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_locking.h"
   8 #include "btree_update.h"
   9 #include "debug.h"
  10 #include "extents.h"
  11
  12 #include <linux/prefetch.h>
  13 #include <trace/events/bcachefs.h>
  14
  15 #define BTREE_ITER_NO_NODE_GET_LOCKS    ((struct btree *) 1)
  16 #define BTREE_ITER_NO_NODE_DROP         ((struct btree *) 2)
  17 #define BTREE_ITER_NO_NODE_LOCK_ROOT    ((struct btree *) 3)
  18 #define BTREE_ITER_NO_NODE_UP           ((struct btree *) 4)
  19 #define BTREE_ITER_NO_NODE_DOWN         ((struct btree *) 5)
  20 #define BTREE_ITER_NO_NODE_INIT         ((struct btree *) 6)
  21 #define BTREE_ITER_NO_NODE_ERROR        ((struct btree *) 7)
  22
  23 static inline bool is_btree_node(struct btree_iter *iter, unsigned l)
  24 {
  25         return l < BTREE_MAX_DEPTH &&
  26                 (unsigned long) iter->l[l].b >= 128;
  27 }
  28
  29 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
  30 {
  31         struct bpos pos = iter->pos;
  32
  33         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
  34             bkey_cmp(pos, POS_MAX))
  35                 pos = bkey_successor(pos);
  36         return pos;
  37 }
  38
  39 static inline bool btree_iter_pos_before_node(struct btree_iter *iter,
  40                                               struct btree *b)
  41 {
  42         return bkey_cmp(btree_iter_search_key(iter), b->data->min_key) < 0;
  43 }
  44
  45 static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
  46                                              struct btree *b)
  47 {
  48         return bkey_cmp(b->key.k.p, btree_iter_search_key(iter)) < 0;
  49 }
  50
  51 static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
  52                                           struct btree *b)
  53 {
  54         return iter->btree_id == b->btree_id &&
  55                 !btree_iter_pos_before_node(iter, b) &&
  56                 !btree_iter_pos_after_node(iter, b);
  57 }
  58
  59 /* Btree node locking: */
  60
  61 void bch2_btree_node_unlock_write(struct btree *b, struct btree_iter *iter)
  62 {
  63         bch2_btree_node_unlock_write_inlined(b, iter);
  64 }
  65
  66 void __bch2_btree_node_lock_write(struct btree *b, struct btree_iter *iter)
  67 {
  68         struct btree_iter *linked;
  69         unsigned readers = 0;
  70
  71         EBUG_ON(!btree_node_intent_locked(iter, b->level));
  72
  73         trans_for_each_iter(iter->trans, linked)
  74                 if (linked->l[b->level].b == b &&
  75                     btree_node_read_locked(linked, b->level))
  76                         readers++;
  77
  78         /*
  79          * Must drop our read locks before calling six_lock_write() -
  80          * six_unlock() won't do wakeups until the reader count
  81          * goes to 0, and it's safe because we have the node intent
  82          * locked:
  83          */
  84         atomic64_sub(__SIX_VAL(read_lock, readers),
  85                      &b->lock.state.counter);
  86         btree_node_lock_type(iter->trans->c, b, SIX_LOCK_write);
  87         atomic64_add(__SIX_VAL(read_lock, readers),
  88                      &b->lock.state.counter);
  89 }
  90
  91 bool __bch2_btree_node_relock(struct btree_iter *iter, unsigned level)
  92 {
  93         struct btree *b = btree_iter_node(iter, level);
  94         int want = __btree_lock_want(iter, level);
  95
  96         if (!is_btree_node(iter, level))
  97                 return false;
  98
  99         if (race_fault())
 100                 return false;
 101
 102         if (six_relock_type(&b->lock, want, iter->l[level].lock_seq) ||
 103             (btree_node_lock_seq_matches(iter, b, level) &&
 104              btree_node_lock_increment(iter, b, level, want))) {
 105                 mark_btree_node_locked(iter, level, want);
 106                 return true;
 107         } else {
 108                 return false;
 109         }
 110 }
 111
 112 static bool bch2_btree_node_upgrade(struct btree_iter *iter, unsigned level)
 113 {
 114         struct btree *b = iter->l[level].b;
 115
 116         EBUG_ON(btree_lock_want(iter, level) != BTREE_NODE_INTENT_LOCKED);
 117
 118         if (!is_btree_node(iter, level))
 119                 return false;
 120
 121         if (btree_node_intent_locked(iter, level))
 122                 return true;
 123
 124         if (race_fault())
 125                 return false;
 126
 127         if (btree_node_locked(iter, level)
 128             ? six_lock_tryupgrade(&b->lock)
 129             : six_relock_type(&b->lock, SIX_LOCK_intent, iter->l[level].lock_seq))
 130                 goto success;
 131
 132         if (btree_node_lock_seq_matches(iter, b, level) &&
 133             btree_node_lock_increment(iter, b, level, BTREE_NODE_INTENT_LOCKED)) {
 134                 btree_node_unlock(iter, level);
 135                 goto success;
 136         }
 137
 138         return false;
 139 success:
 140         mark_btree_node_intent_locked(iter, level);
 141         return true;
 142 }
 143
 144 static inline bool btree_iter_get_locks(struct btree_iter *iter,
 145                                         bool upgrade, bool trace)
 146 {
 147         unsigned l = iter->level;
 148         int fail_idx = -1;
 149
 150         do {
 151                 if (!btree_iter_node(iter, l))
 152                         break;
 153
 154                 if (!(upgrade
 155                       ? bch2_btree_node_upgrade(iter, l)
 156                       : bch2_btree_node_relock(iter, l))) {
 157                         if (trace)
 158                                 (upgrade
 159                                  ? trace_node_upgrade_fail
 160                                  : trace_node_relock_fail)(l, iter->l[l].lock_seq,
 161                                                 is_btree_node(iter, l)
 162                                                 ? 0
 163                                                 : (unsigned long) iter->l[l].b,
 164                                                 is_btree_node(iter, l)
 165                                                 ? iter->l[l].b->lock.state.seq
 166                                                 : 0);
 167
 168                         fail_idx = l;
 169                         btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
 170                 }
 171
 172                 l++;
 173         } while (l < iter->locks_want);
 174
 175         /*
 176          * When we fail to get a lock, we have to ensure that any child nodes
 177          * can't be relocked so bch2_btree_iter_traverse has to walk back up to
 178          * the node that we failed to relock:
 179          */
 180         while (fail_idx >= 0) {
 181                 btree_node_unlock(iter, fail_idx);
 182                 iter->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS;
 183                 --fail_idx;
 184         }
 185
 186         if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
 187                 iter->uptodate = BTREE_ITER_NEED_PEEK;
 188
 189         bch2_btree_trans_verify_locks(iter->trans);
 190
 191         return iter->uptodate < BTREE_ITER_NEED_RELOCK;
 192 }
 193
 194 /* Slowpath: */
 195 bool __bch2_btree_node_lock(struct btree *b, struct bpos pos,
 196                            unsigned level,
 197                            struct btree_iter *iter,
 198                            enum six_lock_type type)
 199 {
 200         struct btree_iter *linked;
 201         bool ret = true;
 202
 203         /* Check if it's safe to block: */
 204         trans_for_each_iter(iter->trans, linked) {
 205                 if (!linked->nodes_locked)
 206                         continue;
 207
 208                 /* Must lock btree nodes in key order: */
 209                 if ((cmp_int(iter->btree_id, linked->btree_id) ?:
 210                      bkey_cmp(pos, linked->pos)) < 0)
 211                         ret = false;
 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 (!(iter->trans->nounlock)) {
 228                                 linked->locks_want = max_t(unsigned,
 229                                                 linked->locks_want,
 230                                                 __fls(linked->nodes_locked) + 1);
 231                                 btree_iter_get_locks(linked, true, false);
 232                         }
 233                         ret = false;
 234                 }
 235
 236                 /*
 237                  * Interior nodes must be locked before their descendants: if
 238                  * another iterator has possible descendants locked of the node
 239                  * we're about to lock, it must have the ancestors locked too:
 240                  */
 241                 if (linked->btree_id == iter->btree_id &&
 242                     level > __fls(linked->nodes_locked)) {
 243                         if (!(iter->trans->nounlock)) {
 244                                 linked->locks_want =
 245                                         max(level + 1, max_t(unsigned,
 246                                             linked->locks_want,
 247                                             iter->locks_want));
 248                                 btree_iter_get_locks(linked, true, false);
 249                         }
 250                         ret = false;
 251                 }
 252         }
 253
 254         if (unlikely(!ret)) {
 255                 trace_trans_restart_would_deadlock(iter->trans->ip);
 256                 return false;
 257         }
 258
 259         __btree_node_lock_type(iter->trans->c, b, type);
 260         return true;
 261 }
 262
 263 /* Btree iterator locking: */
 264
 265 #ifdef CONFIG_BCACHEFS_DEBUG
 266 static void bch2_btree_iter_verify_locks(struct btree_iter *iter)
 267 {
 268         unsigned l;
 269
 270         for (l = 0; btree_iter_node(iter, l); l++) {
 271                 if (iter->uptodate >= BTREE_ITER_NEED_RELOCK &&
 272                     !btree_node_locked(iter, l))
 273                         continue;
 274
 275                 BUG_ON(btree_lock_want(iter, l) !=
 276                        btree_node_locked_type(iter, l));
 277         }
 278 }
 279
 280 void bch2_btree_trans_verify_locks(struct btree_trans *trans)
 281 {
 282         struct btree_iter *iter;
 283
 284         trans_for_each_iter(trans, iter)
 285                 bch2_btree_iter_verify_locks(iter);
 286 }
 287 #else
 288 static inline void bch2_btree_iter_verify_locks(struct btree_iter *iter) {}
 289 #endif
 290
 291 __flatten
 292 static bool bch2_btree_iter_relock(struct btree_iter *iter, bool trace)
 293 {
 294         return btree_iter_get_locks(iter, false, trace);
 295 }
 296
 297 bool __bch2_btree_iter_upgrade(struct btree_iter *iter,
 298                                unsigned new_locks_want)
 299 {
 300         struct btree_iter *linked;
 301
 302         EBUG_ON(iter->locks_want >= new_locks_want);
 303
 304         iter->locks_want = new_locks_want;
 305
 306         if (btree_iter_get_locks(iter, true, true))
 307                 return true;
 308
 309         /*
 310          * Ancestor nodes must be locked before child nodes, so set locks_want
 311          * on iterators that might lock ancestors before us to avoid getting
 312          * -EINTR later:
 313          */
 314         trans_for_each_iter(iter->trans, linked)
 315                 if (linked != iter &&
 316                     linked->btree_id == iter->btree_id &&
 317                     linked->locks_want < new_locks_want) {
 318                         linked->locks_want = new_locks_want;
 319                         btree_iter_get_locks(linked, true, false);
 320                 }
 321
 322         return false;
 323 }
 324
 325 bool __bch2_btree_iter_upgrade_nounlock(struct btree_iter *iter,
 326                                         unsigned new_locks_want)
 327 {
 328         unsigned l = iter->level;
 329
 330         EBUG_ON(iter->locks_want >= new_locks_want);
 331
 332         iter->locks_want = new_locks_want;
 333
 334         do {
 335                 if (!btree_iter_node(iter, l))
 336                         break;
 337
 338                 if (!bch2_btree_node_upgrade(iter, l)) {
 339                         iter->locks_want = l;
 340                         return false;
 341                 }
 342
 343                 l++;
 344         } while (l < iter->locks_want);
 345
 346         return true;
 347 }
 348
 349 void __bch2_btree_iter_downgrade(struct btree_iter *iter,
 350                                  unsigned downgrade_to)
 351 {
 352         struct btree_iter *linked;
 353         unsigned l;
 354
 355         /*
 356          * We downgrade linked iterators as well because btree_iter_upgrade
 357          * might have had to modify locks_want on linked iterators due to lock
 358          * ordering:
 359          */
 360         trans_for_each_iter(iter->trans, linked) {
 361                 unsigned new_locks_want = downgrade_to ?:
 362                         (linked->flags & BTREE_ITER_INTENT ? 1 : 0);
 363
 364                 if (linked->locks_want <= new_locks_want)
 365                         continue;
 366
 367                 linked->locks_want = new_locks_want;
 368
 369                 while (linked->nodes_locked &&
 370                        (l = __fls(linked->nodes_locked)) >= linked->locks_want) {
 371                         if (l > linked->level) {
 372                                 btree_node_unlock(linked, l);
 373                         } else {
 374                                 if (btree_node_intent_locked(linked, l)) {
 375                                         six_lock_downgrade(&linked->l[l].b->lock);
 376                                         linked->nodes_intent_locked ^= 1 << l;
 377                                 }
 378                                 break;
 379                         }
 380                 }
 381         }
 382
 383         bch2_btree_trans_verify_locks(iter->trans);
 384 }
 385
 386 /* Btree transaction locking: */
 387
 388 bool bch2_trans_relock(struct btree_trans *trans)
 389 {
 390         struct btree_iter *iter;
 391         bool ret = true;
 392
 393         trans_for_each_iter(trans, iter)
 394                 if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
 395                         ret &= bch2_btree_iter_relock(iter, true);
 396
 397         return ret;
 398 }
 399
 400 void bch2_trans_unlock(struct btree_trans *trans)
 401 {
 402         struct btree_iter *iter;
 403
 404         trans_for_each_iter(trans, iter)
 405                 __bch2_btree_iter_unlock(iter);
 406 }
 407
 408 /* Btree iterator: */
 409
 410 #ifdef CONFIG_BCACHEFS_DEBUG
 411
 412 static void bch2_btree_iter_verify_level(struct btree_iter *iter,
 413                                          unsigned level)
 414 {
 415         struct bpos pos = btree_iter_search_key(iter);
 416         struct btree_iter_level *l = &iter->l[level];
 417         struct btree_node_iter tmp = l->iter;
 418         bool locked = btree_node_locked(iter, level);
 419         struct bkey_packed *p, *k;
 420         char buf1[100], buf2[100];
 421         const char *msg;
 422
 423         if (!debug_check_iterators(iter->trans->c))
 424                 return;
 425
 426         BUG_ON(iter->level < iter->min_depth);
 427
 428         if (!btree_iter_node(iter, level))
 429                 return;
 430
 431         if (!bch2_btree_node_relock(iter, level))
 432                 return;
 433
 434         /*
 435          * Ideally this invariant would always be true, and hopefully in the
 436          * future it will be, but for now set_pos_same_leaf() breaks it:
 437          */
 438         BUG_ON(iter->uptodate < BTREE_ITER_NEED_TRAVERSE &&
 439                !btree_iter_pos_in_node(iter, l->b));
 440
 441         /*
 442          * node iterators don't use leaf node iterator:
 443          */
 444         if (btree_iter_type(iter) == BTREE_ITER_NODES &&
 445             level <= iter->min_depth)
 446                 goto unlock;
 447
 448         bch2_btree_node_iter_verify(&l->iter, l->b);
 449
 450         /*
 451          * For interior nodes, the iterator will have skipped past
 452          * deleted keys:
 453          *
 454          * For extents, the iterator may have skipped past deleted keys (but not
 455          * whiteouts)
 456          */
 457         p = level || btree_node_type_is_extents(iter->btree_id)
 458                 ? bch2_btree_node_iter_prev_filter(&tmp, l->b, KEY_TYPE_discard)
 459                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 460         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 461
 462         if (p && bkey_iter_pos_cmp(l->b, p, &pos) >= 0) {
 463                 msg = "before";
 464                 goto err;
 465         }
 466
 467         if (k && bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
 468                 msg = "after";
 469                 goto err;
 470         }
 471 unlock:
 472         if (!locked)
 473                 btree_node_unlock(iter, level);
 474         return;
 475 err:
 476         strcpy(buf1, "(none)");
 477         strcpy(buf2, "(none)");
 478
 479         if (p) {
 480                 struct bkey uk = bkey_unpack_key(l->b, p);
 481                 bch2_bkey_to_text(&PBUF(buf1), &uk);
 482         }
 483
 484         if (k) {
 485                 struct bkey uk = bkey_unpack_key(l->b, k);
 486                 bch2_bkey_to_text(&PBUF(buf2), &uk);
 487         }
 488
 489         panic("iterator should be %s key at level %u:\n"
 490               "iter pos %s %llu:%llu\n"
 491               "prev key %s\n"
 492               "cur  key %s\n",
 493               msg, level,
 494               iter->flags & BTREE_ITER_IS_EXTENTS ? ">" : "=>",
 495               iter->pos.inode, iter->pos.offset,
 496               buf1, buf2);
 497 }
 498
 499 static void bch2_btree_iter_verify(struct btree_iter *iter)
 500 {
 501         unsigned i;
 502
 503         bch2_btree_trans_verify_locks(iter->trans);
 504
 505         for (i = 0; i < BTREE_MAX_DEPTH; i++)
 506                 bch2_btree_iter_verify_level(iter, i);
 507 }
 508
 509 void bch2_btree_trans_verify_iters(struct btree_trans *trans, struct btree *b)
 510 {
 511         struct btree_iter *iter;
 512
 513         if (!debug_check_iterators(trans->c))
 514                 return;
 515
 516         trans_for_each_iter_with_node(trans, b, iter)
 517                 bch2_btree_iter_verify_level(iter, b->level);
 518 }
 519
 520 #else
 521
 522 static inline void bch2_btree_iter_verify_level(struct btree_iter *iter, unsigned l) {}
 523 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 524
 525 #endif
 526
 527 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 528                                         struct btree *b,
 529                                         struct bset_tree *t,
 530                                         struct bkey_packed *k)
 531 {
 532         struct btree_node_iter_set *set;
 533
 534         btree_node_iter_for_each(iter, set)
 535                 if (set->end == t->end_offset) {
 536                         set->k = __btree_node_key_to_offset(b, k);
 537                         bch2_btree_node_iter_sort(iter, b);
 538                         return;
 539                 }
 540
 541         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 542 }
 543
 544 static void __bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 545                                                struct btree *b,
 546                                                struct bkey_packed *where)
 547 {
 548         struct btree_iter_level *l = &iter->l[b->level];
 549         struct bpos pos = btree_iter_search_key(iter);
 550
 551         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 552                 return;
 553
 554         if (bkey_iter_pos_cmp(l->b, where, &pos) < 0)
 555                 bch2_btree_node_iter_advance(&l->iter, l->b);
 556
 557         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 558 }
 559
 560 void bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 561                                       struct btree *b,
 562                                       struct bkey_packed *where)
 563 {
 564         struct btree_iter *linked;
 565
 566         trans_for_each_iter_with_node(iter->trans, b, linked) {
 567                 __bch2_btree_iter_fix_key_modified(linked, b, where);
 568                 bch2_btree_iter_verify_level(linked, b->level);
 569         }
 570 }
 571
 572 static void __bch2_btree_node_iter_fix(struct btree_iter *iter,
 573                                       struct btree *b,
 574                                       struct btree_node_iter *node_iter,
 575                                       struct bset_tree *t,
 576                                       struct bkey_packed *where,
 577                                       unsigned clobber_u64s,
 578                                       unsigned new_u64s)
 579 {
 580         const struct bkey_packed *end = btree_bkey_last(b, t);
 581         struct btree_node_iter_set *set;
 582         unsigned offset = __btree_node_key_to_offset(b, where);
 583         int shift = new_u64s - clobber_u64s;
 584         unsigned old_end = t->end_offset - shift;
 585         unsigned orig_iter_pos = node_iter->data[0].k;
 586         bool iter_current_key_modified =
 587                 orig_iter_pos >= offset &&
 588                 orig_iter_pos <= offset + clobber_u64s;
 589         struct bpos iter_pos = btree_iter_search_key(iter);
 590
 591         btree_node_iter_for_each(node_iter, set)
 592                 if (set->end == old_end)
 593                         goto found;
 594
 595         /* didn't find the bset in the iterator - might have to readd it: */
 596         if (new_u64s &&
 597             bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
 598                 bch2_btree_node_iter_push(node_iter, b, where, end);
 599                 goto fixup_done;
 600         } else {
 601                 /* Iterator is after key that changed */
 602                 return;
 603         }
 604 found:
 605         set->end = t->end_offset;
 606
 607         /* Iterator hasn't gotten to the key that changed yet: */
 608         if (set->k < offset)
 609                 return;
 610
 611         if (new_u64s &&
 612             bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
 613                 set->k = offset;
 614         } else if (set->k < offset + clobber_u64s) {
 615                 set->k = offset + new_u64s;
 616                 if (set->k == set->end)
 617                         bch2_btree_node_iter_set_drop(node_iter, set);
 618         } else {
 619                 /* Iterator is after key that changed */
 620                 set->k = (int) set->k + shift;
 621                 return;
 622         }
 623
 624         bch2_btree_node_iter_sort(node_iter, b);
 625 fixup_done:
 626         if (node_iter->data[0].k != orig_iter_pos)
 627                 iter_current_key_modified = true;
 628
 629         /*
 630          * When a new key is added, and the node iterator now points to that
 631          * key, the iterator might have skipped past deleted keys that should
 632          * come after the key the iterator now points to. We have to rewind to
 633          * before those deleted keys - otherwise
 634          * bch2_btree_node_iter_prev_all() breaks:
 635          */
 636         if (!bch2_btree_node_iter_end(node_iter) &&
 637             iter_current_key_modified &&
 638             (b->level ||
 639              btree_node_type_is_extents(iter->btree_id))) {
 640                 struct bset_tree *t;
 641                 struct bkey_packed *k, *k2, *p;
 642
 643                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 644
 645                 for_each_bset(b, t) {
 646                         bool set_pos = false;
 647
 648                         if (node_iter->data[0].end == t->end_offset)
 649                                 continue;
 650
 651                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 652
 653                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 654                                bkey_iter_cmp(b, k, p) < 0) {
 655                                 k2 = p;
 656                                 set_pos = true;
 657                         }
 658
 659                         if (set_pos)
 660                                 btree_node_iter_set_set_pos(node_iter,
 661                                                             b, t, k2);
 662                 }
 663         }
 664
 665         if (!b->level &&
 666             node_iter == &iter->l[0].iter &&
 667             iter_current_key_modified)
 668                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 669 }
 670
 671 void bch2_btree_node_iter_fix(struct btree_iter *iter,
 672                               struct btree *b,
 673                               struct btree_node_iter *node_iter,
 674                               struct bkey_packed *where,
 675                               unsigned clobber_u64s,
 676                               unsigned new_u64s)
 677 {
 678         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 679         struct btree_iter *linked;
 680
 681         if (node_iter != &iter->l[b->level].iter) {
 682                 __bch2_btree_node_iter_fix(iter, b, node_iter, t,
 683                                            where, clobber_u64s, new_u64s);
 684
 685                 if (debug_check_iterators(iter->trans->c))
 686                         bch2_btree_node_iter_verify(node_iter, b);
 687         }
 688
 689         trans_for_each_iter_with_node(iter->trans, b, linked) {
 690                 __bch2_btree_node_iter_fix(linked, b,
 691                                            &linked->l[b->level].iter, t,
 692                                            where, clobber_u64s, new_u64s);
 693                 bch2_btree_iter_verify_level(linked, b->level);
 694         }
 695 }
 696
 697 static inline struct bkey_s_c __btree_iter_unpack(struct btree_iter *iter,
 698                                                   struct btree_iter_level *l,
 699                                                   struct bkey *u,
 700                                                   struct bkey_packed *k)
 701 {
 702         struct bkey_s_c ret;
 703
 704         if (unlikely(!k)) {
 705                 /*
 706                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 707                  * a hole
 708                  */
 709                 u->type = KEY_TYPE_deleted;
 710                 return bkey_s_c_null;
 711         }
 712
 713         ret = bkey_disassemble(l->b, k, u);
 714
 715         if (debug_check_bkeys(iter->trans->c))
 716                 bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
 717
 718         return ret;
 719 }
 720
 721 /* peek_all() doesn't skip deleted keys */
 722 static inline struct bkey_s_c __btree_iter_peek_all(struct btree_iter *iter,
 723                                                     struct btree_iter_level *l,
 724                                                     struct bkey *u)
 725 {
 726         return __btree_iter_unpack(iter, l, u,
 727                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 728 }
 729
 730 static inline struct bkey_s_c __btree_iter_peek(struct btree_iter *iter,
 731                                                 struct btree_iter_level *l)
 732 {
 733         return __btree_iter_unpack(iter, l, &iter->k,
 734                         bch2_btree_node_iter_peek(&l->iter, l->b));
 735 }
 736
 737 static inline struct bkey_s_c __btree_iter_prev(struct btree_iter *iter,
 738                                                 struct btree_iter_level *l)
 739 {
 740         return __btree_iter_unpack(iter, l, &iter->k,
 741                         bch2_btree_node_iter_prev(&l->iter, l->b));
 742 }
 743
 744 static inline bool btree_iter_advance_to_pos(struct btree_iter *iter,
 745                                              struct btree_iter_level *l,
 746                                              int max_advance)
 747 {
 748         struct bpos pos = btree_iter_search_key(iter);
 749         struct bkey_packed *k;
 750         int nr_advanced = 0;
 751
 752         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 753                bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
 754                 if (max_advance > 0 && nr_advanced >= max_advance)
 755                         return false;
 756
 757                 bch2_btree_node_iter_advance(&l->iter, l->b);
 758                 nr_advanced++;
 759         }
 760
 761         return true;
 762 }
 763
 764 /*
 765  * Verify that iterator for parent node points to child node:
 766  */
 767 static void btree_iter_verify_new_node(struct btree_iter *iter, struct btree *b)
 768 {
 769         struct btree_iter_level *l;
 770         unsigned plevel;
 771         bool parent_locked;
 772         struct bkey_packed *k;
 773
 774         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 775                 return;
 776
 777         plevel = b->level + 1;
 778         if (!btree_iter_node(iter, plevel))
 779                 return;
 780
 781         parent_locked = btree_node_locked(iter, plevel);
 782
 783         if (!bch2_btree_node_relock(iter, plevel))
 784                 return;
 785
 786         l = &iter->l[plevel];
 787         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 788         if (!k ||
 789             bkey_deleted(k) ||
 790             bkey_cmp_left_packed(l->b, k, &b->key.k.p)) {
 791                 char buf[100];
 792                 struct bkey uk = bkey_unpack_key(b, k);
 793
 794                 bch2_bkey_to_text(&PBUF(buf), &uk);
 795                 panic("parent iter doesn't point to new node:\n%s\n%llu:%llu\n",
 796                       buf, b->key.k.p.inode, b->key.k.p.offset);
 797         }
 798
 799         if (!parent_locked)
 800                 btree_node_unlock(iter, b->level + 1);
 801 }
 802
 803 static inline void __btree_iter_init(struct btree_iter *iter,
 804                                      unsigned level)
 805 {
 806         struct bpos pos = btree_iter_search_key(iter);
 807         struct btree_iter_level *l = &iter->l[level];
 808
 809         bch2_btree_node_iter_init(&l->iter, l->b, &pos);
 810
 811         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 812 }
 813
 814 static inline void btree_iter_node_set(struct btree_iter *iter,
 815                                        struct btree *b)
 816 {
 817         btree_iter_verify_new_node(iter, b);
 818
 819         EBUG_ON(!btree_iter_pos_in_node(iter, b));
 820         EBUG_ON(b->lock.state.seq & 1);
 821
 822         iter->l[b->level].lock_seq = b->lock.state.seq;
 823         iter->l[b->level].b = b;
 824         __btree_iter_init(iter, b->level);
 825 }
 826
 827 /*
 828  * A btree node is being replaced - update the iterator to point to the new
 829  * node:
 830  */
 831 void bch2_btree_iter_node_replace(struct btree_iter *iter, struct btree *b)
 832 {
 833         enum btree_node_locked_type t;
 834         struct btree_iter *linked;
 835
 836         trans_for_each_iter(iter->trans, linked)
 837                 if (btree_iter_pos_in_node(linked, b)) {
 838                         /*
 839                          * bch2_btree_iter_node_drop() has already been called -
 840                          * the old node we're replacing has already been
 841                          * unlocked and the pointer invalidated
 842                          */
 843                         BUG_ON(btree_node_locked(linked, b->level));
 844
 845                         t = btree_lock_want(linked, b->level);
 846                         if (t != BTREE_NODE_UNLOCKED) {
 847                                 six_lock_increment(&b->lock, t);
 848                                 mark_btree_node_locked(linked, b->level, t);
 849                         }
 850
 851                         btree_iter_node_set(linked, b);
 852                 }
 853 }
 854
 855 void bch2_btree_iter_node_drop(struct btree_iter *iter, struct btree *b)
 856 {
 857         struct btree_iter *linked;
 858         unsigned level = b->level;
 859
 860         trans_for_each_iter(iter->trans, linked)
 861                 if (linked->l[level].b == b) {
 862                         __btree_node_unlock(linked, level);
 863                         linked->l[level].b = BTREE_ITER_NO_NODE_DROP;
 864                 }
 865 }
 866
 867 /*
 868  * A btree node has been modified in such a way as to invalidate iterators - fix
 869  * them:
 870  */
 871 void bch2_btree_iter_reinit_node(struct btree_iter *iter, struct btree *b)
 872 {
 873         struct btree_iter *linked;
 874
 875         trans_for_each_iter_with_node(iter->trans, b, linked)
 876                 __btree_iter_init(linked, b->level);
 877 }
 878
 879 static inline int btree_iter_lock_root(struct btree_iter *iter,
 880                                        unsigned depth_want)
 881 {
 882         struct bch_fs *c = iter->trans->c;
 883         struct btree *b;
 884         enum six_lock_type lock_type;
 885         unsigned i;
 886
 887         EBUG_ON(iter->nodes_locked);
 888
 889         while (1) {
 890                 b = READ_ONCE(c->btree_roots[iter->btree_id].b);
 891                 iter->level = READ_ONCE(b->level);
 892
 893                 if (unlikely(iter->level < depth_want)) {
 894                         /*
 895                          * the root is at a lower depth than the depth we want:
 896                          * got to the end of the btree, or we're walking nodes
 897                          * greater than some depth and there are no nodes >=
 898                          * that depth
 899                          */
 900                         iter->level = depth_want;
 901                         for (i = iter->level; i < BTREE_MAX_DEPTH; i++)
 902                                 iter->l[i].b = NULL;
 903                         return 1;
 904                 }
 905
 906                 lock_type = __btree_lock_want(iter, iter->level);
 907                 if (unlikely(!btree_node_lock(b, POS_MAX, iter->level,
 908                                               iter, lock_type)))
 909                         return -EINTR;
 910
 911                 if (likely(b == c->btree_roots[iter->btree_id].b &&
 912                            b->level == iter->level &&
 913                            !race_fault())) {
 914                         for (i = 0; i < iter->level; i++)
 915                                 iter->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
 916                         iter->l[iter->level].b = b;
 917                         for (i = iter->level + 1; i < BTREE_MAX_DEPTH; i++)
 918                                 iter->l[i].b = NULL;
 919
 920                         mark_btree_node_locked(iter, iter->level, lock_type);
 921                         btree_iter_node_set(iter, b);
 922                         return 0;
 923                 }
 924
 925                 six_unlock_type(&b->lock, lock_type);
 926         }
 927 }
 928
 929 noinline
 930 static void btree_iter_prefetch(struct btree_iter *iter)
 931 {
 932         struct bch_fs *c = iter->trans->c;
 933         struct btree_iter_level *l = &iter->l[iter->level];
 934         struct btree_node_iter node_iter = l->iter;
 935         struct bkey_packed *k;
 936         BKEY_PADDED(k) tmp;
 937         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 938                 ? (iter->level > 1 ? 0 :  2)
 939                 : (iter->level > 1 ? 1 : 16);
 940         bool was_locked = btree_node_locked(iter, iter->level);
 941
 942         while (nr) {
 943                 if (!bch2_btree_node_relock(iter, iter->level))
 944                         return;
 945
 946                 bch2_btree_node_iter_advance(&node_iter, l->b);
 947                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
 948                 if (!k)
 949                         break;
 950
 951                 bch2_bkey_unpack(l->b, &tmp.k, k);
 952                 bch2_btree_node_prefetch(c, iter, &tmp.k, iter->level - 1);
 953         }
 954
 955         if (!was_locked)
 956                 btree_node_unlock(iter, iter->level);
 957 }
 958
 959 static noinline void btree_node_mem_ptr_set(struct btree_iter *iter,
 960                                             unsigned plevel, struct btree *b)
 961 {
 962         struct btree_iter_level *l = &iter->l[plevel];
 963         bool locked = btree_node_locked(iter, plevel);
 964         struct bkey_packed *k;
 965         struct bch_btree_ptr_v2 *bp;
 966
 967         if (!bch2_btree_node_relock(iter, plevel))
 968                 return;
 969
 970         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 971         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
 972
 973         bp = (void *) bkeyp_val(&l->b->format, k);
 974         bp->mem_ptr = (unsigned long)b;
 975
 976         if (!locked)
 977                 btree_node_unlock(iter, plevel);
 978 }
 979
 980 static __always_inline int btree_iter_down(struct btree_iter *iter)
 981 {
 982         struct bch_fs *c = iter->trans->c;
 983         struct btree_iter_level *l = &iter->l[iter->level];
 984         struct btree *b;
 985         unsigned level = iter->level - 1;
 986         enum six_lock_type lock_type = __btree_lock_want(iter, level);
 987         BKEY_PADDED(k) tmp;
 988
 989         EBUG_ON(!btree_node_locked(iter, iter->level));
 990
 991         bch2_bkey_unpack(l->b, &tmp.k,
 992                          bch2_btree_node_iter_peek(&l->iter, l->b));
 993
 994         b = bch2_btree_node_get(c, iter, &tmp.k, level, lock_type);
 995         if (unlikely(IS_ERR(b)))
 996                 return PTR_ERR(b);
 997
 998         mark_btree_node_locked(iter, level, lock_type);
 999         btree_iter_node_set(iter, b);
1000
1001         if (tmp.k.k.type == KEY_TYPE_btree_ptr_v2 &&
1002             unlikely(b != btree_node_mem_ptr(&tmp.k)))
1003                 btree_node_mem_ptr_set(iter, level + 1, b);
1004
1005         if (iter->flags & BTREE_ITER_PREFETCH)
1006                 btree_iter_prefetch(iter);
1007
1008         iter->level = level;
1009
1010         return 0;
1011 }
1012
1013 static void btree_iter_up(struct btree_iter *iter)
1014 {
1015         btree_node_unlock(iter, iter->level++);
1016 }
1017
1018 static int btree_iter_traverse_one(struct btree_iter *);
1019
1020 static int __btree_iter_traverse_all(struct btree_trans *trans,
1021                                    struct btree_iter *orig_iter, int ret)
1022 {
1023         struct bch_fs *c = trans->c;
1024         struct btree_iter *iter;
1025         u8 sorted[BTREE_ITER_MAX];
1026         unsigned i, nr_sorted = 0;
1027
1028         trans_for_each_iter(trans, iter)
1029                 sorted[nr_sorted++] = iter - trans->iters;
1030
1031 #define btree_iter_cmp_by_idx(_l, _r)                           \
1032                 btree_iter_cmp(&trans->iters[_l], &trans->iters[_r])
1033
1034         bubble_sort(sorted, nr_sorted, btree_iter_cmp_by_idx);
1035 #undef btree_iter_cmp_by_idx
1036
1037 retry_all:
1038         bch2_trans_unlock(trans);
1039
1040         if (unlikely(ret == -ENOMEM)) {
1041                 struct closure cl;
1042
1043                 closure_init_stack(&cl);
1044
1045                 do {
1046                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1047                         closure_sync(&cl);
1048                 } while (ret);
1049         }
1050
1051         if (unlikely(ret == -EIO)) {
1052                 trans->error = true;
1053                 if (orig_iter) {
1054                         orig_iter->flags |= BTREE_ITER_ERROR;
1055                         orig_iter->l[orig_iter->level].b =
1056                                 BTREE_ITER_NO_NODE_ERROR;
1057                 }
1058                 goto out;
1059         }
1060
1061         BUG_ON(ret && ret != -EINTR);
1062
1063         /* Now, redo traversals in correct order: */
1064         for (i = 0; i < nr_sorted; i++) {
1065                 iter = &trans->iters[sorted[i]];
1066
1067                 ret = btree_iter_traverse_one(iter);
1068                 if (ret)
1069                         goto retry_all;
1070         }
1071
1072         if (hweight64(trans->iters_live) > 1)
1073                 ret = -EINTR;
1074         else
1075                 trans_for_each_iter(trans, iter)
1076                         if (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) {
1077                                 ret = -EINTR;
1078                                 break;
1079                         }
1080 out:
1081         bch2_btree_cache_cannibalize_unlock(c);
1082         return ret;
1083 }
1084
1085 int bch2_btree_iter_traverse_all(struct btree_trans *trans)
1086 {
1087         return __btree_iter_traverse_all(trans, NULL, 0);
1088 }
1089
1090 static inline bool btree_iter_good_node(struct btree_iter *iter,
1091                                         unsigned l, int check_pos)
1092 {
1093         if (!is_btree_node(iter, l) ||
1094             !bch2_btree_node_relock(iter, l))
1095                 return false;
1096
1097         if (check_pos <= 0 && btree_iter_pos_before_node(iter, iter->l[l].b))
1098                 return false;
1099         if (check_pos >= 0 && btree_iter_pos_after_node(iter, iter->l[l].b))
1100                 return false;
1101         return true;
1102 }
1103
1104 static inline unsigned btree_iter_up_until_good_node(struct btree_iter *iter,
1105                                                      int check_pos)
1106 {
1107         unsigned l = iter->level;
1108
1109         while (btree_iter_node(iter, l) &&
1110                !btree_iter_good_node(iter, l, check_pos)) {
1111                 btree_node_unlock(iter, l);
1112                 iter->l[l].b = BTREE_ITER_NO_NODE_UP;
1113                 l++;
1114         }
1115
1116         return l;
1117 }
1118
1119 /*
1120  * This is the main state machine for walking down the btree - walks down to a
1121  * specified depth
1122  *
1123  * Returns 0 on success, -EIO on error (error reading in a btree node).
1124  *
1125  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1126  * stashed in the iterator and returned from bch2_trans_exit().
1127  */
1128 static int btree_iter_traverse_one(struct btree_iter *iter)
1129 {
1130         unsigned depth_want = iter->level;
1131
1132         if (unlikely(iter->level >= BTREE_MAX_DEPTH))
1133                 return 0;
1134
1135         /*
1136          * if we need interior nodes locked, call btree_iter_relock() to make
1137          * sure we walk back up enough that we lock them:
1138          */
1139         if (iter->uptodate == BTREE_ITER_NEED_RELOCK ||
1140             iter->locks_want > 1)
1141                 bch2_btree_iter_relock(iter, false);
1142
1143         if (iter->uptodate < BTREE_ITER_NEED_RELOCK)
1144                 return 0;
1145
1146         /*
1147          * XXX: correctly using BTREE_ITER_UPTODATE should make using check_pos
1148          * here unnecessary
1149          */
1150         iter->level = btree_iter_up_until_good_node(iter, 0);
1151
1152         /*
1153          * If we've got a btree node locked (i.e. we aren't about to relock the
1154          * root) - advance its node iterator if necessary:
1155          *
1156          * XXX correctly using BTREE_ITER_UPTODATE should make this unnecessary
1157          */
1158         if (btree_iter_node(iter, iter->level)) {
1159                 BUG_ON(!btree_iter_pos_in_node(iter, iter->l[iter->level].b));
1160
1161                 btree_iter_advance_to_pos(iter, &iter->l[iter->level], -1);
1162         }
1163
1164         /*
1165          * Note: iter->nodes[iter->level] may be temporarily NULL here - that
1166          * would indicate to other code that we got to the end of the btree,
1167          * here it indicates that relocking the root failed - it's critical that
1168          * btree_iter_lock_root() comes next and that it can't fail
1169          */
1170         while (iter->level > depth_want) {
1171                 int ret = btree_iter_node(iter, iter->level)
1172                         ? btree_iter_down(iter)
1173                         : btree_iter_lock_root(iter, depth_want);
1174                 if (unlikely(ret)) {
1175                         if (ret == 1)
1176                                 return 0;
1177
1178                         iter->level = depth_want;
1179                         iter->l[iter->level].b = BTREE_ITER_NO_NODE_DOWN;
1180                         return ret;
1181                 }
1182         }
1183
1184         iter->uptodate = BTREE_ITER_NEED_PEEK;
1185
1186         bch2_btree_iter_verify(iter);
1187         return 0;
1188 }
1189
1190 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
1191 {
1192         int ret;
1193
1194         ret =   bch2_trans_cond_resched(iter->trans) ?:
1195                 btree_iter_traverse_one(iter);
1196         if (unlikely(ret))
1197                 ret = __btree_iter_traverse_all(iter->trans, iter, ret);
1198
1199         return ret;
1200 }
1201
1202 static inline void bch2_btree_iter_checks(struct btree_iter *iter,
1203                                           enum btree_iter_type type)
1204 {
1205         EBUG_ON(iter->btree_id >= BTREE_ID_NR);
1206         EBUG_ON(btree_iter_type(iter) != type);
1207
1208         BUG_ON(type == BTREE_ITER_KEYS &&
1209                (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
1210                 bkey_cmp(iter->pos, iter->k.p) > 0));
1211
1212         bch2_btree_iter_verify_locks(iter);
1213         bch2_btree_iter_verify_level(iter, iter->level);
1214 }
1215
1216 /* Iterate across nodes (leaf and interior nodes) */
1217
1218 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1219 {
1220         struct btree *b;
1221         int ret;
1222
1223         bch2_btree_iter_checks(iter, BTREE_ITER_NODES);
1224
1225         if (iter->uptodate == BTREE_ITER_UPTODATE)
1226                 return iter->l[iter->level].b;
1227
1228         ret = bch2_btree_iter_traverse(iter);
1229         if (ret)
1230                 return NULL;
1231
1232         b = btree_iter_node(iter, iter->level);
1233         if (!b)
1234                 return NULL;
1235
1236         BUG_ON(bkey_cmp(b->key.k.p, iter->pos) < 0);
1237
1238         iter->pos = b->key.k.p;
1239         iter->uptodate = BTREE_ITER_UPTODATE;
1240
1241         bch2_btree_iter_verify(iter);
1242
1243         return b;
1244 }
1245
1246 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1247 {
1248         struct btree *b;
1249         int ret;
1250
1251         bch2_btree_iter_checks(iter, BTREE_ITER_NODES);
1252
1253         /* already got to end? */
1254         if (!btree_iter_node(iter, iter->level))
1255                 return NULL;
1256
1257         bch2_trans_cond_resched(iter->trans);
1258
1259         btree_iter_up(iter);
1260
1261         if (!bch2_btree_node_relock(iter, iter->level))
1262                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK);
1263
1264         ret = bch2_btree_iter_traverse(iter);
1265         if (ret)
1266                 return NULL;
1267
1268         /* got to end? */
1269         b = btree_iter_node(iter, iter->level);
1270         if (!b)
1271                 return NULL;
1272
1273         if (bkey_cmp(iter->pos, b->key.k.p) < 0) {
1274                 /*
1275                  * Haven't gotten to the end of the parent node: go back down to
1276                  * the next child node
1277                  */
1278
1279                 /*
1280                  * We don't really want to be unlocking here except we can't
1281                  * directly tell btree_iter_traverse() "traverse to this level"
1282                  * except by setting iter->level, so we have to unlock so we
1283                  * don't screw up our lock invariants:
1284                  */
1285                 if (btree_node_read_locked(iter, iter->level))
1286                         btree_node_unlock(iter, iter->level);
1287
1288                 iter->pos       = bkey_successor(iter->pos);
1289                 iter->level     = iter->min_depth;
1290
1291                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1292                 ret = bch2_btree_iter_traverse(iter);
1293                 if (ret)
1294                         return NULL;
1295
1296                 b = iter->l[iter->level].b;
1297         }
1298
1299         iter->pos = b->key.k.p;
1300         iter->uptodate = BTREE_ITER_UPTODATE;
1301
1302         bch2_btree_iter_verify(iter);
1303
1304         return b;
1305 }
1306
1307 /* Iterate across keys (in leaf nodes only) */
1308
1309 void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_pos)
1310 {
1311         struct btree_iter_level *l = &iter->l[0];
1312
1313         EBUG_ON(iter->level != 0);
1314         EBUG_ON(bkey_cmp(new_pos, iter->pos) < 0);
1315         EBUG_ON(!btree_node_locked(iter, 0));
1316         EBUG_ON(bkey_cmp(new_pos, l->b->key.k.p) > 0);
1317
1318         bkey_init(&iter->k);
1319         iter->k.p = iter->pos = new_pos;
1320         btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
1321
1322         btree_iter_advance_to_pos(iter, l, -1);
1323
1324         /*
1325          * XXX:
1326          * keeping a node locked that's outside (even just outside) iter->pos
1327          * breaks __bch2_btree_node_lock(). This seems to only affect
1328          * bch2_btree_node_get_sibling so for now it's fixed there, but we
1329          * should try to get rid of this corner case.
1330          *
1331          * (this behaviour is currently needed for BTREE_INSERT_NOUNLOCK)
1332          */
1333
1334         if (bch2_btree_node_iter_end(&l->iter) &&
1335             btree_iter_pos_after_node(iter, l->b))
1336                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1337 }
1338
1339 static void btree_iter_pos_changed(struct btree_iter *iter, int cmp)
1340 {
1341         unsigned l = iter->level;
1342
1343         if (!cmp)
1344                 goto out;
1345
1346         l = btree_iter_up_until_good_node(iter, cmp);
1347
1348         if (btree_iter_node(iter, l)) {
1349                 /*
1350                  * We might have to skip over many keys, or just a few: try
1351                  * advancing the node iterator, and if we have to skip over too
1352                  * many keys just reinit it (or if we're rewinding, since that
1353                  * is expensive).
1354                  */
1355                 if (cmp < 0 ||
1356                     !btree_iter_advance_to_pos(iter, &iter->l[l], 8))
1357                         __btree_iter_init(iter, l);
1358
1359                 /* Don't leave it locked if we're not supposed to: */
1360                 if (btree_lock_want(iter, l) == BTREE_NODE_UNLOCKED)
1361                         btree_node_unlock(iter, l);
1362         }
1363 out:
1364         if (l != iter->level)
1365                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
1366         else
1367                 btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
1368 }
1369
1370 void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos,
1371                                bool strictly_greater)
1372 {
1373         struct bpos old = btree_iter_search_key(iter);
1374         int cmp;
1375
1376         iter->flags &= ~BTREE_ITER_IS_EXTENTS;
1377         iter->flags |= strictly_greater ? BTREE_ITER_IS_EXTENTS : 0;
1378
1379         bkey_init(&iter->k);
1380         iter->k.p = iter->pos = new_pos;
1381
1382         cmp = bkey_cmp(btree_iter_search_key(iter), old);
1383
1384         btree_iter_pos_changed(iter, cmp);
1385 }
1386
1387 void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
1388 {
1389         int cmp = bkey_cmp(new_pos, iter->pos);
1390
1391         bkey_init(&iter->k);
1392         iter->k.p = iter->pos = new_pos;
1393
1394         btree_iter_pos_changed(iter, cmp);
1395 }
1396
1397 static inline bool btree_iter_set_pos_to_next_leaf(struct btree_iter *iter)
1398 {
1399         struct btree_iter_level *l = &iter->l[0];
1400         bool ret;
1401
1402         bkey_init(&iter->k);
1403         iter->k.p = iter->pos = l->b->key.k.p;
1404
1405         ret = bkey_cmp(iter->pos, POS_MAX) != 0;
1406         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1407                 iter->k.p = iter->pos = bkey_successor(iter->pos);
1408
1409         btree_iter_pos_changed(iter, 1);
1410         return ret;
1411 }
1412
1413 static inline bool btree_iter_set_pos_to_prev_leaf(struct btree_iter *iter)
1414 {
1415         struct btree_iter_level *l = &iter->l[0];
1416         bool ret;
1417
1418         bkey_init(&iter->k);
1419         iter->k.p = iter->pos = l->b->data->min_key;
1420         iter->uptodate  = BTREE_ITER_NEED_TRAVERSE;
1421
1422         ret = bkey_cmp(iter->pos, POS_MIN) != 0;
1423         if (ret) {
1424                 iter->k.p = iter->pos = bkey_predecessor(iter->pos);
1425
1426                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
1427                         iter->k.p = iter->pos = bkey_predecessor(iter->pos);
1428         }
1429
1430         btree_iter_pos_changed(iter, -1);
1431         return ret;
1432 }
1433
1434 /**
1435  * btree_iter_peek_uptodate - given an iterator that is uptodate, return the key
1436  * it currently points to
1437  */
1438 static inline struct bkey_s_c btree_iter_peek_uptodate(struct btree_iter *iter)
1439 {
1440         struct btree_iter_level *l = &iter->l[0];
1441         struct bkey_s_c ret = { .k = &iter->k };
1442
1443         if (!bkey_deleted(&iter->k)) {
1444                 struct bkey_packed *_k =
1445                         __bch2_btree_node_iter_peek_all(&l->iter, l->b);
1446
1447                 ret.v = bkeyp_val(&l->b->format, _k);
1448
1449                 if (debug_check_iterators(iter->trans->c)) {
1450                         struct bkey k = bkey_unpack_key(l->b, _k);
1451
1452                         BUG_ON(memcmp(&k, &iter->k, sizeof(k)));
1453                 }
1454
1455                 if (debug_check_bkeys(iter->trans->c))
1456                         bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
1457         }
1458
1459         return ret;
1460 }
1461
1462 /**
1463  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
1464  * current position
1465  */
1466 struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
1467 {
1468         struct btree_iter_level *l = &iter->l[0];
1469         struct bkey_s_c k;
1470         int ret;
1471
1472         bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
1473
1474         if (iter->uptodate == BTREE_ITER_UPTODATE &&
1475             !bkey_deleted(&iter->k))
1476                 return btree_iter_peek_uptodate(iter);
1477
1478         while (1) {
1479                 ret = bch2_btree_iter_traverse(iter);
1480                 if (unlikely(ret))
1481                         return bkey_s_c_err(ret);
1482
1483                 k = __btree_iter_peek(iter, l);
1484                 if (likely(k.k))
1485                         break;
1486
1487                 if (!btree_iter_set_pos_to_next_leaf(iter))
1488                         return bkey_s_c_null;
1489         }
1490
1491         /*
1492          * iter->pos should always be equal to the key we just
1493          * returned - except extents can straddle iter->pos:
1494          */
1495         if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
1496             bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
1497                 iter->pos = bkey_start_pos(k.k);
1498
1499         iter->uptodate = BTREE_ITER_UPTODATE;
1500
1501         bch2_btree_iter_verify_level(iter, 0);
1502         return k;
1503 }
1504
1505 /**
1506  * bch2_btree_iter_next: returns first key greater than iterator's current
1507  * position
1508  */
1509 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
1510 {
1511         if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
1512                 return bkey_s_c_null;
1513
1514         bch2_btree_iter_set_pos(iter,
1515                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1516                 ? iter->k.p
1517                 : bkey_successor(iter->k.p));
1518
1519         return bch2_btree_iter_peek(iter);
1520 }
1521
1522 static struct bkey_s_c __btree_trans_updates_peek(struct btree_iter *iter)
1523 {
1524         struct bpos pos = btree_iter_search_key(iter);
1525         struct btree_trans *trans = iter->trans;
1526         struct btree_insert_entry *i;
1527
1528         trans_for_each_update2(trans, i)
1529                 if ((cmp_int(iter->btree_id,    i->iter->btree_id) ?:
1530                      bkey_cmp(pos,              i->k->k.p)) <= 0)
1531                         break;
1532
1533         return i < trans->updates2 + trans->nr_updates2 &&
1534                 iter->btree_id == i->iter->btree_id
1535                 ? bkey_i_to_s_c(i->k)
1536                 : bkey_s_c_null;
1537 }
1538
1539 static struct bkey_s_c __bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
1540 {
1541         struct btree_iter_level *l = &iter->l[0];
1542         struct bkey_s_c k = __btree_iter_peek(iter, l);
1543         struct bkey_s_c u = __btree_trans_updates_peek(iter);
1544
1545         if (k.k && (!u.k || bkey_cmp(k.k->p, u.k->p) < 0))
1546                 return k;
1547         if (u.k && bkey_cmp(u.k->p, l->b->key.k.p) <= 0) {
1548                 iter->k = *u.k;
1549                 return u;
1550         }
1551         return bkey_s_c_null;
1552 }
1553
1554 struct bkey_s_c bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
1555 {
1556         struct bkey_s_c k;
1557         int ret;
1558
1559         bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
1560
1561         while (1) {
1562                 ret = bch2_btree_iter_traverse(iter);
1563                 if (unlikely(ret))
1564                         return bkey_s_c_err(ret);
1565
1566                 k = __bch2_btree_iter_peek_with_updates(iter);
1567
1568                 if (k.k && bkey_deleted(k.k)) {
1569                         bch2_btree_iter_set_pos(iter,
1570                                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1571                                 ? iter->k.p
1572                                 : bkey_successor(iter->k.p));
1573                         continue;
1574                 }
1575
1576                 if (likely(k.k))
1577                         break;
1578
1579                 if (!btree_iter_set_pos_to_next_leaf(iter))
1580                         return bkey_s_c_null;
1581         }
1582
1583         /*
1584          * iter->pos should always be equal to the key we just
1585          * returned - except extents can straddle iter->pos:
1586          */
1587         if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
1588             bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
1589                 iter->pos = bkey_start_pos(k.k);
1590
1591         iter->uptodate = BTREE_ITER_UPTODATE;
1592         return k;
1593 }
1594
1595 struct bkey_s_c bch2_btree_iter_next_with_updates(struct btree_iter *iter)
1596 {
1597         if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
1598                 return bkey_s_c_null;
1599
1600         bch2_btree_iter_set_pos(iter,
1601                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1602                 ? iter->k.p
1603                 : bkey_successor(iter->k.p));
1604
1605         return bch2_btree_iter_peek_with_updates(iter);
1606 }
1607
1608 /**
1609  * bch2_btree_iter_peek_prev: returns first key less than or equal to
1610  * iterator's current position
1611  */
1612 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
1613 {
1614         struct bpos pos = iter->pos;
1615         struct btree_iter_level *l = &iter->l[0];
1616         struct bkey_s_c k;
1617         int ret;
1618
1619         bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
1620
1621         if (iter->uptodate == BTREE_ITER_UPTODATE &&
1622             !bkey_deleted(&iter->k))
1623                 return btree_iter_peek_uptodate(iter);
1624
1625         while (1) {
1626                 ret = bch2_btree_iter_traverse(iter);
1627                 if (unlikely(ret))
1628                         return bkey_s_c_err(ret);
1629
1630                 k = __btree_iter_peek(iter, l);
1631                 if (!k.k || bkey_cmp(bkey_start_pos(k.k), pos) > 0)
1632                         k = __btree_iter_prev(iter, l);
1633
1634                 if (likely(k.k))
1635                         break;
1636
1637                 if (!btree_iter_set_pos_to_prev_leaf(iter))
1638                         return bkey_s_c_null;
1639         }
1640
1641         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), pos) > 0);
1642         iter->pos       = bkey_start_pos(k.k);
1643         iter->uptodate  = BTREE_ITER_UPTODATE;
1644         return k;
1645 }
1646
1647 /**
1648  * bch2_btree_iter_prev: returns first key less than iterator's current
1649  * position
1650  */
1651 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
1652 {
1653         struct bpos pos = bkey_start_pos(&iter->k);
1654
1655         bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
1656
1657         if (unlikely(!bkey_cmp(pos, POS_MIN)))
1658                 return bkey_s_c_null;
1659
1660         bch2_btree_iter_set_pos(iter, bkey_predecessor(pos));
1661
1662         return bch2_btree_iter_peek_prev(iter);
1663 }
1664
1665 static inline struct bkey_s_c
1666 __bch2_btree_iter_peek_slot_extents(struct btree_iter *iter)
1667 {
1668         struct btree_iter_level *l = &iter->l[0];
1669         struct btree_node_iter node_iter;
1670         struct bkey_s_c k;
1671         struct bkey n;
1672         int ret;
1673
1674         /* keys & holes can't span inode numbers: */
1675         if (iter->pos.offset == KEY_OFFSET_MAX) {
1676                 if (iter->pos.inode == KEY_INODE_MAX)
1677                         return bkey_s_c_null;
1678
1679                 bch2_btree_iter_set_pos(iter, bkey_successor(iter->pos));
1680
1681                 ret = bch2_btree_iter_traverse(iter);
1682                 if (unlikely(ret))
1683                         return bkey_s_c_err(ret);
1684         }
1685
1686         /*
1687          * iterator is now at the correct position for inserting at iter->pos,
1688          * but we need to keep iterating until we find the first non whiteout so
1689          * we know how big a hole we have, if any:
1690          */
1691
1692         node_iter = l->iter;
1693         k = __btree_iter_unpack(iter, l, &iter->k,
1694                 bch2_btree_node_iter_peek(&node_iter, l->b));
1695
1696         if (k.k && bkey_cmp(bkey_start_pos(k.k), iter->pos) <= 0) {
1697                 /*
1698                  * We're not setting iter->uptodate because the node iterator
1699                  * doesn't necessarily point at the key we're returning:
1700                  */
1701
1702                 EBUG_ON(bkey_cmp(k.k->p, iter->pos) <= 0);
1703                 bch2_btree_iter_verify_level(iter, 0);
1704                 return k;
1705         }
1706
1707         /* hole */
1708
1709         if (!k.k)
1710                 k.k = &l->b->key.k;
1711
1712         bkey_init(&n);
1713         n.p = iter->pos;
1714         bch2_key_resize(&n,
1715                         min_t(u64, KEY_SIZE_MAX,
1716                               (k.k->p.inode == n.p.inode
1717                                ? bkey_start_offset(k.k)
1718                                : KEY_OFFSET_MAX) -
1719                               n.p.offset));
1720
1721         EBUG_ON(!n.size);
1722
1723         iter->k = n;
1724         iter->uptodate = BTREE_ITER_UPTODATE;
1725
1726         bch2_btree_iter_verify_level(iter, 0);
1727         return (struct bkey_s_c) { &iter->k, NULL };
1728 }
1729
1730 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
1731 {
1732         struct btree_iter_level *l = &iter->l[0];
1733         struct bkey_s_c k;
1734         int ret;
1735
1736         bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
1737
1738         if (iter->uptodate == BTREE_ITER_UPTODATE)
1739                 return btree_iter_peek_uptodate(iter);
1740
1741         ret = bch2_btree_iter_traverse(iter);
1742         if (unlikely(ret))
1743                 return bkey_s_c_err(ret);
1744
1745         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1746                 return __bch2_btree_iter_peek_slot_extents(iter);
1747
1748         k = __btree_iter_peek_all(iter, l, &iter->k);
1749
1750         EBUG_ON(k.k && bkey_deleted(k.k) && bkey_cmp(k.k->p, iter->pos) == 0);
1751
1752         if (!k.k || bkey_cmp(iter->pos, k.k->p)) {
1753                 /* hole */
1754                 bkey_init(&iter->k);
1755                 iter->k.p = iter->pos;
1756                 k = (struct bkey_s_c) { &iter->k, NULL };
1757         }
1758
1759         iter->uptodate = BTREE_ITER_UPTODATE;
1760         bch2_btree_iter_verify_level(iter, 0);
1761         return k;
1762 }
1763
1764 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
1765 {
1766         if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
1767                 return bkey_s_c_null;
1768
1769         bch2_btree_iter_set_pos(iter,
1770                 (iter->flags & BTREE_ITER_IS_EXTENTS)
1771                 ? iter->k.p
1772                 : bkey_successor(iter->k.p));
1773
1774         return bch2_btree_iter_peek_slot(iter);
1775 }
1776
1777 static inline void bch2_btree_iter_init(struct btree_trans *trans,
1778                         struct btree_iter *iter, enum btree_id btree_id,
1779                         struct bpos pos, unsigned flags)
1780 {
1781         struct bch_fs *c = trans->c;
1782         unsigned i;
1783
1784         if (btree_node_type_is_extents(btree_id) &&
1785             !(flags & BTREE_ITER_NODES))
1786                 flags |= BTREE_ITER_IS_EXTENTS;
1787
1788         iter->trans                     = trans;
1789         iter->pos                       = pos;
1790         bkey_init(&iter->k);
1791         iter->k.p                       = pos;
1792         iter->flags                     = flags;
1793         iter->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1794         iter->btree_id                  = btree_id;
1795         iter->level                     = 0;
1796         iter->min_depth                 = 0;
1797         iter->locks_want                = flags & BTREE_ITER_INTENT ? 1 : 0;
1798         iter->nodes_locked              = 0;
1799         iter->nodes_intent_locked       = 0;
1800         for (i = 0; i < ARRAY_SIZE(iter->l); i++)
1801                 iter->l[i].b            = BTREE_ITER_NO_NODE_INIT;
1802
1803         prefetch(c->btree_roots[btree_id].b);
1804 }
1805
1806 /* new transactional stuff: */
1807
1808 static inline void __bch2_trans_iter_free(struct btree_trans *trans,
1809                                           unsigned idx)
1810 {
1811         __bch2_btree_iter_unlock(&trans->iters[idx]);
1812         trans->iters_linked             &= ~(1ULL << idx);
1813         trans->iters_live               &= ~(1ULL << idx);
1814         trans->iters_touched            &= ~(1ULL << idx);
1815 }
1816
1817 int bch2_trans_iter_put(struct btree_trans *trans,
1818                         struct btree_iter *iter)
1819 {
1820         int ret;
1821
1822         if (IS_ERR_OR_NULL(iter))
1823                 return 0;
1824
1825         BUG_ON(trans->iters + iter->idx != iter);
1826
1827         ret = btree_iter_err(iter);
1828
1829         if (!(trans->iters_touched & (1ULL << iter->idx)) &&
1830             !(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT))
1831                 __bch2_trans_iter_free(trans, iter->idx);
1832
1833         trans->iters_live       &= ~(1ULL << iter->idx);
1834         return ret;
1835 }
1836
1837 int bch2_trans_iter_free(struct btree_trans *trans,
1838                          struct btree_iter *iter)
1839 {
1840         if (IS_ERR_OR_NULL(iter))
1841                 return 0;
1842
1843         trans->iters_touched &= ~(1ULL << iter->idx);
1844
1845         return bch2_trans_iter_put(trans, iter);
1846 }
1847
1848 static int bch2_trans_realloc_iters(struct btree_trans *trans,
1849                                     unsigned new_size)
1850 {
1851         void *p, *new_iters, *new_updates, *new_updates2;
1852         size_t iters_bytes;
1853         size_t updates_bytes;
1854
1855         new_size = roundup_pow_of_two(new_size);
1856
1857         BUG_ON(new_size > BTREE_ITER_MAX);
1858
1859         if (new_size <= trans->size)
1860                 return 0;
1861
1862         BUG_ON(trans->used_mempool);
1863
1864         bch2_trans_unlock(trans);
1865
1866         iters_bytes     = sizeof(struct btree_iter) * new_size;
1867         updates_bytes   = sizeof(struct btree_insert_entry) * new_size;
1868
1869         p = kmalloc(iters_bytes +
1870                     updates_bytes +
1871                     updates_bytes, GFP_NOFS);
1872         if (p)
1873                 goto success;
1874
1875         p = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
1876         new_size = BTREE_ITER_MAX;
1877
1878         trans->used_mempool = true;
1879 success:
1880         new_iters       = p; p += iters_bytes;
1881         new_updates     = p; p += updates_bytes;
1882         new_updates2    = p; p += updates_bytes;
1883
1884         memcpy(new_iters, trans->iters,
1885                sizeof(struct btree_iter) * trans->nr_iters);
1886         memcpy(new_updates, trans->updates,
1887                sizeof(struct btree_insert_entry) * trans->nr_updates);
1888         memcpy(new_updates2, trans->updates2,
1889                sizeof(struct btree_insert_entry) * trans->nr_updates2);
1890
1891         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
1892                 memset(trans->iters, POISON_FREE,
1893                        sizeof(struct btree_iter) * trans->nr_iters +
1894                        sizeof(struct btree_insert_entry) * trans->nr_iters);
1895
1896         if (trans->iters != trans->iters_onstack)
1897                 kfree(trans->iters);
1898
1899         trans->iters            = new_iters;
1900         trans->updates          = new_updates;
1901         trans->updates2         = new_updates2;
1902         trans->size             = new_size;
1903
1904         if (trans->iters_live) {
1905                 trace_trans_restart_iters_realloced(trans->ip, trans->size);
1906                 return -EINTR;
1907         }
1908
1909         return 0;
1910 }
1911
1912 static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans)
1913 {
1914         unsigned idx = __ffs64(~trans->iters_linked);
1915
1916         if (idx < trans->nr_iters)
1917                 goto got_slot;
1918
1919         if (trans->nr_iters == trans->size) {
1920                 int ret;
1921
1922                 if (trans->nr_iters >= BTREE_ITER_MAX) {
1923                         struct btree_iter *iter;
1924
1925                         trans_for_each_iter(trans, iter) {
1926                                 pr_err("iter: btree %s pos %llu:%llu%s%s%s %ps",
1927                                        bch2_btree_ids[iter->btree_id],
1928                                        iter->pos.inode,
1929                                        iter->pos.offset,
1930                                        (trans->iters_live & (1ULL << iter->idx)) ? " live" : "",
1931                                        (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "",
1932                                        iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "",
1933                                        (void *) iter->ip_allocated);
1934                         }
1935
1936                         panic("trans iter oveflow\n");
1937                 }
1938
1939                 ret = bch2_trans_realloc_iters(trans, trans->size * 2);
1940                 if (ret)
1941                         return ERR_PTR(ret);
1942         }
1943
1944         idx = trans->nr_iters++;
1945         BUG_ON(trans->nr_iters > trans->size);
1946
1947         trans->iters[idx].idx = idx;
1948 got_slot:
1949         BUG_ON(trans->iters_linked & (1ULL << idx));
1950         trans->iters_linked |= 1ULL << idx;
1951         trans->iters[idx].flags = 0;
1952         return &trans->iters[idx];
1953 }
1954
1955 static inline void btree_iter_copy(struct btree_iter *dst,
1956                                    struct btree_iter *src)
1957 {
1958         unsigned i, idx = dst->idx;
1959
1960         *dst = *src;
1961         dst->idx = idx;
1962
1963         for (i = 0; i < BTREE_MAX_DEPTH; i++)
1964                 if (btree_node_locked(dst, i))
1965                         six_lock_increment(&dst->l[i].b->lock,
1966                                            __btree_lock_want(dst, i));
1967
1968         dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
1969         dst->flags &= ~BTREE_ITER_SET_POS_AFTER_COMMIT;
1970 }
1971
1972 static inline struct bpos bpos_diff(struct bpos l, struct bpos r)
1973 {
1974         if (bkey_cmp(l, r) > 0)
1975                 swap(l, r);
1976
1977         return POS(r.inode - l.inode, r.offset - l.offset);
1978 }
1979
1980 static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans,
1981                                                  unsigned btree_id, struct bpos pos,
1982                                                  unsigned flags)
1983 {
1984         struct btree_iter *iter, *best = NULL;
1985
1986         BUG_ON(trans->nr_iters > BTREE_ITER_MAX);
1987
1988         trans_for_each_iter(trans, iter) {
1989                 if (btree_iter_type(iter) != (flags & BTREE_ITER_TYPE))
1990                         continue;
1991
1992                 if (iter->btree_id != btree_id)
1993                         continue;
1994
1995                 if (best &&
1996                     bkey_cmp(bpos_diff(best->pos, pos),
1997                              bpos_diff(iter->pos, pos)) < 0)
1998                         continue;
1999
2000                 best = iter;
2001         }
2002
2003         if (!best) {
2004                 iter = btree_trans_iter_alloc(trans);
2005                 if (IS_ERR(iter))
2006                         return iter;
2007
2008                 bch2_btree_iter_init(trans, iter, btree_id, pos, flags);
2009         } else if ((trans->iters_live & (1ULL << best->idx)) ||
2010                    (best->flags & BTREE_ITER_KEEP_UNTIL_COMMIT)) {
2011                 iter = btree_trans_iter_alloc(trans);
2012                 if (IS_ERR(iter))
2013                         return iter;
2014
2015                 btree_iter_copy(iter, best);
2016         } else {
2017                 iter = best;
2018         }
2019
2020         iter->flags &= ~(BTREE_ITER_SLOTS|BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
2021         iter->flags |= flags & (BTREE_ITER_SLOTS|BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
2022
2023         if (iter->flags & BTREE_ITER_INTENT)
2024                 bch2_btree_iter_upgrade(iter, 1);
2025         else
2026                 bch2_btree_iter_downgrade(iter);
2027
2028         BUG_ON(iter->btree_id != btree_id);
2029         BUG_ON((iter->flags ^ flags) & BTREE_ITER_TYPE);
2030         BUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
2031         BUG_ON(iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT);
2032         BUG_ON(trans->iters_live & (1ULL << iter->idx));
2033
2034         trans->iters_live       |= 1ULL << iter->idx;
2035         trans->iters_touched    |= 1ULL << iter->idx;
2036
2037         return iter;
2038 }
2039
2040 struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans,
2041                                          enum btree_id btree_id,
2042                                          struct bpos pos, unsigned flags)
2043 {
2044         struct btree_iter *iter =
2045                 __btree_trans_get_iter(trans, btree_id, pos, flags);
2046
2047         if (!IS_ERR(iter))
2048                 __bch2_btree_iter_set_pos(iter, pos,
2049                         btree_node_type_is_extents(btree_id));
2050         return iter;
2051 }
2052
2053 struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *trans,
2054                                             enum btree_id btree_id,
2055                                             struct bpos pos,
2056                                             unsigned locks_want,
2057                                             unsigned depth,
2058                                             unsigned flags)
2059 {
2060         struct btree_iter *iter =
2061                 __btree_trans_get_iter(trans, btree_id, pos,
2062                                        flags|BTREE_ITER_NODES);
2063         unsigned i;
2064
2065         BUG_ON(IS_ERR(iter));
2066         BUG_ON(bkey_cmp(iter->pos, pos));
2067
2068         iter->locks_want = locks_want;
2069         iter->level     = depth;
2070         iter->min_depth = depth;
2071
2072         for (i = 0; i < ARRAY_SIZE(iter->l); i++)
2073                 iter->l[i].b            = NULL;
2074         iter->l[iter->level].b          = BTREE_ITER_NO_NODE_INIT;
2075
2076         return iter;
2077 }
2078
2079 struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *trans,
2080                                         struct btree_iter *src)
2081 {
2082         struct btree_iter *iter;
2083
2084         iter = btree_trans_iter_alloc(trans);
2085         if (IS_ERR(iter))
2086                 return iter;
2087
2088         btree_iter_copy(iter, src);
2089
2090         trans->iters_live |= 1ULL << iter->idx;
2091         /*
2092          * We don't need to preserve this iter since it's cheap to copy it
2093          * again - this will cause trans_iter_put() to free it right away:
2094          */
2095         trans->iters_touched &= ~(1ULL << iter->idx);
2096
2097         return iter;
2098 }
2099
2100 static int bch2_trans_preload_mem(struct btree_trans *trans, size_t size)
2101 {
2102         if (size > trans->mem_bytes) {
2103                 size_t old_bytes = trans->mem_bytes;
2104                 size_t new_bytes = roundup_pow_of_two(size);
2105                 void *new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2106
2107                 if (!new_mem)
2108                         return -ENOMEM;
2109
2110                 trans->mem = new_mem;
2111                 trans->mem_bytes = new_bytes;
2112
2113                 if (old_bytes) {
2114                         trace_trans_restart_mem_realloced(trans->ip, new_bytes);
2115                         return -EINTR;
2116                 }
2117         }
2118
2119         return 0;
2120 }
2121
2122 void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2123 {
2124         void *p;
2125         int ret;
2126
2127         ret = bch2_trans_preload_mem(trans, trans->mem_top + size);
2128         if (ret)
2129                 return ERR_PTR(ret);
2130
2131         p = trans->mem + trans->mem_top;
2132         trans->mem_top += size;
2133         return p;
2134 }
2135
2136 inline void bch2_trans_unlink_iters(struct btree_trans *trans)
2137 {
2138         u64 iters = trans->iters_linked &
2139                 ~trans->iters_touched &
2140                 ~trans->iters_live;
2141
2142         while (iters) {
2143                 unsigned idx = __ffs64(iters);
2144
2145                 iters &= ~(1ULL << idx);
2146                 __bch2_trans_iter_free(trans, idx);
2147         }
2148 }
2149
2150 void bch2_trans_reset(struct btree_trans *trans, unsigned flags)
2151 {
2152         struct btree_iter *iter;
2153
2154         trans_for_each_iter(trans, iter)
2155                 iter->flags &= ~(BTREE_ITER_KEEP_UNTIL_COMMIT|
2156                                  BTREE_ITER_SET_POS_AFTER_COMMIT);
2157
2158         bch2_trans_unlink_iters(trans);
2159
2160         trans->iters_touched &= trans->iters_live;
2161
2162         trans->need_reset               = 0;
2163         trans->nr_updates               = 0;
2164         trans->nr_updates2              = 0;
2165         trans->mem_top                  = 0;
2166
2167         trans->extra_journal_entries    = NULL;
2168         trans->extra_journal_entry_u64s = 0;
2169
2170         if (trans->fs_usage_deltas) {
2171                 trans->fs_usage_deltas->used = 0;
2172                 memset(&trans->fs_usage_deltas->memset_start, 0,
2173                        (void *) &trans->fs_usage_deltas->memset_end -
2174                        (void *) &trans->fs_usage_deltas->memset_start);
2175         }
2176
2177         if (!(flags & TRANS_RESET_NOTRAVERSE))
2178                 bch2_btree_iter_traverse_all(trans);
2179 }
2180
2181 void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
2182                      unsigned expected_nr_iters,
2183                      size_t expected_mem_bytes)
2184 {
2185         memset(trans, 0, offsetof(struct btree_trans, iters_onstack));
2186
2187         /*
2188          * reallocating iterators currently completely breaks
2189          * bch2_trans_iter_put():
2190          */
2191         expected_nr_iters = BTREE_ITER_MAX;
2192
2193         trans->c                = c;
2194         trans->ip               = _RET_IP_;
2195         trans->size             = ARRAY_SIZE(trans->iters_onstack);
2196         trans->iters            = trans->iters_onstack;
2197         trans->updates          = trans->updates_onstack;
2198         trans->updates2         = trans->updates2_onstack;
2199         trans->fs_usage_deltas  = NULL;
2200
2201         if (expected_nr_iters > trans->size)
2202                 bch2_trans_realloc_iters(trans, expected_nr_iters);
2203
2204         if (expected_mem_bytes)
2205                 bch2_trans_preload_mem(trans, expected_mem_bytes);
2206
2207 #ifdef CONFIG_BCACHEFS_DEBUG
2208         trans->pid = current->pid;
2209         mutex_lock(&c->btree_trans_lock);
2210         list_add(&trans->list, &c->btree_trans_list);
2211         mutex_unlock(&c->btree_trans_lock);
2212 #endif
2213 }
2214
2215 int bch2_trans_exit(struct btree_trans *trans)
2216 {
2217         bch2_trans_unlock(trans);
2218
2219 #ifdef CONFIG_BCACHEFS_DEBUG
2220         mutex_lock(&trans->c->btree_trans_lock);
2221         list_del(&trans->list);
2222         mutex_unlock(&trans->c->btree_trans_lock);
2223 #endif
2224
2225         kfree(trans->fs_usage_deltas);
2226         kfree(trans->mem);
2227         if (trans->used_mempool)
2228                 mempool_free(trans->iters, &trans->c->btree_iters_pool);
2229         else if (trans->iters != trans->iters_onstack)
2230                 kfree(trans->iters);
2231         trans->mem      = (void *) 0x1;
2232         trans->iters    = (void *) 0x1;
2233
2234         return trans->error ? -EIO : 0;
2235 }
2236
2237 void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c)
2238 {
2239 #ifdef CONFIG_BCACHEFS_DEBUG
2240         struct btree_trans *trans;
2241         struct btree_iter *iter;
2242         struct btree *b;
2243         unsigned l;
2244
2245         mutex_lock(&c->btree_trans_lock);
2246         list_for_each_entry(trans, &c->btree_trans_list, list) {
2247                 pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip);
2248
2249                 trans_for_each_iter(trans, iter) {
2250                         if (!iter->nodes_locked)
2251                                 continue;
2252
2253                         pr_buf(out, "  iter %s:", bch2_btree_ids[iter->btree_id]);
2254                         bch2_bpos_to_text(out, iter->pos);
2255                         pr_buf(out, "\n");
2256
2257                         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
2258                                 if (btree_node_locked(iter, l)) {
2259                                         b = iter->l[l].b;
2260
2261                                         pr_buf(out, "    %p l=%u %s ",
2262                                                b, l, btree_node_intent_locked(iter, l) ? "i" : "r");
2263                                         bch2_bpos_to_text(out, b->key.k.p);
2264                                         pr_buf(out, "\n");
2265                                 }
2266                         }
2267                 }
2268
2269                 b = READ_ONCE(trans->locking);
2270                 if (b) {
2271                         pr_buf(out, "  locking %px l=%u %s:",
2272                                b, b->level,
2273                                bch2_btree_ids[b->btree_id]);
2274                         bch2_bpos_to_text(out, b->key.k.p);
2275                         pr_buf(out, "\n");
2276                 }
2277         }
2278         mutex_unlock(&c->btree_trans_lock);
2279 #endif
2280 }
2281
2282 void bch2_fs_btree_iter_exit(struct bch_fs *c)
2283 {
2284         mempool_exit(&c->btree_iters_pool);
2285 }
2286
2287 int bch2_fs_btree_iter_init(struct bch_fs *c)
2288 {
2289         unsigned nr = BTREE_ITER_MAX;
2290
2291         INIT_LIST_HEAD(&c->btree_trans_list);
2292         mutex_init(&c->btree_trans_lock);
2293
2294         return mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
2295                         sizeof(struct btree_iter) * nr +
2296                         sizeof(struct btree_insert_entry) * nr +
2297                         sizeof(struct btree_insert_entry) * nr);
2298 }