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_snapshots(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(c, &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(c);
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         /*
1113          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1114          * and re-traverse the path without a transaction restart:
1115          */
1116         if (path->should_be_locked) {
1117                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1118                 goto out;
1119         }
1120
1121         if (path->cached) {
1122                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1123                 goto out;
1124         }
1125
1126         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1127                 goto out;
1128
1129         path->level = btree_path_up_until_good_node(trans, path, 0);
1130
1131         EBUG_ON(btree_path_node(path, path->level) &&
1132                 !btree_node_locked(path, path->level));
1133
1134         /*
1135          * Note: path->nodes[path->level] may be temporarily NULL here - that
1136          * would indicate to other code that we got to the end of the btree,
1137          * here it indicates that relocking the root failed - it's critical that
1138          * btree_path_lock_root() comes next and that it can't fail
1139          */
1140         while (path->level > depth_want) {
1141                 ret = btree_path_node(path, path->level)
1142                         ? btree_path_down(trans, path, flags, trace_ip)
1143                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1144                 if (unlikely(ret)) {
1145                         if (ret == 1) {
1146                                 /*
1147                                  * No nodes at this level - got to the end of
1148                                  * the btree:
1149                                  */
1150                                 ret = 0;
1151                                 goto out;
1152                         }
1153
1154                         __bch2_btree_path_unlock(trans, path);
1155                         path->level = depth_want;
1156                         path->l[path->level].b = ERR_PTR(ret);
1157                         goto out;
1158                 }
1159         }
1160
1161         path->uptodate = BTREE_ITER_UPTODATE;
1162 out:
1163         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1164                 panic("ret %s (%i) trans->restarted %s (%i)\n",
1165                       bch2_err_str(ret), ret,
1166                       bch2_err_str(trans->restarted), trans->restarted);
1167         bch2_btree_path_verify(trans, path);
1168         return ret;
1169 }
1170
1171 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1172                             struct btree_path *src)
1173 {
1174         unsigned i, offset = offsetof(struct btree_path, pos);
1175
1176         memcpy((void *) dst + offset,
1177                (void *) src + offset,
1178                sizeof(struct btree_path) - offset);
1179
1180         for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1181                 unsigned t = btree_node_locked_type(dst, i);
1182
1183                 if (t != BTREE_NODE_UNLOCKED)
1184                         six_lock_increment(&dst->l[i].b->c.lock, t);
1185         }
1186 }
1187
1188 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1189                                            bool intent)
1190 {
1191         struct btree_path *new = btree_path_alloc(trans, src);
1192
1193         btree_path_copy(trans, new, src);
1194         __btree_path_get(new, intent);
1195         return new;
1196 }
1197
1198 __flatten
1199 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *trans,
1200                          struct btree_path *path, bool intent,
1201                          unsigned long ip)
1202 {
1203         __btree_path_put(path, intent);
1204         path = btree_path_clone(trans, path, intent);
1205         path->preserve = false;
1206         return path;
1207 }
1208
1209 struct btree_path * __must_check
1210 __bch2_btree_path_set_pos(struct btree_trans *trans,
1211                    struct btree_path *path, struct bpos new_pos,
1212                    bool intent, unsigned long ip, int cmp)
1213 {
1214         unsigned level = path->level;
1215
1216         bch2_trans_verify_not_in_restart(trans);
1217         EBUG_ON(!path->ref);
1218
1219         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1220
1221         path->pos               = new_pos;
1222         trans->paths_sorted     = false;
1223
1224         if (unlikely(path->cached)) {
1225                 btree_node_unlock(trans, path, 0);
1226                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1227                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1228                 goto out;
1229         }
1230
1231         level = btree_path_up_until_good_node(trans, path, cmp);
1232
1233         if (btree_path_node(path, level)) {
1234                 struct btree_path_level *l = &path->l[level];
1235
1236                 BUG_ON(!btree_node_locked(path, level));
1237                 /*
1238                  * We might have to skip over many keys, or just a few: try
1239                  * advancing the node iterator, and if we have to skip over too
1240                  * many keys just reinit it (or if we're rewinding, since that
1241                  * is expensive).
1242                  */
1243                 if (cmp < 0 ||
1244                     !btree_path_advance_to_pos(path, l, 8))
1245                         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1246
1247                 /*
1248                  * Iterators to interior nodes should always be pointed at the first non
1249                  * whiteout:
1250                  */
1251                 if (unlikely(level))
1252                         bch2_btree_node_iter_peek(&l->iter, l->b);
1253         }
1254
1255         if (unlikely(level != path->level)) {
1256                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1257                 __bch2_btree_path_unlock(trans, path);
1258         }
1259 out:
1260         bch2_btree_path_verify(trans, path);
1261         return path;
1262 }
1263
1264 /* Btree path: main interface: */
1265
1266 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1267 {
1268         struct btree_path *sib;
1269
1270         sib = prev_btree_path(trans, path);
1271         if (sib && !btree_path_cmp(sib, path))
1272                 return sib;
1273
1274         sib = next_btree_path(trans, path);
1275         if (sib && !btree_path_cmp(sib, path))
1276                 return sib;
1277
1278         return NULL;
1279 }
1280
1281 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1282 {
1283         struct btree_path *sib;
1284
1285         sib = prev_btree_path(trans, path);
1286         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1287                 return sib;
1288
1289         sib = next_btree_path(trans, path);
1290         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1291                 return sib;
1292
1293         return NULL;
1294 }
1295
1296 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1297 {
1298         __bch2_btree_path_unlock(trans, path);
1299         btree_path_list_remove(trans, path);
1300         trans->paths_allocated &= ~(1ULL << path->idx);
1301 }
1302
1303 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1304 {
1305         struct btree_path *dup;
1306
1307         EBUG_ON(trans->paths + path->idx != path);
1308         EBUG_ON(!path->ref);
1309
1310         if (!__btree_path_put(path, intent))
1311                 return;
1312
1313         dup = path->preserve
1314                 ? have_path_at_pos(trans, path)
1315                 : have_node_at_pos(trans, path);
1316
1317         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1318                 return;
1319
1320         if (path->should_be_locked &&
1321             !trans->restarted &&
1322             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1323                 return;
1324
1325         if (dup) {
1326                 dup->preserve           |= path->preserve;
1327                 dup->should_be_locked   |= path->should_be_locked;
1328         }
1329
1330         __bch2_path_free(trans, path);
1331 }
1332
1333 static void bch2_path_put_nokeep(struct btree_trans *trans, struct btree_path *path,
1334                                  bool intent)
1335 {
1336         EBUG_ON(trans->paths + path->idx != path);
1337         EBUG_ON(!path->ref);
1338
1339         if (!__btree_path_put(path, intent))
1340                 return;
1341
1342         __bch2_path_free(trans, path);
1343 }
1344
1345 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1346 {
1347         panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1348               trans->restart_count, restart_count,
1349               (void *) trans->last_begin_ip);
1350 }
1351
1352 void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1353 {
1354         panic("in transaction restart: %s, last restarted by %pS\n",
1355               bch2_err_str(trans->restarted),
1356               (void *) trans->last_restarted_ip);
1357 }
1358
1359 noinline __cold
1360 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1361 {
1362         struct btree_insert_entry *i;
1363         struct btree_write_buffered_key *wb;
1364
1365         prt_printf(buf, "transaction updates for %s journal seq %llu",
1366                trans->fn, trans->journal_res.seq);
1367         prt_newline(buf);
1368         printbuf_indent_add(buf, 2);
1369
1370         trans_for_each_update(trans, i) {
1371                 struct bkey_s_c old = { &i->old_k, i->old_v };
1372
1373                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1374                        bch2_btree_id_str(i->btree_id),
1375                        i->cached,
1376                        (void *) i->ip_allocated);
1377                 prt_newline(buf);
1378
1379                 prt_printf(buf, "  old ");
1380                 bch2_bkey_val_to_text(buf, trans->c, old);
1381                 prt_newline(buf);
1382
1383                 prt_printf(buf, "  new ");
1384                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1385                 prt_newline(buf);
1386         }
1387
1388         trans_for_each_wb_update(trans, wb) {
1389                 prt_printf(buf, "update: btree=%s wb=1 %pS",
1390                        bch2_btree_id_str(wb->btree),
1391                        (void *) i->ip_allocated);
1392                 prt_newline(buf);
1393
1394                 prt_printf(buf, "  new ");
1395                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(&wb->k));
1396                 prt_newline(buf);
1397         }
1398
1399         printbuf_indent_sub(buf, 2);
1400 }
1401
1402 noinline __cold
1403 void bch2_dump_trans_updates(struct btree_trans *trans)
1404 {
1405         struct printbuf buf = PRINTBUF;
1406
1407         bch2_trans_updates_to_text(&buf, trans);
1408         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1409         printbuf_exit(&buf);
1410 }
1411
1412 noinline __cold
1413 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1414 {
1415         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1416                    path->idx, path->ref, path->intent_ref,
1417                    path->preserve ? 'P' : ' ',
1418                    path->should_be_locked ? 'S' : ' ',
1419                    bch2_btree_id_str(path->btree_id),
1420                    path->level);
1421         bch2_bpos_to_text(out, path->pos);
1422
1423         prt_printf(out, " locks %u", path->nodes_locked);
1424 #ifdef TRACK_PATH_ALLOCATED
1425         prt_printf(out, " %pS", (void *) path->ip_allocated);
1426 #endif
1427         prt_newline(out);
1428 }
1429
1430 static noinline __cold
1431 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1432                                 bool nosort)
1433 {
1434         struct btree_path *path;
1435         unsigned idx;
1436
1437         if (!nosort)
1438                 btree_trans_sort_paths(trans);
1439
1440         trans_for_each_path_inorder(trans, path, idx)
1441                 bch2_btree_path_to_text(out, path);
1442 }
1443
1444 noinline __cold
1445 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1446 {
1447         __bch2_trans_paths_to_text(out, trans, false);
1448 }
1449
1450 static noinline __cold
1451 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1452 {
1453         struct printbuf buf = PRINTBUF;
1454
1455         __bch2_trans_paths_to_text(&buf, trans, nosort);
1456         bch2_trans_updates_to_text(&buf, trans);
1457
1458         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1459         printbuf_exit(&buf);
1460 }
1461
1462 noinline __cold
1463 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1464 {
1465         __bch2_dump_trans_paths_updates(trans, false);
1466 }
1467
1468 noinline __cold
1469 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1470 {
1471         struct btree_transaction_stats *s = btree_trans_stats(trans);
1472         struct printbuf buf = PRINTBUF;
1473
1474         if (!s)
1475                 return;
1476
1477         bch2_trans_paths_to_text(&buf, trans);
1478
1479         if (!buf.allocation_failure) {
1480                 mutex_lock(&s->lock);
1481                 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1482                         s->nr_max_paths = trans->nr_max_paths =
1483                                 hweight64(trans->paths_allocated);
1484                         swap(s->max_paths_text, buf.buf);
1485                 }
1486                 mutex_unlock(&s->lock);
1487         }
1488
1489         printbuf_exit(&buf);
1490
1491         trans->nr_max_paths = hweight64(trans->paths_allocated);
1492 }
1493
1494 static noinline void btree_path_overflow(struct btree_trans *trans)
1495 {
1496         bch2_dump_trans_paths_updates(trans);
1497         panic("trans path overflow\n");
1498 }
1499
1500 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1501                                                   struct btree_path *pos)
1502 {
1503         struct btree_path *path;
1504         unsigned idx;
1505
1506         if (unlikely(trans->paths_allocated ==
1507                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
1508                 btree_path_overflow(trans);
1509
1510         idx = __ffs64(~trans->paths_allocated);
1511
1512         /*
1513          * Do this before marking the new path as allocated, since it won't be
1514          * initialized yet:
1515          */
1516         if (unlikely(idx > trans->nr_max_paths))
1517                 bch2_trans_update_max_paths(trans);
1518
1519         trans->paths_allocated |= 1ULL << idx;
1520
1521         path = &trans->paths[idx];
1522         path->idx               = idx;
1523         path->ref               = 0;
1524         path->intent_ref        = 0;
1525         path->nodes_locked      = 0;
1526         path->alloc_seq++;
1527
1528         btree_path_list_add(trans, pos, path);
1529         trans->paths_sorted = false;
1530         return path;
1531 }
1532
1533 struct btree_path *bch2_path_get(struct btree_trans *trans,
1534                                  enum btree_id btree_id, struct bpos pos,
1535                                  unsigned locks_want, unsigned level,
1536                                  unsigned flags, unsigned long ip)
1537 {
1538         struct btree_path *path, *path_pos = NULL;
1539         bool cached = flags & BTREE_ITER_CACHED;
1540         bool intent = flags & BTREE_ITER_INTENT;
1541         int i;
1542
1543         bch2_trans_verify_not_in_restart(trans);
1544         bch2_trans_verify_locks(trans);
1545
1546         btree_trans_sort_paths(trans);
1547
1548         trans_for_each_path_inorder(trans, path, i) {
1549                 if (__btree_path_cmp(path,
1550                                      btree_id,
1551                                      cached,
1552                                      pos,
1553                                      level) > 0)
1554                         break;
1555
1556                 path_pos = path;
1557         }
1558
1559         if (path_pos &&
1560             path_pos->cached    == cached &&
1561             path_pos->btree_id  == btree_id &&
1562             path_pos->level     == level) {
1563                 __btree_path_get(path_pos, intent);
1564                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1565         } else {
1566                 path = btree_path_alloc(trans, path_pos);
1567                 path_pos = NULL;
1568
1569                 __btree_path_get(path, intent);
1570                 path->pos                       = pos;
1571                 path->btree_id                  = btree_id;
1572                 path->cached                    = cached;
1573                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1574                 path->should_be_locked          = false;
1575                 path->level                     = level;
1576                 path->locks_want                = locks_want;
1577                 path->nodes_locked              = 0;
1578                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1579                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1580 #ifdef TRACK_PATH_ALLOCATED
1581                 path->ip_allocated              = ip;
1582 #endif
1583                 trans->paths_sorted             = false;
1584         }
1585
1586         if (!(flags & BTREE_ITER_NOPRESERVE))
1587                 path->preserve = true;
1588
1589         if (path->intent_ref)
1590                 locks_want = max(locks_want, level + 1);
1591
1592         /*
1593          * If the path has locks_want greater than requested, we don't downgrade
1594          * it here - on transaction restart because btree node split needs to
1595          * upgrade locks, we might be putting/getting the iterator again.
1596          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1597          * a successful transaction commit.
1598          */
1599
1600         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1601         if (locks_want > path->locks_want)
1602                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
1603
1604         return path;
1605 }
1606
1607 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1608 {
1609
1610         struct btree_path_level *l = path_l(path);
1611         struct bkey_packed *_k;
1612         struct bkey_s_c k;
1613
1614         if (unlikely(!l->b))
1615                 return bkey_s_c_null;
1616
1617         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1618         EBUG_ON(!btree_node_locked(path, path->level));
1619
1620         if (!path->cached) {
1621                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1622                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1623
1624                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1625
1626                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1627                         goto hole;
1628         } else {
1629                 struct bkey_cached *ck = (void *) path->l[0].b;
1630
1631                 EBUG_ON(ck &&
1632                         (path->btree_id != ck->key.btree_id ||
1633                          !bkey_eq(path->pos, ck->key.pos)));
1634                 if (!ck || !ck->valid)
1635                         return bkey_s_c_null;
1636
1637                 *u = ck->k->k;
1638                 k = bkey_i_to_s_c(ck->k);
1639         }
1640
1641         return k;
1642 hole:
1643         bkey_init(u);
1644         u->p = path->pos;
1645         return (struct bkey_s_c) { u, NULL };
1646 }
1647
1648 /* Btree iterators: */
1649
1650 int __must_check
1651 __bch2_btree_iter_traverse(struct btree_iter *iter)
1652 {
1653         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1654 }
1655
1656 int __must_check
1657 bch2_btree_iter_traverse(struct btree_iter *iter)
1658 {
1659         int ret;
1660
1661         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1662                                         btree_iter_search_key(iter),
1663                                         iter->flags & BTREE_ITER_INTENT,
1664                                         btree_iter_ip_allocated(iter));
1665
1666         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1667         if (ret)
1668                 return ret;
1669
1670         btree_path_set_should_be_locked(iter->path);
1671         return 0;
1672 }
1673
1674 /* Iterate across nodes (leaf and interior nodes) */
1675
1676 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1677 {
1678         struct btree_trans *trans = iter->trans;
1679         struct btree *b = NULL;
1680         int ret;
1681
1682         EBUG_ON(iter->path->cached);
1683         bch2_btree_iter_verify(iter);
1684
1685         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1686         if (ret)
1687                 goto err;
1688
1689         b = btree_path_node(iter->path, iter->path->level);
1690         if (!b)
1691                 goto out;
1692
1693         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1694
1695         bkey_init(&iter->k);
1696         iter->k.p = iter->pos = b->key.k.p;
1697
1698         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1699                                         iter->flags & BTREE_ITER_INTENT,
1700                                         btree_iter_ip_allocated(iter));
1701         btree_path_set_should_be_locked(iter->path);
1702 out:
1703         bch2_btree_iter_verify_entry_exit(iter);
1704         bch2_btree_iter_verify(iter);
1705
1706         return b;
1707 err:
1708         b = ERR_PTR(ret);
1709         goto out;
1710 }
1711
1712 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1713 {
1714         struct btree *b;
1715
1716         while (b = bch2_btree_iter_peek_node(iter),
1717                bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1718                 bch2_trans_begin(iter->trans);
1719
1720         return b;
1721 }
1722
1723 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1724 {
1725         struct btree_trans *trans = iter->trans;
1726         struct btree_path *path = iter->path;
1727         struct btree *b = NULL;
1728         int ret;
1729
1730         bch2_trans_verify_not_in_restart(trans);
1731         EBUG_ON(iter->path->cached);
1732         bch2_btree_iter_verify(iter);
1733
1734         /* already at end? */
1735         if (!btree_path_node(path, path->level))
1736                 return NULL;
1737
1738         /* got to end? */
1739         if (!btree_path_node(path, path->level + 1)) {
1740                 btree_path_set_level_up(trans, path);
1741                 return NULL;
1742         }
1743
1744         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1745                 __bch2_btree_path_unlock(trans, path);
1746                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1747                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1748                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1749                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1750                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1751                 goto err;
1752         }
1753
1754         b = btree_path_node(path, path->level + 1);
1755
1756         if (bpos_eq(iter->pos, b->key.k.p)) {
1757                 __btree_path_set_level_up(trans, path, path->level++);
1758         } else {
1759                 /*
1760                  * Haven't gotten to the end of the parent node: go back down to
1761                  * the next child node
1762                  */
1763                 path = iter->path =
1764                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1765                                            iter->flags & BTREE_ITER_INTENT,
1766                                            btree_iter_ip_allocated(iter));
1767
1768                 btree_path_set_level_down(trans, path, iter->min_depth);
1769
1770                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1771                 if (ret)
1772                         goto err;
1773
1774                 b = path->l[path->level].b;
1775         }
1776
1777         bkey_init(&iter->k);
1778         iter->k.p = iter->pos = b->key.k.p;
1779
1780         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1781                                         iter->flags & BTREE_ITER_INTENT,
1782                                         btree_iter_ip_allocated(iter));
1783         btree_path_set_should_be_locked(iter->path);
1784         BUG_ON(iter->path->uptodate);
1785 out:
1786         bch2_btree_iter_verify_entry_exit(iter);
1787         bch2_btree_iter_verify(iter);
1788
1789         return b;
1790 err:
1791         b = ERR_PTR(ret);
1792         goto out;
1793 }
1794
1795 /* Iterate across keys (in leaf nodes only) */
1796
1797 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1798 {
1799         if (likely(!(iter->flags & BTREE_ITER_ALL_LEVELS))) {
1800                 struct bpos pos = iter->k.p;
1801                 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1802                              ? bpos_eq(pos, SPOS_MAX)
1803                              : bkey_eq(pos, SPOS_MAX));
1804
1805                 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1806                         pos = bkey_successor(iter, pos);
1807                 bch2_btree_iter_set_pos(iter, pos);
1808                 return ret;
1809         } else {
1810                 if (!btree_path_node(iter->path, iter->path->level))
1811                         return true;
1812
1813                 iter->advanced = true;
1814                 return false;
1815         }
1816 }
1817
1818 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1819 {
1820         struct bpos pos = bkey_start_pos(&iter->k);
1821         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1822                      ? bpos_eq(pos, POS_MIN)
1823                      : bkey_eq(pos, POS_MIN));
1824
1825         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1826                 pos = bkey_predecessor(iter, pos);
1827         bch2_btree_iter_set_pos(iter, pos);
1828         return ret;
1829 }
1830
1831 static noinline
1832 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1833 {
1834         struct btree_insert_entry *i;
1835         struct bkey_i *ret = NULL;
1836
1837         trans_for_each_update(iter->trans, i) {
1838                 if (i->btree_id < iter->btree_id)
1839                         continue;
1840                 if (i->btree_id > iter->btree_id)
1841                         break;
1842                 if (bpos_lt(i->k->k.p, iter->path->pos))
1843                         continue;
1844                 if (i->key_cache_already_flushed)
1845                         continue;
1846                 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1847                         ret = i->k;
1848         }
1849
1850         return ret;
1851 }
1852
1853 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1854 {
1855         return iter->flags & BTREE_ITER_WITH_UPDATES
1856                 ? __bch2_btree_trans_peek_updates(iter)
1857                 : NULL;
1858 }
1859
1860 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1861                                               struct btree_iter *iter,
1862                                               struct bpos end_pos)
1863 {
1864         struct bkey_i *k;
1865
1866         if (bpos_lt(iter->path->pos, iter->journal_pos))
1867                 iter->journal_idx = 0;
1868
1869         k = bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1870                                         iter->path->level,
1871                                         iter->path->pos,
1872                                         end_pos,
1873                                         &iter->journal_idx);
1874
1875         iter->journal_pos = k ? k->k.p : end_pos;
1876         return k;
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_ALL_LEVELS);
2067         EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2068
2069         if (iter->update_path) {
2070                 bch2_path_put_nokeep(trans, iter->update_path,
2071                                      iter->flags & BTREE_ITER_INTENT);
2072                 iter->update_path = NULL;
2073         }
2074
2075         bch2_btree_iter_verify_entry_exit(iter);
2076
2077         while (1) {
2078                 k = __bch2_btree_iter_peek(iter, search_key);
2079                 if (unlikely(!k.k))
2080                         goto end;
2081                 if (unlikely(bkey_err(k)))
2082                         goto out_no_locked;
2083
2084                 /*
2085                  * iter->pos should be mononotically increasing, and always be
2086                  * equal to the key we just returned - except extents can
2087                  * straddle iter->pos:
2088                  */
2089                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2090                         iter_pos = k.k->p;
2091                 else
2092                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2093
2094                 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2095                              ? bkey_gt(iter_pos, end)
2096                              : bkey_ge(iter_pos, end)))
2097                         goto end;
2098
2099                 if (iter->update_path &&
2100                     !bkey_eq(iter->update_path->pos, k.k->p)) {
2101                         bch2_path_put_nokeep(trans, iter->update_path,
2102                                              iter->flags & BTREE_ITER_INTENT);
2103                         iter->update_path = NULL;
2104                 }
2105
2106                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2107                     (iter->flags & BTREE_ITER_INTENT) &&
2108                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2109                     !iter->update_path) {
2110                         struct bpos pos = k.k->p;
2111
2112                         if (pos.snapshot < iter->snapshot) {
2113                                 search_key = bpos_successor(k.k->p);
2114                                 continue;
2115                         }
2116
2117                         pos.snapshot = iter->snapshot;
2118
2119                         /*
2120                          * advance, same as on exit for iter->path, but only up
2121                          * to snapshot
2122                          */
2123                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2124                         iter->update_path = iter->path;
2125
2126                         iter->update_path = bch2_btree_path_set_pos(trans,
2127                                                 iter->update_path, pos,
2128                                                 iter->flags & BTREE_ITER_INTENT,
2129                                                 _THIS_IP_);
2130                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2131                         if (unlikely(ret)) {
2132                                 k = bkey_s_c_err(ret);
2133                                 goto out_no_locked;
2134                         }
2135                 }
2136
2137                 /*
2138                  * We can never have a key in a leaf node at POS_MAX, so
2139                  * we don't have to check these successor() calls:
2140                  */
2141                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2142                     !bch2_snapshot_is_ancestor(trans->c,
2143                                                iter->snapshot,
2144                                                k.k->p.snapshot)) {
2145                         search_key = bpos_successor(k.k->p);
2146                         continue;
2147                 }
2148
2149                 if (bkey_whiteout(k.k) &&
2150                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2151                         search_key = bkey_successor(iter, k.k->p);
2152                         continue;
2153                 }
2154
2155                 break;
2156         }
2157
2158         iter->pos = iter_pos;
2159
2160         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2161                                 iter->flags & BTREE_ITER_INTENT,
2162                                 btree_iter_ip_allocated(iter));
2163
2164         btree_path_set_should_be_locked(iter->path);
2165 out_no_locked:
2166         if (iter->update_path) {
2167                 ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_);
2168                 if (unlikely(ret))
2169                         k = bkey_s_c_err(ret);
2170                 else
2171                         btree_path_set_should_be_locked(iter->update_path);
2172         }
2173
2174         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2175                 iter->pos.snapshot = iter->snapshot;
2176
2177         ret = bch2_btree_iter_verify_ret(iter, k);
2178         if (unlikely(ret)) {
2179                 bch2_btree_iter_set_pos(iter, iter->pos);
2180                 k = bkey_s_c_err(ret);
2181         }
2182
2183         bch2_btree_iter_verify_entry_exit(iter);
2184
2185         return k;
2186 end:
2187         bch2_btree_iter_set_pos(iter, end);
2188         k = bkey_s_c_null;
2189         goto out_no_locked;
2190 }
2191
2192 /**
2193  * bch2_btree_iter_peek_all_levels() - returns the first key greater than or
2194  * equal to iterator's current position, returning keys from every level of the
2195  * btree. For keys at different levels of the btree that compare equal, the key
2196  * from the lower level (leaf) is returned first.
2197  * @iter:       iterator to peek from
2198  *
2199  * Returns:     key if found, or an error extractable with bkey_err().
2200  */
2201 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *iter)
2202 {
2203         struct btree_trans *trans = iter->trans;
2204         struct bkey_s_c k;
2205         int ret;
2206
2207         EBUG_ON(iter->path->cached);
2208         bch2_btree_iter_verify(iter);
2209         BUG_ON(iter->path->level < iter->min_depth);
2210         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
2211         EBUG_ON(!(iter->flags & BTREE_ITER_ALL_LEVELS));
2212
2213         while (1) {
2214                 iter->path = bch2_btree_path_set_pos(trans, iter->path, iter->pos,
2215                                         iter->flags & BTREE_ITER_INTENT,
2216                                         btree_iter_ip_allocated(iter));
2217
2218                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2219                 if (unlikely(ret)) {
2220                         /* ensure that iter->k is consistent with iter->pos: */
2221                         bch2_btree_iter_set_pos(iter, iter->pos);
2222                         k = bkey_s_c_err(ret);
2223                         goto out_no_locked;
2224                 }
2225
2226                 /* Already at end? */
2227                 if (!btree_path_node(iter->path, iter->path->level)) {
2228                         k = bkey_s_c_null;
2229                         goto out_no_locked;
2230                 }
2231
2232                 k = btree_path_level_peek_all(trans->c,
2233                                 &iter->path->l[iter->path->level], &iter->k);
2234
2235                 /* Check if we should go up to the parent node: */
2236                 if (!k.k ||
2237                     (iter->advanced &&
2238                      bpos_eq(path_l(iter->path)->b->key.k.p, iter->pos))) {
2239                         iter->pos = path_l(iter->path)->b->key.k.p;
2240                         btree_path_set_level_up(trans, iter->path);
2241                         iter->advanced = false;
2242                         continue;
2243                 }
2244
2245                 /*
2246                  * Check if we should go back down to a leaf:
2247                  * If we're not in a leaf node, we only return the current key
2248                  * if it exactly matches iter->pos - otherwise we first have to
2249                  * go back to the leaf:
2250                  */
2251                 if (iter->path->level != iter->min_depth &&
2252                     (iter->advanced ||
2253                      !k.k ||
2254                      !bpos_eq(iter->pos, k.k->p))) {
2255                         btree_path_set_level_down(trans, iter->path, iter->min_depth);
2256                         iter->pos = bpos_successor(iter->pos);
2257                         iter->advanced = false;
2258                         continue;
2259                 }
2260
2261                 /* Check if we should go to the next key: */
2262                 if (iter->path->level == iter->min_depth &&
2263                     iter->advanced &&
2264                     k.k &&
2265                     bpos_eq(iter->pos, k.k->p)) {
2266                         iter->pos = bpos_successor(iter->pos);
2267                         iter->advanced = false;
2268                         continue;
2269                 }
2270
2271                 if (iter->advanced &&
2272                     iter->path->level == iter->min_depth &&
2273                     !bpos_eq(k.k->p, iter->pos))
2274                         iter->advanced = false;
2275
2276                 BUG_ON(iter->advanced);
2277                 BUG_ON(!k.k);
2278                 break;
2279         }
2280
2281         iter->pos = k.k->p;
2282         btree_path_set_should_be_locked(iter->path);
2283 out_no_locked:
2284         bch2_btree_iter_verify(iter);
2285
2286         return k;
2287 }
2288
2289 /**
2290  * bch2_btree_iter_next() - returns first key greater than iterator's current
2291  * position
2292  * @iter:       iterator to peek from
2293  *
2294  * Returns:     key if found, or an error extractable with bkey_err().
2295  */
2296 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2297 {
2298         if (!bch2_btree_iter_advance(iter))
2299                 return bkey_s_c_null;
2300
2301         return bch2_btree_iter_peek(iter);
2302 }
2303
2304 /**
2305  * bch2_btree_iter_peek_prev() - returns first key less than or equal to
2306  * iterator's current position
2307  * @iter:       iterator to peek from
2308  *
2309  * Returns:     key if found, or an error extractable with bkey_err().
2310  */
2311 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2312 {
2313         struct btree_trans *trans = iter->trans;
2314         struct bpos search_key = iter->pos;
2315         struct btree_path *saved_path = NULL;
2316         struct bkey_s_c k;
2317         struct bkey saved_k;
2318         const struct bch_val *saved_v;
2319         int ret;
2320
2321         EBUG_ON(iter->path->cached || iter->path->level);
2322         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2323
2324         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2325                 return bkey_s_c_err(-EIO);
2326
2327         bch2_btree_iter_verify(iter);
2328         bch2_btree_iter_verify_entry_exit(iter);
2329
2330         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2331                 search_key.snapshot = U32_MAX;
2332
2333         while (1) {
2334                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2335                                                 iter->flags & BTREE_ITER_INTENT,
2336                                                 btree_iter_ip_allocated(iter));
2337
2338                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2339                 if (unlikely(ret)) {
2340                         /* ensure that iter->k is consistent with iter->pos: */
2341                         bch2_btree_iter_set_pos(iter, iter->pos);
2342                         k = bkey_s_c_err(ret);
2343                         goto out_no_locked;
2344                 }
2345
2346                 k = btree_path_level_peek(trans, iter->path,
2347                                           &iter->path->l[0], &iter->k);
2348                 if (!k.k ||
2349                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2350                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2351                      : bpos_gt(k.k->p, search_key)))
2352                         k = btree_path_level_prev(trans, iter->path,
2353                                                   &iter->path->l[0], &iter->k);
2354
2355                 if (likely(k.k)) {
2356                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2357                                 if (k.k->p.snapshot == iter->snapshot)
2358                                         goto got_key;
2359
2360                                 /*
2361                                  * If we have a saved candidate, and we're no
2362                                  * longer at the same _key_ (not pos), return
2363                                  * that candidate
2364                                  */
2365                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2366                                         bch2_path_put_nokeep(trans, iter->path,
2367                                                       iter->flags & BTREE_ITER_INTENT);
2368                                         iter->path = saved_path;
2369                                         saved_path = NULL;
2370                                         iter->k = saved_k;
2371                                         k.v     = saved_v;
2372                                         goto got_key;
2373                                 }
2374
2375                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2376                                                               iter->snapshot,
2377                                                               k.k->p.snapshot)) {
2378                                         if (saved_path)
2379                                                 bch2_path_put_nokeep(trans, saved_path,
2380                                                       iter->flags & BTREE_ITER_INTENT);
2381                                         saved_path = btree_path_clone(trans, iter->path,
2382                                                                 iter->flags & BTREE_ITER_INTENT);
2383                                         saved_k = *k.k;
2384                                         saved_v = k.v;
2385                                 }
2386
2387                                 search_key = bpos_predecessor(k.k->p);
2388                                 continue;
2389                         }
2390 got_key:
2391                         if (bkey_whiteout(k.k) &&
2392                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2393                                 search_key = bkey_predecessor(iter, k.k->p);
2394                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2395                                         search_key.snapshot = U32_MAX;
2396                                 continue;
2397                         }
2398
2399                         break;
2400                 } else if (likely(!bpos_eq(iter->path->l[0].b->data->min_key, POS_MIN))) {
2401                         /* Advance to previous leaf node: */
2402                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2403                 } else {
2404                         /* Start of btree: */
2405                         bch2_btree_iter_set_pos(iter, POS_MIN);
2406                         k = bkey_s_c_null;
2407                         goto out_no_locked;
2408                 }
2409         }
2410
2411         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2412
2413         /* Extents can straddle iter->pos: */
2414         if (bkey_lt(k.k->p, iter->pos))
2415                 iter->pos = k.k->p;
2416
2417         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2418                 iter->pos.snapshot = iter->snapshot;
2419
2420         btree_path_set_should_be_locked(iter->path);
2421 out_no_locked:
2422         if (saved_path)
2423                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2424
2425         bch2_btree_iter_verify_entry_exit(iter);
2426         bch2_btree_iter_verify(iter);
2427
2428         return k;
2429 }
2430
2431 /**
2432  * bch2_btree_iter_prev() - returns first key less than iterator's current
2433  * position
2434  * @iter:       iterator to peek from
2435  *
2436  * Returns:     key if found, or an error extractable with bkey_err().
2437  */
2438 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2439 {
2440         if (!bch2_btree_iter_rewind(iter))
2441                 return bkey_s_c_null;
2442
2443         return bch2_btree_iter_peek_prev(iter);
2444 }
2445
2446 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2447 {
2448         struct btree_trans *trans = iter->trans;
2449         struct bpos search_key;
2450         struct bkey_s_c k;
2451         int ret;
2452
2453         bch2_btree_iter_verify(iter);
2454         bch2_btree_iter_verify_entry_exit(iter);
2455         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2456         EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2457
2458         /* extents can't span inode numbers: */
2459         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2460             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2461                 if (iter->pos.inode == KEY_INODE_MAX)
2462                         return bkey_s_c_null;
2463
2464                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2465         }
2466
2467         search_key = btree_iter_search_key(iter);
2468         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2469                                         iter->flags & BTREE_ITER_INTENT,
2470                                         btree_iter_ip_allocated(iter));
2471
2472         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2473         if (unlikely(ret)) {
2474                 k = bkey_s_c_err(ret);
2475                 goto out_no_locked;
2476         }
2477
2478         if ((iter->flags & BTREE_ITER_CACHED) ||
2479             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2480                 struct bkey_i *next_update;
2481
2482                 if ((next_update = btree_trans_peek_updates(iter)) &&
2483                     bpos_eq(next_update->k.p, iter->pos)) {
2484                         iter->k = next_update->k;
2485                         k = bkey_i_to_s_c(next_update);
2486                         goto out;
2487                 }
2488
2489                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2490                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2491                         goto out;
2492
2493                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2494                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2495                         if (!bkey_err(k))
2496                                 iter->k = *k.k;
2497                         /* We're not returning a key from iter->path: */
2498                         goto out_no_locked;
2499                 }
2500
2501                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2502                 if (unlikely(!k.k))
2503                         goto out_no_locked;
2504         } else {
2505                 struct bpos next;
2506                 struct bpos end = iter->pos;
2507
2508                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2509                         end.offset = U64_MAX;
2510
2511                 EBUG_ON(iter->path->level);
2512
2513                 if (iter->flags & BTREE_ITER_INTENT) {
2514                         struct btree_iter iter2;
2515
2516                         bch2_trans_copy_iter(&iter2, iter);
2517                         k = bch2_btree_iter_peek_upto(&iter2, end);
2518
2519                         if (k.k && !bkey_err(k)) {
2520                                 iter->k = iter2.k;
2521                                 k.k = &iter->k;
2522                         }
2523                         bch2_trans_iter_exit(trans, &iter2);
2524                 } else {
2525                         struct bpos pos = iter->pos;
2526
2527                         k = bch2_btree_iter_peek_upto(iter, end);
2528                         if (unlikely(bkey_err(k)))
2529                                 bch2_btree_iter_set_pos(iter, pos);
2530                         else
2531                                 iter->pos = pos;
2532                 }
2533
2534                 if (unlikely(bkey_err(k)))
2535                         goto out_no_locked;
2536
2537                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2538
2539                 if (bkey_lt(iter->pos, next)) {
2540                         bkey_init(&iter->k);
2541                         iter->k.p = iter->pos;
2542
2543                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2544                                 bch2_key_resize(&iter->k,
2545                                                 min_t(u64, KEY_SIZE_MAX,
2546                                                       (next.inode == iter->pos.inode
2547                                                        ? next.offset
2548                                                        : KEY_OFFSET_MAX) -
2549                                                       iter->pos.offset));
2550                                 EBUG_ON(!iter->k.size);
2551                         }
2552
2553                         k = (struct bkey_s_c) { &iter->k, NULL };
2554                 }
2555         }
2556 out:
2557         btree_path_set_should_be_locked(iter->path);
2558 out_no_locked:
2559         bch2_btree_iter_verify_entry_exit(iter);
2560         bch2_btree_iter_verify(iter);
2561         ret = bch2_btree_iter_verify_ret(iter, k);
2562         if (unlikely(ret))
2563                 return bkey_s_c_err(ret);
2564
2565         return k;
2566 }
2567
2568 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2569 {
2570         if (!bch2_btree_iter_advance(iter))
2571                 return bkey_s_c_null;
2572
2573         return bch2_btree_iter_peek_slot(iter);
2574 }
2575
2576 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2577 {
2578         if (!bch2_btree_iter_rewind(iter))
2579                 return bkey_s_c_null;
2580
2581         return bch2_btree_iter_peek_slot(iter);
2582 }
2583
2584 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2585 {
2586         struct bkey_s_c k;
2587
2588         while (btree_trans_too_many_iters(iter->trans) ||
2589                (k = bch2_btree_iter_peek_type(iter, iter->flags),
2590                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2591                 bch2_trans_begin(iter->trans);
2592
2593         return k;
2594 }
2595
2596 /* new transactional stuff: */
2597
2598 #ifdef CONFIG_BCACHEFS_DEBUG
2599 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2600 {
2601         struct btree_path *path;
2602         unsigned i;
2603
2604         BUG_ON(trans->nr_sorted != hweight64(trans->paths_allocated));
2605
2606         trans_for_each_path(trans, path) {
2607                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2608                 BUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2609         }
2610
2611         for (i = 0; i < trans->nr_sorted; i++) {
2612                 unsigned idx = trans->sorted[i];
2613
2614                 EBUG_ON(!(trans->paths_allocated & (1ULL << idx)));
2615                 BUG_ON(trans->paths[idx].sorted_idx != i);
2616         }
2617 }
2618
2619 static void btree_trans_verify_sorted(struct btree_trans *trans)
2620 {
2621         struct btree_path *path, *prev = NULL;
2622         unsigned i;
2623
2624         if (!bch2_debug_check_iterators)
2625                 return;
2626
2627         trans_for_each_path_inorder(trans, path, i) {
2628                 if (prev && btree_path_cmp(prev, path) > 0) {
2629                         __bch2_dump_trans_paths_updates(trans, true);
2630                         panic("trans paths out of order!\n");
2631                 }
2632                 prev = path;
2633         }
2634 }
2635 #else
2636 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2637 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2638 #endif
2639
2640 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2641 {
2642         int i, l = 0, r = trans->nr_sorted, inc = 1;
2643         bool swapped;
2644
2645         btree_trans_verify_sorted_refs(trans);
2646
2647         if (trans->paths_sorted)
2648                 goto out;
2649
2650         /*
2651          * Cocktail shaker sort: this is efficient because iterators will be
2652          * mostly sorted.
2653          */
2654         do {
2655                 swapped = false;
2656
2657                 for (i = inc > 0 ? l : r - 2;
2658                      i + 1 < r && i >= l;
2659                      i += inc) {
2660                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2661                                            trans->paths + trans->sorted[i + 1]) > 0) {
2662                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2663                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2664                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2665                                 swapped = true;
2666                         }
2667                 }
2668
2669                 if (inc > 0)
2670                         --r;
2671                 else
2672                         l++;
2673                 inc = -inc;
2674         } while (swapped);
2675
2676         trans->paths_sorted = true;
2677 out:
2678         btree_trans_verify_sorted(trans);
2679 }
2680
2681 static inline void btree_path_list_remove(struct btree_trans *trans,
2682                                           struct btree_path *path)
2683 {
2684         unsigned i;
2685
2686         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2687 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2688         trans->nr_sorted--;
2689         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2690                                 trans->sorted + path->sorted_idx + 1,
2691                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2692 #else
2693         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2694 #endif
2695         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2696                 trans->paths[trans->sorted[i]].sorted_idx = i;
2697
2698         path->sorted_idx = U8_MAX;
2699 }
2700
2701 static inline void btree_path_list_add(struct btree_trans *trans,
2702                                        struct btree_path *pos,
2703                                        struct btree_path *path)
2704 {
2705         unsigned i;
2706
2707         path->sorted_idx = pos ? pos->sorted_idx + 1 : trans->nr_sorted;
2708
2709 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2710         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2711                               trans->sorted + path->sorted_idx,
2712                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2713         trans->nr_sorted++;
2714         trans->sorted[path->sorted_idx] = path->idx;
2715 #else
2716         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2717 #endif
2718
2719         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2720                 trans->paths[trans->sorted[i]].sorted_idx = i;
2721
2722         btree_trans_verify_sorted_refs(trans);
2723 }
2724
2725 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2726 {
2727         if (iter->update_path)
2728                 bch2_path_put_nokeep(trans, iter->update_path,
2729                               iter->flags & BTREE_ITER_INTENT);
2730         if (iter->path)
2731                 bch2_path_put(trans, iter->path,
2732                               iter->flags & BTREE_ITER_INTENT);
2733         if (iter->key_cache_path)
2734                 bch2_path_put(trans, iter->key_cache_path,
2735                               iter->flags & BTREE_ITER_INTENT);
2736         iter->path = NULL;
2737         iter->update_path = NULL;
2738         iter->key_cache_path = NULL;
2739 }
2740
2741 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2742                           struct btree_iter *iter,
2743                           enum btree_id btree_id, struct bpos pos,
2744                           unsigned flags)
2745 {
2746         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2747                                bch2_btree_iter_flags(trans, btree_id, flags),
2748                                _RET_IP_);
2749 }
2750
2751 void bch2_trans_node_iter_init(struct btree_trans *trans,
2752                                struct btree_iter *iter,
2753                                enum btree_id btree_id,
2754                                struct bpos pos,
2755                                unsigned locks_want,
2756                                unsigned depth,
2757                                unsigned flags)
2758 {
2759         flags |= BTREE_ITER_NOT_EXTENTS;
2760         flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2761         flags |= BTREE_ITER_ALL_SNAPSHOTS;
2762
2763         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2764                                __bch2_btree_iter_flags(trans, btree_id, flags),
2765                                _RET_IP_);
2766
2767         iter->min_depth = depth;
2768
2769         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2770         BUG_ON(iter->path->level        != depth);
2771         BUG_ON(iter->min_depth          != depth);
2772 }
2773
2774 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2775 {
2776         *dst = *src;
2777         if (src->path)
2778                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2779         if (src->update_path)
2780                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2781         dst->key_cache_path = NULL;
2782 }
2783
2784 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2785 {
2786         unsigned new_top = trans->mem_top + size;
2787         size_t old_bytes = trans->mem_bytes;
2788         size_t new_bytes = roundup_pow_of_two(new_top);
2789         int ret;
2790         void *new_mem;
2791         void *p;
2792
2793         trans->mem_max = max(trans->mem_max, new_top);
2794
2795         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2796
2797         new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2798         if (unlikely(!new_mem)) {
2799                 bch2_trans_unlock(trans);
2800
2801                 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2802                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2803                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2804                         new_bytes = BTREE_TRANS_MEM_MAX;
2805                         kfree(trans->mem);
2806                 }
2807
2808                 if (!new_mem)
2809                         return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2810
2811                 trans->mem = new_mem;
2812                 trans->mem_bytes = new_bytes;
2813
2814                 ret = bch2_trans_relock(trans);
2815                 if (ret)
2816                         return ERR_PTR(ret);
2817         }
2818
2819         trans->mem = new_mem;
2820         trans->mem_bytes = new_bytes;
2821
2822         if (old_bytes) {
2823                 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2824                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2825         }
2826
2827         p = trans->mem + trans->mem_top;
2828         trans->mem_top += size;
2829         memset(p, 0, size);
2830         return p;
2831 }
2832
2833 static noinline void bch2_trans_reset_srcu_lock(struct btree_trans *trans)
2834 {
2835         struct bch_fs *c = trans->c;
2836         struct btree_path *path;
2837
2838         trans_for_each_path(trans, path)
2839                 if (path->cached && !btree_node_locked(path, 0))
2840                         path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2841
2842         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2843         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2844         trans->srcu_lock_time   = jiffies;
2845 }
2846
2847 /**
2848  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2849  * @trans: transaction to reset
2850  *
2851  * Returns:     current restart counter, to be used with trans_was_restarted()
2852  *
2853  * While iterating over nodes or updating nodes a attempt to lock a btree node
2854  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2855  * occurs bch2_trans_begin() should be called and the transaction retried.
2856  */
2857 u32 bch2_trans_begin(struct btree_trans *trans)
2858 {
2859         struct btree_path *path;
2860         u64 now;
2861
2862         bch2_trans_reset_updates(trans);
2863
2864         trans->restart_count++;
2865         trans->mem_top                  = 0;
2866
2867         trans_for_each_path(trans, path) {
2868                 path->should_be_locked = false;
2869
2870                 /*
2871                  * If the transaction wasn't restarted, we're presuming to be
2872                  * doing something new: dont keep iterators excpt the ones that
2873                  * are in use - except for the subvolumes btree:
2874                  */
2875                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2876                         path->preserve = false;
2877
2878                 /*
2879                  * XXX: we probably shouldn't be doing this if the transaction
2880                  * was restarted, but currently we still overflow transaction
2881                  * iterators if we do that
2882                  */
2883                 if (!path->ref && !path->preserve)
2884                         __bch2_path_free(trans, path);
2885                 else
2886                         path->preserve = false;
2887         }
2888
2889         now = local_clock();
2890         if (!trans->restarted &&
2891             (need_resched() ||
2892              now - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2893                 drop_locks_do(trans, (cond_resched(), 0));
2894                 now = local_clock();
2895         }
2896         trans->last_begin_time = now;
2897
2898         if (unlikely(time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2899                 bch2_trans_reset_srcu_lock(trans);
2900
2901         trans->last_begin_ip = _RET_IP_;
2902         if (trans->restarted) {
2903                 bch2_btree_path_traverse_all(trans);
2904                 trans->notrace_relock_fail = false;
2905         }
2906
2907         return trans->restart_count;
2908 }
2909
2910 static struct btree_trans *bch2_trans_alloc(struct bch_fs *c)
2911 {
2912         struct btree_trans *trans;
2913
2914         if (IS_ENABLED(__KERNEL__)) {
2915                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
2916                 if (trans)
2917                         return trans;
2918         }
2919
2920         trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
2921         /*
2922          * paths need to be zeroed, bch2_check_for_deadlock looks at
2923          * paths in other threads
2924          */
2925         memset(&trans->paths, 0, sizeof(trans->paths));
2926         return trans;
2927 }
2928
2929 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2930
2931 unsigned bch2_trans_get_fn_idx(const char *fn)
2932 {
2933         unsigned i;
2934
2935         for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2936                 if (!bch2_btree_transaction_fns[i] ||
2937                     bch2_btree_transaction_fns[i] == fn) {
2938                         bch2_btree_transaction_fns[i] = fn;
2939                         return i;
2940                 }
2941
2942         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2943         return i;
2944 }
2945
2946 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
2947         __acquires(&c->btree_trans_barrier)
2948 {
2949         struct btree_trans *trans;
2950         struct btree_transaction_stats *s;
2951
2952         bch2_assert_btree_nodes_not_locked();
2953
2954         trans = bch2_trans_alloc(c);
2955
2956         memset(trans, 0, sizeof(*trans));
2957         trans->c                = c;
2958         trans->fn               = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2959                 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2960         trans->last_begin_time  = local_clock();
2961         trans->fn_idx           = fn_idx;
2962         trans->locking_wait.task = current;
2963         trans->journal_replay_not_finished =
2964                 !test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
2965         closure_init_stack(&trans->ref);
2966
2967         s = btree_trans_stats(trans);
2968         if (s && s->max_mem) {
2969                 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2970
2971                 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2972
2973                 if (!unlikely(trans->mem)) {
2974                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2975                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2976                 } else {
2977                         trans->mem_bytes = expected_mem_bytes;
2978                 }
2979         }
2980
2981         if (s) {
2982                 trans->nr_max_paths = s->nr_max_paths;
2983                 trans->wb_updates_size = s->wb_updates_size;
2984         }
2985
2986         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2987         trans->srcu_lock_time   = jiffies;
2988
2989         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2990                 struct btree_trans *pos;
2991
2992                 seqmutex_lock(&c->btree_trans_lock);
2993                 list_for_each_entry(pos, &c->btree_trans_list, list) {
2994                         /*
2995                          * We'd much prefer to be stricter here and completely
2996                          * disallow multiple btree_trans in the same thread -
2997                          * but the data move path calls bch2_write when we
2998                          * already have a btree_trans initialized.
2999                          */
3000                         BUG_ON(trans->locking_wait.task->pid == pos->locking_wait.task->pid &&
3001                                bch2_trans_locked(pos));
3002
3003                         if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
3004                                 list_add_tail(&trans->list, &pos->list);
3005                                 goto list_add_done;
3006                         }
3007                 }
3008                 list_add_tail(&trans->list, &c->btree_trans_list);
3009 list_add_done:
3010                 seqmutex_unlock(&c->btree_trans_lock);
3011         }
3012
3013         return trans;
3014 }
3015
3016 static void check_btree_paths_leaked(struct btree_trans *trans)
3017 {
3018 #ifdef CONFIG_BCACHEFS_DEBUG
3019         struct bch_fs *c = trans->c;
3020         struct btree_path *path;
3021
3022         trans_for_each_path(trans, path)
3023                 if (path->ref)
3024                         goto leaked;
3025         return;
3026 leaked:
3027         bch_err(c, "btree paths leaked from %s!", trans->fn);
3028         trans_for_each_path(trans, path)
3029                 if (path->ref)
3030                         printk(KERN_ERR "  btree %s %pS\n",
3031                                bch2_btree_id_str(path->btree_id),
3032                                (void *) path->ip_allocated);
3033         /* Be noisy about this: */
3034         bch2_fatal_error(c);
3035 #endif
3036 }
3037
3038 void bch2_trans_put(struct btree_trans *trans)
3039         __releases(&c->btree_trans_barrier)
3040 {
3041         struct btree_insert_entry *i;
3042         struct bch_fs *c = trans->c;
3043         struct btree_transaction_stats *s = btree_trans_stats(trans);
3044
3045         bch2_trans_unlock(trans);
3046
3047         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
3048                 seqmutex_lock(&c->btree_trans_lock);
3049                 list_del(&trans->list);
3050                 seqmutex_unlock(&c->btree_trans_lock);
3051         }
3052
3053         closure_sync(&trans->ref);
3054
3055         if (s)
3056                 s->max_mem = max(s->max_mem, trans->mem_max);
3057
3058         trans_for_each_update(trans, i)
3059                 __btree_path_put(i->path, true);
3060         trans->nr_updates               = 0;
3061
3062         check_btree_paths_leaked(trans);
3063
3064         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3065
3066         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3067
3068         kfree(trans->extra_journal_entries.data);
3069
3070         if (trans->fs_usage_deltas) {
3071                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3072                     REPLICAS_DELTA_LIST_MAX)
3073                         mempool_free(trans->fs_usage_deltas,
3074                                      &c->replicas_delta_pool);
3075                 else
3076                         kfree(trans->fs_usage_deltas);
3077         }
3078
3079         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3080                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3081         else
3082                 kfree(trans->mem);
3083
3084         /* Userspace doesn't have a real percpu implementation: */
3085         if (IS_ENABLED(__KERNEL__))
3086                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
3087         if (trans)
3088                 mempool_free(trans, &c->btree_trans_pool);
3089 }
3090
3091 static void __maybe_unused
3092 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3093                                       struct btree_bkey_cached_common *b)
3094 {
3095         struct six_lock_count c = six_lock_counts(&b->lock);
3096         struct task_struct *owner;
3097         pid_t pid;
3098
3099         rcu_read_lock();
3100         owner = READ_ONCE(b->lock.owner);
3101         pid = owner ? owner->pid : 0;
3102         rcu_read_unlock();
3103
3104         prt_tab(out);
3105         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3106                    b->level, bch2_btree_id_str(b->btree_id));
3107         bch2_bpos_to_text(out, btree_node_pos(b));
3108
3109         prt_tab(out);
3110         prt_printf(out, " locks %u:%u:%u held by pid %u",
3111                    c.n[0], c.n[1], c.n[2], pid);
3112 }
3113
3114 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3115 {
3116         struct btree_path *path;
3117         struct btree_bkey_cached_common *b;
3118         static char lock_types[] = { 'r', 'i', 'w' };
3119         unsigned l, idx;
3120
3121         if (!out->nr_tabstops) {
3122                 printbuf_tabstop_push(out, 16);
3123                 printbuf_tabstop_push(out, 32);
3124         }
3125
3126         prt_printf(out, "%i %s\n", trans->locking_wait.task->pid, trans->fn);
3127
3128         trans_for_each_path_safe(trans, path, idx) {
3129                 if (!path->nodes_locked)
3130                         continue;
3131
3132                 prt_printf(out, "  path %u %c l=%u %s:",
3133                        path->idx,
3134                        path->cached ? 'c' : 'b',
3135                        path->level,
3136                        bch2_btree_id_str(path->btree_id));
3137                 bch2_bpos_to_text(out, path->pos);
3138                 prt_newline(out);
3139
3140                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3141                         if (btree_node_locked(path, l) &&
3142                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3143                                 prt_printf(out, "    %c l=%u ",
3144                                            lock_types[btree_node_locked_type(path, l)], l);
3145                                 bch2_btree_bkey_cached_common_to_text(out, b);
3146                                 prt_newline(out);
3147                         }
3148                 }
3149         }
3150
3151         b = READ_ONCE(trans->locking);
3152         if (b) {
3153                 prt_printf(out, "  blocked for %lluus on",
3154                            div_u64(local_clock() - trans->locking_wait.start_time,
3155                                    1000));
3156                 prt_newline(out);
3157                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3158                 bch2_btree_bkey_cached_common_to_text(out, b);
3159                 prt_newline(out);
3160         }
3161 }
3162
3163 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3164 {
3165         struct btree_transaction_stats *s;
3166         struct btree_trans *trans;
3167         int cpu;
3168
3169         trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3170         if (trans)
3171                 panic("%s leaked btree_trans\n", trans->fn);
3172
3173         if (c->btree_trans_bufs)
3174                 for_each_possible_cpu(cpu)
3175                         kfree(per_cpu_ptr(c->btree_trans_bufs, cpu)->trans);
3176         free_percpu(c->btree_trans_bufs);
3177
3178         for (s = c->btree_transaction_stats;
3179              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3180              s++) {
3181                 kfree(s->max_paths_text);
3182                 bch2_time_stats_exit(&s->lock_hold_times);
3183         }
3184
3185         if (c->btree_trans_barrier_initialized)
3186                 cleanup_srcu_struct(&c->btree_trans_barrier);
3187         mempool_exit(&c->btree_trans_mem_pool);
3188         mempool_exit(&c->btree_trans_pool);
3189 }
3190
3191 int bch2_fs_btree_iter_init(struct bch_fs *c)
3192 {
3193         struct btree_transaction_stats *s;
3194         int ret;
3195
3196         for (s = c->btree_transaction_stats;
3197              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3198              s++) {
3199                 bch2_time_stats_init(&s->lock_hold_times);
3200                 mutex_init(&s->lock);
3201         }
3202
3203         INIT_LIST_HEAD(&c->btree_trans_list);
3204         seqmutex_init(&c->btree_trans_lock);
3205
3206         c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3207         if (!c->btree_trans_bufs)
3208                 return -ENOMEM;
3209
3210         ret   = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3211                                           sizeof(struct btree_trans)) ?:
3212                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3213                                           BTREE_TRANS_MEM_MAX) ?:
3214                 init_srcu_struct(&c->btree_trans_barrier);
3215         if (!ret)
3216                 c->btree_trans_barrier_initialized = true;
3217         return ret;
3218 }