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 #include "trace.h"
  19
  20 #include <linux/random.h>
  21 #include <linux/prefetch.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
 656         path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 657         path->l[b->c.level].b = b;
 658         __btree_path_level_init(path, b->c.level);
 659 }
 660
 661 /* Btree path: fixups after btree node updates: */
 662
 663 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
 664 {
 665         struct bch_fs *c = trans->c;
 666         struct btree_insert_entry *i;
 667
 668         trans_for_each_update(trans, i)
 669                 if (!i->cached &&
 670                     i->level    == b->c.level &&
 671                     i->btree_id == b->c.btree_id &&
 672                     bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
 673                     bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
 674                         i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 675
 676                         if (unlikely(trans->journal_replay_not_finished)) {
 677                                 struct bkey_i *j_k =
 678                                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 679                                                                     i->k->k.p);
 680
 681                                 if (j_k) {
 682                                         i->old_k = j_k->k;
 683                                         i->old_v = &j_k->v;
 684                                 }
 685                         }
 686                 }
 687 }
 688
 689 /*
 690  * A btree node is being replaced - update the iterator to point to the new
 691  * node:
 692  */
 693 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
 694 {
 695         struct btree_path *path;
 696
 697         trans_for_each_path(trans, path)
 698                 if (path->uptodate == BTREE_ITER_UPTODATE &&
 699                     !path->cached &&
 700                     btree_path_pos_in_node(path, b)) {
 701                         enum btree_node_locked_type t =
 702                                 btree_lock_want(path, b->c.level);
 703
 704                         if (t != BTREE_NODE_UNLOCKED) {
 705                                 btree_node_unlock(trans, path, b->c.level);
 706                                 six_lock_increment(&b->c.lock, t);
 707                                 mark_btree_node_locked(trans, path, b->c.level, t);
 708                         }
 709
 710                         bch2_btree_path_level_init(trans, path, b);
 711                 }
 712
 713         bch2_trans_revalidate_updates_in_node(trans, b);
 714 }
 715
 716 /*
 717  * A btree node has been modified in such a way as to invalidate iterators - fix
 718  * them:
 719  */
 720 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
 721 {
 722         struct btree_path *path;
 723
 724         trans_for_each_path_with_node(trans, b, path)
 725                 __btree_path_level_init(path, b->c.level);
 726
 727         bch2_trans_revalidate_updates_in_node(trans, b);
 728 }
 729
 730 /* Btree path: traverse, set_pos: */
 731
 732 static inline int btree_path_lock_root(struct btree_trans *trans,
 733                                        struct btree_path *path,
 734                                        unsigned depth_want,
 735                                        unsigned long trace_ip)
 736 {
 737         struct bch_fs *c = trans->c;
 738         struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
 739         enum six_lock_type lock_type;
 740         unsigned i;
 741         int ret;
 742
 743         EBUG_ON(path->nodes_locked);
 744
 745         while (1) {
 746                 b = READ_ONCE(*rootp);
 747                 path->level = READ_ONCE(b->c.level);
 748
 749                 if (unlikely(path->level < depth_want)) {
 750                         /*
 751                          * the root is at a lower depth than the depth we want:
 752                          * got to the end of the btree, or we're walking nodes
 753                          * greater than some depth and there are no nodes >=
 754                          * that depth
 755                          */
 756                         path->level = depth_want;
 757                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
 758                                 path->l[i].b = NULL;
 759                         return 1;
 760                 }
 761
 762                 lock_type = __btree_lock_want(path, path->level);
 763                 ret = btree_node_lock(trans, path, &b->c,
 764                                       path->level, lock_type, trace_ip);
 765                 if (unlikely(ret)) {
 766                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
 767                                 continue;
 768                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 769                                 return ret;
 770                         BUG();
 771                 }
 772
 773                 if (likely(b == READ_ONCE(*rootp) &&
 774                            b->c.level == path->level &&
 775                            !race_fault())) {
 776                         for (i = 0; i < path->level; i++)
 777                                 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
 778                         path->l[path->level].b = b;
 779                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
 780                                 path->l[i].b = NULL;
 781
 782                         mark_btree_node_locked(trans, path, path->level, lock_type);
 783                         bch2_btree_path_level_init(trans, path, b);
 784                         return 0;
 785                 }
 786
 787                 six_unlock_type(&b->c.lock, lock_type);
 788         }
 789 }
 790
 791 noinline
 792 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
 793 {
 794         struct bch_fs *c = trans->c;
 795         struct btree_path_level *l = path_l(path);
 796         struct btree_node_iter node_iter = l->iter;
 797         struct bkey_packed *k;
 798         struct bkey_buf tmp;
 799         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 800                 ? (path->level > 1 ? 0 :  2)
 801                 : (path->level > 1 ? 1 : 16);
 802         bool was_locked = btree_node_locked(path, path->level);
 803         int ret = 0;
 804
 805         bch2_bkey_buf_init(&tmp);
 806
 807         while (nr-- && !ret) {
 808                 if (!bch2_btree_node_relock(trans, path, path->level))
 809                         break;
 810
 811                 bch2_btree_node_iter_advance(&node_iter, l->b);
 812                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
 813                 if (!k)
 814                         break;
 815
 816                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
 817                 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
 818                                                path->level - 1);
 819         }
 820
 821         if (!was_locked)
 822                 btree_node_unlock(trans, path, path->level);
 823
 824         bch2_bkey_buf_exit(&tmp, c);
 825         return ret;
 826 }
 827
 828 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
 829                                  struct btree_and_journal_iter *jiter)
 830 {
 831         struct bch_fs *c = trans->c;
 832         struct bkey_s_c k;
 833         struct bkey_buf tmp;
 834         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 835                 ? (path->level > 1 ? 0 :  2)
 836                 : (path->level > 1 ? 1 : 16);
 837         bool was_locked = btree_node_locked(path, path->level);
 838         int ret = 0;
 839
 840         bch2_bkey_buf_init(&tmp);
 841
 842         while (nr-- && !ret) {
 843                 if (!bch2_btree_node_relock(trans, path, path->level))
 844                         break;
 845
 846                 bch2_btree_and_journal_iter_advance(jiter);
 847                 k = bch2_btree_and_journal_iter_peek(jiter);
 848                 if (!k.k)
 849                         break;
 850
 851                 bch2_bkey_buf_reassemble(&tmp, c, k);
 852                 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
 853                                                path->level - 1);
 854         }
 855
 856         if (!was_locked)
 857                 btree_node_unlock(trans, path, path->level);
 858
 859         bch2_bkey_buf_exit(&tmp, c);
 860         return ret;
 861 }
 862
 863 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
 864                                             struct btree_path *path,
 865                                             unsigned plevel, struct btree *b)
 866 {
 867         struct btree_path_level *l = &path->l[plevel];
 868         bool locked = btree_node_locked(path, plevel);
 869         struct bkey_packed *k;
 870         struct bch_btree_ptr_v2 *bp;
 871
 872         if (!bch2_btree_node_relock(trans, path, plevel))
 873                 return;
 874
 875         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 876         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
 877
 878         bp = (void *) bkeyp_val(&l->b->format, k);
 879         bp->mem_ptr = (unsigned long)b;
 880
 881         if (!locked)
 882                 btree_node_unlock(trans, path, plevel);
 883 }
 884
 885 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
 886                                                      struct btree_path *path,
 887                                                      unsigned flags,
 888                                                      struct bkey_buf *out)
 889 {
 890         struct bch_fs *c = trans->c;
 891         struct btree_path_level *l = path_l(path);
 892         struct btree_and_journal_iter jiter;
 893         struct bkey_s_c k;
 894         int ret = 0;
 895
 896         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
 897
 898         k = bch2_btree_and_journal_iter_peek(&jiter);
 899
 900         bch2_bkey_buf_reassemble(out, c, k);
 901
 902         if (flags & BTREE_ITER_PREFETCH)
 903                 ret = btree_path_prefetch_j(trans, path, &jiter);
 904
 905         bch2_btree_and_journal_iter_exit(&jiter);
 906         return ret;
 907 }
 908
 909 static __always_inline int btree_path_down(struct btree_trans *trans,
 910                                            struct btree_path *path,
 911                                            unsigned flags,
 912                                            unsigned long trace_ip)
 913 {
 914         struct bch_fs *c = trans->c;
 915         struct btree_path_level *l = path_l(path);
 916         struct btree *b;
 917         unsigned level = path->level - 1;
 918         enum six_lock_type lock_type = __btree_lock_want(path, level);
 919         struct bkey_buf tmp;
 920         int ret;
 921
 922         EBUG_ON(!btree_node_locked(path, path->level));
 923
 924         bch2_bkey_buf_init(&tmp);
 925
 926         if (unlikely(trans->journal_replay_not_finished)) {
 927                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
 928                 if (ret)
 929                         goto err;
 930         } else {
 931                 bch2_bkey_buf_unpack(&tmp, c, l->b,
 932                                  bch2_btree_node_iter_peek(&l->iter, l->b));
 933
 934                 if (flags & BTREE_ITER_PREFETCH) {
 935                         ret = btree_path_prefetch(trans, path);
 936                         if (ret)
 937                                 goto err;
 938                 }
 939         }
 940
 941         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
 942         ret = PTR_ERR_OR_ZERO(b);
 943         if (unlikely(ret))
 944                 goto err;
 945
 946         if (likely(!trans->journal_replay_not_finished &&
 947                    tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
 948             unlikely(b != btree_node_mem_ptr(tmp.k)))
 949                 btree_node_mem_ptr_set(trans, path, level + 1, b);
 950
 951         if (btree_node_read_locked(path, level + 1))
 952                 btree_node_unlock(trans, path, level + 1);
 953
 954         mark_btree_node_locked(trans, path, level, lock_type);
 955         path->level = level;
 956         bch2_btree_path_level_init(trans, path, b);
 957
 958         bch2_btree_path_verify_locks(path);
 959 err:
 960         bch2_bkey_buf_exit(&tmp, c);
 961         return ret;
 962 }
 963
 964
 965 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
 966 {
 967         struct bch_fs *c = trans->c;
 968         struct btree_path *path;
 969         unsigned long trace_ip = _RET_IP_;
 970         int i, ret = 0;
 971
 972         if (trans->in_traverse_all)
 973                 return -BCH_ERR_transaction_restart_in_traverse_all;
 974
 975         trans->in_traverse_all = true;
 976 retry_all:
 977         trans->restarted = 0;
 978         trans->last_restarted_ip = 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                             bch2_err_matches(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_begin_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         struct btree_write_buffered_key *wb;
1378
1379         prt_printf(buf, "transaction updates for %s journal seq %llu",
1380                trans->fn, trans->journal_res.seq);
1381         prt_newline(buf);
1382         printbuf_indent_add(buf, 2);
1383
1384         trans_for_each_update(trans, i) {
1385                 struct bkey_s_c old = { &i->old_k, i->old_v };
1386
1387                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1388                        bch2_btree_ids[i->btree_id],
1389                        i->cached,
1390                        (void *) i->ip_allocated);
1391                 prt_newline(buf);
1392
1393                 prt_printf(buf, "  old ");
1394                 bch2_bkey_val_to_text(buf, trans->c, old);
1395                 prt_newline(buf);
1396
1397                 prt_printf(buf, "  new ");
1398                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1399                 prt_newline(buf);
1400         }
1401
1402         trans_for_each_wb_update(trans, wb) {
1403                 prt_printf(buf, "update: btree=%s wb=1 %pS",
1404                        bch2_btree_ids[wb->btree],
1405                        (void *) i->ip_allocated);
1406                 prt_newline(buf);
1407
1408                 prt_printf(buf, "  new ");
1409                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(&wb->k));
1410                 prt_newline(buf);
1411         }
1412
1413         printbuf_indent_sub(buf, 2);
1414 }
1415
1416 noinline __cold
1417 void bch2_dump_trans_updates(struct btree_trans *trans)
1418 {
1419         struct printbuf buf = PRINTBUF;
1420
1421         bch2_trans_updates_to_text(&buf, trans);
1422         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1423         printbuf_exit(&buf);
1424 }
1425
1426 noinline __cold
1427 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1428 {
1429         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1430                    path->idx, path->ref, path->intent_ref,
1431                    path->preserve ? 'P' : ' ',
1432                    path->should_be_locked ? 'S' : ' ',
1433                    bch2_btree_ids[path->btree_id],
1434                    path->level);
1435         bch2_bpos_to_text(out, path->pos);
1436
1437         prt_printf(out, " locks %u", path->nodes_locked);
1438 #ifdef TRACK_PATH_ALLOCATED
1439         prt_printf(out, " %pS", (void *) path->ip_allocated);
1440 #endif
1441         prt_newline(out);
1442 }
1443
1444 noinline __cold
1445 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1446                                 bool nosort)
1447 {
1448         struct btree_path *path;
1449         unsigned idx;
1450
1451         if (!nosort)
1452                 btree_trans_sort_paths(trans);
1453
1454         trans_for_each_path_inorder(trans, path, idx)
1455                 bch2_btree_path_to_text(out, path);
1456 }
1457
1458 noinline __cold
1459 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1460 {
1461         __bch2_trans_paths_to_text(out, trans, false);
1462 }
1463
1464 noinline __cold
1465 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1466 {
1467         struct printbuf buf = PRINTBUF;
1468
1469         __bch2_trans_paths_to_text(&buf, trans, nosort);
1470         bch2_trans_updates_to_text(&buf, trans);
1471
1472         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1473         printbuf_exit(&buf);
1474 }
1475
1476 noinline __cold
1477 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1478 {
1479         __bch2_dump_trans_paths_updates(trans, false);
1480 }
1481
1482 noinline __cold
1483 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1484 {
1485         struct btree_transaction_stats *s = btree_trans_stats(trans);
1486         struct printbuf buf = PRINTBUF;
1487
1488         if (!s)
1489                 return;
1490
1491         bch2_trans_paths_to_text(&buf, trans);
1492
1493         if (!buf.allocation_failure) {
1494                 mutex_lock(&s->lock);
1495                 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1496                         s->nr_max_paths = trans->nr_max_paths =
1497                                 hweight64(trans->paths_allocated);
1498                         swap(s->max_paths_text, buf.buf);
1499                 }
1500                 mutex_unlock(&s->lock);
1501         }
1502
1503         printbuf_exit(&buf);
1504
1505         trans->nr_max_paths = hweight64(trans->paths_allocated);
1506 }
1507
1508 static noinline void btree_path_overflow(struct btree_trans *trans)
1509 {
1510         bch2_dump_trans_paths_updates(trans);
1511         panic("trans path oveflow\n");
1512 }
1513
1514 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1515                                                   struct btree_path *pos)
1516 {
1517         struct btree_path *path;
1518         unsigned idx;
1519
1520         if (unlikely(trans->paths_allocated ==
1521                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
1522                 btree_path_overflow(trans);
1523
1524         idx = __ffs64(~trans->paths_allocated);
1525
1526         /*
1527          * Do this before marking the new path as allocated, since it won't be
1528          * initialized yet:
1529          */
1530         if (unlikely(idx > trans->nr_max_paths))
1531                 bch2_trans_update_max_paths(trans);
1532
1533         trans->paths_allocated |= 1ULL << idx;
1534
1535         path = &trans->paths[idx];
1536         path->idx               = idx;
1537         path->ref               = 0;
1538         path->intent_ref        = 0;
1539         path->nodes_locked      = 0;
1540
1541         btree_path_list_add(trans, pos, path);
1542         trans->paths_sorted = false;
1543         return path;
1544 }
1545
1546 struct btree_path *bch2_path_get(struct btree_trans *trans,
1547                                  enum btree_id btree_id, struct bpos pos,
1548                                  unsigned locks_want, unsigned level,
1549                                  unsigned flags, unsigned long ip)
1550 {
1551         struct btree_path *path, *path_pos = NULL;
1552         bool cached = flags & BTREE_ITER_CACHED;
1553         bool intent = flags & BTREE_ITER_INTENT;
1554         int i;
1555
1556         bch2_trans_verify_not_in_restart(trans);
1557         bch2_trans_verify_locks(trans);
1558
1559         btree_trans_sort_paths(trans);
1560
1561         trans_for_each_path_inorder(trans, path, i) {
1562                 if (__btree_path_cmp(path,
1563                                      btree_id,
1564                                      cached,
1565                                      pos,
1566                                      level) > 0)
1567                         break;
1568
1569                 path_pos = path;
1570         }
1571
1572         if (path_pos &&
1573             path_pos->cached    == cached &&
1574             path_pos->btree_id  == btree_id &&
1575             path_pos->level     == level) {
1576                 __btree_path_get(path_pos, intent);
1577                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1578         } else {
1579                 path = btree_path_alloc(trans, path_pos);
1580                 path_pos = NULL;
1581
1582                 __btree_path_get(path, intent);
1583                 path->pos                       = pos;
1584                 path->btree_id                  = btree_id;
1585                 path->cached                    = cached;
1586                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1587                 path->should_be_locked          = false;
1588                 path->level                     = level;
1589                 path->locks_want                = locks_want;
1590                 path->nodes_locked              = 0;
1591                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1592                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1593 #ifdef TRACK_PATH_ALLOCATED
1594                 path->ip_allocated              = ip;
1595 #endif
1596                 trans->paths_sorted             = false;
1597         }
1598
1599         if (!(flags & BTREE_ITER_NOPRESERVE))
1600                 path->preserve = true;
1601
1602         if (path->intent_ref)
1603                 locks_want = max(locks_want, level + 1);
1604
1605         /*
1606          * If the path has locks_want greater than requested, we don't downgrade
1607          * it here - on transaction restart because btree node split needs to
1608          * upgrade locks, we might be putting/getting the iterator again.
1609          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1610          * a successful transaction commit.
1611          */
1612
1613         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1614         if (locks_want > path->locks_want)
1615                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want);
1616
1617         return path;
1618 }
1619
1620 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1621 {
1622
1623         struct btree_path_level *l = path_l(path);
1624         struct bkey_packed *_k;
1625         struct bkey_s_c k;
1626
1627         if (unlikely(!l->b))
1628                 return bkey_s_c_null;
1629
1630         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1631         EBUG_ON(!btree_node_locked(path, path->level));
1632
1633         if (!path->cached) {
1634                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1635                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1636
1637                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1638
1639                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1640                         goto hole;
1641         } else {
1642                 struct bkey_cached *ck = (void *) path->l[0].b;
1643
1644                 EBUG_ON(ck &&
1645                         (path->btree_id != ck->key.btree_id ||
1646                          !bkey_eq(path->pos, ck->key.pos)));
1647                 if (!ck || !ck->valid)
1648                         return bkey_s_c_null;
1649
1650                 *u = ck->k->k;
1651                 k = bkey_i_to_s_c(ck->k);
1652         }
1653
1654         return k;
1655 hole:
1656         bkey_init(u);
1657         u->p = path->pos;
1658         return (struct bkey_s_c) { u, NULL };
1659 }
1660
1661 /* Btree iterators: */
1662
1663 int __must_check
1664 __bch2_btree_iter_traverse(struct btree_iter *iter)
1665 {
1666         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1667 }
1668
1669 int __must_check
1670 bch2_btree_iter_traverse(struct btree_iter *iter)
1671 {
1672         int ret;
1673
1674         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1675                                         btree_iter_search_key(iter),
1676                                         iter->flags & BTREE_ITER_INTENT,
1677                                         btree_iter_ip_allocated(iter));
1678
1679         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1680         if (ret)
1681                 return ret;
1682
1683         btree_path_set_should_be_locked(iter->path);
1684         return 0;
1685 }
1686
1687 /* Iterate across nodes (leaf and interior nodes) */
1688
1689 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1690 {
1691         struct btree_trans *trans = iter->trans;
1692         struct btree *b = NULL;
1693         int ret;
1694
1695         EBUG_ON(iter->path->cached);
1696         bch2_btree_iter_verify(iter);
1697
1698         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1699         if (ret)
1700                 goto err;
1701
1702         b = btree_path_node(iter->path, iter->path->level);
1703         if (!b)
1704                 goto out;
1705
1706         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1707
1708         bkey_init(&iter->k);
1709         iter->k.p = iter->pos = b->key.k.p;
1710
1711         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1712                                         iter->flags & BTREE_ITER_INTENT,
1713                                         btree_iter_ip_allocated(iter));
1714         btree_path_set_should_be_locked(iter->path);
1715 out:
1716         bch2_btree_iter_verify_entry_exit(iter);
1717         bch2_btree_iter_verify(iter);
1718
1719         return b;
1720 err:
1721         b = ERR_PTR(ret);
1722         goto out;
1723 }
1724
1725 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1726 {
1727         struct btree *b;
1728
1729         while (b = bch2_btree_iter_peek_node(iter),
1730                bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1731                 bch2_trans_begin(iter->trans);
1732
1733         return b;
1734 }
1735
1736 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1737 {
1738         struct btree_trans *trans = iter->trans;
1739         struct btree_path *path = iter->path;
1740         struct btree *b = NULL;
1741         int ret;
1742
1743         bch2_trans_verify_not_in_restart(trans);
1744         EBUG_ON(iter->path->cached);
1745         bch2_btree_iter_verify(iter);
1746
1747         /* already at end? */
1748         if (!btree_path_node(path, path->level))
1749                 return NULL;
1750
1751         /* got to end? */
1752         if (!btree_path_node(path, path->level + 1)) {
1753                 btree_path_set_level_up(trans, path);
1754                 return NULL;
1755         }
1756
1757         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1758                 __bch2_btree_path_unlock(trans, path);
1759                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1760                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1761                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1762                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1763                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1764                 goto err;
1765         }
1766
1767         b = btree_path_node(path, path->level + 1);
1768
1769         if (bpos_eq(iter->pos, b->key.k.p)) {
1770                 __btree_path_set_level_up(trans, path, path->level++);
1771         } else {
1772                 /*
1773                  * Haven't gotten to the end of the parent node: go back down to
1774                  * the next child node
1775                  */
1776                 path = iter->path =
1777                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1778                                            iter->flags & BTREE_ITER_INTENT,
1779                                            btree_iter_ip_allocated(iter));
1780
1781                 btree_path_set_level_down(trans, path, iter->min_depth);
1782
1783                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1784                 if (ret)
1785                         goto err;
1786
1787                 b = path->l[path->level].b;
1788         }
1789
1790         bkey_init(&iter->k);
1791         iter->k.p = iter->pos = b->key.k.p;
1792
1793         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1794                                         iter->flags & BTREE_ITER_INTENT,
1795                                         btree_iter_ip_allocated(iter));
1796         btree_path_set_should_be_locked(iter->path);
1797         BUG_ON(iter->path->uptodate);
1798 out:
1799         bch2_btree_iter_verify_entry_exit(iter);
1800         bch2_btree_iter_verify(iter);
1801
1802         return b;
1803 err:
1804         b = ERR_PTR(ret);
1805         goto out;
1806 }
1807
1808 /* Iterate across keys (in leaf nodes only) */
1809
1810 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1811 {
1812         if (likely(!(iter->flags & BTREE_ITER_ALL_LEVELS))) {
1813                 struct bpos pos = iter->k.p;
1814                 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1815                              ? bpos_eq(pos, SPOS_MAX)
1816                              : bkey_eq(pos, SPOS_MAX));
1817
1818                 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1819                         pos = bkey_successor(iter, pos);
1820                 bch2_btree_iter_set_pos(iter, pos);
1821                 return ret;
1822         } else {
1823                 if (!btree_path_node(iter->path, iter->path->level))
1824                         return true;
1825
1826                 iter->advanced = true;
1827                 return false;
1828         }
1829 }
1830
1831 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1832 {
1833         struct bpos pos = bkey_start_pos(&iter->k);
1834         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1835                      ? bpos_eq(pos, POS_MIN)
1836                      : bkey_eq(pos, POS_MIN));
1837
1838         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1839                 pos = bkey_predecessor(iter, pos);
1840         bch2_btree_iter_set_pos(iter, pos);
1841         return ret;
1842 }
1843
1844 static noinline
1845 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1846 {
1847         struct btree_insert_entry *i;
1848         struct bkey_i *ret = NULL;
1849
1850         trans_for_each_update(iter->trans, i) {
1851                 if (i->btree_id < iter->btree_id)
1852                         continue;
1853                 if (i->btree_id > iter->btree_id)
1854                         break;
1855                 if (bpos_lt(i->k->k.p, iter->path->pos))
1856                         continue;
1857                 if (i->key_cache_already_flushed)
1858                         continue;
1859                 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1860                         ret = i->k;
1861         }
1862
1863         return ret;
1864 }
1865
1866 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1867 {
1868         return iter->flags & BTREE_ITER_WITH_UPDATES
1869                 ? __bch2_btree_trans_peek_updates(iter)
1870                 : NULL;
1871 }
1872
1873 struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1874                                        struct btree_iter *iter,
1875                                        struct bpos end_pos)
1876 {
1877         struct bkey_i *k;
1878
1879         if (bpos_lt(iter->path->pos, iter->journal_pos))
1880                 iter->journal_idx = 0;
1881
1882         k = bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1883                                         iter->path->level,
1884                                         iter->path->pos,
1885                                         end_pos,
1886                                         &iter->journal_idx);
1887
1888         iter->journal_pos = k ? k->k.p : end_pos;
1889         return k;
1890 }
1891
1892 static noinline
1893 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1894                                               struct btree_iter *iter)
1895 {
1896         struct bkey_i *k = bch2_btree_journal_peek(trans, iter, iter->path->pos);
1897
1898         if (k) {
1899                 iter->k = k->k;
1900                 return bkey_i_to_s_c(k);
1901         } else {
1902                 return bkey_s_c_null;
1903         }
1904 }
1905
1906 static noinline
1907 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1908                                          struct btree_iter *iter,
1909                                          struct bkey_s_c k)
1910 {
1911         struct bkey_i *next_journal =
1912                 bch2_btree_journal_peek(trans, iter,
1913                                 k.k ? k.k->p : path_l(iter->path)->b->key.k.p);
1914
1915         if (next_journal) {
1916                 iter->k = next_journal->k;
1917                 k = bkey_i_to_s_c(next_journal);
1918         }
1919
1920         return k;
1921 }
1922
1923 /*
1924  * Checks btree key cache for key at iter->pos and returns it if present, or
1925  * bkey_s_c_null:
1926  */
1927 static noinline
1928 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1929 {
1930         struct btree_trans *trans = iter->trans;
1931         struct bch_fs *c = trans->c;
1932         struct bkey u;
1933         struct bkey_s_c k;
1934         int ret;
1935
1936         if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1937             bpos_eq(iter->pos, pos))
1938                 return bkey_s_c_null;
1939
1940         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1941                 return bkey_s_c_null;
1942
1943         if (!iter->key_cache_path)
1944                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1945                                                      iter->flags & BTREE_ITER_INTENT, 0,
1946                                                      iter->flags|BTREE_ITER_CACHED|
1947                                                      BTREE_ITER_CACHED_NOFILL,
1948                                                      _THIS_IP_);
1949
1950         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1951                                         iter->flags & BTREE_ITER_INTENT,
1952                                         btree_iter_ip_allocated(iter));
1953
1954         ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
1955                                          iter->flags|BTREE_ITER_CACHED) ?:
1956                 bch2_btree_path_relock(trans, iter->path, _THIS_IP_);
1957         if (unlikely(ret))
1958                 return bkey_s_c_err(ret);
1959
1960         btree_path_set_should_be_locked(iter->key_cache_path);
1961
1962         k = bch2_btree_path_peek_slot(iter->key_cache_path, &u);
1963         if (k.k && !bkey_err(k)) {
1964                 iter->k = u;
1965                 k.k = &iter->k;
1966         }
1967         return k;
1968 }
1969
1970 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1971 {
1972         struct btree_trans *trans = iter->trans;
1973         struct bkey_i *next_update;
1974         struct bkey_s_c k, k2;
1975         int ret;
1976
1977         EBUG_ON(iter->path->cached);
1978         bch2_btree_iter_verify(iter);
1979
1980         while (1) {
1981                 struct btree_path_level *l;
1982
1983                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
1984                                         iter->flags & BTREE_ITER_INTENT,
1985                                         btree_iter_ip_allocated(iter));
1986
1987                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1988                 if (unlikely(ret)) {
1989                         /* ensure that iter->k is consistent with iter->pos: */
1990                         bch2_btree_iter_set_pos(iter, iter->pos);
1991                         k = bkey_s_c_err(ret);
1992                         goto out;
1993                 }
1994
1995                 l = path_l(iter->path);
1996
1997                 if (unlikely(!l->b)) {
1998                         /* No btree nodes at requested level: */
1999                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2000                         k = bkey_s_c_null;
2001                         goto out;
2002                 }
2003
2004                 btree_path_set_should_be_locked(iter->path);
2005
2006                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
2007
2008                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2009                     k.k &&
2010                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2011                         k = k2;
2012                         ret = bkey_err(k);
2013                         if (ret) {
2014                                 bch2_btree_iter_set_pos(iter, iter->pos);
2015                                 goto out;
2016                         }
2017                 }
2018
2019                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2020                         k = btree_trans_peek_journal(trans, iter, k);
2021
2022                 next_update = btree_trans_peek_updates(iter);
2023
2024                 if (next_update &&
2025                     bpos_le(next_update->k.p,
2026                             k.k ? k.k->p : l->b->key.k.p)) {
2027                         iter->k = next_update->k;
2028                         k = bkey_i_to_s_c(next_update);
2029                 }
2030
2031                 if (k.k && bkey_deleted(k.k)) {
2032                         /*
2033                          * If we've got a whiteout, and it's after the search
2034                          * key, advance the search key to the whiteout instead
2035                          * of just after the whiteout - it might be a btree
2036                          * whiteout, with a real key at the same position, since
2037                          * in the btree deleted keys sort before non deleted.
2038                          */
2039                         search_key = !bpos_eq(search_key, k.k->p)
2040                                 ? k.k->p
2041                                 : bpos_successor(k.k->p);
2042                         continue;
2043                 }
2044
2045                 if (likely(k.k)) {
2046                         break;
2047                 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2048                         /* Advance to next leaf node: */
2049                         search_key = bpos_successor(l->b->key.k.p);
2050                 } else {
2051                         /* End of btree: */
2052                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2053                         k = bkey_s_c_null;
2054                         goto out;
2055                 }
2056         }
2057 out:
2058         bch2_btree_iter_verify(iter);
2059
2060         return k;
2061 }
2062
2063 /**
2064  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2065  * current position
2066  */
2067 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2068 {
2069         struct btree_trans *trans = iter->trans;
2070         struct bpos search_key = btree_iter_search_key(iter);
2071         struct bkey_s_c k;
2072         struct bpos iter_pos;
2073         int ret;
2074
2075         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2076         EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2077
2078         if (iter->update_path) {
2079                 bch2_path_put_nokeep(trans, iter->update_path,
2080                                      iter->flags & BTREE_ITER_INTENT);
2081                 iter->update_path = NULL;
2082         }
2083
2084         bch2_btree_iter_verify_entry_exit(iter);
2085
2086         while (1) {
2087                 k = __bch2_btree_iter_peek(iter, search_key);
2088                 if (unlikely(!k.k))
2089                         goto end;
2090                 if (unlikely(bkey_err(k)))
2091                         goto out_no_locked;
2092
2093                 /*
2094                  * iter->pos should be mononotically increasing, and always be
2095                  * equal to the key we just returned - except extents can
2096                  * straddle iter->pos:
2097                  */
2098                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2099                         iter_pos = k.k->p;
2100                 else
2101                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2102
2103                 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2104                              ? bkey_gt(iter_pos, end)
2105                              : bkey_ge(iter_pos, end)))
2106                         goto end;
2107
2108                 if (iter->update_path &&
2109                     !bkey_eq(iter->update_path->pos, k.k->p)) {
2110                         bch2_path_put_nokeep(trans, iter->update_path,
2111                                              iter->flags & BTREE_ITER_INTENT);
2112                         iter->update_path = NULL;
2113                 }
2114
2115                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2116                     (iter->flags & BTREE_ITER_INTENT) &&
2117                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2118                     !iter->update_path) {
2119                         struct bpos pos = k.k->p;
2120
2121                         if (pos.snapshot < iter->snapshot) {
2122                                 search_key = bpos_successor(k.k->p);
2123                                 continue;
2124                         }
2125
2126                         pos.snapshot = iter->snapshot;
2127
2128                         /*
2129                          * advance, same as on exit for iter->path, but only up
2130                          * to snapshot
2131                          */
2132                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2133                         iter->update_path = iter->path;
2134
2135                         iter->update_path = bch2_btree_path_set_pos(trans,
2136                                                 iter->update_path, pos,
2137                                                 iter->flags & BTREE_ITER_INTENT,
2138                                                 _THIS_IP_);
2139                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2140                         if (unlikely(ret)) {
2141                                 k = bkey_s_c_err(ret);
2142                                 goto out_no_locked;
2143                         }
2144                 }
2145
2146                 /*
2147                  * We can never have a key in a leaf node at POS_MAX, so
2148                  * we don't have to check these successor() calls:
2149                  */
2150                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2151                     !bch2_snapshot_is_ancestor(trans->c,
2152                                                iter->snapshot,
2153                                                k.k->p.snapshot)) {
2154                         search_key = bpos_successor(k.k->p);
2155                         continue;
2156                 }
2157
2158                 if (bkey_whiteout(k.k) &&
2159                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2160                         search_key = bkey_successor(iter, k.k->p);
2161                         continue;
2162                 }
2163
2164                 break;
2165         }
2166
2167         iter->pos = iter_pos;
2168
2169         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2170                                 iter->flags & BTREE_ITER_INTENT,
2171                                 btree_iter_ip_allocated(iter));
2172
2173         btree_path_set_should_be_locked(iter->path);
2174 out_no_locked:
2175         if (iter->update_path) {
2176                 ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_);
2177                 if (unlikely(ret))
2178                         k = bkey_s_c_err(ret);
2179                 else
2180                         btree_path_set_should_be_locked(iter->update_path);
2181         }
2182
2183         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2184                 iter->pos.snapshot = iter->snapshot;
2185
2186         ret = bch2_btree_iter_verify_ret(iter, k);
2187         if (unlikely(ret)) {
2188                 bch2_btree_iter_set_pos(iter, iter->pos);
2189                 k = bkey_s_c_err(ret);
2190         }
2191
2192         bch2_btree_iter_verify_entry_exit(iter);
2193
2194         return k;
2195 end:
2196         bch2_btree_iter_set_pos(iter, end);
2197         k = bkey_s_c_null;
2198         goto out_no_locked;
2199 }
2200
2201 /**
2202  * bch2_btree_iter_peek_all_levels: returns the first key greater than or equal
2203  * to iterator's current position, returning keys from every level of the btree.
2204  * For keys at different levels of the btree that compare equal, the key from
2205  * the lower level (leaf) is returned first.
2206  */
2207 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *iter)
2208 {
2209         struct btree_trans *trans = iter->trans;
2210         struct bkey_s_c k;
2211         int ret;
2212
2213         EBUG_ON(iter->path->cached);
2214         bch2_btree_iter_verify(iter);
2215         BUG_ON(iter->path->level < iter->min_depth);
2216         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
2217         EBUG_ON(!(iter->flags & BTREE_ITER_ALL_LEVELS));
2218
2219         while (1) {
2220                 iter->path = bch2_btree_path_set_pos(trans, iter->path, iter->pos,
2221                                         iter->flags & BTREE_ITER_INTENT,
2222                                         btree_iter_ip_allocated(iter));
2223
2224                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2225                 if (unlikely(ret)) {
2226                         /* ensure that iter->k is consistent with iter->pos: */
2227                         bch2_btree_iter_set_pos(iter, iter->pos);
2228                         k = bkey_s_c_err(ret);
2229                         goto out_no_locked;
2230                 }
2231
2232                 /* Already at end? */
2233                 if (!btree_path_node(iter->path, iter->path->level)) {
2234                         k = bkey_s_c_null;
2235                         goto out_no_locked;
2236                 }
2237
2238                 k = btree_path_level_peek_all(trans->c,
2239                                 &iter->path->l[iter->path->level], &iter->k);
2240
2241                 /* Check if we should go up to the parent node: */
2242                 if (!k.k ||
2243                     (iter->advanced &&
2244                      bpos_eq(path_l(iter->path)->b->key.k.p, iter->pos))) {
2245                         iter->pos = path_l(iter->path)->b->key.k.p;
2246                         btree_path_set_level_up(trans, iter->path);
2247                         iter->advanced = false;
2248                         continue;
2249                 }
2250
2251                 /*
2252                  * Check if we should go back down to a leaf:
2253                  * If we're not in a leaf node, we only return the current key
2254                  * if it exactly matches iter->pos - otherwise we first have to
2255                  * go back to the leaf:
2256                  */
2257                 if (iter->path->level != iter->min_depth &&
2258                     (iter->advanced ||
2259                      !k.k ||
2260                      !bpos_eq(iter->pos, k.k->p))) {
2261                         btree_path_set_level_down(trans, iter->path, iter->min_depth);
2262                         iter->pos = bpos_successor(iter->pos);
2263                         iter->advanced = false;
2264                         continue;
2265                 }
2266
2267                 /* Check if we should go to the next key: */
2268                 if (iter->path->level == iter->min_depth &&
2269                     iter->advanced &&
2270                     k.k &&
2271                     bpos_eq(iter->pos, k.k->p)) {
2272                         iter->pos = bpos_successor(iter->pos);
2273                         iter->advanced = false;
2274                         continue;
2275                 }
2276
2277                 if (iter->advanced &&
2278                     iter->path->level == iter->min_depth &&
2279                     !bpos_eq(k.k->p, iter->pos))
2280                         iter->advanced = false;
2281
2282                 BUG_ON(iter->advanced);
2283                 BUG_ON(!k.k);
2284                 break;
2285         }
2286
2287         iter->pos = k.k->p;
2288         btree_path_set_should_be_locked(iter->path);
2289 out_no_locked:
2290         bch2_btree_iter_verify(iter);
2291
2292         return k;
2293 }
2294
2295 /**
2296  * bch2_btree_iter_next: returns first key greater than iterator's current
2297  * position
2298  */
2299 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2300 {
2301         if (!bch2_btree_iter_advance(iter))
2302                 return bkey_s_c_null;
2303
2304         return bch2_btree_iter_peek(iter);
2305 }
2306
2307 /**
2308  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2309  * iterator's current position
2310  */
2311 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2312 {
2313         struct btree_trans *trans = iter->trans;
2314         struct bpos search_key = iter->pos;
2315         struct btree_path *saved_path = NULL;
2316         struct bkey_s_c k;
2317         struct bkey saved_k;
2318         const struct bch_val *saved_v;
2319         int ret;
2320
2321         EBUG_ON(iter->path->cached || iter->path->level);
2322         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2323
2324         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2325                 return bkey_s_c_err(-EIO);
2326
2327         bch2_btree_iter_verify(iter);
2328         bch2_btree_iter_verify_entry_exit(iter);
2329
2330         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2331                 search_key.snapshot = U32_MAX;
2332
2333         while (1) {
2334                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2335                                                 iter->flags & BTREE_ITER_INTENT,
2336                                                 btree_iter_ip_allocated(iter));
2337
2338                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2339                 if (unlikely(ret)) {
2340                         /* ensure that iter->k is consistent with iter->pos: */
2341                         bch2_btree_iter_set_pos(iter, iter->pos);
2342                         k = bkey_s_c_err(ret);
2343                         goto out_no_locked;
2344                 }
2345
2346                 k = btree_path_level_peek(trans, iter->path,
2347                                           &iter->path->l[0], &iter->k);
2348                 if (!k.k ||
2349                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2350                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2351                      : bpos_gt(k.k->p, search_key)))
2352                         k = btree_path_level_prev(trans, iter->path,
2353                                                   &iter->path->l[0], &iter->k);
2354
2355                 if (likely(k.k)) {
2356                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2357                                 if (k.k->p.snapshot == iter->snapshot)
2358                                         goto got_key;
2359
2360                                 /*
2361                                  * If we have a saved candidate, and we're no
2362                                  * longer at the same _key_ (not pos), return
2363                                  * that candidate
2364                                  */
2365                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2366                                         bch2_path_put_nokeep(trans, iter->path,
2367                                                       iter->flags & BTREE_ITER_INTENT);
2368                                         iter->path = saved_path;
2369                                         saved_path = NULL;
2370                                         iter->k = saved_k;
2371                                         k.v     = saved_v;
2372                                         goto got_key;
2373                                 }
2374
2375                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2376                                                               iter->snapshot,
2377                                                               k.k->p.snapshot)) {
2378                                         if (saved_path)
2379                                                 bch2_path_put_nokeep(trans, saved_path,
2380                                                       iter->flags & BTREE_ITER_INTENT);
2381                                         saved_path = btree_path_clone(trans, iter->path,
2382                                                                 iter->flags & BTREE_ITER_INTENT);
2383                                         saved_k = *k.k;
2384                                         saved_v = k.v;
2385                                 }
2386
2387                                 search_key = bpos_predecessor(k.k->p);
2388                                 continue;
2389                         }
2390 got_key:
2391                         if (bkey_whiteout(k.k) &&
2392                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2393                                 search_key = bkey_predecessor(iter, k.k->p);
2394                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2395                                         search_key.snapshot = U32_MAX;
2396                                 continue;
2397                         }
2398
2399                         break;
2400                 } else if (likely(!bpos_eq(iter->path->l[0].b->data->min_key, POS_MIN))) {
2401                         /* Advance to previous leaf node: */
2402                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2403                 } else {
2404                         /* Start of btree: */
2405                         bch2_btree_iter_set_pos(iter, POS_MIN);
2406                         k = bkey_s_c_null;
2407                         goto out_no_locked;
2408                 }
2409         }
2410
2411         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2412
2413         /* Extents can straddle iter->pos: */
2414         if (bkey_lt(k.k->p, iter->pos))
2415                 iter->pos = k.k->p;
2416
2417         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2418                 iter->pos.snapshot = iter->snapshot;
2419
2420         btree_path_set_should_be_locked(iter->path);
2421 out_no_locked:
2422         if (saved_path)
2423                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2424
2425         bch2_btree_iter_verify_entry_exit(iter);
2426         bch2_btree_iter_verify(iter);
2427
2428         return k;
2429 }
2430
2431 /**
2432  * bch2_btree_iter_prev: returns first key less than iterator's current
2433  * position
2434  */
2435 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2436 {
2437         if (!bch2_btree_iter_rewind(iter))
2438                 return bkey_s_c_null;
2439
2440         return bch2_btree_iter_peek_prev(iter);
2441 }
2442
2443 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2444 {
2445         struct btree_trans *trans = iter->trans;
2446         struct bpos search_key;
2447         struct bkey_s_c k;
2448         int ret;
2449
2450         bch2_btree_iter_verify(iter);
2451         bch2_btree_iter_verify_entry_exit(iter);
2452         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2453         EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2454
2455         /* extents can't span inode numbers: */
2456         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2457             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2458                 if (iter->pos.inode == KEY_INODE_MAX)
2459                         return bkey_s_c_null;
2460
2461                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2462         }
2463
2464         search_key = btree_iter_search_key(iter);
2465         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2466                                         iter->flags & BTREE_ITER_INTENT,
2467                                         btree_iter_ip_allocated(iter));
2468
2469         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2470         if (unlikely(ret)) {
2471                 k = bkey_s_c_err(ret);
2472                 goto out_no_locked;
2473         }
2474
2475         if ((iter->flags & BTREE_ITER_CACHED) ||
2476             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2477                 struct bkey_i *next_update;
2478
2479                 if ((next_update = btree_trans_peek_updates(iter)) &&
2480                     bpos_eq(next_update->k.p, iter->pos)) {
2481                         iter->k = next_update->k;
2482                         k = bkey_i_to_s_c(next_update);
2483                         goto out;
2484                 }
2485
2486                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2487                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2488                         goto out;
2489
2490                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2491                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2492                         if (!bkey_err(k))
2493                                 iter->k = *k.k;
2494                         /* We're not returning a key from iter->path: */
2495                         goto out_no_locked;
2496                 }
2497
2498                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2499                 if (unlikely(!k.k))
2500                         goto out_no_locked;
2501         } else {
2502                 struct bpos next;
2503                 struct bpos end = iter->pos;
2504
2505                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2506                         end.offset = U64_MAX;
2507
2508                 EBUG_ON(iter->path->level);
2509
2510                 if (iter->flags & BTREE_ITER_INTENT) {
2511                         struct btree_iter iter2;
2512
2513                         bch2_trans_copy_iter(&iter2, iter);
2514                         k = bch2_btree_iter_peek_upto(&iter2, end);
2515
2516                         if (k.k && !bkey_err(k)) {
2517                                 iter->k = iter2.k;
2518                                 k.k = &iter->k;
2519                         }
2520                         bch2_trans_iter_exit(trans, &iter2);
2521                 } else {
2522                         struct bpos pos = iter->pos;
2523
2524                         k = bch2_btree_iter_peek_upto(iter, end);
2525                         if (unlikely(bkey_err(k)))
2526                                 bch2_btree_iter_set_pos(iter, pos);
2527                         else
2528                                 iter->pos = pos;
2529                 }
2530
2531                 if (unlikely(bkey_err(k)))
2532                         goto out_no_locked;
2533
2534                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2535
2536                 if (bkey_lt(iter->pos, next)) {
2537                         bkey_init(&iter->k);
2538                         iter->k.p = iter->pos;
2539
2540                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2541                                 bch2_key_resize(&iter->k,
2542                                                 min_t(u64, KEY_SIZE_MAX,
2543                                                       (next.inode == iter->pos.inode
2544                                                        ? next.offset
2545                                                        : KEY_OFFSET_MAX) -
2546                                                       iter->pos.offset));
2547                                 EBUG_ON(!iter->k.size);
2548                         }
2549
2550                         k = (struct bkey_s_c) { &iter->k, NULL };
2551                 }
2552         }
2553 out:
2554         btree_path_set_should_be_locked(iter->path);
2555 out_no_locked:
2556         bch2_btree_iter_verify_entry_exit(iter);
2557         bch2_btree_iter_verify(iter);
2558         ret = bch2_btree_iter_verify_ret(iter, k);
2559         if (unlikely(ret))
2560                 return bkey_s_c_err(ret);
2561
2562         return k;
2563 }
2564
2565 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2566 {
2567         if (!bch2_btree_iter_advance(iter))
2568                 return bkey_s_c_null;
2569
2570         return bch2_btree_iter_peek_slot(iter);
2571 }
2572
2573 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2574 {
2575         if (!bch2_btree_iter_rewind(iter))
2576                 return bkey_s_c_null;
2577
2578         return bch2_btree_iter_peek_slot(iter);
2579 }
2580
2581 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2582 {
2583         struct bkey_s_c k;
2584
2585         while (btree_trans_too_many_iters(iter->trans) ||
2586                (k = bch2_btree_iter_peek_type(iter, iter->flags),
2587                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2588                 bch2_trans_begin(iter->trans);
2589
2590         return k;
2591 }
2592
2593 /* new transactional stuff: */
2594
2595 #ifdef CONFIG_BCACHEFS_DEBUG
2596 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2597 {
2598         struct btree_path *path;
2599         unsigned i;
2600
2601         BUG_ON(trans->nr_sorted != hweight64(trans->paths_allocated));
2602
2603         trans_for_each_path(trans, path) {
2604                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2605                 BUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2606         }
2607
2608         for (i = 0; i < trans->nr_sorted; i++) {
2609                 unsigned idx = trans->sorted[i];
2610
2611                 EBUG_ON(!(trans->paths_allocated & (1ULL << idx)));
2612                 BUG_ON(trans->paths[idx].sorted_idx != i);
2613         }
2614 }
2615
2616 static void btree_trans_verify_sorted(struct btree_trans *trans)
2617 {
2618         struct btree_path *path, *prev = NULL;
2619         unsigned i;
2620
2621         if (!bch2_debug_check_iterators)
2622                 return;
2623
2624         trans_for_each_path_inorder(trans, path, i) {
2625                 if (prev && btree_path_cmp(prev, path) > 0) {
2626                         __bch2_dump_trans_paths_updates(trans, true);
2627                         panic("trans paths out of order!\n");
2628                 }
2629                 prev = path;
2630         }
2631 }
2632 #else
2633 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2634 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2635 #endif
2636
2637 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2638 {
2639         int i, l = 0, r = trans->nr_sorted, inc = 1;
2640         bool swapped;
2641
2642         btree_trans_verify_sorted_refs(trans);
2643
2644         if (trans->paths_sorted)
2645                 goto out;
2646
2647         /*
2648          * Cocktail shaker sort: this is efficient because iterators will be
2649          * mostly sorted.
2650          */
2651         do {
2652                 swapped = false;
2653
2654                 for (i = inc > 0 ? l : r - 2;
2655                      i + 1 < r && i >= l;
2656                      i += inc) {
2657                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2658                                            trans->paths + trans->sorted[i + 1]) > 0) {
2659                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2660                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2661                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2662                                 swapped = true;
2663                         }
2664                 }
2665
2666                 if (inc > 0)
2667                         --r;
2668                 else
2669                         l++;
2670                 inc = -inc;
2671         } while (swapped);
2672
2673         trans->paths_sorted = true;
2674 out:
2675         btree_trans_verify_sorted(trans);
2676 }
2677
2678 static inline void btree_path_list_remove(struct btree_trans *trans,
2679                                           struct btree_path *path)
2680 {
2681         unsigned i;
2682
2683         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2684 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2685         trans->nr_sorted--;
2686         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2687                                 trans->sorted + path->sorted_idx + 1,
2688                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2689 #else
2690         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2691 #endif
2692         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2693                 trans->paths[trans->sorted[i]].sorted_idx = i;
2694
2695         path->sorted_idx = U8_MAX;
2696 }
2697
2698 static inline void btree_path_list_add(struct btree_trans *trans,
2699                                        struct btree_path *pos,
2700                                        struct btree_path *path)
2701 {
2702         unsigned i;
2703
2704         path->sorted_idx = pos ? pos->sorted_idx + 1 : trans->nr_sorted;
2705
2706 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2707         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2708                               trans->sorted + path->sorted_idx,
2709                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2710         trans->nr_sorted++;
2711         trans->sorted[path->sorted_idx] = path->idx;
2712 #else
2713         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2714 #endif
2715
2716         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2717                 trans->paths[trans->sorted[i]].sorted_idx = i;
2718
2719         btree_trans_verify_sorted_refs(trans);
2720 }
2721
2722 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2723 {
2724         if (iter->update_path)
2725                 bch2_path_put_nokeep(trans, iter->update_path,
2726                               iter->flags & BTREE_ITER_INTENT);
2727         if (iter->path)
2728                 bch2_path_put(trans, iter->path,
2729                               iter->flags & BTREE_ITER_INTENT);
2730         if (iter->key_cache_path)
2731                 bch2_path_put(trans, iter->key_cache_path,
2732                               iter->flags & BTREE_ITER_INTENT);
2733         iter->path = NULL;
2734         iter->update_path = NULL;
2735         iter->key_cache_path = NULL;
2736 }
2737
2738 static inline void bch2_trans_iter_init_inlined(struct btree_trans *trans,
2739                           struct btree_iter *iter,
2740                           unsigned btree_id, struct bpos pos,
2741                           unsigned flags)
2742 {
2743         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2744                                bch2_btree_iter_flags(trans, btree_id, flags),
2745                                _RET_IP_);
2746 }
2747
2748 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2749                           struct btree_iter *iter,
2750                           unsigned btree_id, struct bpos pos,
2751                           unsigned flags)
2752 {
2753         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2754                                bch2_btree_iter_flags(trans, btree_id, flags),
2755                                _RET_IP_);
2756 }
2757
2758 void bch2_trans_node_iter_init(struct btree_trans *trans,
2759                                struct btree_iter *iter,
2760                                enum btree_id btree_id,
2761                                struct bpos pos,
2762                                unsigned locks_want,
2763                                unsigned depth,
2764                                unsigned flags)
2765 {
2766        flags |= BTREE_ITER_NOT_EXTENTS;
2767        flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2768        flags |= BTREE_ITER_ALL_SNAPSHOTS;
2769
2770         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2771                                __bch2_btree_iter_flags(trans, btree_id, flags),
2772                                _RET_IP_);
2773
2774         iter->min_depth = depth;
2775
2776         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2777         BUG_ON(iter->path->level        != depth);
2778         BUG_ON(iter->min_depth          != depth);
2779 }
2780
2781 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2782 {
2783         *dst = *src;
2784         if (src->path)
2785                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2786         if (src->update_path)
2787                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2788         dst->key_cache_path = NULL;
2789 }
2790
2791 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2792 {
2793         unsigned new_top = trans->mem_top + size;
2794         size_t old_bytes = trans->mem_bytes;
2795         size_t new_bytes = roundup_pow_of_two(new_top);
2796         void *new_mem;
2797         void *p;
2798
2799         trans->mem_max = max(trans->mem_max, new_top);
2800
2801         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2802
2803         new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2804         if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2805                 new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2806                 new_bytes = BTREE_TRANS_MEM_MAX;
2807                 kfree(trans->mem);
2808         }
2809
2810         if (!new_mem)
2811                 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2812
2813         trans->mem = new_mem;
2814         trans->mem_bytes = new_bytes;
2815
2816         if (old_bytes) {
2817                 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2818                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2819         }
2820
2821         p = trans->mem + trans->mem_top;
2822         trans->mem_top += size;
2823         memset(p, 0, size);
2824         return p;
2825 }
2826
2827 static noinline void bch2_trans_reset_srcu_lock(struct btree_trans *trans)
2828 {
2829         struct bch_fs *c = trans->c;
2830         struct btree_path *path;
2831
2832         trans_for_each_path(trans, path)
2833                 if (path->cached && !btree_node_locked(path, 0))
2834                         path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2835
2836         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2837         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2838         trans->srcu_lock_time   = jiffies;
2839 }
2840
2841 /**
2842  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2843  * @trans: transaction to reset
2844  *
2845  * While iterating over nodes or updating nodes a attempt to lock a btree node
2846  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2847  * occurs bch2_trans_begin() should be called and the transaction retried.
2848  */
2849 u32 bch2_trans_begin(struct btree_trans *trans)
2850 {
2851         struct btree_path *path;
2852
2853         bch2_trans_reset_updates(trans);
2854
2855         trans->restart_count++;
2856         trans->mem_top                  = 0;
2857
2858         trans_for_each_path(trans, path) {
2859                 path->should_be_locked = false;
2860
2861                 /*
2862                  * If the transaction wasn't restarted, we're presuming to be
2863                  * doing something new: dont keep iterators excpt the ones that
2864                  * are in use - except for the subvolumes btree:
2865                  */
2866                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2867                         path->preserve = false;
2868
2869                 /*
2870                  * XXX: we probably shouldn't be doing this if the transaction
2871                  * was restarted, but currently we still overflow transaction
2872                  * iterators if we do that
2873                  */
2874                 if (!path->ref && !path->preserve)
2875                         __bch2_path_free(trans, path);
2876                 else
2877                         path->preserve = false;
2878         }
2879
2880         if (!trans->restarted &&
2881             (need_resched() ||
2882              local_clock() - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2883                 bch2_trans_unlock(trans);
2884                 cond_resched();
2885                 bch2_trans_relock(trans);
2886         }
2887
2888         if (unlikely(time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2889                 bch2_trans_reset_srcu_lock(trans);
2890
2891         trans->last_begin_ip = _RET_IP_;
2892         if (trans->restarted) {
2893                 bch2_btree_path_traverse_all(trans);
2894                 trans->notrace_relock_fail = false;
2895         }
2896
2897         trans->last_begin_time = local_clock();
2898         return trans->restart_count;
2899 }
2900
2901 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
2902 {
2903         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
2904         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2905         void *p = NULL;
2906
2907         BUG_ON(trans->used_mempool);
2908
2909 #ifdef __KERNEL__
2910         p = this_cpu_xchg(c->btree_paths_bufs->path, NULL);
2911 #endif
2912         if (!p)
2913                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
2914
2915         trans->paths            = p; p += paths_bytes;
2916         trans->updates          = p; p += updates_bytes;
2917 }
2918
2919 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2920
2921 unsigned bch2_trans_get_fn_idx(const char *fn)
2922 {
2923         unsigned i;
2924
2925         for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2926                 if (!bch2_btree_transaction_fns[i] ||
2927                     bch2_btree_transaction_fns[i] == fn) {
2928                         bch2_btree_transaction_fns[i] = fn;
2929                         return i;
2930                 }
2931
2932         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2933         return i;
2934 }
2935
2936 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, unsigned fn_idx)
2937         __acquires(&c->btree_trans_barrier)
2938 {
2939         struct btree_transaction_stats *s;
2940
2941         bch2_assert_btree_nodes_not_locked();
2942
2943         memset(trans, 0, sizeof(*trans));
2944         trans->c                = c;
2945         trans->fn               = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2946                 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2947         trans->last_begin_time  = local_clock();
2948         trans->fn_idx           = fn_idx;
2949         trans->locking_wait.task = current;
2950         trans->journal_replay_not_finished =
2951                 !test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
2952         closure_init_stack(&trans->ref);
2953
2954         bch2_trans_alloc_paths(trans, c);
2955
2956         s = btree_trans_stats(trans);
2957         if (s && s->max_mem) {
2958                 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2959
2960                 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2961
2962                 if (!unlikely(trans->mem)) {
2963                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2964                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2965                 } else {
2966                         trans->mem_bytes = expected_mem_bytes;
2967                 }
2968         }
2969
2970         if (s) {
2971                 trans->nr_max_paths = s->nr_max_paths;
2972                 trans->wb_updates_size = s->wb_updates_size;
2973         }
2974
2975         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2976         trans->srcu_lock_time   = jiffies;
2977
2978         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2979                 struct btree_trans *pos;
2980
2981                 mutex_lock(&c->btree_trans_lock);
2982                 list_for_each_entry(pos, &c->btree_trans_list, list) {
2983                         /*
2984                          * We'd much prefer to be stricter here and completely
2985                          * disallow multiple btree_trans in the same thread -
2986                          * but the data move path calls bch2_write when we
2987                          * already have a btree_trans initialized.
2988                          */
2989                         BUG_ON(trans->locking_wait.task->pid == pos->locking_wait.task->pid &&
2990                                bch2_trans_locked(pos));
2991
2992                         if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
2993                                 list_add_tail(&trans->list, &pos->list);
2994                                 goto list_add_done;
2995                         }
2996                 }
2997                 list_add_tail(&trans->list, &c->btree_trans_list);
2998 list_add_done:
2999                 mutex_unlock(&c->btree_trans_lock);
3000         }
3001 }
3002
3003 static void check_btree_paths_leaked(struct btree_trans *trans)
3004 {
3005 #ifdef CONFIG_BCACHEFS_DEBUG
3006         struct bch_fs *c = trans->c;
3007         struct btree_path *path;
3008
3009         trans_for_each_path(trans, path)
3010                 if (path->ref)
3011                         goto leaked;
3012         return;
3013 leaked:
3014         bch_err(c, "btree paths leaked from %s!", trans->fn);
3015         trans_for_each_path(trans, path)
3016                 if (path->ref)
3017                         printk(KERN_ERR "  btree %s %pS\n",
3018                                bch2_btree_ids[path->btree_id],
3019                                (void *) path->ip_allocated);
3020         /* Be noisy about this: */
3021         bch2_fatal_error(c);
3022 #endif
3023 }
3024
3025 void bch2_trans_exit(struct btree_trans *trans)
3026         __releases(&c->btree_trans_barrier)
3027 {
3028         struct btree_insert_entry *i;
3029         struct bch_fs *c = trans->c;
3030         struct btree_transaction_stats *s = btree_trans_stats(trans);
3031
3032         bch2_trans_unlock(trans);
3033
3034         closure_sync(&trans->ref);
3035
3036         if (s)
3037                 s->max_mem = max(s->max_mem, trans->mem_max);
3038
3039         trans_for_each_update(trans, i)
3040                 __btree_path_put(i->path, true);
3041         trans->nr_updates               = 0;
3042
3043         check_btree_paths_leaked(trans);
3044
3045         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
3046                 mutex_lock(&c->btree_trans_lock);
3047                 list_del(&trans->list);
3048                 mutex_unlock(&c->btree_trans_lock);
3049         }
3050
3051         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3052
3053         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3054
3055         kfree(trans->extra_journal_entries.data);
3056
3057         if (trans->fs_usage_deltas) {
3058                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3059                     REPLICAS_DELTA_LIST_MAX)
3060                         mempool_free(trans->fs_usage_deltas,
3061                                      &c->replicas_delta_pool);
3062                 else
3063                         kfree(trans->fs_usage_deltas);
3064         }
3065
3066         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3067                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3068         else
3069                 kfree(trans->mem);
3070
3071 #ifdef __KERNEL__
3072         /*
3073          * Userspace doesn't have a real percpu implementation:
3074          */
3075         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3076 #endif
3077
3078         if (trans->paths)
3079                 mempool_free(trans->paths, &c->btree_paths_pool);
3080
3081         trans->mem      = (void *) 0x1;
3082         trans->paths    = (void *) 0x1;
3083 }
3084
3085 static void __maybe_unused
3086 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3087                                       struct btree_bkey_cached_common *b)
3088 {
3089         struct six_lock_count c = six_lock_counts(&b->lock);
3090         struct task_struct *owner;
3091         pid_t pid;
3092
3093         rcu_read_lock();
3094         owner = READ_ONCE(b->lock.owner);
3095         pid = owner ? owner->pid : 0;
3096         rcu_read_unlock();
3097
3098         prt_tab(out);
3099         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3100                    b->level, bch2_btree_ids[b->btree_id]);
3101         bch2_bpos_to_text(out, btree_node_pos(b));
3102
3103         prt_tab(out);
3104         prt_printf(out, " locks %u:%u:%u held by pid %u",
3105                    c.n[0], c.n[1], c.n[2], pid);
3106 }
3107
3108 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3109 {
3110         struct btree_path *path;
3111         struct btree_bkey_cached_common *b;
3112         static char lock_types[] = { 'r', 'i', 'w' };
3113         unsigned l;
3114
3115         if (!out->nr_tabstops) {
3116                 printbuf_tabstop_push(out, 16);
3117                 printbuf_tabstop_push(out, 32);
3118         }
3119
3120         prt_printf(out, "%i %s\n", trans->locking_wait.task->pid, trans->fn);
3121
3122         trans_for_each_path(trans, path) {
3123                 if (!path->nodes_locked)
3124                         continue;
3125
3126                 prt_printf(out, "  path %u %c l=%u %s:",
3127                        path->idx,
3128                        path->cached ? 'c' : 'b',
3129                        path->level,
3130                        bch2_btree_ids[path->btree_id]);
3131                 bch2_bpos_to_text(out, path->pos);
3132                 prt_newline(out);
3133
3134                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3135                         if (btree_node_locked(path, l) &&
3136                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3137                                 prt_printf(out, "    %c l=%u ",
3138                                            lock_types[btree_node_locked_type(path, l)], l);
3139                                 bch2_btree_bkey_cached_common_to_text(out, b);
3140                                 prt_newline(out);
3141                         }
3142                 }
3143         }
3144
3145         b = READ_ONCE(trans->locking);
3146         if (b) {
3147                 prt_printf(out, "  blocked for %lluus on",
3148                            div_u64(local_clock() - trans->locking_wait.start_time,
3149                                    1000));
3150                 prt_newline(out);
3151                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3152                 bch2_btree_bkey_cached_common_to_text(out, b);
3153                 prt_newline(out);
3154         }
3155 }
3156
3157 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3158 {
3159         struct btree_transaction_stats *s;
3160
3161         for (s = c->btree_transaction_stats;
3162              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3163              s++) {
3164                 kfree(s->max_paths_text);
3165                 bch2_time_stats_exit(&s->lock_hold_times);
3166         }
3167
3168         if (c->btree_trans_barrier_initialized)
3169                 cleanup_srcu_struct(&c->btree_trans_barrier);
3170         mempool_exit(&c->btree_trans_mem_pool);
3171         mempool_exit(&c->btree_paths_pool);
3172 }
3173
3174 int bch2_fs_btree_iter_init(struct bch_fs *c)
3175 {
3176         struct btree_transaction_stats *s;
3177         unsigned nr = BTREE_ITER_MAX;
3178         int ret;
3179
3180         for (s = c->btree_transaction_stats;
3181              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3182              s++) {
3183                 bch2_time_stats_init(&s->lock_hold_times);
3184                 mutex_init(&s->lock);
3185         }
3186
3187         INIT_LIST_HEAD(&c->btree_trans_list);
3188         mutex_init(&c->btree_trans_lock);
3189
3190         ret   = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3191                         sizeof(struct btree_path) * nr +
3192                         sizeof(struct btree_insert_entry) * nr) ?:
3193                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3194                                           BTREE_TRANS_MEM_MAX) ?:
3195                 init_srcu_struct(&c->btree_trans_barrier);
3196         if (!ret)
3197                 c->btree_trans_barrier_initialized = true;
3198         return ret;
3199 }