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_journal_iter.h"
   9 #include "btree_key_cache.h"
  10 #include "btree_locking.h"
  11 #include "btree_update.h"
  12 #include "debug.h"
  13 #include "error.h"
  14 #include "extents.h"
  15 #include "journal.h"
  16 #include "replicas.h"
  17 #include "snapshot.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 static inline int __btree_path_cmp(const struct btree_path *l,
  39                                    enum btree_id        r_btree_id,
  40                                    bool                 r_cached,
  41                                    struct bpos          r_pos,
  42                                    unsigned             r_level)
  43 {
  44         /*
  45          * Must match lock ordering as defined by __bch2_btree_node_lock:
  46          */
  47         return   cmp_int(l->btree_id,   r_btree_id) ?:
  48                  cmp_int((int) l->cached,       (int) r_cached) ?:
  49                  bpos_cmp(l->pos,       r_pos) ?:
  50                 -cmp_int(l->level,      r_level);
  51 }
  52
  53 static inline int btree_path_cmp(const struct btree_path *l,
  54                                  const struct btree_path *r)
  55 {
  56         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  57 }
  58
  59 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  60 {
  61         /* Are we iterating over keys in all snapshots? */
  62         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  63                 p = bpos_successor(p);
  64         } else {
  65                 p = bpos_nosnap_successor(p);
  66                 p.snapshot = iter->snapshot;
  67         }
  68
  69         return p;
  70 }
  71
  72 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  73 {
  74         /* Are we iterating over keys in all snapshots? */
  75         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  76                 p = bpos_predecessor(p);
  77         } else {
  78                 p = bpos_nosnap_predecessor(p);
  79                 p.snapshot = iter->snapshot;
  80         }
  81
  82         return p;
  83 }
  84
  85 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
  86 {
  87         struct bpos pos = iter->pos;
  88
  89         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
  90             !bkey_eq(pos, POS_MAX))
  91                 pos = bkey_successor(iter, pos);
  92         return pos;
  93 }
  94
  95 static inline bool btree_path_pos_before_node(struct btree_path *path,
  96                                               struct btree *b)
  97 {
  98         return bpos_lt(path->pos, b->data->min_key);
  99 }
 100
 101 static inline bool btree_path_pos_after_node(struct btree_path *path,
 102                                              struct btree *b)
 103 {
 104         return bpos_gt(path->pos, b->key.k.p);
 105 }
 106
 107 static inline bool btree_path_pos_in_node(struct btree_path *path,
 108                                           struct btree *b)
 109 {
 110         return path->btree_id == b->c.btree_id &&
 111                 !btree_path_pos_before_node(path, b) &&
 112                 !btree_path_pos_after_node(path, b);
 113 }
 114
 115 /* Btree iterator: */
 116
 117 #ifdef CONFIG_BCACHEFS_DEBUG
 118
 119 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 120                                           struct btree_path *path)
 121 {
 122         struct bkey_cached *ck;
 123         bool locked = btree_node_locked(path, 0);
 124
 125         if (!bch2_btree_node_relock(trans, path, 0))
 126                 return;
 127
 128         ck = (void *) path->l[0].b;
 129         BUG_ON(ck->key.btree_id != path->btree_id ||
 130                !bkey_eq(ck->key.pos, path->pos));
 131
 132         if (!locked)
 133                 btree_node_unlock(trans, path, 0);
 134 }
 135
 136 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 137                                 struct btree_path *path, unsigned level)
 138 {
 139         struct btree_path_level *l;
 140         struct btree_node_iter tmp;
 141         bool locked;
 142         struct bkey_packed *p, *k;
 143         struct printbuf buf1 = PRINTBUF;
 144         struct printbuf buf2 = PRINTBUF;
 145         struct printbuf buf3 = PRINTBUF;
 146         const char *msg;
 147
 148         if (!bch2_debug_check_iterators)
 149                 return;
 150
 151         l       = &path->l[level];
 152         tmp     = l->iter;
 153         locked  = btree_node_locked(path, level);
 154
 155         if (path->cached) {
 156                 if (!level)
 157                         bch2_btree_path_verify_cached(trans, path);
 158                 return;
 159         }
 160
 161         if (!btree_path_node(path, level))
 162                 return;
 163
 164         if (!bch2_btree_node_relock_notrace(trans, path, level))
 165                 return;
 166
 167         BUG_ON(!btree_path_pos_in_node(path, l->b));
 168
 169         bch2_btree_node_iter_verify(&l->iter, l->b);
 170
 171         /*
 172          * For interior nodes, the iterator will have skipped past deleted keys:
 173          */
 174         p = level
 175                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 176                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 177         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 178
 179         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 180                 msg = "before";
 181                 goto err;
 182         }
 183
 184         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 185                 msg = "after";
 186                 goto err;
 187         }
 188
 189         if (!locked)
 190                 btree_node_unlock(trans, path, level);
 191         return;
 192 err:
 193         bch2_bpos_to_text(&buf1, path->pos);
 194
 195         if (p) {
 196                 struct bkey uk = bkey_unpack_key(l->b, p);
 197
 198                 bch2_bkey_to_text(&buf2, &uk);
 199         } else {
 200                 prt_printf(&buf2, "(none)");
 201         }
 202
 203         if (k) {
 204                 struct bkey uk = bkey_unpack_key(l->b, k);
 205
 206                 bch2_bkey_to_text(&buf3, &uk);
 207         } else {
 208                 prt_printf(&buf3, "(none)");
 209         }
 210
 211         panic("path should be %s key at level %u:\n"
 212               "path pos %s\n"
 213               "prev key %s\n"
 214               "cur  key %s\n",
 215               msg, level, buf1.buf, buf2.buf, buf3.buf);
 216 }
 217
 218 static void bch2_btree_path_verify(struct btree_trans *trans,
 219                                    struct btree_path *path)
 220 {
 221         struct bch_fs *c = trans->c;
 222         unsigned i;
 223
 224         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 225
 226         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 227                 if (!path->l[i].b) {
 228                         BUG_ON(!path->cached &&
 229                                bch2_btree_id_root(c, path->btree_id)->b->c.level > i);
 230                         break;
 231                 }
 232
 233                 bch2_btree_path_verify_level(trans, path, i);
 234         }
 235
 236         bch2_btree_path_verify_locks(path);
 237 }
 238
 239 void bch2_trans_verify_paths(struct btree_trans *trans)
 240 {
 241         struct btree_path *path;
 242
 243         trans_for_each_path(trans, path)
 244                 bch2_btree_path_verify(trans, path);
 245 }
 246
 247 static void bch2_btree_iter_verify(struct btree_iter *iter)
 248 {
 249         struct btree_trans *trans = iter->trans;
 250
 251         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 252
 253         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
 254
 255         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 256                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 257
 258         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 259                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 260                !btree_type_has_snapshot_field(iter->btree_id));
 261
 262         if (iter->update_path)
 263                 bch2_btree_path_verify(trans, iter->update_path);
 264         bch2_btree_path_verify(trans, iter->path);
 265 }
 266
 267 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 268 {
 269         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 270                !iter->pos.snapshot);
 271
 272         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 273                iter->pos.snapshot != iter->snapshot);
 274
 275         BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
 276                bkey_gt(iter->pos, iter->k.p));
 277 }
 278
 279 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 280 {
 281         struct btree_trans *trans = iter->trans;
 282         struct btree_iter copy;
 283         struct bkey_s_c prev;
 284         int ret = 0;
 285
 286         if (!bch2_debug_check_iterators)
 287                 return 0;
 288
 289         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 290                 return 0;
 291
 292         if (bkey_err(k) || !k.k)
 293                 return 0;
 294
 295         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 296                                           iter->snapshot,
 297                                           k.k->p.snapshot));
 298
 299         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 300                              BTREE_ITER_NOPRESERVE|
 301                              BTREE_ITER_ALL_SNAPSHOTS);
 302         prev = bch2_btree_iter_prev(&copy);
 303         if (!prev.k)
 304                 goto out;
 305
 306         ret = bkey_err(prev);
 307         if (ret)
 308                 goto out;
 309
 310         if (bkey_eq(prev.k->p, k.k->p) &&
 311             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 312                                       prev.k->p.snapshot) > 0) {
 313                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 314
 315                 bch2_bkey_to_text(&buf1, k.k);
 316                 bch2_bkey_to_text(&buf2, prev.k);
 317
 318                 panic("iter snap %u\n"
 319                       "k    %s\n"
 320                       "prev %s\n",
 321                       iter->snapshot,
 322                       buf1.buf, buf2.buf);
 323         }
 324 out:
 325         bch2_trans_iter_exit(trans, &copy);
 326         return ret;
 327 }
 328
 329 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 330                             struct bpos pos, bool key_cache)
 331 {
 332         struct btree_path *path;
 333         unsigned idx;
 334         struct printbuf buf = PRINTBUF;
 335
 336         btree_trans_sort_paths(trans);
 337
 338         trans_for_each_path_inorder(trans, path, idx) {
 339                 int cmp = cmp_int(path->btree_id, id) ?:
 340                         cmp_int(path->cached, key_cache);
 341
 342                 if (cmp > 0)
 343                         break;
 344                 if (cmp < 0)
 345                         continue;
 346
 347                 if (!btree_node_locked(path, 0) ||
 348                     !path->should_be_locked)
 349                         continue;
 350
 351                 if (!key_cache) {
 352                         if (bkey_ge(pos, path->l[0].b->data->min_key) &&
 353                             bkey_le(pos, path->l[0].b->key.k.p))
 354                                 return;
 355                 } else {
 356                         if (bkey_eq(pos, path->pos))
 357                                 return;
 358                 }
 359         }
 360
 361         bch2_dump_trans_paths_updates(trans);
 362         bch2_bpos_to_text(&buf, pos);
 363
 364         panic("not locked: %s %s%s\n",
 365               bch2_btree_id_str(id), buf.buf,
 366               key_cache ? " cached" : "");
 367 }
 368
 369 #else
 370
 371 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 372                                                 struct btree_path *path, unsigned l) {}
 373 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 374                                           struct btree_path *path) {}
 375 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 376 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 377 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 378
 379 #endif
 380
 381 /* Btree path: fixups after btree updates */
 382
 383 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 384                                         struct btree *b,
 385                                         struct bset_tree *t,
 386                                         struct bkey_packed *k)
 387 {
 388         struct btree_node_iter_set *set;
 389
 390         btree_node_iter_for_each(iter, set)
 391                 if (set->end == t->end_offset) {
 392                         set->k = __btree_node_key_to_offset(b, k);
 393                         bch2_btree_node_iter_sort(iter, b);
 394                         return;
 395                 }
 396
 397         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 398 }
 399
 400 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 401                                                struct btree *b,
 402                                                struct bkey_packed *where)
 403 {
 404         struct btree_path_level *l = &path->l[b->c.level];
 405
 406         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 407                 return;
 408
 409         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 410                 bch2_btree_node_iter_advance(&l->iter, l->b);
 411 }
 412
 413 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 414                                       struct btree *b,
 415                                       struct bkey_packed *where)
 416 {
 417         struct btree_path *path;
 418
 419         trans_for_each_path_with_node(trans, b, path) {
 420                 __bch2_btree_path_fix_key_modified(path, b, where);
 421                 bch2_btree_path_verify_level(trans, path, b->c.level);
 422         }
 423 }
 424
 425 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 426                                        struct btree *b,
 427                                        struct btree_node_iter *node_iter,
 428                                        struct bset_tree *t,
 429                                        struct bkey_packed *where,
 430                                        unsigned clobber_u64s,
 431                                        unsigned new_u64s)
 432 {
 433         const struct bkey_packed *end = btree_bkey_last(b, t);
 434         struct btree_node_iter_set *set;
 435         unsigned offset = __btree_node_key_to_offset(b, where);
 436         int shift = new_u64s - clobber_u64s;
 437         unsigned old_end = t->end_offset - shift;
 438         unsigned orig_iter_pos = node_iter->data[0].k;
 439         bool iter_current_key_modified =
 440                 orig_iter_pos >= offset &&
 441                 orig_iter_pos <= offset + clobber_u64s;
 442
 443         btree_node_iter_for_each(node_iter, set)
 444                 if (set->end == old_end)
 445                         goto found;
 446
 447         /* didn't find the bset in the iterator - might have to readd it: */
 448         if (new_u64s &&
 449             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 450                 bch2_btree_node_iter_push(node_iter, b, where, end);
 451                 goto fixup_done;
 452         } else {
 453                 /* Iterator is after key that changed */
 454                 return;
 455         }
 456 found:
 457         set->end = t->end_offset;
 458
 459         /* Iterator hasn't gotten to the key that changed yet: */
 460         if (set->k < offset)
 461                 return;
 462
 463         if (new_u64s &&
 464             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 465                 set->k = offset;
 466         } else if (set->k < offset + clobber_u64s) {
 467                 set->k = offset + new_u64s;
 468                 if (set->k == set->end)
 469                         bch2_btree_node_iter_set_drop(node_iter, set);
 470         } else {
 471                 /* Iterator is after key that changed */
 472                 set->k = (int) set->k + shift;
 473                 return;
 474         }
 475
 476         bch2_btree_node_iter_sort(node_iter, b);
 477 fixup_done:
 478         if (node_iter->data[0].k != orig_iter_pos)
 479                 iter_current_key_modified = true;
 480
 481         /*
 482          * When a new key is added, and the node iterator now points to that
 483          * key, the iterator might have skipped past deleted keys that should
 484          * come after the key the iterator now points to. We have to rewind to
 485          * before those deleted keys - otherwise
 486          * bch2_btree_node_iter_prev_all() breaks:
 487          */
 488         if (!bch2_btree_node_iter_end(node_iter) &&
 489             iter_current_key_modified &&
 490             b->c.level) {
 491                 struct bkey_packed *k, *k2, *p;
 492
 493                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 494
 495                 for_each_bset(b, t) {
 496                         bool set_pos = false;
 497
 498                         if (node_iter->data[0].end == t->end_offset)
 499                                 continue;
 500
 501                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 502
 503                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 504                                bkey_iter_cmp(b, k, p) < 0) {
 505                                 k2 = p;
 506                                 set_pos = true;
 507                         }
 508
 509                         if (set_pos)
 510                                 btree_node_iter_set_set_pos(node_iter,
 511                                                             b, t, k2);
 512                 }
 513         }
 514 }
 515
 516 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 517                               struct btree_path *path,
 518                               struct btree *b,
 519                               struct btree_node_iter *node_iter,
 520                               struct bkey_packed *where,
 521                               unsigned clobber_u64s,
 522                               unsigned new_u64s)
 523 {
 524         struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
 525         struct btree_path *linked;
 526
 527         if (node_iter != &path->l[b->c.level].iter) {
 528                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 529                                            where, clobber_u64s, new_u64s);
 530
 531                 if (bch2_debug_check_iterators)
 532                         bch2_btree_node_iter_verify(node_iter, b);
 533         }
 534
 535         trans_for_each_path_with_node(trans, b, linked) {
 536                 __bch2_btree_node_iter_fix(linked, b,
 537                                            &linked->l[b->c.level].iter, t,
 538                                            where, clobber_u64s, new_u64s);
 539                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 540         }
 541 }
 542
 543 /* Btree path level: pointer to a particular btree node and node iter */
 544
 545 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
 546                                                   struct btree_path_level *l,
 547                                                   struct bkey *u,
 548                                                   struct bkey_packed *k)
 549 {
 550         if (unlikely(!k)) {
 551                 /*
 552                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 553                  * a hole
 554                  */
 555                 u->type = KEY_TYPE_deleted;
 556                 return bkey_s_c_null;
 557         }
 558
 559         return bkey_disassemble(l->b, k, u);
 560 }
 561
 562 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
 563                                                         struct btree_path_level *l,
 564                                                         struct bkey *u)
 565 {
 566         return __btree_iter_unpack(c, l, u,
 567                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 568 }
 569
 570 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
 571                                                     struct btree_path *path,
 572                                                     struct btree_path_level *l,
 573                                                     struct bkey *u)
 574 {
 575         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 576                         bch2_btree_node_iter_peek(&l->iter, l->b));
 577
 578         path->pos = k.k ? k.k->p : l->b->key.k.p;
 579         trans->paths_sorted = false;
 580         bch2_btree_path_verify_level(trans, path, l - path->l);
 581         return k;
 582 }
 583
 584 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
 585                                                     struct btree_path *path,
 586                                                     struct btree_path_level *l,
 587                                                     struct bkey *u)
 588 {
 589         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 590                         bch2_btree_node_iter_prev(&l->iter, l->b));
 591
 592         path->pos = k.k ? k.k->p : l->b->data->min_key;
 593         trans->paths_sorted = false;
 594         bch2_btree_path_verify_level(trans, path, l - path->l);
 595         return k;
 596 }
 597
 598 static inline bool btree_path_advance_to_pos(struct btree_path *path,
 599                                              struct btree_path_level *l,
 600                                              int max_advance)
 601 {
 602         struct bkey_packed *k;
 603         int nr_advanced = 0;
 604
 605         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 606                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 607                 if (max_advance > 0 && nr_advanced >= max_advance)
 608                         return false;
 609
 610                 bch2_btree_node_iter_advance(&l->iter, l->b);
 611                 nr_advanced++;
 612         }
 613
 614         return true;
 615 }
 616
 617 static inline void __btree_path_level_init(struct btree_path *path,
 618                                            unsigned level)
 619 {
 620         struct btree_path_level *l = &path->l[level];
 621
 622         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
 623
 624         /*
 625          * Iterators to interior nodes should always be pointed at the first non
 626          * whiteout:
 627          */
 628         if (level)
 629                 bch2_btree_node_iter_peek(&l->iter, l->b);
 630 }
 631
 632 void bch2_btree_path_level_init(struct btree_trans *trans,
 633                                 struct btree_path *path,
 634                                 struct btree *b)
 635 {
 636         BUG_ON(path->cached);
 637
 638         EBUG_ON(!btree_path_pos_in_node(path, b));
 639
 640         path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 641         path->l[b->c.level].b = b;
 642         __btree_path_level_init(path, b->c.level);
 643 }
 644
 645 /* Btree path: fixups after btree node updates: */
 646
 647 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
 648 {
 649         struct bch_fs *c = trans->c;
 650         struct btree_insert_entry *i;
 651
 652         trans_for_each_update(trans, i)
 653                 if (!i->cached &&
 654                     i->level    == b->c.level &&
 655                     i->btree_id == b->c.btree_id &&
 656                     bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
 657                     bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
 658                         i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 659
 660                         if (unlikely(trans->journal_replay_not_finished)) {
 661                                 struct bkey_i *j_k =
 662                                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 663                                                                     i->k->k.p);
 664
 665                                 if (j_k) {
 666                                         i->old_k = j_k->k;
 667                                         i->old_v = &j_k->v;
 668                                 }
 669                         }
 670                 }
 671 }
 672
 673 /*
 674  * A btree node is being replaced - update the iterator to point to the new
 675  * node:
 676  */
 677 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
 678 {
 679         struct btree_path *path;
 680
 681         trans_for_each_path(trans, path)
 682                 if (path->uptodate == BTREE_ITER_UPTODATE &&
 683                     !path->cached &&
 684                     btree_path_pos_in_node(path, b)) {
 685                         enum btree_node_locked_type t =
 686                                 btree_lock_want(path, b->c.level);
 687
 688                         if (t != BTREE_NODE_UNLOCKED) {
 689                                 btree_node_unlock(trans, path, b->c.level);
 690                                 six_lock_increment(&b->c.lock, (enum six_lock_type) t);
 691                                 mark_btree_node_locked(trans, path, b->c.level, t);
 692                         }
 693
 694                         bch2_btree_path_level_init(trans, path, b);
 695                 }
 696
 697         bch2_trans_revalidate_updates_in_node(trans, b);
 698 }
 699
 700 /*
 701  * A btree node has been modified in such a way as to invalidate iterators - fix
 702  * them:
 703  */
 704 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
 705 {
 706         struct btree_path *path;
 707
 708         trans_for_each_path_with_node(trans, b, path)
 709                 __btree_path_level_init(path, b->c.level);
 710
 711         bch2_trans_revalidate_updates_in_node(trans, b);
 712 }
 713
 714 /* Btree path: traverse, set_pos: */
 715
 716 static inline int btree_path_lock_root(struct btree_trans *trans,
 717                                        struct btree_path *path,
 718                                        unsigned depth_want,
 719                                        unsigned long trace_ip)
 720 {
 721         struct bch_fs *c = trans->c;
 722         struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b;
 723         enum six_lock_type lock_type;
 724         unsigned i;
 725         int ret;
 726
 727         EBUG_ON(path->nodes_locked);
 728
 729         while (1) {
 730                 b = READ_ONCE(*rootp);
 731                 path->level = READ_ONCE(b->c.level);
 732
 733                 if (unlikely(path->level < depth_want)) {
 734                         /*
 735                          * the root is at a lower depth than the depth we want:
 736                          * got to the end of the btree, or we're walking nodes
 737                          * greater than some depth and there are no nodes >=
 738                          * that depth
 739                          */
 740                         path->level = depth_want;
 741                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
 742                                 path->l[i].b = NULL;
 743                         return 1;
 744                 }
 745
 746                 lock_type = __btree_lock_want(path, path->level);
 747                 ret = btree_node_lock(trans, path, &b->c,
 748                                       path->level, lock_type, trace_ip);
 749                 if (unlikely(ret)) {
 750                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
 751                                 continue;
 752                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 753                                 return ret;
 754                         BUG();
 755                 }
 756
 757                 if (likely(b == READ_ONCE(*rootp) &&
 758                            b->c.level == path->level &&
 759                            !race_fault())) {
 760                         for (i = 0; i < path->level; i++)
 761                                 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
 762                         path->l[path->level].b = b;
 763                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
 764                                 path->l[i].b = NULL;
 765
 766                         mark_btree_node_locked(trans, path, path->level,
 767                                                (enum btree_node_locked_type) lock_type);
 768                         bch2_btree_path_level_init(trans, path, b);
 769                         return 0;
 770                 }
 771
 772                 six_unlock_type(&b->c.lock, lock_type);
 773         }
 774 }
 775
 776 noinline
 777 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
 778 {
 779         struct bch_fs *c = trans->c;
 780         struct btree_path_level *l = path_l(path);
 781         struct btree_node_iter node_iter = l->iter;
 782         struct bkey_packed *k;
 783         struct bkey_buf tmp;
 784         unsigned nr = test_bit(BCH_FS_started, &c->flags)
 785                 ? (path->level > 1 ? 0 :  2)
 786                 : (path->level > 1 ? 1 : 16);
 787         bool was_locked = btree_node_locked(path, path->level);
 788         int ret = 0;
 789
 790         bch2_bkey_buf_init(&tmp);
 791
 792         while (nr-- && !ret) {
 793                 if (!bch2_btree_node_relock(trans, path, path->level))
 794                         break;
 795
 796                 bch2_btree_node_iter_advance(&node_iter, l->b);
 797                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
 798                 if (!k)
 799                         break;
 800
 801                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
 802                 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
 803                                                path->level - 1);
 804         }
 805
 806         if (!was_locked)
 807                 btree_node_unlock(trans, path, path->level);
 808
 809         bch2_bkey_buf_exit(&tmp, c);
 810         return ret;
 811 }
 812
 813 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
 814                                  struct btree_and_journal_iter *jiter)
 815 {
 816         struct bch_fs *c = trans->c;
 817         struct bkey_s_c k;
 818         struct bkey_buf tmp;
 819         unsigned nr = test_bit(BCH_FS_started, &c->flags)
 820                 ? (path->level > 1 ? 0 :  2)
 821                 : (path->level > 1 ? 1 : 16);
 822         bool was_locked = btree_node_locked(path, path->level);
 823         int ret = 0;
 824
 825         bch2_bkey_buf_init(&tmp);
 826
 827         while (nr-- && !ret) {
 828                 if (!bch2_btree_node_relock(trans, path, path->level))
 829                         break;
 830
 831                 bch2_btree_and_journal_iter_advance(jiter);
 832                 k = bch2_btree_and_journal_iter_peek(jiter);
 833                 if (!k.k)
 834                         break;
 835
 836                 bch2_bkey_buf_reassemble(&tmp, c, k);
 837                 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
 838                                                path->level - 1);
 839         }
 840
 841         if (!was_locked)
 842                 btree_node_unlock(trans, path, path->level);
 843
 844         bch2_bkey_buf_exit(&tmp, c);
 845         return ret;
 846 }
 847
 848 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
 849                                             struct btree_path *path,
 850                                             unsigned plevel, struct btree *b)
 851 {
 852         struct btree_path_level *l = &path->l[plevel];
 853         bool locked = btree_node_locked(path, plevel);
 854         struct bkey_packed *k;
 855         struct bch_btree_ptr_v2 *bp;
 856
 857         if (!bch2_btree_node_relock(trans, path, plevel))
 858                 return;
 859
 860         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 861         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
 862
 863         bp = (void *) bkeyp_val(&l->b->format, k);
 864         bp->mem_ptr = (unsigned long)b;
 865
 866         if (!locked)
 867                 btree_node_unlock(trans, path, plevel);
 868 }
 869
 870 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
 871                                                      struct btree_path *path,
 872                                                      unsigned flags,
 873                                                      struct bkey_buf *out)
 874 {
 875         struct bch_fs *c = trans->c;
 876         struct btree_path_level *l = path_l(path);
 877         struct btree_and_journal_iter jiter;
 878         struct bkey_s_c k;
 879         int ret = 0;
 880
 881         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
 882
 883         k = bch2_btree_and_journal_iter_peek(&jiter);
 884
 885         bch2_bkey_buf_reassemble(out, c, k);
 886
 887         if (flags & BTREE_ITER_PREFETCH)
 888                 ret = btree_path_prefetch_j(trans, path, &jiter);
 889
 890         bch2_btree_and_journal_iter_exit(&jiter);
 891         return ret;
 892 }
 893
 894 static __always_inline int btree_path_down(struct btree_trans *trans,
 895                                            struct btree_path *path,
 896                                            unsigned flags,
 897                                            unsigned long trace_ip)
 898 {
 899         struct bch_fs *c = trans->c;
 900         struct btree_path_level *l = path_l(path);
 901         struct btree *b;
 902         unsigned level = path->level - 1;
 903         enum six_lock_type lock_type = __btree_lock_want(path, level);
 904         struct bkey_buf tmp;
 905         int ret;
 906
 907         EBUG_ON(!btree_node_locked(path, path->level));
 908
 909         bch2_bkey_buf_init(&tmp);
 910
 911         if (unlikely(trans->journal_replay_not_finished)) {
 912                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
 913                 if (ret)
 914                         goto err;
 915         } else {
 916                 bch2_bkey_buf_unpack(&tmp, c, l->b,
 917                                  bch2_btree_node_iter_peek(&l->iter, l->b));
 918
 919                 if (flags & BTREE_ITER_PREFETCH) {
 920                         ret = btree_path_prefetch(trans, path);
 921                         if (ret)
 922                                 goto err;
 923                 }
 924         }
 925
 926         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
 927         ret = PTR_ERR_OR_ZERO(b);
 928         if (unlikely(ret))
 929                 goto err;
 930
 931         if (likely(!trans->journal_replay_not_finished &&
 932                    tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
 933             unlikely(b != btree_node_mem_ptr(tmp.k)))
 934                 btree_node_mem_ptr_set(trans, path, level + 1, b);
 935
 936         if (btree_node_read_locked(path, level + 1))
 937                 btree_node_unlock(trans, path, level + 1);
 938
 939         mark_btree_node_locked(trans, path, level,
 940                                (enum btree_node_locked_type) lock_type);
 941         path->level = level;
 942         bch2_btree_path_level_init(trans, path, b);
 943
 944         bch2_btree_path_verify_locks(path);
 945 err:
 946         bch2_bkey_buf_exit(&tmp, c);
 947         return ret;
 948 }
 949
 950
 951 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
 952 {
 953         struct bch_fs *c = trans->c;
 954         struct btree_path *path;
 955         unsigned long trace_ip = _RET_IP_;
 956         int i, ret = 0;
 957
 958         if (trans->in_traverse_all)
 959                 return -BCH_ERR_transaction_restart_in_traverse_all;
 960
 961         trans->in_traverse_all = true;
 962 retry_all:
 963         trans->restarted = 0;
 964         trans->last_restarted_ip = 0;
 965
 966         trans_for_each_path(trans, path)
 967                 path->should_be_locked = false;
 968
 969         btree_trans_sort_paths(trans);
 970
 971         bch2_trans_unlock(trans);
 972         cond_resched();
 973
 974         if (unlikely(trans->memory_allocation_failure)) {
 975                 struct closure cl;
 976
 977                 closure_init_stack(&cl);
 978
 979                 do {
 980                         ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
 981                         closure_sync(&cl);
 982                 } while (ret);
 983         }
 984
 985         /* Now, redo traversals in correct order: */
 986         i = 0;
 987         while (i < trans->nr_sorted) {
 988                 path = trans->paths + trans->sorted[i];
 989
 990                 /*
 991                  * Traversing a path can cause another path to be added at about
 992                  * the same position:
 993                  */
 994                 if (path->uptodate) {
 995                         __btree_path_get(path, false);
 996                         ret = bch2_btree_path_traverse_one(trans, path, 0, _THIS_IP_);
 997                         __btree_path_put(path, false);
 998
 999                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1000                             bch2_err_matches(ret, ENOMEM))
1001                                 goto retry_all;
1002                         if (ret)
1003                                 goto err;
1004                 } else {
1005                         i++;
1006                 }
1007         }
1008
1009         /*
1010          * We used to assert that all paths had been traversed here
1011          * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1012          * path->should_be_locked is not set yet, we might have unlocked and
1013          * then failed to relock a path - that's fine.
1014          */
1015 err:
1016         bch2_btree_cache_cannibalize_unlock(trans);
1017
1018         trans->in_traverse_all = false;
1019
1020         trace_and_count(c, trans_traverse_all, trans, trace_ip);
1021         return ret;
1022 }
1023
1024 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1025                                                 unsigned l, int check_pos)
1026 {
1027         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1028                 return false;
1029         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1030                 return false;
1031         return true;
1032 }
1033
1034 static inline bool btree_path_good_node(struct btree_trans *trans,
1035                                         struct btree_path *path,
1036                                         unsigned l, int check_pos)
1037 {
1038         return is_btree_node(path, l) &&
1039                 bch2_btree_node_relock(trans, path, l) &&
1040                 btree_path_check_pos_in_node(path, l, check_pos);
1041 }
1042
1043 static void btree_path_set_level_down(struct btree_trans *trans,
1044                                       struct btree_path *path,
1045                                       unsigned new_level)
1046 {
1047         unsigned l;
1048
1049         path->level = new_level;
1050
1051         for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1052                 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1053                         btree_node_unlock(trans, path, l);
1054
1055         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1056         bch2_btree_path_verify(trans, path);
1057 }
1058
1059 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1060                                                          struct btree_path *path,
1061                                                          int check_pos)
1062 {
1063         unsigned i, l = path->level;
1064 again:
1065         while (btree_path_node(path, l) &&
1066                !btree_path_good_node(trans, path, l, check_pos))
1067                 __btree_path_set_level_up(trans, path, l++);
1068
1069         /* If we need intent locks, take them too: */
1070         for (i = l + 1;
1071              i < path->locks_want && btree_path_node(path, i);
1072              i++)
1073                 if (!bch2_btree_node_relock(trans, path, i)) {
1074                         while (l <= i)
1075                                 __btree_path_set_level_up(trans, path, l++);
1076                         goto again;
1077                 }
1078
1079         return l;
1080 }
1081
1082 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1083                                                      struct btree_path *path,
1084                                                      int check_pos)
1085 {
1086         return likely(btree_node_locked(path, path->level) &&
1087                       btree_path_check_pos_in_node(path, path->level, check_pos))
1088                 ? path->level
1089                 : __btree_path_up_until_good_node(trans, path, check_pos);
1090 }
1091
1092 /*
1093  * This is the main state machine for walking down the btree - walks down to a
1094  * specified depth
1095  *
1096  * Returns 0 on success, -EIO on error (error reading in a btree node).
1097  *
1098  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1099  * stashed in the iterator and returned from bch2_trans_exit().
1100  */
1101 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1102                                  struct btree_path *path,
1103                                  unsigned flags,
1104                                  unsigned long trace_ip)
1105 {
1106         unsigned depth_want = path->level;
1107         int ret = -((int) trans->restarted);
1108
1109         if (unlikely(ret))
1110                 goto out;
1111
1112         if (unlikely(!trans->srcu_held))
1113                 bch2_trans_srcu_lock(trans);
1114
1115         /*
1116          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1117          * and re-traverse the path without a transaction restart:
1118          */
1119         if (path->should_be_locked) {
1120                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1121                 goto out;
1122         }
1123
1124         if (path->cached) {
1125                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1126                 goto out;
1127         }
1128
1129         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1130                 goto out;
1131
1132         path->level = btree_path_up_until_good_node(trans, path, 0);
1133
1134         EBUG_ON(btree_path_node(path, path->level) &&
1135                 !btree_node_locked(path, path->level));
1136
1137         /*
1138          * Note: path->nodes[path->level] may be temporarily NULL here - that
1139          * would indicate to other code that we got to the end of the btree,
1140          * here it indicates that relocking the root failed - it's critical that
1141          * btree_path_lock_root() comes next and that it can't fail
1142          */
1143         while (path->level > depth_want) {
1144                 ret = btree_path_node(path, path->level)
1145                         ? btree_path_down(trans, path, flags, trace_ip)
1146                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1147                 if (unlikely(ret)) {
1148                         if (ret == 1) {
1149                                 /*
1150                                  * No nodes at this level - got to the end of
1151                                  * the btree:
1152                                  */
1153                                 ret = 0;
1154                                 goto out;
1155                         }
1156
1157                         __bch2_btree_path_unlock(trans, path);
1158                         path->level = depth_want;
1159                         path->l[path->level].b = ERR_PTR(ret);
1160                         goto out;
1161                 }
1162         }
1163
1164         path->uptodate = BTREE_ITER_UPTODATE;
1165 out:
1166         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1167                 panic("ret %s (%i) trans->restarted %s (%i)\n",
1168                       bch2_err_str(ret), ret,
1169                       bch2_err_str(trans->restarted), trans->restarted);
1170         bch2_btree_path_verify(trans, path);
1171         return ret;
1172 }
1173
1174 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1175                             struct btree_path *src)
1176 {
1177         unsigned i, offset = offsetof(struct btree_path, pos);
1178
1179         memcpy((void *) dst + offset,
1180                (void *) src + offset,
1181                sizeof(struct btree_path) - offset);
1182
1183         for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1184                 unsigned t = btree_node_locked_type(dst, i);
1185
1186                 if (t != BTREE_NODE_UNLOCKED)
1187                         six_lock_increment(&dst->l[i].b->c.lock, t);
1188         }
1189 }
1190
1191 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1192                                            bool intent)
1193 {
1194         struct btree_path *new = btree_path_alloc(trans, src);
1195
1196         btree_path_copy(trans, new, src);
1197         __btree_path_get(new, intent);
1198         return new;
1199 }
1200
1201 __flatten
1202 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *trans,
1203                          struct btree_path *path, bool intent,
1204                          unsigned long ip)
1205 {
1206         __btree_path_put(path, intent);
1207         path = btree_path_clone(trans, path, intent);
1208         path->preserve = false;
1209         return path;
1210 }
1211
1212 struct btree_path * __must_check
1213 __bch2_btree_path_set_pos(struct btree_trans *trans,
1214                    struct btree_path *path, struct bpos new_pos,
1215                    bool intent, unsigned long ip, int cmp)
1216 {
1217         bch2_trans_verify_not_in_restart(trans);
1218         EBUG_ON(!path->ref);
1219
1220         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1221
1222         path->pos               = new_pos;
1223         trans->paths_sorted     = false;
1224
1225         if (unlikely(path->cached)) {
1226                 btree_node_unlock(trans, path, 0);
1227                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1228                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1229                 goto out;
1230         }
1231
1232         unsigned level = btree_path_up_until_good_node(trans, path, cmp);
1233
1234         if (btree_path_node(path, level)) {
1235                 struct btree_path_level *l = &path->l[level];
1236
1237                 BUG_ON(!btree_node_locked(path, level));
1238                 /*
1239                  * We might have to skip over many keys, or just a few: try
1240                  * advancing the node iterator, and if we have to skip over too
1241                  * many keys just reinit it (or if we're rewinding, since that
1242                  * is expensive).
1243                  */
1244                 if (cmp < 0 ||
1245                     !btree_path_advance_to_pos(path, l, 8))
1246                         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1247
1248                 /*
1249                  * Iterators to interior nodes should always be pointed at the first non
1250                  * whiteout:
1251                  */
1252                 if (unlikely(level))
1253                         bch2_btree_node_iter_peek(&l->iter, l->b);
1254         }
1255
1256         if (unlikely(level != path->level)) {
1257                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1258                 __bch2_btree_path_unlock(trans, path);
1259         }
1260 out:
1261         bch2_btree_path_verify(trans, path);
1262         return path;
1263 }
1264
1265 /* Btree path: main interface: */
1266
1267 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1268 {
1269         struct btree_path *sib;
1270
1271         sib = prev_btree_path(trans, path);
1272         if (sib && !btree_path_cmp(sib, path))
1273                 return sib;
1274
1275         sib = next_btree_path(trans, path);
1276         if (sib && !btree_path_cmp(sib, path))
1277                 return sib;
1278
1279         return NULL;
1280 }
1281
1282 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1283 {
1284         struct btree_path *sib;
1285
1286         sib = prev_btree_path(trans, path);
1287         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1288                 return sib;
1289
1290         sib = next_btree_path(trans, path);
1291         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1292                 return sib;
1293
1294         return NULL;
1295 }
1296
1297 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1298 {
1299         __bch2_btree_path_unlock(trans, path);
1300         btree_path_list_remove(trans, path);
1301         __clear_bit(path->idx, trans->paths_allocated);
1302 }
1303
1304 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1305 {
1306         struct btree_path *dup;
1307
1308         EBUG_ON(trans->paths + path->idx != path);
1309         EBUG_ON(!path->ref);
1310
1311         if (!__btree_path_put(path, intent))
1312                 return;
1313
1314         dup = path->preserve
1315                 ? have_path_at_pos(trans, path)
1316                 : have_node_at_pos(trans, path);
1317
1318         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1319                 return;
1320
1321         if (path->should_be_locked &&
1322             !trans->restarted &&
1323             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1324                 return;
1325
1326         if (dup) {
1327                 dup->preserve           |= path->preserve;
1328                 dup->should_be_locked   |= path->should_be_locked;
1329         }
1330
1331         __bch2_path_free(trans, path);
1332 }
1333
1334 static void bch2_path_put_nokeep(struct btree_trans *trans, struct btree_path *path,
1335                                  bool intent)
1336 {
1337         EBUG_ON(trans->paths + path->idx != path);
1338         EBUG_ON(!path->ref);
1339
1340         if (!__btree_path_put(path, intent))
1341                 return;
1342
1343         __bch2_path_free(trans, path);
1344 }
1345
1346 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1347 {
1348         panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1349               trans->restart_count, restart_count,
1350               (void *) trans->last_begin_ip);
1351 }
1352
1353 void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1354 {
1355         panic("in transaction restart: %s, last restarted by %pS\n",
1356               bch2_err_str(trans->restarted),
1357               (void *) trans->last_restarted_ip);
1358 }
1359
1360 noinline __cold
1361 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1362 {
1363         struct btree_insert_entry *i;
1364         struct btree_write_buffered_key *wb;
1365
1366         prt_printf(buf, "transaction updates for %s journal seq %llu",
1367                trans->fn, trans->journal_res.seq);
1368         prt_newline(buf);
1369         printbuf_indent_add(buf, 2);
1370
1371         trans_for_each_update(trans, i) {
1372                 struct bkey_s_c old = { &i->old_k, i->old_v };
1373
1374                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1375                        bch2_btree_id_str(i->btree_id),
1376                        i->cached,
1377                        (void *) i->ip_allocated);
1378                 prt_newline(buf);
1379
1380                 prt_printf(buf, "  old ");
1381                 bch2_bkey_val_to_text(buf, trans->c, old);
1382                 prt_newline(buf);
1383
1384                 prt_printf(buf, "  new ");
1385                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1386                 prt_newline(buf);
1387         }
1388
1389         trans_for_each_wb_update(trans, wb) {
1390                 prt_printf(buf, "update: btree=%s wb=1 %pS",
1391                        bch2_btree_id_str(wb->btree),
1392                        (void *) i->ip_allocated);
1393                 prt_newline(buf);
1394
1395                 prt_printf(buf, "  new ");
1396                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(&wb->k));
1397                 prt_newline(buf);
1398         }
1399
1400         printbuf_indent_sub(buf, 2);
1401 }
1402
1403 noinline __cold
1404 void bch2_dump_trans_updates(struct btree_trans *trans)
1405 {
1406         struct printbuf buf = PRINTBUF;
1407
1408         bch2_trans_updates_to_text(&buf, trans);
1409         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1410         printbuf_exit(&buf);
1411 }
1412
1413 noinline __cold
1414 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1415 {
1416         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1417                    path->idx, path->ref, path->intent_ref,
1418                    path->preserve ? 'P' : ' ',
1419                    path->should_be_locked ? 'S' : ' ',
1420                    bch2_btree_id_str(path->btree_id),
1421                    path->level);
1422         bch2_bpos_to_text(out, path->pos);
1423
1424         prt_printf(out, " locks %u", path->nodes_locked);
1425 #ifdef TRACK_PATH_ALLOCATED
1426         prt_printf(out, " %pS", (void *) path->ip_allocated);
1427 #endif
1428         prt_newline(out);
1429 }
1430
1431 static noinline __cold
1432 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1433                                 bool nosort)
1434 {
1435         struct btree_path *path;
1436         unsigned idx;
1437
1438         if (!nosort)
1439                 btree_trans_sort_paths(trans);
1440
1441         trans_for_each_path_inorder(trans, path, idx)
1442                 bch2_btree_path_to_text(out, path);
1443 }
1444
1445 noinline __cold
1446 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1447 {
1448         __bch2_trans_paths_to_text(out, trans, false);
1449 }
1450
1451 static noinline __cold
1452 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1453 {
1454         struct printbuf buf = PRINTBUF;
1455
1456         __bch2_trans_paths_to_text(&buf, trans, nosort);
1457         bch2_trans_updates_to_text(&buf, trans);
1458
1459         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1460         printbuf_exit(&buf);
1461 }
1462
1463 noinline __cold
1464 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1465 {
1466         __bch2_dump_trans_paths_updates(trans, false);
1467 }
1468
1469 noinline __cold
1470 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1471 {
1472         struct btree_transaction_stats *s = btree_trans_stats(trans);
1473         struct printbuf buf = PRINTBUF;
1474         size_t nr = bitmap_weight(trans->paths_allocated, BTREE_ITER_MAX);
1475
1476         if (!s)
1477                 return;
1478
1479         bch2_trans_paths_to_text(&buf, trans);
1480
1481         if (!buf.allocation_failure) {
1482                 mutex_lock(&s->lock);
1483                 if (nr > s->nr_max_paths) {
1484                         s->nr_max_paths = nr;
1485                         swap(s->max_paths_text, buf.buf);
1486                 }
1487                 mutex_unlock(&s->lock);
1488         }
1489
1490         printbuf_exit(&buf);
1491
1492         trans->nr_max_paths = nr;
1493 }
1494
1495 noinline __cold
1496 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
1497 {
1498         if (trace_trans_restart_too_many_iters_enabled()) {
1499                 struct printbuf buf = PRINTBUF;
1500
1501                 bch2_trans_paths_to_text(&buf, trans);
1502                 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
1503                 printbuf_exit(&buf);
1504         }
1505
1506         count_event(trans->c, trans_restart_too_many_iters);
1507
1508         return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
1509 }
1510
1511 static noinline void btree_path_overflow(struct btree_trans *trans)
1512 {
1513         bch2_dump_trans_paths_updates(trans);
1514         panic("trans path overflow\n");
1515 }
1516
1517 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1518                                                   struct btree_path *pos)
1519 {
1520         struct btree_path *path;
1521         size_t idx = find_first_zero_bit(trans->paths_allocated, BTREE_ITER_MAX);
1522
1523         if (unlikely(idx == BTREE_ITER_MAX))
1524                 btree_path_overflow(trans);
1525
1526         BUG_ON(idx > BTREE_ITER_MAX);
1527
1528         /*
1529          * Do this before marking the new path as allocated, since it won't be
1530          * initialized yet:
1531          */
1532         if (unlikely(idx > trans->nr_max_paths))
1533                 bch2_trans_update_max_paths(trans);
1534
1535         __set_bit(idx, trans->paths_allocated);
1536
1537         path = &trans->paths[idx];
1538         path->idx               = idx;
1539         path->ref               = 0;
1540         path->intent_ref        = 0;
1541         path->nodes_locked      = 0;
1542         path->alloc_seq++;
1543
1544         btree_path_list_add(trans, pos, path);
1545         trans->paths_sorted = false;
1546         return path;
1547 }
1548
1549 struct btree_path *bch2_path_get(struct btree_trans *trans,
1550                                  enum btree_id btree_id, struct bpos pos,
1551                                  unsigned locks_want, unsigned level,
1552                                  unsigned flags, unsigned long ip)
1553 {
1554         struct btree_path *path, *path_pos = NULL;
1555         bool cached = flags & BTREE_ITER_CACHED;
1556         bool intent = flags & BTREE_ITER_INTENT;
1557         int i;
1558
1559         bch2_trans_verify_not_in_restart(trans);
1560         bch2_trans_verify_locks(trans);
1561
1562         btree_trans_sort_paths(trans);
1563
1564         trans_for_each_path_inorder(trans, path, i) {
1565                 if (__btree_path_cmp(path,
1566                                      btree_id,
1567                                      cached,
1568                                      pos,
1569                                      level) > 0)
1570                         break;
1571
1572                 path_pos = path;
1573         }
1574
1575         if (path_pos &&
1576             path_pos->cached    == cached &&
1577             path_pos->btree_id  == btree_id &&
1578             path_pos->level     == level) {
1579                 __btree_path_get(path_pos, intent);
1580                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1581         } else {
1582                 path = btree_path_alloc(trans, path_pos);
1583                 path_pos = NULL;
1584
1585                 __btree_path_get(path, intent);
1586                 path->pos                       = pos;
1587                 path->btree_id                  = btree_id;
1588                 path->cached                    = cached;
1589                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1590                 path->should_be_locked          = false;
1591                 path->level                     = level;
1592                 path->locks_want                = locks_want;
1593                 path->nodes_locked              = 0;
1594                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1595                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1596 #ifdef TRACK_PATH_ALLOCATED
1597                 path->ip_allocated              = ip;
1598 #endif
1599                 trans->paths_sorted             = false;
1600         }
1601
1602         if (!(flags & BTREE_ITER_NOPRESERVE))
1603                 path->preserve = true;
1604
1605         if (path->intent_ref)
1606                 locks_want = max(locks_want, level + 1);
1607
1608         /*
1609          * If the path has locks_want greater than requested, we don't downgrade
1610          * it here - on transaction restart because btree node split needs to
1611          * upgrade locks, we might be putting/getting the iterator again.
1612          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1613          * a successful transaction commit.
1614          */
1615
1616         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1617         if (locks_want > path->locks_want)
1618                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
1619
1620         return path;
1621 }
1622
1623 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1624 {
1625
1626         struct btree_path_level *l = path_l(path);
1627         struct bkey_packed *_k;
1628         struct bkey_s_c k;
1629
1630         if (unlikely(!l->b))
1631                 return bkey_s_c_null;
1632
1633         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1634         EBUG_ON(!btree_node_locked(path, path->level));
1635
1636         if (!path->cached) {
1637                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1638                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1639
1640                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1641
1642                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1643                         goto hole;
1644         } else {
1645                 struct bkey_cached *ck = (void *) path->l[0].b;
1646
1647                 EBUG_ON(ck &&
1648                         (path->btree_id != ck->key.btree_id ||
1649                          !bkey_eq(path->pos, ck->key.pos)));
1650                 if (!ck || !ck->valid)
1651                         return bkey_s_c_null;
1652
1653                 *u = ck->k->k;
1654                 k = bkey_i_to_s_c(ck->k);
1655         }
1656
1657         return k;
1658 hole:
1659         bkey_init(u);
1660         u->p = path->pos;
1661         return (struct bkey_s_c) { u, NULL };
1662 }
1663
1664 /* Btree iterators: */
1665
1666 int __must_check
1667 __bch2_btree_iter_traverse(struct btree_iter *iter)
1668 {
1669         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1670 }
1671
1672 int __must_check
1673 bch2_btree_iter_traverse(struct btree_iter *iter)
1674 {
1675         int ret;
1676
1677         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1678                                         btree_iter_search_key(iter),
1679                                         iter->flags & BTREE_ITER_INTENT,
1680                                         btree_iter_ip_allocated(iter));
1681
1682         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1683         if (ret)
1684                 return ret;
1685
1686         btree_path_set_should_be_locked(iter->path);
1687         return 0;
1688 }
1689
1690 /* Iterate across nodes (leaf and interior nodes) */
1691
1692 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1693 {
1694         struct btree_trans *trans = iter->trans;
1695         struct btree *b = NULL;
1696         int ret;
1697
1698         EBUG_ON(iter->path->cached);
1699         bch2_btree_iter_verify(iter);
1700
1701         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1702         if (ret)
1703                 goto err;
1704
1705         b = btree_path_node(iter->path, iter->path->level);
1706         if (!b)
1707                 goto out;
1708
1709         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1710
1711         bkey_init(&iter->k);
1712         iter->k.p = iter->pos = b->key.k.p;
1713
1714         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1715                                         iter->flags & BTREE_ITER_INTENT,
1716                                         btree_iter_ip_allocated(iter));
1717         btree_path_set_should_be_locked(iter->path);
1718 out:
1719         bch2_btree_iter_verify_entry_exit(iter);
1720         bch2_btree_iter_verify(iter);
1721
1722         return b;
1723 err:
1724         b = ERR_PTR(ret);
1725         goto out;
1726 }
1727
1728 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1729 {
1730         struct btree *b;
1731
1732         while (b = bch2_btree_iter_peek_node(iter),
1733                bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1734                 bch2_trans_begin(iter->trans);
1735
1736         return b;
1737 }
1738
1739 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1740 {
1741         struct btree_trans *trans = iter->trans;
1742         struct btree_path *path = iter->path;
1743         struct btree *b = NULL;
1744         int ret;
1745
1746         bch2_trans_verify_not_in_restart(trans);
1747         EBUG_ON(iter->path->cached);
1748         bch2_btree_iter_verify(iter);
1749
1750         /* already at end? */
1751         if (!btree_path_node(path, path->level))
1752                 return NULL;
1753
1754         /* got to end? */
1755         if (!btree_path_node(path, path->level + 1)) {
1756                 btree_path_set_level_up(trans, path);
1757                 return NULL;
1758         }
1759
1760         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1761                 __bch2_btree_path_unlock(trans, path);
1762                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1763                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1764                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1765                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1766                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1767                 goto err;
1768         }
1769
1770         b = btree_path_node(path, path->level + 1);
1771
1772         if (bpos_eq(iter->pos, b->key.k.p)) {
1773                 __btree_path_set_level_up(trans, path, path->level++);
1774         } else {
1775                 /*
1776                  * Haven't gotten to the end of the parent node: go back down to
1777                  * the next child node
1778                  */
1779                 path = iter->path =
1780                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1781                                            iter->flags & BTREE_ITER_INTENT,
1782                                            btree_iter_ip_allocated(iter));
1783
1784                 btree_path_set_level_down(trans, path, iter->min_depth);
1785
1786                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1787                 if (ret)
1788                         goto err;
1789
1790                 b = path->l[path->level].b;
1791         }
1792
1793         bkey_init(&iter->k);
1794         iter->k.p = iter->pos = b->key.k.p;
1795
1796         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1797                                         iter->flags & BTREE_ITER_INTENT,
1798                                         btree_iter_ip_allocated(iter));
1799         btree_path_set_should_be_locked(iter->path);
1800         BUG_ON(iter->path->uptodate);
1801 out:
1802         bch2_btree_iter_verify_entry_exit(iter);
1803         bch2_btree_iter_verify(iter);
1804
1805         return b;
1806 err:
1807         b = ERR_PTR(ret);
1808         goto out;
1809 }
1810
1811 /* Iterate across keys (in leaf nodes only) */
1812
1813 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1814 {
1815         struct bpos pos = iter->k.p;
1816         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1817                      ? bpos_eq(pos, SPOS_MAX)
1818                      : bkey_eq(pos, SPOS_MAX));
1819
1820         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1821                 pos = bkey_successor(iter, pos);
1822         bch2_btree_iter_set_pos(iter, pos);
1823         return ret;
1824 }
1825
1826 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1827 {
1828         struct bpos pos = bkey_start_pos(&iter->k);
1829         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1830                      ? bpos_eq(pos, POS_MIN)
1831                      : bkey_eq(pos, POS_MIN));
1832
1833         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1834                 pos = bkey_predecessor(iter, pos);
1835         bch2_btree_iter_set_pos(iter, pos);
1836         return ret;
1837 }
1838
1839 static noinline
1840 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1841 {
1842         struct btree_insert_entry *i;
1843         struct bkey_i *ret = NULL;
1844
1845         trans_for_each_update(iter->trans, i) {
1846                 if (i->btree_id < iter->btree_id)
1847                         continue;
1848                 if (i->btree_id > iter->btree_id)
1849                         break;
1850                 if (bpos_lt(i->k->k.p, iter->path->pos))
1851                         continue;
1852                 if (i->key_cache_already_flushed)
1853                         continue;
1854                 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1855                         ret = i->k;
1856         }
1857
1858         return ret;
1859 }
1860
1861 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1862 {
1863         return iter->flags & BTREE_ITER_WITH_UPDATES
1864                 ? __bch2_btree_trans_peek_updates(iter)
1865                 : NULL;
1866 }
1867
1868 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1869                                               struct btree_iter *iter,
1870                                               struct bpos end_pos)
1871 {
1872         return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1873                                            iter->path->level,
1874                                            iter->path->pos,
1875                                            end_pos,
1876                                            &iter->journal_idx);
1877 }
1878
1879 static noinline
1880 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1881                                               struct btree_iter *iter)
1882 {
1883         struct bkey_i *k = bch2_btree_journal_peek(trans, iter, iter->path->pos);
1884
1885         if (k) {
1886                 iter->k = k->k;
1887                 return bkey_i_to_s_c(k);
1888         } else {
1889                 return bkey_s_c_null;
1890         }
1891 }
1892
1893 static noinline
1894 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1895                                          struct btree_iter *iter,
1896                                          struct bkey_s_c k)
1897 {
1898         struct bkey_i *next_journal =
1899                 bch2_btree_journal_peek(trans, iter,
1900                                 k.k ? k.k->p : path_l(iter->path)->b->key.k.p);
1901
1902         if (next_journal) {
1903                 iter->k = next_journal->k;
1904                 k = bkey_i_to_s_c(next_journal);
1905         }
1906
1907         return k;
1908 }
1909
1910 /*
1911  * Checks btree key cache for key at iter->pos and returns it if present, or
1912  * bkey_s_c_null:
1913  */
1914 static noinline
1915 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1916 {
1917         struct btree_trans *trans = iter->trans;
1918         struct bch_fs *c = trans->c;
1919         struct bkey u;
1920         struct bkey_s_c k;
1921         int ret;
1922
1923         if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1924             bpos_eq(iter->pos, pos))
1925                 return bkey_s_c_null;
1926
1927         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1928                 return bkey_s_c_null;
1929
1930         if (!iter->key_cache_path)
1931                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1932                                                      iter->flags & BTREE_ITER_INTENT, 0,
1933                                                      iter->flags|BTREE_ITER_CACHED|
1934                                                      BTREE_ITER_CACHED_NOFILL,
1935                                                      _THIS_IP_);
1936
1937         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1938                                         iter->flags & BTREE_ITER_INTENT,
1939                                         btree_iter_ip_allocated(iter));
1940
1941         ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
1942                                          iter->flags|BTREE_ITER_CACHED) ?:
1943                 bch2_btree_path_relock(trans, iter->path, _THIS_IP_);
1944         if (unlikely(ret))
1945                 return bkey_s_c_err(ret);
1946
1947         btree_path_set_should_be_locked(iter->key_cache_path);
1948
1949         k = bch2_btree_path_peek_slot(iter->key_cache_path, &u);
1950         if (k.k && !bkey_err(k)) {
1951                 iter->k = u;
1952                 k.k = &iter->k;
1953         }
1954         return k;
1955 }
1956
1957 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1958 {
1959         struct btree_trans *trans = iter->trans;
1960         struct bkey_i *next_update;
1961         struct bkey_s_c k, k2;
1962         int ret;
1963
1964         EBUG_ON(iter->path->cached);
1965         bch2_btree_iter_verify(iter);
1966
1967         while (1) {
1968                 struct btree_path_level *l;
1969
1970                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
1971                                         iter->flags & BTREE_ITER_INTENT,
1972                                         btree_iter_ip_allocated(iter));
1973
1974                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1975                 if (unlikely(ret)) {
1976                         /* ensure that iter->k is consistent with iter->pos: */
1977                         bch2_btree_iter_set_pos(iter, iter->pos);
1978                         k = bkey_s_c_err(ret);
1979                         goto out;
1980                 }
1981
1982                 l = path_l(iter->path);
1983
1984                 if (unlikely(!l->b)) {
1985                         /* No btree nodes at requested level: */
1986                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
1987                         k = bkey_s_c_null;
1988                         goto out;
1989                 }
1990
1991                 btree_path_set_should_be_locked(iter->path);
1992
1993                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
1994
1995                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
1996                     k.k &&
1997                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
1998                         k = k2;
1999                         ret = bkey_err(k);
2000                         if (ret) {
2001                                 bch2_btree_iter_set_pos(iter, iter->pos);
2002                                 goto out;
2003                         }
2004                 }
2005
2006                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2007                         k = btree_trans_peek_journal(trans, iter, k);
2008
2009                 next_update = btree_trans_peek_updates(iter);
2010
2011                 if (next_update &&
2012                     bpos_le(next_update->k.p,
2013                             k.k ? k.k->p : l->b->key.k.p)) {
2014                         iter->k = next_update->k;
2015                         k = bkey_i_to_s_c(next_update);
2016                 }
2017
2018                 if (k.k && bkey_deleted(k.k)) {
2019                         /*
2020                          * If we've got a whiteout, and it's after the search
2021                          * key, advance the search key to the whiteout instead
2022                          * of just after the whiteout - it might be a btree
2023                          * whiteout, with a real key at the same position, since
2024                          * in the btree deleted keys sort before non deleted.
2025                          */
2026                         search_key = !bpos_eq(search_key, k.k->p)
2027                                 ? k.k->p
2028                                 : bpos_successor(k.k->p);
2029                         continue;
2030                 }
2031
2032                 if (likely(k.k)) {
2033                         break;
2034                 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2035                         /* Advance to next leaf node: */
2036                         search_key = bpos_successor(l->b->key.k.p);
2037                 } else {
2038                         /* End of btree: */
2039                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2040                         k = bkey_s_c_null;
2041                         goto out;
2042                 }
2043         }
2044 out:
2045         bch2_btree_iter_verify(iter);
2046
2047         return k;
2048 }
2049
2050 /**
2051  * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
2052  * iterator's current position
2053  * @iter:       iterator to peek from
2054  * @end:        search limit: returns keys less than or equal to @end
2055  *
2056  * Returns:     key if found, or an error extractable with bkey_err().
2057  */
2058 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2059 {
2060         struct btree_trans *trans = iter->trans;
2061         struct bpos search_key = btree_iter_search_key(iter);
2062         struct bkey_s_c k;
2063         struct bpos iter_pos;
2064         int ret;
2065
2066         EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2067
2068         if (iter->update_path) {
2069                 bch2_path_put_nokeep(trans, iter->update_path,
2070                                      iter->flags & BTREE_ITER_INTENT);
2071                 iter->update_path = NULL;
2072         }
2073
2074         bch2_btree_iter_verify_entry_exit(iter);
2075
2076         while (1) {
2077                 k = __bch2_btree_iter_peek(iter, search_key);
2078                 if (unlikely(!k.k))
2079                         goto end;
2080                 if (unlikely(bkey_err(k)))
2081                         goto out_no_locked;
2082
2083                 /*
2084                  * iter->pos should be mononotically increasing, and always be
2085                  * equal to the key we just returned - except extents can
2086                  * straddle iter->pos:
2087                  */
2088                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2089                         iter_pos = k.k->p;
2090                 else
2091                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2092
2093                 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2094                              ? bkey_gt(iter_pos, end)
2095                              : bkey_ge(iter_pos, end)))
2096                         goto end;
2097
2098                 if (iter->update_path &&
2099                     !bkey_eq(iter->update_path->pos, k.k->p)) {
2100                         bch2_path_put_nokeep(trans, iter->update_path,
2101                                              iter->flags & BTREE_ITER_INTENT);
2102                         iter->update_path = NULL;
2103                 }
2104
2105                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2106                     (iter->flags & BTREE_ITER_INTENT) &&
2107                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2108                     !iter->update_path) {
2109                         struct bpos pos = k.k->p;
2110
2111                         if (pos.snapshot < iter->snapshot) {
2112                                 search_key = bpos_successor(k.k->p);
2113                                 continue;
2114                         }
2115
2116                         pos.snapshot = iter->snapshot;
2117
2118                         /*
2119                          * advance, same as on exit for iter->path, but only up
2120                          * to snapshot
2121                          */
2122                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2123                         iter->update_path = iter->path;
2124
2125                         iter->update_path = bch2_btree_path_set_pos(trans,
2126                                                 iter->update_path, pos,
2127                                                 iter->flags & BTREE_ITER_INTENT,
2128                                                 _THIS_IP_);
2129                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2130                         if (unlikely(ret)) {
2131                                 k = bkey_s_c_err(ret);
2132                                 goto out_no_locked;
2133                         }
2134                 }
2135
2136                 /*
2137                  * We can never have a key in a leaf node at POS_MAX, so
2138                  * we don't have to check these successor() calls:
2139                  */
2140                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2141                     !bch2_snapshot_is_ancestor(trans->c,
2142                                                iter->snapshot,
2143                                                k.k->p.snapshot)) {
2144                         search_key = bpos_successor(k.k->p);
2145                         continue;
2146                 }
2147
2148                 if (bkey_whiteout(k.k) &&
2149                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2150                         search_key = bkey_successor(iter, k.k->p);
2151                         continue;
2152                 }
2153
2154                 break;
2155         }
2156
2157         iter->pos = iter_pos;
2158
2159         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2160                                 iter->flags & BTREE_ITER_INTENT,
2161                                 btree_iter_ip_allocated(iter));
2162
2163         btree_path_set_should_be_locked(iter->path);
2164 out_no_locked:
2165         if (iter->update_path) {
2166                 ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_);
2167                 if (unlikely(ret))
2168                         k = bkey_s_c_err(ret);
2169                 else
2170                         btree_path_set_should_be_locked(iter->update_path);
2171         }
2172
2173         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2174                 iter->pos.snapshot = iter->snapshot;
2175
2176         ret = bch2_btree_iter_verify_ret(iter, k);
2177         if (unlikely(ret)) {
2178                 bch2_btree_iter_set_pos(iter, iter->pos);
2179                 k = bkey_s_c_err(ret);
2180         }
2181
2182         bch2_btree_iter_verify_entry_exit(iter);
2183
2184         return k;
2185 end:
2186         bch2_btree_iter_set_pos(iter, end);
2187         k = bkey_s_c_null;
2188         goto out_no_locked;
2189 }
2190
2191 /**
2192  * bch2_btree_iter_next() - returns first key greater than iterator's current
2193  * position
2194  * @iter:       iterator to peek from
2195  *
2196  * Returns:     key if found, or an error extractable with bkey_err().
2197  */
2198 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2199 {
2200         if (!bch2_btree_iter_advance(iter))
2201                 return bkey_s_c_null;
2202
2203         return bch2_btree_iter_peek(iter);
2204 }
2205
2206 /**
2207  * bch2_btree_iter_peek_prev() - returns first key less than or equal to
2208  * iterator's current position
2209  * @iter:       iterator to peek from
2210  *
2211  * Returns:     key if found, or an error extractable with bkey_err().
2212  */
2213 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2214 {
2215         struct btree_trans *trans = iter->trans;
2216         struct bpos search_key = iter->pos;
2217         struct btree_path *saved_path = NULL;
2218         struct bkey_s_c k;
2219         struct bkey saved_k;
2220         const struct bch_val *saved_v;
2221         int ret;
2222
2223         EBUG_ON(iter->path->cached || iter->path->level);
2224         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2225
2226         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2227                 return bkey_s_c_err(-EIO);
2228
2229         bch2_btree_iter_verify(iter);
2230         bch2_btree_iter_verify_entry_exit(iter);
2231
2232         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2233                 search_key.snapshot = U32_MAX;
2234
2235         while (1) {
2236                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2237                                                 iter->flags & BTREE_ITER_INTENT,
2238                                                 btree_iter_ip_allocated(iter));
2239
2240                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2241                 if (unlikely(ret)) {
2242                         /* ensure that iter->k is consistent with iter->pos: */
2243                         bch2_btree_iter_set_pos(iter, iter->pos);
2244                         k = bkey_s_c_err(ret);
2245                         goto out_no_locked;
2246                 }
2247
2248                 k = btree_path_level_peek(trans, iter->path,
2249                                           &iter->path->l[0], &iter->k);
2250                 if (!k.k ||
2251                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2252                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2253                      : bpos_gt(k.k->p, search_key)))
2254                         k = btree_path_level_prev(trans, iter->path,
2255                                                   &iter->path->l[0], &iter->k);
2256
2257                 if (likely(k.k)) {
2258                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2259                                 if (k.k->p.snapshot == iter->snapshot)
2260                                         goto got_key;
2261
2262                                 /*
2263                                  * If we have a saved candidate, and we're no
2264                                  * longer at the same _key_ (not pos), return
2265                                  * that candidate
2266                                  */
2267                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2268                                         bch2_path_put_nokeep(trans, iter->path,
2269                                                       iter->flags & BTREE_ITER_INTENT);
2270                                         iter->path = saved_path;
2271                                         saved_path = NULL;
2272                                         iter->k = saved_k;
2273                                         k.v     = saved_v;
2274                                         goto got_key;
2275                                 }
2276
2277                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2278                                                               iter->snapshot,
2279                                                               k.k->p.snapshot)) {
2280                                         if (saved_path)
2281                                                 bch2_path_put_nokeep(trans, saved_path,
2282                                                       iter->flags & BTREE_ITER_INTENT);
2283                                         saved_path = btree_path_clone(trans, iter->path,
2284                                                                 iter->flags & BTREE_ITER_INTENT);
2285                                         saved_k = *k.k;
2286                                         saved_v = k.v;
2287                                 }
2288
2289                                 search_key = bpos_predecessor(k.k->p);
2290                                 continue;
2291                         }
2292 got_key:
2293                         if (bkey_whiteout(k.k) &&
2294                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2295                                 search_key = bkey_predecessor(iter, k.k->p);
2296                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2297                                         search_key.snapshot = U32_MAX;
2298                                 continue;
2299                         }
2300
2301                         break;
2302                 } else if (likely(!bpos_eq(iter->path->l[0].b->data->min_key, POS_MIN))) {
2303                         /* Advance to previous leaf node: */
2304                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2305                 } else {
2306                         /* Start of btree: */
2307                         bch2_btree_iter_set_pos(iter, POS_MIN);
2308                         k = bkey_s_c_null;
2309                         goto out_no_locked;
2310                 }
2311         }
2312
2313         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2314
2315         /* Extents can straddle iter->pos: */
2316         if (bkey_lt(k.k->p, iter->pos))
2317                 iter->pos = k.k->p;
2318
2319         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2320                 iter->pos.snapshot = iter->snapshot;
2321
2322         btree_path_set_should_be_locked(iter->path);
2323 out_no_locked:
2324         if (saved_path)
2325                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2326
2327         bch2_btree_iter_verify_entry_exit(iter);
2328         bch2_btree_iter_verify(iter);
2329
2330         return k;
2331 }
2332
2333 /**
2334  * bch2_btree_iter_prev() - returns first key less than iterator's current
2335  * position
2336  * @iter:       iterator to peek from
2337  *
2338  * Returns:     key if found, or an error extractable with bkey_err().
2339  */
2340 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2341 {
2342         if (!bch2_btree_iter_rewind(iter))
2343                 return bkey_s_c_null;
2344
2345         return bch2_btree_iter_peek_prev(iter);
2346 }
2347
2348 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2349 {
2350         struct btree_trans *trans = iter->trans;
2351         struct bpos search_key;
2352         struct bkey_s_c k;
2353         int ret;
2354
2355         bch2_btree_iter_verify(iter);
2356         bch2_btree_iter_verify_entry_exit(iter);
2357         EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2358
2359         /* extents can't span inode numbers: */
2360         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2361             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2362                 if (iter->pos.inode == KEY_INODE_MAX)
2363                         return bkey_s_c_null;
2364
2365                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2366         }
2367
2368         search_key = btree_iter_search_key(iter);
2369         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2370                                         iter->flags & BTREE_ITER_INTENT,
2371                                         btree_iter_ip_allocated(iter));
2372
2373         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2374         if (unlikely(ret)) {
2375                 k = bkey_s_c_err(ret);
2376                 goto out_no_locked;
2377         }
2378
2379         if ((iter->flags & BTREE_ITER_CACHED) ||
2380             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2381                 struct bkey_i *next_update;
2382
2383                 if ((next_update = btree_trans_peek_updates(iter)) &&
2384                     bpos_eq(next_update->k.p, iter->pos)) {
2385                         iter->k = next_update->k;
2386                         k = bkey_i_to_s_c(next_update);
2387                         goto out;
2388                 }
2389
2390                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2391                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2392                         goto out;
2393
2394                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2395                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2396                         if (!bkey_err(k))
2397                                 iter->k = *k.k;
2398                         /* We're not returning a key from iter->path: */
2399                         goto out_no_locked;
2400                 }
2401
2402                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2403                 if (unlikely(!k.k))
2404                         goto out_no_locked;
2405         } else {
2406                 struct bpos next;
2407                 struct bpos end = iter->pos;
2408
2409                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2410                         end.offset = U64_MAX;
2411
2412                 EBUG_ON(iter->path->level);
2413
2414                 if (iter->flags & BTREE_ITER_INTENT) {
2415                         struct btree_iter iter2;
2416
2417                         bch2_trans_copy_iter(&iter2, iter);
2418                         k = bch2_btree_iter_peek_upto(&iter2, end);
2419
2420                         if (k.k && !bkey_err(k)) {
2421                                 iter->k = iter2.k;
2422                                 k.k = &iter->k;
2423                         }
2424                         bch2_trans_iter_exit(trans, &iter2);
2425                 } else {
2426                         struct bpos pos = iter->pos;
2427
2428                         k = bch2_btree_iter_peek_upto(iter, end);
2429                         if (unlikely(bkey_err(k)))
2430                                 bch2_btree_iter_set_pos(iter, pos);
2431                         else
2432                                 iter->pos = pos;
2433                 }
2434
2435                 if (unlikely(bkey_err(k)))
2436                         goto out_no_locked;
2437
2438                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2439
2440                 if (bkey_lt(iter->pos, next)) {
2441                         bkey_init(&iter->k);
2442                         iter->k.p = iter->pos;
2443
2444                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2445                                 bch2_key_resize(&iter->k,
2446                                                 min_t(u64, KEY_SIZE_MAX,
2447                                                       (next.inode == iter->pos.inode
2448                                                        ? next.offset
2449                                                        : KEY_OFFSET_MAX) -
2450                                                       iter->pos.offset));
2451                                 EBUG_ON(!iter->k.size);
2452                         }
2453
2454                         k = (struct bkey_s_c) { &iter->k, NULL };
2455                 }
2456         }
2457 out:
2458         btree_path_set_should_be_locked(iter->path);
2459 out_no_locked:
2460         bch2_btree_iter_verify_entry_exit(iter);
2461         bch2_btree_iter_verify(iter);
2462         ret = bch2_btree_iter_verify_ret(iter, k);
2463         if (unlikely(ret))
2464                 return bkey_s_c_err(ret);
2465
2466         return k;
2467 }
2468
2469 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2470 {
2471         if (!bch2_btree_iter_advance(iter))
2472                 return bkey_s_c_null;
2473
2474         return bch2_btree_iter_peek_slot(iter);
2475 }
2476
2477 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2478 {
2479         if (!bch2_btree_iter_rewind(iter))
2480                 return bkey_s_c_null;
2481
2482         return bch2_btree_iter_peek_slot(iter);
2483 }
2484
2485 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2486 {
2487         struct bkey_s_c k;
2488
2489         while (btree_trans_too_many_iters(iter->trans) ||
2490                (k = bch2_btree_iter_peek_type(iter, iter->flags),
2491                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2492                 bch2_trans_begin(iter->trans);
2493
2494         return k;
2495 }
2496
2497 /* new transactional stuff: */
2498
2499 #ifdef CONFIG_BCACHEFS_DEBUG
2500 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2501 {
2502         struct btree_path *path;
2503         unsigned i;
2504
2505         BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, BTREE_ITER_MAX));
2506
2507         trans_for_each_path(trans, path) {
2508                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2509                 BUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2510         }
2511
2512         for (i = 0; i < trans->nr_sorted; i++) {
2513                 unsigned idx = trans->sorted[i];
2514
2515                 BUG_ON(!test_bit(idx, trans->paths_allocated));
2516                 BUG_ON(trans->paths[idx].sorted_idx != i);
2517         }
2518 }
2519
2520 static void btree_trans_verify_sorted(struct btree_trans *trans)
2521 {
2522         struct btree_path *path, *prev = NULL;
2523         unsigned i;
2524
2525         if (!bch2_debug_check_iterators)
2526                 return;
2527
2528         trans_for_each_path_inorder(trans, path, i) {
2529                 if (prev && btree_path_cmp(prev, path) > 0) {
2530                         __bch2_dump_trans_paths_updates(trans, true);
2531                         panic("trans paths out of order!\n");
2532                 }
2533                 prev = path;
2534         }
2535 }
2536 #else
2537 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2538 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2539 #endif
2540
2541 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2542 {
2543         int i, l = 0, r = trans->nr_sorted, inc = 1;
2544         bool swapped;
2545
2546         btree_trans_verify_sorted_refs(trans);
2547
2548         if (trans->paths_sorted)
2549                 goto out;
2550
2551         /*
2552          * Cocktail shaker sort: this is efficient because iterators will be
2553          * mostly sorted.
2554          */
2555         do {
2556                 swapped = false;
2557
2558                 for (i = inc > 0 ? l : r - 2;
2559                      i + 1 < r && i >= l;
2560                      i += inc) {
2561                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2562                                            trans->paths + trans->sorted[i + 1]) > 0) {
2563                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2564                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2565                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2566                                 swapped = true;
2567                         }
2568                 }
2569
2570                 if (inc > 0)
2571                         --r;
2572                 else
2573                         l++;
2574                 inc = -inc;
2575         } while (swapped);
2576
2577         trans->paths_sorted = true;
2578 out:
2579         btree_trans_verify_sorted(trans);
2580 }
2581
2582 static inline void btree_path_list_remove(struct btree_trans *trans,
2583                                           struct btree_path *path)
2584 {
2585         unsigned i;
2586
2587         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2588 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2589         trans->nr_sorted--;
2590         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2591                                 trans->sorted + path->sorted_idx + 1,
2592                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2593 #else
2594         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2595 #endif
2596         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2597                 trans->paths[trans->sorted[i]].sorted_idx = i;
2598
2599         path->sorted_idx = U8_MAX;
2600 }
2601
2602 static inline void btree_path_list_add(struct btree_trans *trans,
2603                                        struct btree_path *pos,
2604                                        struct btree_path *path)
2605 {
2606         unsigned i;
2607
2608         path->sorted_idx = pos ? pos->sorted_idx + 1 : trans->nr_sorted;
2609
2610 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2611         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2612                               trans->sorted + path->sorted_idx,
2613                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2614         trans->nr_sorted++;
2615         trans->sorted[path->sorted_idx] = path->idx;
2616 #else
2617         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2618 #endif
2619
2620         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2621                 trans->paths[trans->sorted[i]].sorted_idx = i;
2622
2623         btree_trans_verify_sorted_refs(trans);
2624 }
2625
2626 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2627 {
2628         if (iter->update_path)
2629                 bch2_path_put_nokeep(trans, iter->update_path,
2630                               iter->flags & BTREE_ITER_INTENT);
2631         if (iter->path)
2632                 bch2_path_put(trans, iter->path,
2633                               iter->flags & BTREE_ITER_INTENT);
2634         if (iter->key_cache_path)
2635                 bch2_path_put(trans, iter->key_cache_path,
2636                               iter->flags & BTREE_ITER_INTENT);
2637         iter->path = NULL;
2638         iter->update_path = NULL;
2639         iter->key_cache_path = NULL;
2640 }
2641
2642 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2643                           struct btree_iter *iter,
2644                           enum btree_id btree_id, struct bpos pos,
2645                           unsigned flags)
2646 {
2647         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2648                                bch2_btree_iter_flags(trans, btree_id, flags),
2649                                _RET_IP_);
2650 }
2651
2652 void bch2_trans_node_iter_init(struct btree_trans *trans,
2653                                struct btree_iter *iter,
2654                                enum btree_id btree_id,
2655                                struct bpos pos,
2656                                unsigned locks_want,
2657                                unsigned depth,
2658                                unsigned flags)
2659 {
2660         flags |= BTREE_ITER_NOT_EXTENTS;
2661         flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2662         flags |= BTREE_ITER_ALL_SNAPSHOTS;
2663
2664         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2665                                __bch2_btree_iter_flags(trans, btree_id, flags),
2666                                _RET_IP_);
2667
2668         iter->min_depth = depth;
2669
2670         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2671         BUG_ON(iter->path->level        != depth);
2672         BUG_ON(iter->min_depth          != depth);
2673 }
2674
2675 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2676 {
2677         *dst = *src;
2678         if (src->path)
2679                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2680         if (src->update_path)
2681                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2682         dst->key_cache_path = NULL;
2683 }
2684
2685 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2686 {
2687         unsigned new_top = trans->mem_top + size;
2688         size_t old_bytes = trans->mem_bytes;
2689         size_t new_bytes = roundup_pow_of_two(new_top);
2690         int ret;
2691         void *new_mem;
2692         void *p;
2693
2694         trans->mem_max = max(trans->mem_max, new_top);
2695
2696         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2697
2698         new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2699         if (unlikely(!new_mem)) {
2700                 bch2_trans_unlock(trans);
2701
2702                 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2703                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2704                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2705                         new_bytes = BTREE_TRANS_MEM_MAX;
2706                         kfree(trans->mem);
2707                 }
2708
2709                 if (!new_mem)
2710                         return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2711
2712                 trans->mem = new_mem;
2713                 trans->mem_bytes = new_bytes;
2714
2715                 ret = bch2_trans_relock(trans);
2716                 if (ret)
2717                         return ERR_PTR(ret);
2718         }
2719
2720         trans->mem = new_mem;
2721         trans->mem_bytes = new_bytes;
2722
2723         if (old_bytes) {
2724                 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2725                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2726         }
2727
2728         p = trans->mem + trans->mem_top;
2729         trans->mem_top += size;
2730         memset(p, 0, size);
2731         return p;
2732 }
2733
2734 static inline void check_srcu_held_too_long(struct btree_trans *trans)
2735 {
2736         WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
2737              "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
2738              (jiffies - trans->srcu_lock_time) / HZ);
2739 }
2740
2741 void bch2_trans_srcu_unlock(struct btree_trans *trans)
2742 {
2743         if (trans->srcu_held) {
2744                 struct bch_fs *c = trans->c;
2745                 struct btree_path *path;
2746
2747                 trans_for_each_path(trans, path)
2748                         if (path->cached && !btree_node_locked(path, 0))
2749                                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2750
2751                 check_srcu_held_too_long(trans);
2752                 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2753                 trans->srcu_held = false;
2754         }
2755 }
2756
2757 void bch2_trans_srcu_lock(struct btree_trans *trans)
2758 {
2759         if (!trans->srcu_held) {
2760                 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
2761                 trans->srcu_lock_time   = jiffies;
2762                 trans->srcu_held = true;
2763         }
2764 }
2765
2766 /**
2767  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2768  * @trans: transaction to reset
2769  *
2770  * Returns:     current restart counter, to be used with trans_was_restarted()
2771  *
2772  * While iterating over nodes or updating nodes a attempt to lock a btree node
2773  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2774  * occurs bch2_trans_begin() should be called and the transaction retried.
2775  */
2776 u32 bch2_trans_begin(struct btree_trans *trans)
2777 {
2778         struct btree_path *path;
2779         u64 now;
2780
2781         bch2_trans_reset_updates(trans);
2782
2783         trans->restart_count++;
2784         trans->mem_top                  = 0;
2785
2786         trans_for_each_path(trans, path) {
2787                 path->should_be_locked = false;
2788
2789                 /*
2790                  * If the transaction wasn't restarted, we're presuming to be
2791                  * doing something new: dont keep iterators excpt the ones that
2792                  * are in use - except for the subvolumes btree:
2793                  */
2794                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2795                         path->preserve = false;
2796
2797                 /*
2798                  * XXX: we probably shouldn't be doing this if the transaction
2799                  * was restarted, but currently we still overflow transaction
2800                  * iterators if we do that
2801                  */
2802                 if (!path->ref && !path->preserve)
2803                         __bch2_path_free(trans, path);
2804                 else
2805                         path->preserve = false;
2806         }
2807
2808         now = local_clock();
2809         if (!trans->restarted &&
2810             (need_resched() ||
2811              now - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2812                 drop_locks_do(trans, (cond_resched(), 0));
2813                 now = local_clock();
2814         }
2815         trans->last_begin_time = now;
2816
2817         if (unlikely(trans->srcu_held &&
2818                      time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2819                 bch2_trans_srcu_unlock(trans);
2820
2821         trans->last_begin_ip = _RET_IP_;
2822         if (trans->restarted) {
2823                 bch2_btree_path_traverse_all(trans);
2824                 trans->notrace_relock_fail = false;
2825         }
2826
2827         return trans->restart_count;
2828 }
2829
2830 static struct btree_trans *bch2_trans_alloc(struct bch_fs *c)
2831 {
2832         struct btree_trans *trans;
2833
2834         if (IS_ENABLED(__KERNEL__)) {
2835                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
2836                 if (trans)
2837                         return trans;
2838         }
2839
2840         trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
2841         /*
2842          * paths need to be zeroed, bch2_check_for_deadlock looks at
2843          * paths in other threads
2844          */
2845         memset(&trans->paths, 0, sizeof(trans->paths));
2846         return trans;
2847 }
2848
2849 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2850
2851 unsigned bch2_trans_get_fn_idx(const char *fn)
2852 {
2853         unsigned i;
2854
2855         for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2856                 if (!bch2_btree_transaction_fns[i] ||
2857                     bch2_btree_transaction_fns[i] == fn) {
2858                         bch2_btree_transaction_fns[i] = fn;
2859                         return i;
2860                 }
2861
2862         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2863         return i;
2864 }
2865
2866 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
2867         __acquires(&c->btree_trans_barrier)
2868 {
2869         struct btree_trans *trans;
2870         struct btree_transaction_stats *s;
2871
2872         trans = bch2_trans_alloc(c);
2873
2874         memset(trans, 0, sizeof(*trans));
2875         trans->c                = c;
2876         trans->fn               = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2877                 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2878         trans->last_begin_time  = local_clock();
2879         trans->fn_idx           = fn_idx;
2880         trans->locking_wait.task = current;
2881         trans->journal_replay_not_finished =
2882                 unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) &&
2883                 atomic_inc_not_zero(&c->journal_keys.ref);
2884         closure_init_stack(&trans->ref);
2885
2886         s = btree_trans_stats(trans);
2887         if (s && s->max_mem) {
2888                 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2889
2890                 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2891
2892                 if (!unlikely(trans->mem)) {
2893                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2894                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2895                 } else {
2896                         trans->mem_bytes = expected_mem_bytes;
2897                 }
2898         }
2899
2900         if (s) {
2901                 trans->nr_max_paths = s->nr_max_paths;
2902                 trans->wb_updates_size = s->wb_updates_size;
2903         }
2904
2905         trans->srcu_idx         = srcu_read_lock(&c->btree_trans_barrier);
2906         trans->srcu_lock_time   = jiffies;
2907         trans->srcu_held        = true;
2908
2909         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2910                 struct btree_trans *pos;
2911
2912                 seqmutex_lock(&c->btree_trans_lock);
2913                 list_for_each_entry(pos, &c->btree_trans_list, list) {
2914                         /*
2915                          * We'd much prefer to be stricter here and completely
2916                          * disallow multiple btree_trans in the same thread -
2917                          * but the data move path calls bch2_write when we
2918                          * already have a btree_trans initialized.
2919                          */
2920                         BUG_ON(trans->locking_wait.task->pid == pos->locking_wait.task->pid &&
2921                                bch2_trans_locked(pos));
2922
2923                         if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
2924                                 list_add_tail(&trans->list, &pos->list);
2925                                 goto list_add_done;
2926                         }
2927                 }
2928                 list_add_tail(&trans->list, &c->btree_trans_list);
2929 list_add_done:
2930                 seqmutex_unlock(&c->btree_trans_lock);
2931         }
2932
2933         return trans;
2934 }
2935
2936 static void check_btree_paths_leaked(struct btree_trans *trans)
2937 {
2938 #ifdef CONFIG_BCACHEFS_DEBUG
2939         struct bch_fs *c = trans->c;
2940         struct btree_path *path;
2941
2942         trans_for_each_path(trans, path)
2943                 if (path->ref)
2944                         goto leaked;
2945         return;
2946 leaked:
2947         bch_err(c, "btree paths leaked from %s!", trans->fn);
2948         trans_for_each_path(trans, path)
2949                 if (path->ref)
2950                         printk(KERN_ERR "  btree %s %pS\n",
2951                                bch2_btree_id_str(path->btree_id),
2952                                (void *) path->ip_allocated);
2953         /* Be noisy about this: */
2954         bch2_fatal_error(c);
2955 #endif
2956 }
2957
2958 void bch2_trans_put(struct btree_trans *trans)
2959         __releases(&c->btree_trans_barrier)
2960 {
2961         struct btree_insert_entry *i;
2962         struct bch_fs *c = trans->c;
2963         struct btree_transaction_stats *s = btree_trans_stats(trans);
2964
2965         bch2_trans_unlock(trans);
2966
2967         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2968                 seqmutex_lock(&c->btree_trans_lock);
2969                 list_del(&trans->list);
2970                 seqmutex_unlock(&c->btree_trans_lock);
2971         }
2972
2973         closure_sync(&trans->ref);
2974
2975         if (s)
2976                 s->max_mem = max(s->max_mem, trans->mem_max);
2977
2978         trans_for_each_update(trans, i)
2979                 __btree_path_put(i->path, true);
2980         trans->nr_updates               = 0;
2981
2982         check_btree_paths_leaked(trans);
2983
2984         if (trans->srcu_held) {
2985                 check_srcu_held_too_long(trans);
2986                 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2987         }
2988
2989         kfree(trans->extra_journal_entries.data);
2990
2991         if (trans->fs_usage_deltas) {
2992                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
2993                     REPLICAS_DELTA_LIST_MAX)
2994                         mempool_free(trans->fs_usage_deltas,
2995                                      &c->replicas_delta_pool);
2996                 else
2997                         kfree(trans->fs_usage_deltas);
2998         }
2999
3000         if (unlikely(trans->journal_replay_not_finished))
3001                 bch2_journal_keys_put(c);
3002
3003         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3004                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3005         else
3006                 kfree(trans->mem);
3007
3008         /* Userspace doesn't have a real percpu implementation: */
3009         if (IS_ENABLED(__KERNEL__))
3010                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
3011         if (trans)
3012                 mempool_free(trans, &c->btree_trans_pool);
3013 }
3014
3015 static void __maybe_unused
3016 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3017                                       struct btree_bkey_cached_common *b)
3018 {
3019         struct six_lock_count c = six_lock_counts(&b->lock);
3020         struct task_struct *owner;
3021         pid_t pid;
3022
3023         rcu_read_lock();
3024         owner = READ_ONCE(b->lock.owner);
3025         pid = owner ? owner->pid : 0;
3026         rcu_read_unlock();
3027
3028         prt_tab(out);
3029         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3030                    b->level, bch2_btree_id_str(b->btree_id));
3031         bch2_bpos_to_text(out, btree_node_pos(b));
3032
3033         prt_tab(out);
3034         prt_printf(out, " locks %u:%u:%u held by pid %u",
3035                    c.n[0], c.n[1], c.n[2], pid);
3036 }
3037
3038 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3039 {
3040         struct btree_path *path;
3041         struct btree_bkey_cached_common *b;
3042         static char lock_types[] = { 'r', 'i', 'w' };
3043         struct task_struct *task = READ_ONCE(trans->locking_wait.task);
3044         unsigned l, idx;
3045
3046         if (!out->nr_tabstops) {
3047                 printbuf_tabstop_push(out, 16);
3048                 printbuf_tabstop_push(out, 32);
3049         }
3050
3051         prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);
3052
3053         trans_for_each_path_safe(trans, path, idx) {
3054                 if (!path->nodes_locked)
3055                         continue;
3056
3057                 prt_printf(out, "  path %u %c l=%u %s:",
3058                        path->idx,
3059                        path->cached ? 'c' : 'b',
3060                        path->level,
3061                        bch2_btree_id_str(path->btree_id));
3062                 bch2_bpos_to_text(out, path->pos);
3063                 prt_newline(out);
3064
3065                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3066                         if (btree_node_locked(path, l) &&
3067                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3068                                 prt_printf(out, "    %c l=%u ",
3069                                            lock_types[btree_node_locked_type(path, l)], l);
3070                                 bch2_btree_bkey_cached_common_to_text(out, b);
3071                                 prt_newline(out);
3072                         }
3073                 }
3074         }
3075
3076         b = READ_ONCE(trans->locking);
3077         if (b) {
3078                 prt_printf(out, "  blocked for %lluus on",
3079                            div_u64(local_clock() - trans->locking_wait.start_time,
3080                                    1000));
3081                 prt_newline(out);
3082                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3083                 bch2_btree_bkey_cached_common_to_text(out, b);
3084                 prt_newline(out);
3085         }
3086 }
3087
3088 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3089 {
3090         struct btree_transaction_stats *s;
3091         struct btree_trans *trans;
3092         int cpu;
3093
3094         trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3095         if (trans)
3096                 panic("%s leaked btree_trans\n", trans->fn);
3097
3098         if (c->btree_trans_bufs)
3099                 for_each_possible_cpu(cpu)
3100                         kfree(per_cpu_ptr(c->btree_trans_bufs, cpu)->trans);
3101         free_percpu(c->btree_trans_bufs);
3102
3103         for (s = c->btree_transaction_stats;
3104              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3105              s++) {
3106                 kfree(s->max_paths_text);
3107                 bch2_time_stats_exit(&s->lock_hold_times);
3108         }
3109
3110         if (c->btree_trans_barrier_initialized)
3111                 cleanup_srcu_struct(&c->btree_trans_barrier);
3112         mempool_exit(&c->btree_trans_mem_pool);
3113         mempool_exit(&c->btree_trans_pool);
3114 }
3115
3116 int bch2_fs_btree_iter_init(struct bch_fs *c)
3117 {
3118         struct btree_transaction_stats *s;
3119         int ret;
3120
3121         for (s = c->btree_transaction_stats;
3122              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3123              s++) {
3124                 bch2_time_stats_init(&s->lock_hold_times);
3125                 mutex_init(&s->lock);
3126         }
3127
3128         INIT_LIST_HEAD(&c->btree_trans_list);
3129         seqmutex_init(&c->btree_trans_lock);
3130
3131         c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3132         if (!c->btree_trans_bufs)
3133                 return -ENOMEM;
3134
3135         ret   = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3136                                           sizeof(struct btree_trans)) ?:
3137                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3138                                           BTREE_TRANS_MEM_MAX) ?:
3139                 init_srcu_struct(&c->btree_trans_barrier);
3140         if (!ret)
3141                 c->btree_trans_barrier_initialized = true;
3142         return ret;
3143 }