git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_iter.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_buf.h"
   6 #include "btree_cache.h"
   7 #include "btree_iter.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "debug.h"
  12 #include "error.h"
  13 #include "extents.h"
  14 #include "journal.h"
  15 #include "recovery.h"
  16 #include "replicas.h"
  17 #include "subvolume.h"
  18
  19 #include <linux/prandom.h>
  20 #include <linux/prefetch.h>
  21 #include <trace/events/bcachefs.h>
  22
  23 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
  24 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
  25                                        struct btree_path *);
  26
  27 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
  28 {
  29 #ifdef TRACK_PATH_ALLOCATED
  30         return iter->ip_allocated;
  31 #else
  32         return 0;
  33 #endif
  34 }
  35
  36 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
  37
  38 /*
  39  * Unlocks before scheduling
  40  * Note: does not revalidate iterator
  41  */
  42 static inline int bch2_trans_cond_resched(struct btree_trans *trans)
  43 {
  44         if (need_resched() || race_fault()) {
  45                 bch2_trans_unlock(trans);
  46                 schedule();
  47                 return bch2_trans_relock(trans);
  48         } else {
  49                 return 0;
  50         }
  51 }
  52
  53 static inline int __btree_path_cmp(const struct btree_path *l,
  54                                    enum btree_id        r_btree_id,
  55                                    bool                 r_cached,
  56                                    struct bpos          r_pos,
  57                                    unsigned             r_level)
  58 {
  59         /*
  60          * Must match lock ordering as defined by __bch2_btree_node_lock:
  61          */
  62         return   cmp_int(l->btree_id,   r_btree_id) ?:
  63                  cmp_int((int) l->cached,       (int) r_cached) ?:
  64                  bpos_cmp(l->pos,       r_pos) ?:
  65                 -cmp_int(l->level,      r_level);
  66 }
  67
  68 static inline int btree_path_cmp(const struct btree_path *l,
  69                                  const struct btree_path *r)
  70 {
  71         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  72 }
  73
  74 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  75 {
  76         /* Are we iterating over keys in all snapshots? */
  77         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  78                 p = bpos_successor(p);
  79         } else {
  80                 p = bpos_nosnap_successor(p);
  81                 p.snapshot = iter->snapshot;
  82         }
  83
  84         return p;
  85 }
  86
  87 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  88 {
  89         /* Are we iterating over keys in all snapshots? */
  90         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  91                 p = bpos_predecessor(p);
  92         } else {
  93                 p = bpos_nosnap_predecessor(p);
  94                 p.snapshot = iter->snapshot;
  95         }
  96
  97         return p;
  98 }
  99
 100 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
 101 {
 102         struct bpos pos = iter->pos;
 103
 104         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 105             !bkey_eq(pos, POS_MAX))
 106                 pos = bkey_successor(iter, pos);
 107         return pos;
 108 }
 109
 110 static inline bool btree_path_pos_before_node(struct btree_path *path,
 111                                               struct btree *b)
 112 {
 113         return bpos_lt(path->pos, b->data->min_key);
 114 }
 115
 116 static inline bool btree_path_pos_after_node(struct btree_path *path,
 117                                              struct btree *b)
 118 {
 119         return bpos_gt(path->pos, b->key.k.p);
 120 }
 121
 122 static inline bool btree_path_pos_in_node(struct btree_path *path,
 123                                           struct btree *b)
 124 {
 125         return path->btree_id == b->c.btree_id &&
 126                 !btree_path_pos_before_node(path, b) &&
 127                 !btree_path_pos_after_node(path, b);
 128 }
 129
 130 /* Btree iterator: */
 131
 132 #ifdef CONFIG_BCACHEFS_DEBUG
 133
 134 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 135                                           struct btree_path *path)
 136 {
 137         struct bkey_cached *ck;
 138         bool locked = btree_node_locked(path, 0);
 139
 140         if (!bch2_btree_node_relock(trans, path, 0))
 141                 return;
 142
 143         ck = (void *) path->l[0].b;
 144         BUG_ON(ck->key.btree_id != path->btree_id ||
 145                !bkey_eq(ck->key.pos, path->pos));
 146
 147         if (!locked)
 148                 btree_node_unlock(trans, path, 0);
 149 }
 150
 151 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 152                                 struct btree_path *path, unsigned level)
 153 {
 154         struct btree_path_level *l;
 155         struct btree_node_iter tmp;
 156         bool locked;
 157         struct bkey_packed *p, *k;
 158         struct printbuf buf1 = PRINTBUF;
 159         struct printbuf buf2 = PRINTBUF;
 160         struct printbuf buf3 = PRINTBUF;
 161         const char *msg;
 162
 163         if (!bch2_debug_check_iterators)
 164                 return;
 165
 166         l       = &path->l[level];
 167         tmp     = l->iter;
 168         locked  = btree_node_locked(path, level);
 169
 170         if (path->cached) {
 171                 if (!level)
 172                         bch2_btree_path_verify_cached(trans, path);
 173                 return;
 174         }
 175
 176         if (!btree_path_node(path, level))
 177                 return;
 178
 179         if (!bch2_btree_node_relock_notrace(trans, path, level))
 180                 return;
 181
 182         BUG_ON(!btree_path_pos_in_node(path, l->b));
 183
 184         bch2_btree_node_iter_verify(&l->iter, l->b);
 185
 186         /*
 187          * For interior nodes, the iterator will have skipped past deleted keys:
 188          */
 189         p = level
 190                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 191                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 192         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 193
 194         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 195                 msg = "before";
 196                 goto err;
 197         }
 198
 199         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 200                 msg = "after";
 201                 goto err;
 202         }
 203
 204         if (!locked)
 205                 btree_node_unlock(trans, path, level);
 206         return;
 207 err:
 208         bch2_bpos_to_text(&buf1, path->pos);
 209
 210         if (p) {
 211                 struct bkey uk = bkey_unpack_key(l->b, p);
 212
 213                 bch2_bkey_to_text(&buf2, &uk);
 214         } else {
 215                 prt_printf(&buf2, "(none)");
 216         }
 217
 218         if (k) {
 219                 struct bkey uk = bkey_unpack_key(l->b, k);
 220
 221                 bch2_bkey_to_text(&buf3, &uk);
 222         } else {
 223                 prt_printf(&buf3, "(none)");
 224         }
 225
 226         panic("path should be %s key at level %u:\n"
 227               "path pos %s\n"
 228               "prev key %s\n"
 229               "cur  key %s\n",
 230               msg, level, buf1.buf, buf2.buf, buf3.buf);
 231 }
 232
 233 static void bch2_btree_path_verify(struct btree_trans *trans,
 234                                    struct btree_path *path)
 235 {
 236         struct bch_fs *c = trans->c;
 237         unsigned i;
 238
 239         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 240
 241         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 242                 if (!path->l[i].b) {
 243                         BUG_ON(!path->cached &&
 244                                c->btree_roots[path->btree_id].b->c.level > i);
 245                         break;
 246                 }
 247
 248                 bch2_btree_path_verify_level(trans, path, i);
 249         }
 250
 251         bch2_btree_path_verify_locks(path);
 252 }
 253
 254 void bch2_trans_verify_paths(struct btree_trans *trans)
 255 {
 256         struct btree_path *path;
 257
 258         trans_for_each_path(trans, path)
 259                 bch2_btree_path_verify(trans, path);
 260 }
 261
 262 static void bch2_btree_iter_verify(struct btree_iter *iter)
 263 {
 264         struct btree_trans *trans = iter->trans;
 265
 266         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 267
 268         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
 269
 270         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 271                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 272
 273         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 274                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 275                !btree_type_has_snapshots(iter->btree_id));
 276
 277         if (iter->update_path)
 278                 bch2_btree_path_verify(trans, iter->update_path);
 279         bch2_btree_path_verify(trans, iter->path);
 280 }
 281
 282 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 283 {
 284         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 285                !iter->pos.snapshot);
 286
 287         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 288                iter->pos.snapshot != iter->snapshot);
 289
 290         BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
 291                bkey_gt(iter->pos, iter->k.p));
 292 }
 293
 294 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 295 {
 296         struct btree_trans *trans = iter->trans;
 297         struct btree_iter copy;
 298         struct bkey_s_c prev;
 299         int ret = 0;
 300
 301         if (!bch2_debug_check_iterators)
 302                 return 0;
 303
 304         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 305                 return 0;
 306
 307         if (bkey_err(k) || !k.k)
 308                 return 0;
 309
 310         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 311                                           iter->snapshot,
 312                                           k.k->p.snapshot));
 313
 314         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 315                              BTREE_ITER_NOPRESERVE|
 316                              BTREE_ITER_ALL_SNAPSHOTS);
 317         prev = bch2_btree_iter_prev(&copy);
 318         if (!prev.k)
 319                 goto out;
 320
 321         ret = bkey_err(prev);
 322         if (ret)
 323                 goto out;
 324
 325         if (bkey_eq(prev.k->p, k.k->p) &&
 326             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 327                                       prev.k->p.snapshot) > 0) {
 328                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 329
 330                 bch2_bkey_to_text(&buf1, k.k);
 331                 bch2_bkey_to_text(&buf2, prev.k);
 332
 333                 panic("iter snap %u\n"
 334                       "k    %s\n"
 335                       "prev %s\n",
 336                       iter->snapshot,
 337                       buf1.buf, buf2.buf);
 338         }
 339 out:
 340         bch2_trans_iter_exit(trans, &copy);
 341         return ret;
 342 }
 343
 344 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 345                             struct bpos pos, bool key_cache)
 346 {
 347         struct btree_path *path;
 348         unsigned idx;
 349         struct printbuf buf = PRINTBUF;
 350
 351         btree_trans_sort_paths(trans);
 352
 353         trans_for_each_path_inorder(trans, path, idx) {
 354                 int cmp = cmp_int(path->btree_id, id) ?:
 355                         cmp_int(path->cached, key_cache);
 356
 357                 if (cmp > 0)
 358                         break;
 359                 if (cmp < 0)
 360                         continue;
 361
 362                 if (!btree_node_locked(path, 0) ||
 363                     !path->should_be_locked)
 364                         continue;
 365
 366                 if (!key_cache) {
 367                         if (bkey_ge(pos, path->l[0].b->data->min_key) &&
 368                             bkey_le(pos, path->l[0].b->key.k.p))
 369                                 return;
 370                 } else {
 371                         if (bkey_eq(pos, path->pos))
 372                                 return;
 373                 }
 374         }
 375
 376         bch2_dump_trans_paths_updates(trans);
 377         bch2_bpos_to_text(&buf, pos);
 378
 379         panic("not locked: %s %s%s\n",
 380               bch2_btree_ids[id], buf.buf,
 381               key_cache ? " cached" : "");
 382 }
 383
 384 #else
 385
 386 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 387                                                 struct btree_path *path, unsigned l) {}
 388 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 389                                           struct btree_path *path) {}
 390 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 391 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 392 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 393
 394 #endif
 395
 396 /* Btree path: fixups after btree updates */
 397
 398 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 399                                         struct btree *b,
 400                                         struct bset_tree *t,
 401                                         struct bkey_packed *k)
 402 {
 403         struct btree_node_iter_set *set;
 404
 405         btree_node_iter_for_each(iter, set)
 406                 if (set->end == t->end_offset) {
 407                         set->k = __btree_node_key_to_offset(b, k);
 408                         bch2_btree_node_iter_sort(iter, b);
 409                         return;
 410                 }
 411
 412         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 413 }
 414
 415 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 416                                                struct btree *b,
 417                                                struct bkey_packed *where)
 418 {
 419         struct btree_path_level *l = &path->l[b->c.level];
 420
 421         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 422                 return;
 423
 424         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 425                 bch2_btree_node_iter_advance(&l->iter, l->b);
 426 }
 427
 428 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 429                                       struct btree *b,
 430                                       struct bkey_packed *where)
 431 {
 432         struct btree_path *path;
 433
 434         trans_for_each_path_with_node(trans, b, path) {
 435                 __bch2_btree_path_fix_key_modified(path, b, where);
 436                 bch2_btree_path_verify_level(trans, path, b->c.level);
 437         }
 438 }
 439
 440 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 441                                        struct btree *b,
 442                                        struct btree_node_iter *node_iter,
 443                                        struct bset_tree *t,
 444                                        struct bkey_packed *where,
 445                                        unsigned clobber_u64s,
 446                                        unsigned new_u64s)
 447 {
 448         const struct bkey_packed *end = btree_bkey_last(b, t);
 449         struct btree_node_iter_set *set;
 450         unsigned offset = __btree_node_key_to_offset(b, where);
 451         int shift = new_u64s - clobber_u64s;
 452         unsigned old_end = t->end_offset - shift;
 453         unsigned orig_iter_pos = node_iter->data[0].k;
 454         bool iter_current_key_modified =
 455                 orig_iter_pos >= offset &&
 456                 orig_iter_pos <= offset + clobber_u64s;
 457
 458         btree_node_iter_for_each(node_iter, set)
 459                 if (set->end == old_end)
 460                         goto found;
 461
 462         /* didn't find the bset in the iterator - might have to readd it: */
 463         if (new_u64s &&
 464             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 465                 bch2_btree_node_iter_push(node_iter, b, where, end);
 466                 goto fixup_done;
 467         } else {
 468                 /* Iterator is after key that changed */
 469                 return;
 470         }
 471 found:
 472         set->end = t->end_offset;
 473
 474         /* Iterator hasn't gotten to the key that changed yet: */
 475         if (set->k < offset)
 476                 return;
 477
 478         if (new_u64s &&
 479             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 480                 set->k = offset;
 481         } else if (set->k < offset + clobber_u64s) {
 482                 set->k = offset + new_u64s;
 483                 if (set->k == set->end)
 484                         bch2_btree_node_iter_set_drop(node_iter, set);
 485         } else {
 486                 /* Iterator is after key that changed */
 487                 set->k = (int) set->k + shift;
 488                 return;
 489         }
 490
 491         bch2_btree_node_iter_sort(node_iter, b);
 492 fixup_done:
 493         if (node_iter->data[0].k != orig_iter_pos)
 494                 iter_current_key_modified = true;
 495
 496         /*
 497          * When a new key is added, and the node iterator now points to that
 498          * key, the iterator might have skipped past deleted keys that should
 499          * come after the key the iterator now points to. We have to rewind to
 500          * before those deleted keys - otherwise
 501          * bch2_btree_node_iter_prev_all() breaks:
 502          */
 503         if (!bch2_btree_node_iter_end(node_iter) &&
 504             iter_current_key_modified &&
 505             b->c.level) {
 506                 struct bset_tree *t;
 507                 struct bkey_packed *k, *k2, *p;
 508
 509                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 510
 511                 for_each_bset(b, t) {
 512                         bool set_pos = false;
 513
 514                         if (node_iter->data[0].end == t->end_offset)
 515                                 continue;
 516
 517                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 518
 519                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 520                                bkey_iter_cmp(b, k, p) < 0) {
 521                                 k2 = p;
 522                                 set_pos = true;
 523                         }
 524
 525                         if (set_pos)
 526                                 btree_node_iter_set_set_pos(node_iter,
 527                                                             b, t, k2);
 528                 }
 529         }
 530 }
 531
 532 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 533                               struct btree_path *path,
 534                               struct btree *b,
 535                               struct btree_node_iter *node_iter,
 536                               struct bkey_packed *where,
 537                               unsigned clobber_u64s,
 538                               unsigned new_u64s)
 539 {
 540         struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
 541         struct btree_path *linked;
 542
 543         if (node_iter != &path->l[b->c.level].iter) {
 544                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 545                                            where, clobber_u64s, new_u64s);
 546
 547                 if (bch2_debug_check_iterators)
 548                         bch2_btree_node_iter_verify(node_iter, b);
 549         }
 550
 551         trans_for_each_path_with_node(trans, b, linked) {
 552                 __bch2_btree_node_iter_fix(linked, b,
 553                                            &linked->l[b->c.level].iter, t,
 554                                            where, clobber_u64s, new_u64s);
 555                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 556         }
 557 }
 558
 559 /* Btree path level: pointer to a particular btree node and node iter */
 560
 561 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
 562                                                   struct btree_path_level *l,
 563                                                   struct bkey *u,
 564                                                   struct bkey_packed *k)
 565 {
 566         if (unlikely(!k)) {
 567                 /*
 568                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 569                  * a hole
 570                  */
 571                 u->type = KEY_TYPE_deleted;
 572                 return bkey_s_c_null;
 573         }
 574
 575         return bkey_disassemble(l->b, k, u);
 576 }
 577
 578 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
 579                                                         struct btree_path_level *l,
 580                                                         struct bkey *u)
 581 {
 582         return __btree_iter_unpack(c, l, u,
 583                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 584 }
 585
 586 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
 587                                                     struct btree_path *path,
 588                                                     struct btree_path_level *l,
 589                                                     struct bkey *u)
 590 {
 591         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 592                         bch2_btree_node_iter_peek(&l->iter, l->b));
 593
 594         path->pos = k.k ? k.k->p : l->b->key.k.p;
 595         trans->paths_sorted = false;
 596         bch2_btree_path_verify_level(trans, path, l - path->l);
 597         return k;
 598 }
 599
 600 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
 601                                                     struct btree_path *path,
 602                                                     struct btree_path_level *l,
 603                                                     struct bkey *u)
 604 {
 605         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 606                         bch2_btree_node_iter_prev(&l->iter, l->b));
 607
 608         path->pos = k.k ? k.k->p : l->b->data->min_key;
 609         trans->paths_sorted = false;
 610         bch2_btree_path_verify_level(trans, path, l - path->l);
 611         return k;
 612 }
 613
 614 static inline bool btree_path_advance_to_pos(struct btree_path *path,
 615                                              struct btree_path_level *l,
 616                                              int max_advance)
 617 {
 618         struct bkey_packed *k;
 619         int nr_advanced = 0;
 620
 621         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 622                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 623                 if (max_advance > 0 && nr_advanced >= max_advance)
 624                         return false;
 625
 626                 bch2_btree_node_iter_advance(&l->iter, l->b);
 627                 nr_advanced++;
 628         }
 629
 630         return true;
 631 }
 632
 633 static inline void __btree_path_level_init(struct btree_path *path,
 634                                            unsigned level)
 635 {
 636         struct btree_path_level *l = &path->l[level];
 637
 638         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
 639
 640         /*
 641          * Iterators to interior nodes should always be pointed at the first non
 642          * whiteout:
 643          */
 644         if (level)
 645                 bch2_btree_node_iter_peek(&l->iter, l->b);
 646 }
 647
 648 void bch2_btree_path_level_init(struct btree_trans *trans,
 649                                 struct btree_path *path,
 650                                 struct btree *b)
 651 {
 652         BUG_ON(path->cached);
 653
 654         EBUG_ON(!btree_path_pos_in_node(path, b));
 655         EBUG_ON(b->c.lock.state.seq & 1);
 656
 657         path->l[b->c.level].lock_seq = b->c.lock.state.seq;
 658         path->l[b->c.level].b = b;
 659         __btree_path_level_init(path, b->c.level);
 660 }
 661
 662 /* Btree path: fixups after btree node updates: */
 663
 664 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
 665 {
 666         struct bch_fs *c = trans->c;
 667         struct btree_insert_entry *i;
 668
 669         trans_for_each_update(trans, i)
 670                 if (!i->cached &&
 671                     i->level    == b->c.level &&
 672                     i->btree_id == b->c.btree_id &&
 673                     bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
 674                     bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
 675                         i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 676
 677                         if (unlikely(trans->journal_replay_not_finished)) {
 678                                 struct bkey_i *j_k =
 679                                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 680                                                                     i->k->k.p);
 681
 682                                 if (j_k) {
 683                                         i->old_k = j_k->k;
 684                                         i->old_v = &j_k->v;
 685                                 }
 686                         }
 687                 }
 688 }
 689
 690 /*
 691  * A btree node is being replaced - update the iterator to point to the new
 692  * node:
 693  */
 694 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
 695 {
 696         struct btree_path *path;
 697
 698         trans_for_each_path(trans, path)
 699                 if (path->uptodate == BTREE_ITER_UPTODATE &&
 700                     !path->cached &&
 701                     btree_path_pos_in_node(path, b)) {
 702                         enum btree_node_locked_type t =
 703                                 btree_lock_want(path, b->c.level);
 704
 705                         if (t != BTREE_NODE_UNLOCKED) {
 706                                 btree_node_unlock(trans, path, b->c.level);
 707                                 six_lock_increment(&b->c.lock, t);
 708                                 mark_btree_node_locked(trans, path, b->c.level, t);
 709                         }
 710
 711                         bch2_btree_path_level_init(trans, path, b);
 712                 }
 713
 714         bch2_trans_revalidate_updates_in_node(trans, b);
 715 }
 716
 717 /*
 718  * A btree node has been modified in such a way as to invalidate iterators - fix
 719  * them:
 720  */
 721 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
 722 {
 723         struct btree_path *path;
 724
 725         trans_for_each_path_with_node(trans, b, path)
 726                 __btree_path_level_init(path, b->c.level);
 727
 728         bch2_trans_revalidate_updates_in_node(trans, b);
 729 }
 730
 731 /* Btree path: traverse, set_pos: */
 732
 733 static inline int btree_path_lock_root(struct btree_trans *trans,
 734                                        struct btree_path *path,
 735                                        unsigned depth_want,
 736                                        unsigned long trace_ip)
 737 {
 738         struct bch_fs *c = trans->c;
 739         struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
 740         enum six_lock_type lock_type;
 741         unsigned i;
 742         int ret;
 743
 744         EBUG_ON(path->nodes_locked);
 745
 746         while (1) {
 747                 b = READ_ONCE(*rootp);
 748                 path->level = READ_ONCE(b->c.level);
 749
 750                 if (unlikely(path->level < depth_want)) {
 751                         /*
 752                          * the root is at a lower depth than the depth we want:
 753                          * got to the end of the btree, or we're walking nodes
 754                          * greater than some depth and there are no nodes >=
 755                          * that depth
 756                          */
 757                         path->level = depth_want;
 758                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
 759                                 path->l[i].b = NULL;
 760                         return 1;
 761                 }
 762
 763                 lock_type = __btree_lock_want(path, path->level);
 764                 ret = btree_node_lock(trans, path, &b->c,
 765                                       path->level, lock_type, trace_ip);
 766                 if (unlikely(ret)) {
 767                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
 768                                 continue;
 769                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 770                                 return ret;
 771                         BUG();
 772                 }
 773
 774                 if (likely(b == READ_ONCE(*rootp) &&
 775                            b->c.level == path->level &&
 776                            !race_fault())) {
 777                         for (i = 0; i < path->level; i++)
 778                                 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
 779                         path->l[path->level].b = b;
 780                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
 781                                 path->l[i].b = NULL;
 782
 783                         mark_btree_node_locked(trans, path, path->level, lock_type);
 784                         bch2_btree_path_level_init(trans, path, b);
 785                         return 0;
 786                 }
 787
 788                 six_unlock_type(&b->c.lock, lock_type);
 789         }
 790 }
 791
 792 noinline
 793 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
 794 {
 795         struct bch_fs *c = trans->c;
 796         struct btree_path_level *l = path_l(path);
 797         struct btree_node_iter node_iter = l->iter;
 798         struct bkey_packed *k;
 799         struct bkey_buf tmp;
 800         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 801                 ? (path->level > 1 ? 0 :  2)
 802                 : (path->level > 1 ? 1 : 16);
 803         bool was_locked = btree_node_locked(path, path->level);
 804         int ret = 0;
 805
 806         bch2_bkey_buf_init(&tmp);
 807
 808         while (nr-- && !ret) {
 809                 if (!bch2_btree_node_relock(trans, path, path->level))
 810                         break;
 811
 812                 bch2_btree_node_iter_advance(&node_iter, l->b);
 813                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
 814                 if (!k)
 815                         break;
 816
 817                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
 818                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
 819                                                path->level - 1);
 820         }
 821
 822         if (!was_locked)
 823                 btree_node_unlock(trans, path, path->level);
 824
 825         bch2_bkey_buf_exit(&tmp, c);
 826         return ret;
 827 }
 828
 829 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
 830                                  struct btree_and_journal_iter *jiter)
 831 {
 832         struct bch_fs *c = trans->c;
 833         struct bkey_s_c k;
 834         struct bkey_buf tmp;
 835         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 836                 ? (path->level > 1 ? 0 :  2)
 837                 : (path->level > 1 ? 1 : 16);
 838         bool was_locked = btree_node_locked(path, path->level);
 839         int ret = 0;
 840
 841         bch2_bkey_buf_init(&tmp);
 842
 843         while (nr-- && !ret) {
 844                 if (!bch2_btree_node_relock(trans, path, path->level))
 845                         break;
 846
 847                 bch2_btree_and_journal_iter_advance(jiter);
 848                 k = bch2_btree_and_journal_iter_peek(jiter);
 849                 if (!k.k)
 850                         break;
 851
 852                 bch2_bkey_buf_reassemble(&tmp, c, k);
 853                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
 854                                                path->level - 1);
 855         }
 856
 857         if (!was_locked)
 858                 btree_node_unlock(trans, path, path->level);
 859
 860         bch2_bkey_buf_exit(&tmp, c);
 861         return ret;
 862 }
 863
 864 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
 865                                             struct btree_path *path,
 866                                             unsigned plevel, struct btree *b)
 867 {
 868         struct btree_path_level *l = &path->l[plevel];
 869         bool locked = btree_node_locked(path, plevel);
 870         struct bkey_packed *k;
 871         struct bch_btree_ptr_v2 *bp;
 872
 873         if (!bch2_btree_node_relock(trans, path, plevel))
 874                 return;
 875
 876         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 877         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
 878
 879         bp = (void *) bkeyp_val(&l->b->format, k);
 880         bp->mem_ptr = (unsigned long)b;
 881
 882         if (!locked)
 883                 btree_node_unlock(trans, path, plevel);
 884 }
 885
 886 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
 887                                                      struct btree_path *path,
 888                                                      unsigned flags,
 889                                                      struct bkey_buf *out)
 890 {
 891         struct bch_fs *c = trans->c;
 892         struct btree_path_level *l = path_l(path);
 893         struct btree_and_journal_iter jiter;
 894         struct bkey_s_c k;
 895         int ret = 0;
 896
 897         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
 898
 899         k = bch2_btree_and_journal_iter_peek(&jiter);
 900
 901         bch2_bkey_buf_reassemble(out, c, k);
 902
 903         if (flags & BTREE_ITER_PREFETCH)
 904                 ret = btree_path_prefetch_j(trans, path, &jiter);
 905
 906         bch2_btree_and_journal_iter_exit(&jiter);
 907         return ret;
 908 }
 909
 910 static __always_inline int btree_path_down(struct btree_trans *trans,
 911                                            struct btree_path *path,
 912                                            unsigned flags,
 913                                            unsigned long trace_ip)
 914 {
 915         struct bch_fs *c = trans->c;
 916         struct btree_path_level *l = path_l(path);
 917         struct btree *b;
 918         unsigned level = path->level - 1;
 919         enum six_lock_type lock_type = __btree_lock_want(path, level);
 920         struct bkey_buf tmp;
 921         int ret;
 922
 923         EBUG_ON(!btree_node_locked(path, path->level));
 924
 925         bch2_bkey_buf_init(&tmp);
 926
 927         if (unlikely(trans->journal_replay_not_finished)) {
 928                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
 929                 if (ret)
 930                         goto err;
 931         } else {
 932                 bch2_bkey_buf_unpack(&tmp, c, l->b,
 933                                  bch2_btree_node_iter_peek(&l->iter, l->b));
 934
 935                 if (flags & BTREE_ITER_PREFETCH) {
 936                         ret = btree_path_prefetch(trans, path);
 937                         if (ret)
 938                                 goto err;
 939                 }
 940         }
 941
 942         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
 943         ret = PTR_ERR_OR_ZERO(b);
 944         if (unlikely(ret))
 945                 goto err;
 946
 947         if (likely(!trans->journal_replay_not_finished &&
 948                    tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
 949             unlikely(b != btree_node_mem_ptr(tmp.k)))
 950                 btree_node_mem_ptr_set(trans, path, level + 1, b);
 951
 952         if (btree_node_read_locked(path, level + 1))
 953                 btree_node_unlock(trans, path, level + 1);
 954
 955         mark_btree_node_locked(trans, path, level, lock_type);
 956         path->level = level;
 957         bch2_btree_path_level_init(trans, path, b);
 958
 959         bch2_btree_path_verify_locks(path);
 960 err:
 961         bch2_bkey_buf_exit(&tmp, c);
 962         return ret;
 963 }
 964
 965
 966 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
 967 {
 968         struct bch_fs *c = trans->c;
 969         struct btree_path *path;
 970         unsigned long trace_ip = _RET_IP_;
 971         int i, ret = 0;
 972
 973         if (trans->in_traverse_all)
 974                 return -BCH_ERR_transaction_restart_in_traverse_all;
 975
 976         trans->in_traverse_all = true;
 977 retry_all:
 978         trans->restarted = 0;
 979
 980         trans_for_each_path(trans, path)
 981                 path->should_be_locked = false;
 982
 983         btree_trans_sort_paths(trans);
 984
 985         bch2_trans_unlock(trans);
 986         cond_resched();
 987
 988         if (unlikely(trans->memory_allocation_failure)) {
 989                 struct closure cl;
 990
 991                 closure_init_stack(&cl);
 992
 993                 do {
 994                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 995                         closure_sync(&cl);
 996                 } while (ret);
 997         }
 998
 999         /* Now, redo traversals in correct order: */
1000         i = 0;
1001         while (i < trans->nr_sorted) {
1002                 path = trans->paths + trans->sorted[i];
1003
1004                 /*
1005                  * Traversing a path can cause another path to be added at about
1006                  * the same position:
1007                  */
1008                 if (path->uptodate) {
1009                         __btree_path_get(path, false);
1010                         ret = bch2_btree_path_traverse_one(trans, path, 0, _THIS_IP_);
1011                         __btree_path_put(path, false);
1012
1013                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1014                             ret == -ENOMEM)
1015                                 goto retry_all;
1016                         if (ret)
1017                                 goto err;
1018                 } else {
1019                         i++;
1020                 }
1021         }
1022
1023         /*
1024          * We used to assert that all paths had been traversed here
1025          * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1026          * path->Should_be_locked is not set yet, we we might have unlocked and
1027          * then failed to relock a path - that's fine.
1028          */
1029 err:
1030         bch2_btree_cache_cannibalize_unlock(c);
1031
1032         trans->in_traverse_all = false;
1033
1034         trace_and_count(c, trans_traverse_all, trans, trace_ip);
1035         return ret;
1036 }
1037
1038 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1039                                                 unsigned l, int check_pos)
1040 {
1041         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1042                 return false;
1043         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1044                 return false;
1045         return true;
1046 }
1047
1048 static inline bool btree_path_good_node(struct btree_trans *trans,
1049                                         struct btree_path *path,
1050                                         unsigned l, int check_pos)
1051 {
1052         return is_btree_node(path, l) &&
1053                 bch2_btree_node_relock(trans, path, l) &&
1054                 btree_path_check_pos_in_node(path, l, check_pos);
1055 }
1056
1057 static void btree_path_set_level_down(struct btree_trans *trans,
1058                                       struct btree_path *path,
1059                                       unsigned new_level)
1060 {
1061         unsigned l;
1062
1063         path->level = new_level;
1064
1065         for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1066                 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1067                         btree_node_unlock(trans, path, l);
1068
1069         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1070         bch2_btree_path_verify(trans, path);
1071 }
1072
1073 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1074                                                          struct btree_path *path,
1075                                                          int check_pos)
1076 {
1077         unsigned i, l = path->level;
1078 again:
1079         while (btree_path_node(path, l) &&
1080                !btree_path_good_node(trans, path, l, check_pos))
1081                 __btree_path_set_level_up(trans, path, l++);
1082
1083         /* If we need intent locks, take them too: */
1084         for (i = l + 1;
1085              i < path->locks_want && btree_path_node(path, i);
1086              i++)
1087                 if (!bch2_btree_node_relock(trans, path, i)) {
1088                         while (l <= i)
1089                                 __btree_path_set_level_up(trans, path, l++);
1090                         goto again;
1091                 }
1092
1093         return l;
1094 }
1095
1096 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1097                                                      struct btree_path *path,
1098                                                      int check_pos)
1099 {
1100         return likely(btree_node_locked(path, path->level) &&
1101                       btree_path_check_pos_in_node(path, path->level, check_pos))
1102                 ? path->level
1103                 : __btree_path_up_until_good_node(trans, path, check_pos);
1104 }
1105
1106 /*
1107  * This is the main state machine for walking down the btree - walks down to a
1108  * specified depth
1109  *
1110  * Returns 0 on success, -EIO on error (error reading in a btree node).
1111  *
1112  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1113  * stashed in the iterator and returned from bch2_trans_exit().
1114  */
1115 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1116                                  struct btree_path *path,
1117                                  unsigned flags,
1118                                  unsigned long trace_ip)
1119 {
1120         unsigned depth_want = path->level;
1121         int ret = -((int) trans->restarted);
1122
1123         if (unlikely(ret))
1124                 goto out;
1125
1126         /*
1127          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1128          * and re-traverse the path without a transaction restart:
1129          */
1130         if (path->should_be_locked) {
1131                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1132                 goto out;
1133         }
1134
1135         if (path->cached) {
1136                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1137                 goto out;
1138         }
1139
1140         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1141                 goto out;
1142
1143         path->level = btree_path_up_until_good_node(trans, path, 0);
1144
1145         EBUG_ON(btree_path_node(path, path->level) &&
1146                 !btree_node_locked(path, path->level));
1147
1148         /*
1149          * Note: path->nodes[path->level] may be temporarily NULL here - that
1150          * would indicate to other code that we got to the end of the btree,
1151          * here it indicates that relocking the root failed - it's critical that
1152          * btree_path_lock_root() comes next and that it can't fail
1153          */
1154         while (path->level > depth_want) {
1155                 ret = btree_path_node(path, path->level)
1156                         ? btree_path_down(trans, path, flags, trace_ip)
1157                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1158                 if (unlikely(ret)) {
1159                         if (ret == 1) {
1160                                 /*
1161                                  * No nodes at this level - got to the end of
1162                                  * the btree:
1163                                  */
1164                                 ret = 0;
1165                                 goto out;
1166                         }
1167
1168                         __bch2_btree_path_unlock(trans, path);
1169                         path->level = depth_want;
1170                         path->l[path->level].b = ERR_PTR(ret);
1171                         goto out;
1172                 }
1173         }
1174
1175         path->uptodate = BTREE_ITER_UPTODATE;
1176 out:
1177         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1178                 panic("ret %s (%i) trans->restarted %s (%i)\n",
1179                       bch2_err_str(ret), ret,
1180                       bch2_err_str(trans->restarted), trans->restarted);
1181         bch2_btree_path_verify(trans, path);
1182         return ret;
1183 }
1184
1185 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1186                             struct btree_path *src)
1187 {
1188         unsigned i, offset = offsetof(struct btree_path, pos);
1189
1190         memcpy((void *) dst + offset,
1191                (void *) src + offset,
1192                sizeof(struct btree_path) - offset);
1193
1194         for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1195                 unsigned t = btree_node_locked_type(dst, i);
1196
1197                 if (t != BTREE_NODE_UNLOCKED)
1198                         six_lock_increment(&dst->l[i].b->c.lock, t);
1199         }
1200 }
1201
1202 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1203                                            bool intent)
1204 {
1205         struct btree_path *new = btree_path_alloc(trans, src);
1206
1207         btree_path_copy(trans, new, src);
1208         __btree_path_get(new, intent);
1209         return new;
1210 }
1211
1212 __flatten
1213 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *trans,
1214                          struct btree_path *path, bool intent,
1215                          unsigned long ip)
1216 {
1217         __btree_path_put(path, intent);
1218         path = btree_path_clone(trans, path, intent);
1219         path->preserve = false;
1220         return path;
1221 }
1222
1223 struct btree_path * __must_check
1224 __bch2_btree_path_set_pos(struct btree_trans *trans,
1225                    struct btree_path *path, struct bpos new_pos,
1226                    bool intent, unsigned long ip, int cmp)
1227 {
1228         unsigned level = path->level;
1229
1230         bch2_trans_verify_not_in_restart(trans);
1231         EBUG_ON(!path->ref);
1232
1233         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1234
1235         path->pos               = new_pos;
1236         trans->paths_sorted     = false;
1237
1238         if (unlikely(path->cached)) {
1239                 btree_node_unlock(trans, path, 0);
1240                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1241                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1242                 goto out;
1243         }
1244
1245         level = btree_path_up_until_good_node(trans, path, cmp);
1246
1247         if (btree_path_node(path, level)) {
1248                 struct btree_path_level *l = &path->l[level];
1249
1250                 BUG_ON(!btree_node_locked(path, level));
1251                 /*
1252                  * We might have to skip over many keys, or just a few: try
1253                  * advancing the node iterator, and if we have to skip over too
1254                  * many keys just reinit it (or if we're rewinding, since that
1255                  * is expensive).
1256                  */
1257                 if (cmp < 0 ||
1258                     !btree_path_advance_to_pos(path, l, 8))
1259                         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1260
1261                 /*
1262                  * Iterators to interior nodes should always be pointed at the first non
1263                  * whiteout:
1264                  */
1265                 if (unlikely(level))
1266                         bch2_btree_node_iter_peek(&l->iter, l->b);
1267         }
1268
1269         if (unlikely(level != path->level)) {
1270                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1271                 __bch2_btree_path_unlock(trans, path);
1272         }
1273 out:
1274         bch2_btree_path_verify(trans, path);
1275         return path;
1276 }
1277
1278 /* Btree path: main interface: */
1279
1280 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1281 {
1282         struct btree_path *sib;
1283
1284         sib = prev_btree_path(trans, path);
1285         if (sib && !btree_path_cmp(sib, path))
1286                 return sib;
1287
1288         sib = next_btree_path(trans, path);
1289         if (sib && !btree_path_cmp(sib, path))
1290                 return sib;
1291
1292         return NULL;
1293 }
1294
1295 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1296 {
1297         struct btree_path *sib;
1298
1299         sib = prev_btree_path(trans, path);
1300         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1301                 return sib;
1302
1303         sib = next_btree_path(trans, path);
1304         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1305                 return sib;
1306
1307         return NULL;
1308 }
1309
1310 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1311 {
1312         __bch2_btree_path_unlock(trans, path);
1313         btree_path_list_remove(trans, path);
1314         trans->paths_allocated &= ~(1ULL << path->idx);
1315 }
1316
1317 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1318 {
1319         struct btree_path *dup;
1320
1321         EBUG_ON(trans->paths + path->idx != path);
1322         EBUG_ON(!path->ref);
1323
1324         if (!__btree_path_put(path, intent))
1325                 return;
1326
1327         dup = path->preserve
1328                 ? have_path_at_pos(trans, path)
1329                 : have_node_at_pos(trans, path);
1330
1331         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1332                 return;
1333
1334         if (path->should_be_locked &&
1335             !trans->restarted &&
1336             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1337                 return;
1338
1339         if (dup) {
1340                 dup->preserve           |= path->preserve;
1341                 dup->should_be_locked   |= path->should_be_locked;
1342         }
1343
1344         __bch2_path_free(trans, path);
1345 }
1346
1347 static void bch2_path_put_nokeep(struct btree_trans *trans, struct btree_path *path,
1348                                  bool intent)
1349 {
1350         EBUG_ON(trans->paths + path->idx != path);
1351         EBUG_ON(!path->ref);
1352
1353         if (!__btree_path_put(path, intent))
1354                 return;
1355
1356         __bch2_path_free(trans, path);
1357 }
1358
1359 void bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1360 {
1361         panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1362               trans->restart_count, restart_count,
1363               (void *) trans->last_restarted_ip);
1364 }
1365
1366 void bch2_trans_in_restart_error(struct btree_trans *trans)
1367 {
1368         panic("in transaction restart: %s, last restarted by %pS\n",
1369               bch2_err_str(trans->restarted),
1370               (void *) trans->last_restarted_ip);
1371 }
1372
1373 noinline __cold
1374 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1375 {
1376         struct btree_insert_entry *i;
1377
1378         prt_printf(buf, "transaction updates for %s journal seq %llu",
1379                trans->fn, trans->journal_res.seq);
1380         prt_newline(buf);
1381         printbuf_indent_add(buf, 2);
1382
1383         trans_for_each_update(trans, i) {
1384                 struct bkey_s_c old = { &i->old_k, i->old_v };
1385
1386                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1387                        bch2_btree_ids[i->btree_id],
1388                        i->cached,
1389                        (void *) i->ip_allocated);
1390                 prt_newline(buf);
1391
1392                 prt_printf(buf, "  old ");
1393                 bch2_bkey_val_to_text(buf, trans->c, old);
1394                 prt_newline(buf);
1395
1396                 prt_printf(buf, "  new ");
1397                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1398                 prt_newline(buf);
1399         }
1400
1401         printbuf_indent_sub(buf, 2);
1402 }
1403
1404 noinline __cold
1405 void bch2_dump_trans_updates(struct btree_trans *trans)
1406 {
1407         struct printbuf buf = PRINTBUF;
1408
1409         bch2_trans_updates_to_text(&buf, trans);
1410         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1411         printbuf_exit(&buf);
1412 }
1413
1414 noinline __cold
1415 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1416 {
1417         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1418                    path->idx, path->ref, path->intent_ref,
1419                    path->preserve ? 'P' : ' ',
1420                    path->should_be_locked ? 'S' : ' ',
1421                    bch2_btree_ids[path->btree_id],
1422                    path->level);
1423         bch2_bpos_to_text(out, path->pos);
1424
1425         prt_printf(out, " locks %u", path->nodes_locked);
1426 #ifdef TRACK_PATH_ALLOCATED
1427         prt_printf(out, " %pS", (void *) path->ip_allocated);
1428 #endif
1429         prt_newline(out);
1430 }
1431
1432 noinline __cold
1433 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1434                                 bool nosort)
1435 {
1436         struct btree_path *path;
1437         unsigned idx;
1438
1439         if (!nosort)
1440                 btree_trans_sort_paths(trans);
1441
1442         trans_for_each_path_inorder(trans, path, idx)
1443                 bch2_btree_path_to_text(out, path);
1444 }
1445
1446 noinline __cold
1447 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1448 {
1449         __bch2_trans_paths_to_text(out, trans, false);
1450 }
1451
1452 noinline __cold
1453 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1454 {
1455         struct printbuf buf = PRINTBUF;
1456
1457         __bch2_trans_paths_to_text(&buf, trans, nosort);
1458         bch2_trans_updates_to_text(&buf, trans);
1459
1460         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1461         printbuf_exit(&buf);
1462 }
1463
1464 noinline __cold
1465 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1466 {
1467         __bch2_dump_trans_paths_updates(trans, false);
1468 }
1469
1470 noinline __cold
1471 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1472 {
1473         struct btree_transaction_stats *s = btree_trans_stats(trans);
1474         struct printbuf buf = PRINTBUF;
1475
1476         if (!s)
1477                 return;
1478
1479         bch2_trans_paths_to_text(&buf, trans);
1480
1481         if (!buf.allocation_failure) {
1482                 mutex_lock(&s->lock);
1483                 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1484                         s->nr_max_paths = trans->nr_max_paths =
1485                                 hweight64(trans->paths_allocated);
1486                         swap(s->max_paths_text, buf.buf);
1487                 }
1488                 mutex_unlock(&s->lock);
1489         }
1490
1491         printbuf_exit(&buf);
1492
1493         trans->nr_max_paths = hweight64(trans->paths_allocated);
1494 }
1495
1496 static noinline void btree_path_overflow(struct btree_trans *trans)
1497 {
1498         bch2_dump_trans_paths_updates(trans);
1499         panic("trans path oveflow\n");
1500 }
1501
1502 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1503                                                   struct btree_path *pos)
1504 {
1505         struct btree_path *path;
1506         unsigned idx;
1507
1508         if (unlikely(trans->paths_allocated ==
1509                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
1510                 btree_path_overflow(trans);
1511
1512         idx = __ffs64(~trans->paths_allocated);
1513
1514         /*
1515          * Do this before marking the new path as allocated, since it won't be
1516          * initialized yet:
1517          */
1518         if (unlikely(idx > trans->nr_max_paths))
1519                 bch2_trans_update_max_paths(trans);
1520
1521         trans->paths_allocated |= 1ULL << idx;
1522
1523         path = &trans->paths[idx];
1524         path->idx               = idx;
1525         path->ref               = 0;
1526         path->intent_ref        = 0;
1527         path->nodes_locked      = 0;
1528
1529         btree_path_list_add(trans, pos, path);
1530         trans->paths_sorted = false;
1531         return path;
1532 }
1533
1534 struct btree_path *bch2_path_get(struct btree_trans *trans,
1535                                  enum btree_id btree_id, struct bpos pos,
1536                                  unsigned locks_want, unsigned level,
1537                                  unsigned flags, unsigned long ip)
1538 {
1539         struct btree_path *path, *path_pos = NULL;
1540         bool cached = flags & BTREE_ITER_CACHED;
1541         bool intent = flags & BTREE_ITER_INTENT;
1542         int i;
1543
1544         bch2_trans_verify_not_in_restart(trans);
1545         bch2_trans_verify_locks(trans);
1546
1547         btree_trans_sort_paths(trans);
1548
1549         trans_for_each_path_inorder(trans, path, i) {
1550                 if (__btree_path_cmp(path,
1551                                      btree_id,
1552                                      cached,
1553                                      pos,
1554                                      level) > 0)
1555                         break;
1556
1557                 path_pos = path;
1558         }
1559
1560         if (path_pos &&
1561             path_pos->cached    == cached &&
1562             path_pos->btree_id  == btree_id &&
1563             path_pos->level     == level) {
1564                 __btree_path_get(path_pos, intent);
1565                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1566         } else {
1567                 path = btree_path_alloc(trans, path_pos);
1568                 path_pos = NULL;
1569
1570                 __btree_path_get(path, intent);
1571                 path->pos                       = pos;
1572                 path->btree_id                  = btree_id;
1573                 path->cached                    = cached;
1574                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1575                 path->should_be_locked          = false;
1576                 path->level                     = level;
1577                 path->locks_want                = locks_want;
1578                 path->nodes_locked              = 0;
1579                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1580                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1581 #ifdef TRACK_PATH_ALLOCATED
1582                 path->ip_allocated              = ip;
1583 #endif
1584                 trans->paths_sorted             = false;
1585         }
1586
1587         if (!(flags & BTREE_ITER_NOPRESERVE))
1588                 path->preserve = true;
1589
1590         if (path->intent_ref)
1591                 locks_want = max(locks_want, level + 1);
1592
1593         /*
1594          * If the path has locks_want greater than requested, we don't downgrade
1595          * it here - on transaction restart because btree node split needs to
1596          * upgrade locks, we might be putting/getting the iterator again.
1597          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1598          * a successful transaction commit.
1599          */
1600
1601         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1602         if (locks_want > path->locks_want)
1603                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want);
1604
1605         return path;
1606 }
1607
1608 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1609 {
1610
1611         struct btree_path_level *l = path_l(path);
1612         struct bkey_packed *_k;
1613         struct bkey_s_c k;
1614
1615         if (unlikely(!l->b))
1616                 return bkey_s_c_null;
1617
1618         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1619         EBUG_ON(!btree_node_locked(path, path->level));
1620
1621         if (!path->cached) {
1622                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1623                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1624
1625                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1626
1627                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1628                         goto hole;
1629         } else {
1630                 struct bkey_cached *ck = (void *) path->l[0].b;
1631
1632                 EBUG_ON(ck &&
1633                         (path->btree_id != ck->key.btree_id ||
1634                          !bkey_eq(path->pos, ck->key.pos)));
1635                 if (!ck || !ck->valid)
1636                         return bkey_s_c_null;
1637
1638                 *u = ck->k->k;
1639                 k = bkey_i_to_s_c(ck->k);
1640         }
1641
1642         return k;
1643 hole:
1644         bkey_init(u);
1645         u->p = path->pos;
1646         return (struct bkey_s_c) { u, NULL };
1647 }
1648
1649 /* Btree iterators: */
1650
1651 int __must_check
1652 __bch2_btree_iter_traverse(struct btree_iter *iter)
1653 {
1654         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1655 }
1656
1657 int __must_check
1658 bch2_btree_iter_traverse(struct btree_iter *iter)
1659 {
1660         int ret;
1661
1662         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1663                                         btree_iter_search_key(iter),
1664                                         iter->flags & BTREE_ITER_INTENT,
1665                                         btree_iter_ip_allocated(iter));
1666
1667         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1668         if (ret)
1669                 return ret;
1670
1671         btree_path_set_should_be_locked(iter->path);
1672         return 0;
1673 }
1674
1675 /* Iterate across nodes (leaf and interior nodes) */
1676
1677 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1678 {
1679         struct btree_trans *trans = iter->trans;
1680         struct btree *b = NULL;
1681         int ret;
1682
1683         EBUG_ON(iter->path->cached);
1684         bch2_btree_iter_verify(iter);
1685
1686         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1687         if (ret)
1688                 goto err;
1689
1690         b = btree_path_node(iter->path, iter->path->level);
1691         if (!b)
1692                 goto out;
1693
1694         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1695
1696         bkey_init(&iter->k);
1697         iter->k.p = iter->pos = b->key.k.p;
1698
1699         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1700                                         iter->flags & BTREE_ITER_INTENT,
1701                                         btree_iter_ip_allocated(iter));
1702         btree_path_set_should_be_locked(iter->path);
1703 out:
1704         bch2_btree_iter_verify_entry_exit(iter);
1705         bch2_btree_iter_verify(iter);
1706
1707         return b;
1708 err:
1709         b = ERR_PTR(ret);
1710         goto out;
1711 }
1712
1713 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1714 {
1715         struct btree_trans *trans = iter->trans;
1716         struct btree_path *path = iter->path;
1717         struct btree *b = NULL;
1718         int ret;
1719
1720         bch2_trans_verify_not_in_restart(trans);
1721         EBUG_ON(iter->path->cached);
1722         bch2_btree_iter_verify(iter);
1723
1724         /* already at end? */
1725         if (!btree_path_node(path, path->level))
1726                 return NULL;
1727
1728         /* got to end? */
1729         if (!btree_path_node(path, path->level + 1)) {
1730                 btree_path_set_level_up(trans, path);
1731                 return NULL;
1732         }
1733
1734         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1735                 __bch2_btree_path_unlock(trans, path);
1736                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1737                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1738                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1739                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1740                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1741                 goto err;
1742         }
1743
1744         b = btree_path_node(path, path->level + 1);
1745
1746         if (bpos_eq(iter->pos, b->key.k.p)) {
1747                 __btree_path_set_level_up(trans, path, path->level++);
1748         } else {
1749                 /*
1750                  * Haven't gotten to the end of the parent node: go back down to
1751                  * the next child node
1752                  */
1753                 path = iter->path =
1754                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1755                                            iter->flags & BTREE_ITER_INTENT,
1756                                            btree_iter_ip_allocated(iter));
1757
1758                 btree_path_set_level_down(trans, path, iter->min_depth);
1759
1760                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1761                 if (ret)
1762                         goto err;
1763
1764                 b = path->l[path->level].b;
1765         }
1766
1767         bkey_init(&iter->k);
1768         iter->k.p = iter->pos = b->key.k.p;
1769
1770         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1771                                         iter->flags & BTREE_ITER_INTENT,
1772                                         btree_iter_ip_allocated(iter));
1773         btree_path_set_should_be_locked(iter->path);
1774         BUG_ON(iter->path->uptodate);
1775 out:
1776         bch2_btree_iter_verify_entry_exit(iter);
1777         bch2_btree_iter_verify(iter);
1778
1779         return b;
1780 err:
1781         b = ERR_PTR(ret);
1782         goto out;
1783 }
1784
1785 /* Iterate across keys (in leaf nodes only) */
1786
1787 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1788 {
1789         if (likely(!(iter->flags & BTREE_ITER_ALL_LEVELS))) {
1790                 struct bpos pos = iter->k.p;
1791                 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1792                              ? bpos_eq(pos, SPOS_MAX)
1793                              : bkey_eq(pos, SPOS_MAX));
1794
1795                 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1796                         pos = bkey_successor(iter, pos);
1797                 bch2_btree_iter_set_pos(iter, pos);
1798                 return ret;
1799         } else {
1800                 if (!btree_path_node(iter->path, iter->path->level))
1801                         return true;
1802
1803                 iter->advanced = true;
1804                 return false;
1805         }
1806 }
1807
1808 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1809 {
1810         struct bpos pos = bkey_start_pos(&iter->k);
1811         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1812                      ? bpos_eq(pos, POS_MIN)
1813                      : bkey_eq(pos, POS_MIN));
1814
1815         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1816                 pos = bkey_predecessor(iter, pos);
1817         bch2_btree_iter_set_pos(iter, pos);
1818         return ret;
1819 }
1820
1821 static noinline
1822 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1823 {
1824         struct btree_insert_entry *i;
1825         struct bkey_i *ret = NULL;
1826
1827         trans_for_each_update(iter->trans, i) {
1828                 if (i->btree_id < iter->btree_id)
1829                         continue;
1830                 if (i->btree_id > iter->btree_id)
1831                         break;
1832                 if (bpos_lt(i->k->k.p, iter->path->pos))
1833                         continue;
1834                 if (i->key_cache_already_flushed)
1835                         continue;
1836                 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1837                         ret = i->k;
1838         }
1839
1840         return ret;
1841 }
1842
1843 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1844 {
1845         return iter->flags & BTREE_ITER_WITH_UPDATES
1846                 ? __bch2_btree_trans_peek_updates(iter)
1847                 : NULL;
1848 }
1849
1850 struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1851                                        struct btree_iter *iter,
1852                                        struct bpos end_pos)
1853 {
1854         struct bkey_i *k;
1855
1856         if (bpos_lt(iter->path->pos, iter->journal_pos))
1857                 iter->journal_idx = 0;
1858
1859         k = bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1860                                         iter->path->level,
1861                                         iter->path->pos,
1862                                         end_pos,
1863                                         &iter->journal_idx);
1864
1865         iter->journal_pos = k ? k->k.p : end_pos;
1866         return k;
1867 }
1868
1869 static noinline
1870 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1871                                               struct btree_iter *iter)
1872 {
1873         struct bkey_i *k = bch2_btree_journal_peek(trans, iter, iter->path->pos);
1874
1875         if (k) {
1876                 iter->k = k->k;
1877                 return bkey_i_to_s_c(k);
1878         } else {
1879                 return bkey_s_c_null;
1880         }
1881 }
1882
1883 static noinline
1884 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1885                                          struct btree_iter *iter,
1886                                          struct bkey_s_c k)
1887 {
1888         struct bkey_i *next_journal =
1889                 bch2_btree_journal_peek(trans, iter,
1890                                 k.k ? k.k->p : path_l(iter->path)->b->key.k.p);
1891
1892         if (next_journal) {
1893                 iter->k = next_journal->k;
1894                 k = bkey_i_to_s_c(next_journal);
1895         }
1896
1897         return k;
1898 }
1899
1900 /*
1901  * Checks btree key cache for key at iter->pos and returns it if present, or
1902  * bkey_s_c_null:
1903  */
1904 static noinline
1905 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1906 {
1907         struct btree_trans *trans = iter->trans;
1908         struct bch_fs *c = trans->c;
1909         struct bkey u;
1910         struct bkey_s_c k;
1911         int ret;
1912
1913         if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1914             bpos_eq(iter->pos, pos))
1915                 return bkey_s_c_null;
1916
1917         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1918                 return bkey_s_c_null;
1919
1920         if (!iter->key_cache_path)
1921                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1922                                                      iter->flags & BTREE_ITER_INTENT, 0,
1923                                                      iter->flags|BTREE_ITER_CACHED|
1924                                                      BTREE_ITER_CACHED_NOFILL,
1925                                                      _THIS_IP_);
1926
1927         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1928                                         iter->flags & BTREE_ITER_INTENT,
1929                                         btree_iter_ip_allocated(iter));
1930
1931         ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
1932                                          iter->flags|BTREE_ITER_CACHED) ?:
1933                 bch2_btree_path_relock(trans, iter->path, _THIS_IP_);
1934         if (unlikely(ret))
1935                 return bkey_s_c_err(ret);
1936
1937         btree_path_set_should_be_locked(iter->key_cache_path);
1938
1939         k = bch2_btree_path_peek_slot(iter->key_cache_path, &u);
1940         if (k.k && !bkey_err(k)) {
1941                 iter->k = u;
1942                 k.k = &iter->k;
1943         }
1944         return k;
1945 }
1946
1947 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1948 {
1949         struct btree_trans *trans = iter->trans;
1950         struct bkey_i *next_update;
1951         struct bkey_s_c k, k2;
1952         int ret;
1953
1954         EBUG_ON(iter->path->cached);
1955         bch2_btree_iter_verify(iter);
1956
1957         while (1) {
1958                 struct btree_path_level *l;
1959
1960                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
1961                                         iter->flags & BTREE_ITER_INTENT,
1962                                         btree_iter_ip_allocated(iter));
1963
1964                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1965                 if (unlikely(ret)) {
1966                         /* ensure that iter->k is consistent with iter->pos: */
1967                         bch2_btree_iter_set_pos(iter, iter->pos);
1968                         k = bkey_s_c_err(ret);
1969                         goto out;
1970                 }
1971
1972                 l = path_l(iter->path);
1973
1974                 if (unlikely(!l->b)) {
1975                         /* No btree nodes at requested level: */
1976                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
1977                         k = bkey_s_c_null;
1978                         goto out;
1979                 }
1980
1981                 btree_path_set_should_be_locked(iter->path);
1982
1983                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
1984
1985                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
1986                     k.k &&
1987                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
1988                         k = k2;
1989                         ret = bkey_err(k);
1990                         if (ret) {
1991                                 bch2_btree_iter_set_pos(iter, iter->pos);
1992                                 goto out;
1993                         }
1994                 }
1995
1996                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
1997                         k = btree_trans_peek_journal(trans, iter, k);
1998
1999                 next_update = btree_trans_peek_updates(iter);
2000
2001                 if (next_update &&
2002                     bpos_le(next_update->k.p,
2003                             k.k ? k.k->p : l->b->key.k.p)) {
2004                         iter->k = next_update->k;
2005                         k = bkey_i_to_s_c(next_update);
2006                 }
2007
2008                 if (k.k && bkey_deleted(k.k)) {
2009                         /*
2010                          * If we've got a whiteout, and it's after the search
2011                          * key, advance the search key to the whiteout instead
2012                          * of just after the whiteout - it might be a btree
2013                          * whiteout, with a real key at the same position, since
2014                          * in the btree deleted keys sort before non deleted.
2015                          */
2016                         search_key = !bpos_eq(search_key, k.k->p)
2017                                 ? k.k->p
2018                                 : bpos_successor(k.k->p);
2019                         continue;
2020                 }
2021
2022                 if (likely(k.k)) {
2023                         break;
2024                 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2025                         /* Advance to next leaf node: */
2026                         search_key = bpos_successor(l->b->key.k.p);
2027                 } else {
2028                         /* End of btree: */
2029                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2030                         k = bkey_s_c_null;
2031                         goto out;
2032                 }
2033         }
2034 out:
2035         bch2_btree_iter_verify(iter);
2036
2037         return k;
2038 }
2039
2040 /**
2041  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2042  * current position
2043  */
2044 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2045 {
2046         struct btree_trans *trans = iter->trans;
2047         struct bpos search_key = btree_iter_search_key(iter);
2048         struct bkey_s_c k;
2049         struct bpos iter_pos;
2050         int ret;
2051
2052         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2053         EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2054
2055         if (iter->update_path) {
2056                 bch2_path_put_nokeep(trans, iter->update_path,
2057                                      iter->flags & BTREE_ITER_INTENT);
2058                 iter->update_path = NULL;
2059         }
2060
2061         bch2_btree_iter_verify_entry_exit(iter);
2062
2063         while (1) {
2064                 k = __bch2_btree_iter_peek(iter, search_key);
2065                 if (unlikely(!k.k))
2066                         goto end;
2067                 if (unlikely(bkey_err(k)))
2068                         goto out_no_locked;
2069
2070                 /*
2071                  * iter->pos should be mononotically increasing, and always be
2072                  * equal to the key we just returned - except extents can
2073                  * straddle iter->pos:
2074                  */
2075                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2076                         iter_pos = k.k->p;
2077                 else
2078                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2079
2080                 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2081                              ? bkey_gt(iter_pos, end)
2082                              : bkey_ge(iter_pos, end)))
2083                         goto end;
2084
2085                 if (iter->update_path &&
2086                     !bkey_eq(iter->update_path->pos, k.k->p)) {
2087                         bch2_path_put_nokeep(trans, iter->update_path,
2088                                              iter->flags & BTREE_ITER_INTENT);
2089                         iter->update_path = NULL;
2090                 }
2091
2092                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2093                     (iter->flags & BTREE_ITER_INTENT) &&
2094                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2095                     !iter->update_path) {
2096                         struct bpos pos = k.k->p;
2097
2098                         if (pos.snapshot < iter->snapshot) {
2099                                 search_key = bpos_successor(k.k->p);
2100                                 continue;
2101                         }
2102
2103                         pos.snapshot = iter->snapshot;
2104
2105                         /*
2106                          * advance, same as on exit for iter->path, but only up
2107                          * to snapshot
2108                          */
2109                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2110                         iter->update_path = iter->path;
2111
2112                         iter->update_path = bch2_btree_path_set_pos(trans,
2113                                                 iter->update_path, pos,
2114                                                 iter->flags & BTREE_ITER_INTENT,
2115                                                 _THIS_IP_);
2116                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2117                         if (unlikely(ret)) {
2118                                 k = bkey_s_c_err(ret);
2119                                 goto out_no_locked;
2120                         }
2121                 }
2122
2123                 /*
2124                  * We can never have a key in a leaf node at POS_MAX, so
2125                  * we don't have to check these successor() calls:
2126                  */
2127                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2128                     !bch2_snapshot_is_ancestor(trans->c,
2129                                                iter->snapshot,
2130                                                k.k->p.snapshot)) {
2131                         search_key = bpos_successor(k.k->p);
2132                         continue;
2133                 }
2134
2135                 if (bkey_whiteout(k.k) &&
2136                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2137                         search_key = bkey_successor(iter, k.k->p);
2138                         continue;
2139                 }
2140
2141                 break;
2142         }
2143
2144         iter->pos = iter_pos;
2145
2146         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2147                                 iter->flags & BTREE_ITER_INTENT,
2148                                 btree_iter_ip_allocated(iter));
2149
2150         btree_path_set_should_be_locked(iter->path);
2151 out_no_locked:
2152         if (iter->update_path) {
2153                 ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_);
2154                 if (unlikely(ret))
2155                         k = bkey_s_c_err(ret);
2156                 else
2157                         btree_path_set_should_be_locked(iter->update_path);
2158         }
2159
2160         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2161                 iter->pos.snapshot = iter->snapshot;
2162
2163         ret = bch2_btree_iter_verify_ret(iter, k);
2164         if (unlikely(ret)) {
2165                 bch2_btree_iter_set_pos(iter, iter->pos);
2166                 k = bkey_s_c_err(ret);
2167         }
2168
2169         bch2_btree_iter_verify_entry_exit(iter);
2170
2171         return k;
2172 end:
2173         bch2_btree_iter_set_pos(iter, end);
2174         k = bkey_s_c_null;
2175         goto out_no_locked;
2176 }
2177
2178 /**
2179  * bch2_btree_iter_peek_all_levels: returns the first key greater than or equal
2180  * to iterator's current position, returning keys from every level of the btree.
2181  * For keys at different levels of the btree that compare equal, the key from
2182  * the lower level (leaf) is returned first.
2183  */
2184 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *iter)
2185 {
2186         struct btree_trans *trans = iter->trans;
2187         struct bkey_s_c k;
2188         int ret;
2189
2190         EBUG_ON(iter->path->cached);
2191         bch2_btree_iter_verify(iter);
2192         BUG_ON(iter->path->level < iter->min_depth);
2193         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
2194         EBUG_ON(!(iter->flags & BTREE_ITER_ALL_LEVELS));
2195
2196         while (1) {
2197                 iter->path = bch2_btree_path_set_pos(trans, iter->path, iter->pos,
2198                                         iter->flags & BTREE_ITER_INTENT,
2199                                         btree_iter_ip_allocated(iter));
2200
2201                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2202                 if (unlikely(ret)) {
2203                         /* ensure that iter->k is consistent with iter->pos: */
2204                         bch2_btree_iter_set_pos(iter, iter->pos);
2205                         k = bkey_s_c_err(ret);
2206                         goto out_no_locked;
2207                 }
2208
2209                 /* Already at end? */
2210                 if (!btree_path_node(iter->path, iter->path->level)) {
2211                         k = bkey_s_c_null;
2212                         goto out_no_locked;
2213                 }
2214
2215                 k = btree_path_level_peek_all(trans->c,
2216                                 &iter->path->l[iter->path->level], &iter->k);
2217
2218                 /* Check if we should go up to the parent node: */
2219                 if (!k.k ||
2220                     (iter->advanced &&
2221                      bpos_eq(path_l(iter->path)->b->key.k.p, iter->pos))) {
2222                         iter->pos = path_l(iter->path)->b->key.k.p;
2223                         btree_path_set_level_up(trans, iter->path);
2224                         iter->advanced = false;
2225                         continue;
2226                 }
2227
2228                 /*
2229                  * Check if we should go back down to a leaf:
2230                  * If we're not in a leaf node, we only return the current key
2231                  * if it exactly matches iter->pos - otherwise we first have to
2232                  * go back to the leaf:
2233                  */
2234                 if (iter->path->level != iter->min_depth &&
2235                     (iter->advanced ||
2236                      !k.k ||
2237                      !bpos_eq(iter->pos, k.k->p))) {
2238                         btree_path_set_level_down(trans, iter->path, iter->min_depth);
2239                         iter->pos = bpos_successor(iter->pos);
2240                         iter->advanced = false;
2241                         continue;
2242                 }
2243
2244                 /* Check if we should go to the next key: */
2245                 if (iter->path->level == iter->min_depth &&
2246                     iter->advanced &&
2247                     k.k &&
2248                     bpos_eq(iter->pos, k.k->p)) {
2249                         iter->pos = bpos_successor(iter->pos);
2250                         iter->advanced = false;
2251                         continue;
2252                 }
2253
2254                 if (iter->advanced &&
2255                     iter->path->level == iter->min_depth &&
2256                     !bpos_eq(k.k->p, iter->pos))
2257                         iter->advanced = false;
2258
2259                 BUG_ON(iter->advanced);
2260                 BUG_ON(!k.k);
2261                 break;
2262         }
2263
2264         iter->pos = k.k->p;
2265         btree_path_set_should_be_locked(iter->path);
2266 out_no_locked:
2267         bch2_btree_iter_verify(iter);
2268
2269         return k;
2270 }
2271
2272 /**
2273  * bch2_btree_iter_next: returns first key greater than iterator's current
2274  * position
2275  */
2276 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2277 {
2278         if (!bch2_btree_iter_advance(iter))
2279                 return bkey_s_c_null;
2280
2281         return bch2_btree_iter_peek(iter);
2282 }
2283
2284 /**
2285  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2286  * iterator's current position
2287  */
2288 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2289 {
2290         struct btree_trans *trans = iter->trans;
2291         struct bpos search_key = iter->pos;
2292         struct btree_path *saved_path = NULL;
2293         struct bkey_s_c k;
2294         struct bkey saved_k;
2295         const struct bch_val *saved_v;
2296         int ret;
2297
2298         EBUG_ON(iter->path->cached || iter->path->level);
2299         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2300
2301         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2302                 return bkey_s_c_err(-EIO);
2303
2304         bch2_btree_iter_verify(iter);
2305         bch2_btree_iter_verify_entry_exit(iter);
2306
2307         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2308                 search_key.snapshot = U32_MAX;
2309
2310         while (1) {
2311                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2312                                                 iter->flags & BTREE_ITER_INTENT,
2313                                                 btree_iter_ip_allocated(iter));
2314
2315                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2316                 if (unlikely(ret)) {
2317                         /* ensure that iter->k is consistent with iter->pos: */
2318                         bch2_btree_iter_set_pos(iter, iter->pos);
2319                         k = bkey_s_c_err(ret);
2320                         goto out_no_locked;
2321                 }
2322
2323                 k = btree_path_level_peek(trans, iter->path,
2324                                           &iter->path->l[0], &iter->k);
2325                 if (!k.k ||
2326                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2327                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2328                      : bpos_gt(k.k->p, search_key)))
2329                         k = btree_path_level_prev(trans, iter->path,
2330                                                   &iter->path->l[0], &iter->k);
2331
2332                 if (likely(k.k)) {
2333                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2334                                 if (k.k->p.snapshot == iter->snapshot)
2335                                         goto got_key;
2336
2337                                 /*
2338                                  * If we have a saved candidate, and we're no
2339                                  * longer at the same _key_ (not pos), return
2340                                  * that candidate
2341                                  */
2342                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2343                                         bch2_path_put_nokeep(trans, iter->path,
2344                                                       iter->flags & BTREE_ITER_INTENT);
2345                                         iter->path = saved_path;
2346                                         saved_path = NULL;
2347                                         iter->k = saved_k;
2348                                         k.v     = saved_v;
2349                                         goto got_key;
2350                                 }
2351
2352                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2353                                                               iter->snapshot,
2354                                                               k.k->p.snapshot)) {
2355                                         if (saved_path)
2356                                                 bch2_path_put_nokeep(trans, saved_path,
2357                                                       iter->flags & BTREE_ITER_INTENT);
2358                                         saved_path = btree_path_clone(trans, iter->path,
2359                                                                 iter->flags & BTREE_ITER_INTENT);
2360                                         saved_k = *k.k;
2361                                         saved_v = k.v;
2362                                 }
2363
2364                                 search_key = bpos_predecessor(k.k->p);
2365                                 continue;
2366                         }
2367 got_key:
2368                         if (bkey_whiteout(k.k) &&
2369                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2370                                 search_key = bkey_predecessor(iter, k.k->p);
2371                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2372                                         search_key.snapshot = U32_MAX;
2373                                 continue;
2374                         }
2375
2376                         break;
2377                 } else if (likely(!bpos_eq(iter->path->l[0].b->data->min_key, POS_MIN))) {
2378                         /* Advance to previous leaf node: */
2379                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2380                 } else {
2381                         /* Start of btree: */
2382                         bch2_btree_iter_set_pos(iter, POS_MIN);
2383                         k = bkey_s_c_null;
2384                         goto out_no_locked;
2385                 }
2386         }
2387
2388         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2389
2390         /* Extents can straddle iter->pos: */
2391         if (bkey_lt(k.k->p, iter->pos))
2392                 iter->pos = k.k->p;
2393
2394         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2395                 iter->pos.snapshot = iter->snapshot;
2396
2397         btree_path_set_should_be_locked(iter->path);
2398 out_no_locked:
2399         if (saved_path)
2400                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2401
2402         bch2_btree_iter_verify_entry_exit(iter);
2403         bch2_btree_iter_verify(iter);
2404
2405         return k;
2406 }
2407
2408 /**
2409  * bch2_btree_iter_prev: returns first key less than iterator's current
2410  * position
2411  */
2412 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2413 {
2414         if (!bch2_btree_iter_rewind(iter))
2415                 return bkey_s_c_null;
2416
2417         return bch2_btree_iter_peek_prev(iter);
2418 }
2419
2420 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2421 {
2422         struct btree_trans *trans = iter->trans;
2423         struct bpos search_key;
2424         struct bkey_s_c k;
2425         int ret;
2426
2427         bch2_btree_iter_verify(iter);
2428         bch2_btree_iter_verify_entry_exit(iter);
2429         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2430         EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2431
2432         /* extents can't span inode numbers: */
2433         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2434             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2435                 if (iter->pos.inode == KEY_INODE_MAX)
2436                         return bkey_s_c_null;
2437
2438                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2439         }
2440
2441         search_key = btree_iter_search_key(iter);
2442         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2443                                         iter->flags & BTREE_ITER_INTENT,
2444                                         btree_iter_ip_allocated(iter));
2445
2446         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2447         if (unlikely(ret)) {
2448                 k = bkey_s_c_err(ret);
2449                 goto out_no_locked;
2450         }
2451
2452         if ((iter->flags & BTREE_ITER_CACHED) ||
2453             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2454                 struct bkey_i *next_update;
2455
2456                 if ((next_update = btree_trans_peek_updates(iter)) &&
2457                     bpos_eq(next_update->k.p, iter->pos)) {
2458                         iter->k = next_update->k;
2459                         k = bkey_i_to_s_c(next_update);
2460                         goto out;
2461                 }
2462
2463                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2464                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2465                         goto out;
2466
2467                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2468                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2469                         if (!bkey_err(k))
2470                                 iter->k = *k.k;
2471                         /* We're not returning a key from iter->path: */
2472                         goto out_no_locked;
2473                 }
2474
2475                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2476                 if (unlikely(!k.k))
2477                         goto out_no_locked;
2478         } else {
2479                 struct bpos next;
2480                 struct bpos end = iter->pos;
2481
2482                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2483                         end.offset = U64_MAX;
2484
2485                 EBUG_ON(iter->path->level);
2486
2487                 if (iter->flags & BTREE_ITER_INTENT) {
2488                         struct btree_iter iter2;
2489
2490                         bch2_trans_copy_iter(&iter2, iter);
2491                         k = bch2_btree_iter_peek_upto(&iter2, end);
2492
2493                         if (k.k && !bkey_err(k)) {
2494                                 iter->k = iter2.k;
2495                                 k.k = &iter->k;
2496                         }
2497                         bch2_trans_iter_exit(trans, &iter2);
2498                 } else {
2499                         struct bpos pos = iter->pos;
2500
2501                         k = bch2_btree_iter_peek_upto(iter, end);
2502                         if (unlikely(bkey_err(k)))
2503                                 bch2_btree_iter_set_pos(iter, pos);
2504                         else
2505                                 iter->pos = pos;
2506                 }
2507
2508                 if (unlikely(bkey_err(k)))
2509                         goto out_no_locked;
2510
2511                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2512
2513                 if (bkey_lt(iter->pos, next)) {
2514                         bkey_init(&iter->k);
2515                         iter->k.p = iter->pos;
2516
2517                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2518                                 bch2_key_resize(&iter->k,
2519                                                 min_t(u64, KEY_SIZE_MAX,
2520                                                       (next.inode == iter->pos.inode
2521                                                        ? next.offset
2522                                                        : KEY_OFFSET_MAX) -
2523                                                       iter->pos.offset));
2524                                 EBUG_ON(!iter->k.size);
2525                         }
2526
2527                         k = (struct bkey_s_c) { &iter->k, NULL };
2528                 }
2529         }
2530 out:
2531         btree_path_set_should_be_locked(iter->path);
2532 out_no_locked:
2533         bch2_btree_iter_verify_entry_exit(iter);
2534         bch2_btree_iter_verify(iter);
2535         ret = bch2_btree_iter_verify_ret(iter, k);
2536         if (unlikely(ret))
2537                 return bkey_s_c_err(ret);
2538
2539         return k;
2540 }
2541
2542 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2543 {
2544         if (!bch2_btree_iter_advance(iter))
2545                 return bkey_s_c_null;
2546
2547         return bch2_btree_iter_peek_slot(iter);
2548 }
2549
2550 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2551 {
2552         if (!bch2_btree_iter_rewind(iter))
2553                 return bkey_s_c_null;
2554
2555         return bch2_btree_iter_peek_slot(iter);
2556 }
2557
2558 /* new transactional stuff: */
2559
2560 #ifdef CONFIG_BCACHEFS_DEBUG
2561 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2562 {
2563         struct btree_path *path;
2564         unsigned i;
2565
2566         BUG_ON(trans->nr_sorted != hweight64(trans->paths_allocated));
2567
2568         trans_for_each_path(trans, path) {
2569                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2570                 BUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2571         }
2572
2573         for (i = 0; i < trans->nr_sorted; i++) {
2574                 unsigned idx = trans->sorted[i];
2575
2576                 EBUG_ON(!(trans->paths_allocated & (1ULL << idx)));
2577                 BUG_ON(trans->paths[idx].sorted_idx != i);
2578         }
2579 }
2580
2581 static void btree_trans_verify_sorted(struct btree_trans *trans)
2582 {
2583         struct btree_path *path, *prev = NULL;
2584         unsigned i;
2585
2586         if (!bch2_debug_check_iterators)
2587                 return;
2588
2589         trans_for_each_path_inorder(trans, path, i) {
2590                 if (prev && btree_path_cmp(prev, path) > 0) {
2591                         __bch2_dump_trans_paths_updates(trans, true);
2592                         panic("trans paths out of order!\n");
2593                 }
2594                 prev = path;
2595         }
2596 }
2597 #else
2598 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2599 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2600 #endif
2601
2602 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2603 {
2604         int i, l = 0, r = trans->nr_sorted, inc = 1;
2605         bool swapped;
2606
2607         btree_trans_verify_sorted_refs(trans);
2608
2609         if (trans->paths_sorted)
2610                 goto out;
2611
2612         /*
2613          * Cocktail shaker sort: this is efficient because iterators will be
2614          * mostly sorted.
2615          */
2616         do {
2617                 swapped = false;
2618
2619                 for (i = inc > 0 ? l : r - 2;
2620                      i + 1 < r && i >= l;
2621                      i += inc) {
2622                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2623                                            trans->paths + trans->sorted[i + 1]) > 0) {
2624                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2625                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2626                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2627                                 swapped = true;
2628                         }
2629                 }
2630
2631                 if (inc > 0)
2632                         --r;
2633                 else
2634                         l++;
2635                 inc = -inc;
2636         } while (swapped);
2637
2638         trans->paths_sorted = true;
2639 out:
2640         btree_trans_verify_sorted(trans);
2641 }
2642
2643 static inline void btree_path_list_remove(struct btree_trans *trans,
2644                                           struct btree_path *path)
2645 {
2646         unsigned i;
2647
2648         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2649 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2650         trans->nr_sorted--;
2651         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2652                                 trans->sorted + path->sorted_idx + 1,
2653                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2654 #else
2655         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2656 #endif
2657         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2658                 trans->paths[trans->sorted[i]].sorted_idx = i;
2659
2660         path->sorted_idx = U8_MAX;
2661 }
2662
2663 static inline void btree_path_list_add(struct btree_trans *trans,
2664                                        struct btree_path *pos,
2665                                        struct btree_path *path)
2666 {
2667         unsigned i;
2668
2669         path->sorted_idx = pos ? pos->sorted_idx + 1 : trans->nr_sorted;
2670
2671 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2672         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2673                               trans->sorted + path->sorted_idx,
2674                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2675         trans->nr_sorted++;
2676         trans->sorted[path->sorted_idx] = path->idx;
2677 #else
2678         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2679 #endif
2680
2681         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2682                 trans->paths[trans->sorted[i]].sorted_idx = i;
2683
2684         btree_trans_verify_sorted_refs(trans);
2685 }
2686
2687 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2688 {
2689         if (iter->path)
2690                 bch2_path_put(trans, iter->path,
2691                               iter->flags & BTREE_ITER_INTENT);
2692         if (iter->update_path)
2693                 bch2_path_put_nokeep(trans, iter->update_path,
2694                               iter->flags & BTREE_ITER_INTENT);
2695         if (iter->key_cache_path)
2696                 bch2_path_put(trans, iter->key_cache_path,
2697                               iter->flags & BTREE_ITER_INTENT);
2698         iter->path = NULL;
2699         iter->update_path = NULL;
2700         iter->key_cache_path = NULL;
2701 }
2702
2703 static inline void bch2_trans_iter_init_inlined(struct btree_trans *trans,
2704                           struct btree_iter *iter,
2705                           unsigned btree_id, struct bpos pos,
2706                           unsigned flags)
2707 {
2708         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2709                                bch2_btree_iter_flags(trans, btree_id, flags),
2710                                _RET_IP_);
2711 }
2712
2713 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2714                           struct btree_iter *iter,
2715                           unsigned btree_id, struct bpos pos,
2716                           unsigned flags)
2717 {
2718         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2719                                bch2_btree_iter_flags(trans, btree_id, flags),
2720                                _RET_IP_);
2721 }
2722
2723 void bch2_trans_node_iter_init(struct btree_trans *trans,
2724                                struct btree_iter *iter,
2725                                enum btree_id btree_id,
2726                                struct bpos pos,
2727                                unsigned locks_want,
2728                                unsigned depth,
2729                                unsigned flags)
2730 {
2731        flags |= BTREE_ITER_NOT_EXTENTS;
2732        flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2733        flags |= BTREE_ITER_ALL_SNAPSHOTS;
2734
2735         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2736                                __bch2_btree_iter_flags(trans, btree_id, flags),
2737                                _RET_IP_);
2738
2739         iter->min_depth = depth;
2740
2741         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2742         BUG_ON(iter->path->level        != depth);
2743         BUG_ON(iter->min_depth          != depth);
2744 }
2745
2746 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2747 {
2748         *dst = *src;
2749         if (src->path)
2750                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2751         if (src->update_path)
2752                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2753         dst->key_cache_path = NULL;
2754 }
2755
2756 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2757 {
2758         unsigned new_top = trans->mem_top + size;
2759         size_t old_bytes = trans->mem_bytes;
2760         size_t new_bytes = roundup_pow_of_two(new_top);
2761         void *new_mem;
2762         void *p;
2763
2764         trans->mem_max = max(trans->mem_max, new_top);
2765
2766         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2767
2768         new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2769         if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2770                 new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2771                 new_bytes = BTREE_TRANS_MEM_MAX;
2772                 kfree(trans->mem);
2773         }
2774
2775         if (!new_mem)
2776                 return ERR_PTR(-ENOMEM);
2777
2778         trans->mem = new_mem;
2779         trans->mem_bytes = new_bytes;
2780
2781         if (old_bytes) {
2782                 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2783                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2784         }
2785
2786         p = trans->mem + trans->mem_top;
2787         trans->mem_top += size;
2788         memset(p, 0, size);
2789         return p;
2790 }
2791
2792 static noinline void bch2_trans_reset_srcu_lock(struct btree_trans *trans)
2793 {
2794         struct bch_fs *c = trans->c;
2795         struct btree_path *path;
2796
2797         trans_for_each_path(trans, path)
2798                 if (path->cached && !btree_node_locked(path, 0))
2799                         path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2800
2801         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2802         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2803         trans->srcu_lock_time   = jiffies;
2804 }
2805
2806 /**
2807  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2808  * @trans: transaction to reset
2809  *
2810  * While iterating over nodes or updating nodes a attempt to lock a btree node
2811  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2812  * occurs bch2_trans_begin() should be called and the transaction retried.
2813  */
2814 u32 bch2_trans_begin(struct btree_trans *trans)
2815 {
2816         struct btree_path *path;
2817
2818         bch2_trans_reset_updates(trans);
2819
2820         trans->restart_count++;
2821         trans->mem_top                  = 0;
2822
2823         trans_for_each_path(trans, path) {
2824                 path->should_be_locked = false;
2825
2826                 /*
2827                  * If the transaction wasn't restarted, we're presuming to be
2828                  * doing something new: dont keep iterators excpt the ones that
2829                  * are in use - except for the subvolumes btree:
2830                  */
2831                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2832                         path->preserve = false;
2833
2834                 /*
2835                  * XXX: we probably shouldn't be doing this if the transaction
2836                  * was restarted, but currently we still overflow transaction
2837                  * iterators if we do that
2838                  */
2839                 if (!path->ref && !path->preserve)
2840                         __bch2_path_free(trans, path);
2841                 else
2842                         path->preserve = false;
2843         }
2844
2845         if (!trans->restarted &&
2846             (need_resched() ||
2847              local_clock() - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2848                 bch2_trans_unlock(trans);
2849                 cond_resched();
2850                 bch2_trans_relock(trans);
2851         }
2852
2853         if (unlikely(time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2854                 bch2_trans_reset_srcu_lock(trans);
2855
2856         trans->last_restarted_ip = _RET_IP_;
2857         if (trans->restarted) {
2858                 bch2_btree_path_traverse_all(trans);
2859                 trans->notrace_relock_fail = false;
2860         }
2861
2862         trans->last_begin_time = local_clock();
2863         return trans->restart_count;
2864 }
2865
2866 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
2867 {
2868         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
2869         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2870         void *p = NULL;
2871
2872         BUG_ON(trans->used_mempool);
2873
2874 #ifdef __KERNEL__
2875         p = this_cpu_xchg(c->btree_paths_bufs->path, NULL);
2876 #endif
2877         if (!p)
2878                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
2879
2880         trans->paths            = p; p += paths_bytes;
2881         trans->updates          = p; p += updates_bytes;
2882 }
2883
2884 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2885
2886 unsigned bch2_trans_get_fn_idx(const char *fn)
2887 {
2888         unsigned i;
2889
2890         for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2891                 if (!bch2_btree_transaction_fns[i] ||
2892                     bch2_btree_transaction_fns[i] == fn) {
2893                         bch2_btree_transaction_fns[i] = fn;
2894                         return i;
2895                 }
2896
2897         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2898         return i;
2899 }
2900
2901 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, unsigned fn_idx)
2902         __acquires(&c->btree_trans_barrier)
2903 {
2904         struct btree_transaction_stats *s;
2905
2906         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
2907
2908         memset(trans, 0, sizeof(*trans));
2909         trans->c                = c;
2910         trans->fn               = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2911                 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2912         trans->last_begin_time  = local_clock();
2913         trans->fn_idx           = fn_idx;
2914         trans->locking_wait.task = current;
2915         trans->journal_replay_not_finished =
2916                 !test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
2917         closure_init_stack(&trans->ref);
2918
2919         bch2_trans_alloc_paths(trans, c);
2920
2921         s = btree_trans_stats(trans);
2922         if (s && s->max_mem) {
2923                 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2924
2925                 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2926
2927                 if (!unlikely(trans->mem)) {
2928                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2929                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2930                 } else {
2931                         trans->mem_bytes = expected_mem_bytes;
2932                 }
2933         }
2934         if (s)
2935                 trans->nr_max_paths = s->nr_max_paths;
2936
2937         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2938         trans->srcu_lock_time   = jiffies;
2939
2940         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2941                 struct btree_trans *pos;
2942
2943                 mutex_lock(&c->btree_trans_lock);
2944                 list_for_each_entry(pos, &c->btree_trans_list, list) {
2945                         if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
2946                                 list_add_tail(&trans->list, &pos->list);
2947                                 goto list_add_done;
2948                         }
2949                 }
2950                 list_add_tail(&trans->list, &c->btree_trans_list);
2951 list_add_done:
2952                 mutex_unlock(&c->btree_trans_lock);
2953         }
2954 }
2955
2956 static void check_btree_paths_leaked(struct btree_trans *trans)
2957 {
2958 #ifdef CONFIG_BCACHEFS_DEBUG
2959         struct bch_fs *c = trans->c;
2960         struct btree_path *path;
2961
2962         trans_for_each_path(trans, path)
2963                 if (path->ref)
2964                         goto leaked;
2965         return;
2966 leaked:
2967         bch_err(c, "btree paths leaked from %s!", trans->fn);
2968         trans_for_each_path(trans, path)
2969                 if (path->ref)
2970                         printk(KERN_ERR "  btree %s %pS\n",
2971                                bch2_btree_ids[path->btree_id],
2972                                (void *) path->ip_allocated);
2973         /* Be noisy about this: */
2974         bch2_fatal_error(c);
2975 #endif
2976 }
2977
2978 void bch2_trans_exit(struct btree_trans *trans)
2979         __releases(&c->btree_trans_barrier)
2980 {
2981         struct btree_insert_entry *i;
2982         struct bch_fs *c = trans->c;
2983         struct btree_transaction_stats *s = btree_trans_stats(trans);
2984
2985         bch2_trans_unlock(trans);
2986
2987         closure_sync(&trans->ref);
2988
2989         if (s)
2990                 s->max_mem = max(s->max_mem, trans->mem_max);
2991
2992         trans_for_each_update(trans, i)
2993                 __btree_path_put(i->path, true);
2994         trans->nr_updates               = 0;
2995
2996         check_btree_paths_leaked(trans);
2997
2998         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2999                 mutex_lock(&c->btree_trans_lock);
3000                 list_del(&trans->list);
3001                 mutex_unlock(&c->btree_trans_lock);
3002         }
3003
3004         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3005
3006         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3007
3008         kfree(trans->extra_journal_entries.data);
3009
3010         if (trans->fs_usage_deltas) {
3011                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3012                     REPLICAS_DELTA_LIST_MAX)
3013                         mempool_free(trans->fs_usage_deltas,
3014                                      &c->replicas_delta_pool);
3015                 else
3016                         kfree(trans->fs_usage_deltas);
3017         }
3018
3019         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3020                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3021         else
3022                 kfree(trans->mem);
3023
3024 #ifdef __KERNEL__
3025         /*
3026          * Userspace doesn't have a real percpu implementation:
3027          */
3028         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3029 #endif
3030
3031         if (trans->paths)
3032                 mempool_free(trans->paths, &c->btree_paths_pool);
3033
3034         trans->mem      = (void *) 0x1;
3035         trans->paths    = (void *) 0x1;
3036 }
3037
3038 static void __maybe_unused
3039 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3040                                       struct btree_bkey_cached_common *b)
3041 {
3042         struct six_lock_count c = six_lock_counts(&b->lock);
3043         struct task_struct *owner;
3044         pid_t pid;
3045
3046         rcu_read_lock();
3047         owner = READ_ONCE(b->lock.owner);
3048         pid = owner ? owner->pid : 0;
3049         rcu_read_unlock();
3050
3051         prt_tab(out);
3052         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3053                    b->level, bch2_btree_ids[b->btree_id]);
3054         bch2_bpos_to_text(out, btree_node_pos(b));
3055
3056         prt_tab(out);
3057         prt_printf(out, " locks %u:%u:%u held by pid %u",
3058                    c.n[0], c.n[1], c.n[2], pid);
3059 }
3060
3061 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3062 {
3063         struct btree_path *path;
3064         struct btree_bkey_cached_common *b;
3065         static char lock_types[] = { 'r', 'i', 'w' };
3066         unsigned l;
3067
3068         if (!out->nr_tabstops) {
3069                 printbuf_tabstop_push(out, 16);
3070                 printbuf_tabstop_push(out, 32);
3071         }
3072
3073         prt_printf(out, "%i %s\n", trans->locking_wait.task->pid, trans->fn);
3074
3075         trans_for_each_path(trans, path) {
3076                 if (!path->nodes_locked)
3077                         continue;
3078
3079                 prt_printf(out, "  path %u %c l=%u %s:",
3080                        path->idx,
3081                        path->cached ? 'c' : 'b',
3082                        path->level,
3083                        bch2_btree_ids[path->btree_id]);
3084                 bch2_bpos_to_text(out, path->pos);
3085                 prt_newline(out);
3086
3087                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3088                         if (btree_node_locked(path, l) &&
3089                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3090                                 prt_printf(out, "    %c l=%u ",
3091                                            lock_types[btree_node_locked_type(path, l)], l);
3092                                 bch2_btree_bkey_cached_common_to_text(out, b);
3093                                 prt_newline(out);
3094                         }
3095                 }
3096         }
3097
3098         b = READ_ONCE(trans->locking);
3099         if (b) {
3100                 prt_printf(out, "  blocked for %lluus on",
3101                            div_u64(local_clock() - trans->locking_wait.start_time,
3102                                    1000));
3103                 prt_newline(out);
3104                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3105                 bch2_btree_bkey_cached_common_to_text(out, b);
3106                 prt_newline(out);
3107         }
3108 }
3109
3110 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3111 {
3112         struct btree_transaction_stats *s;
3113
3114         for (s = c->btree_transaction_stats;
3115              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3116              s++) {
3117                 kfree(s->max_paths_text);
3118                 bch2_time_stats_exit(&s->lock_hold_times);
3119         }
3120
3121         if (c->btree_trans_barrier_initialized)
3122                 cleanup_srcu_struct(&c->btree_trans_barrier);
3123         mempool_exit(&c->btree_trans_mem_pool);
3124         mempool_exit(&c->btree_paths_pool);
3125 }
3126
3127 int bch2_fs_btree_iter_init(struct bch_fs *c)
3128 {
3129         struct btree_transaction_stats *s;
3130         unsigned nr = BTREE_ITER_MAX;
3131         int ret;
3132
3133         for (s = c->btree_transaction_stats;
3134              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3135              s++) {
3136                 bch2_time_stats_init(&s->lock_hold_times);
3137                 mutex_init(&s->lock);
3138         }
3139
3140         INIT_LIST_HEAD(&c->btree_trans_list);
3141         mutex_init(&c->btree_trans_lock);
3142
3143         ret   = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3144                         sizeof(struct btree_path) * nr +
3145                         sizeof(struct btree_insert_entry) * nr) ?:
3146                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3147                                           BTREE_TRANS_MEM_MAX) ?:
3148                 init_srcu_struct(&c->btree_trans_barrier);
3149         if (!ret)
3150                 c->btree_trans_barrier_initialized = true;
3151         return ret;
3152 }