trans, expected_nr_iters);
2330
2331         if (expected_mem_bytes)
2332                 bch2_trans_preload_mem(trans, expected_mem_bytes);
2333
2334 #ifdef CONFIG_BCACHEFS_DEBUG
2335         trans->pid = current->pid;
2336         mutex_lock(&c->btree_trans_lock);
2337         list_add(&trans->list, &c->btree_trans_list);
2338         mutex_unlock(&c->btree_trans_lock);
2339 #endif
2340 }
2341
2342 int bch2_trans_exit(struct btree_trans *trans)
2343 {
2344         bch2_trans_unlock(trans);
2345
2346 #ifdef CONFIG_BCACHEFS_DEBUG
2347         mutex_lock(&trans->c->btree_trans_lock);
2348         list_del(&trans->list);
2349         mutex_unlock(&trans->c->btree_trans_lock);
2350 #endif
2351
2352         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
2353
2354         kfree(trans->fs_usage_deltas);
2355         kfree(trans->mem);
2356         if (trans->used_mempool)
2357                 mempool_free(trans->iters, &trans->c->btree_iters_pool);
2358         else if (trans->iters != trans->iters_onstack)
2359                 kfree(trans->iters);
2360         trans->mem      = (void *) 0x1;
2361         trans->iters    = (void *) 0x1;
2362
2363         return trans->error ? -EIO : 0;
2364 }
2365
2366 static void bch2_btree_iter_node_to_text(struct printbuf *out,
2367                                  struct btree_bkey_cached_common *_b,
2368                                  enum btree_iter_type type)
2369 {
2370         pr_buf(out, "    %px l=%u %s:",
2371                _b, _b->level, bch2_btree_ids[_b->btree_id]);
2372         bch2_bpos_to_text(out, btree_node_pos(_b, type));
2373 }
2374
2375 void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
2376 {
2377 #ifdef CONFIG_BCACHEFS_DEBUG
2378         struct btree_trans *trans;
2379         struct btree_iter *iter;
2380         struct btree *b;
2381         unsigned l;
2382
2383         mutex_lock(&c->btree_trans_lock);
2384         list_for_each_entry(trans, &c->btree_trans_list, list) {
2385                 pr_buf(out, "%i %px %ps\n", trans->pid, trans, (void *) trans->ip);
2386
2387                 trans_for_each_iter(trans, iter) {
2388                         if (!iter->nodes_locked)
2389                                 continue;
2390
2391                         pr_buf(out, "  iter %u %s:",
2392                                iter->idx,
2393                                bch2_btree_ids[iter->btree_id]);
2394                         bch2_bpos_to_text(out, iter->pos);
2395                         pr_buf(out, "\n");
2396
2397                         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
2398                                 if (btree_node_locked(iter, l)) {
2399                                         pr_buf(out, "    %s l=%u ",
2400                                                btree_node_intent_locked(iter, l) ? "i" : "r", l);
2401                                         bch2_btree_iter_node_to_text(out,
2402                                                         (void *) iter->l[l].b,
2403                                                         btree_iter_type(iter));
2404                                         pr_buf(out, "\n");
2405                                 }
2406                         }
2407                 }
2408
2409                 b = READ_ONCE(trans->locking);
2410                 if (b) {
2411                         pr_buf(out, "  locking iter %u l=%u %s:",
2412                                trans->locking_iter_idx,
2413                                trans->locking_level,
2414                                bch2_btree_ids[trans->locking_btree_id]);
2415                         bch2_bpos_to_text(out, trans->locking_pos);
2416
2417
2418                         pr_buf(out, " node ");
2419                         bch2_btree_iter_node_to_text(out,
2420                                         (void *) b,
2421                                         btree_iter_type(&trans->iters[trans->locking_iter_idx]));
2422                         pr_buf(out, "\n");
2423                 }
2424         }
2425         mutex_unlock(&c->btree_trans_lock);
2426 #endif
2427 }
2428
2429 void bch2_fs_btree_iter_exit(struct bch_fs *c)
2430 {
2431         mempool_exit(&c->btree_iters_pool);
2432 }
2433
2434 int bch2_fs_btree_iter_init(struct bch_fs *c)
2435 {
2436         unsigned nr = BTREE_ITER_MAX;
2437
2438         INIT_LIST_HEAD(&c->btree_trans_list);
2439         mutex_init(&c->btree_trans_lock);
2440
2441         return mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
2442                         sizeof(struct btree_iter) * nr +
2443                         sizeof(struct btree_insert_entry) * nr +
2444                         sizeof(struct btree_insert_entry) * nr);
2445 }