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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_buf.h"
   6 #include "btree_cache.h"
   7 #include "btree_iter.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "debug.h"
  12 #include "error.h"
  13 #include "extents.h"
  14 #include "journal.h"
  15 #include "recovery.h"
  16 #include "replicas.h"
  17 #include "subvolume.h"
  18
  19 #include <linux/prandom.h>
  20 #include <linux/prefetch.h>
  21 #include <trace/events/bcachefs.h>
  22
  23 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
  24 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
  25                                        struct btree_path *);
  26
  27 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
  28 {
  29 #ifdef TRACK_PATH_ALLOCATED
  30         return iter->ip_allocated;
  31 #else
  32         return 0;
  33 #endif
  34 }
  35
  36 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
  37
  38 /*
  39  * Unlocks before scheduling
  40  * Note: does not revalidate iterator
  41  */
  42 static inline int bch2_trans_cond_resched(struct btree_trans *trans)
  43 {
  44         if (need_resched() || race_fault()) {
  45                 bch2_trans_unlock(trans);
  46                 schedule();
  47                 return bch2_trans_relock(trans);
  48         } else {
  49                 return 0;
  50         }
  51 }
  52
  53 static inline int __btree_path_cmp(const struct btree_path *l,
  54                                    enum btree_id        r_btree_id,
  55                                    bool                 r_cached,
  56                                    struct bpos          r_pos,
  57                                    unsigned             r_level)
  58 {
  59         /*
  60          * Must match lock ordering as defined by __bch2_btree_node_lock:
  61          */
  62         return   cmp_int(l->btree_id,   r_btree_id) ?:
  63                  cmp_int((int) l->cached,       (int) r_cached) ?:
  64                  bpos_cmp(l->pos,       r_pos) ?:
  65                 -cmp_int(l->level,      r_level);
  66 }
  67
  68 static inline int btree_path_cmp(const struct btree_path *l,
  69                                  const struct btree_path *r)
  70 {
  71         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  72 }
  73
  74 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  75 {
  76         /* Are we iterating over keys in all snapshots? */
  77         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  78                 p = bpos_successor(p);
  79         } else {
  80                 p = bpos_nosnap_successor(p);
  81                 p.snapshot = iter->snapshot;
  82         }
  83
  84         return p;
  85 }
  86
  87 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  88 {
  89         /* Are we iterating over keys in all snapshots? */
  90         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  91                 p = bpos_predecessor(p);
  92         } else {
  93                 p = bpos_nosnap_predecessor(p);
  94                 p.snapshot = iter->snapshot;
  95         }
  96
  97         return p;
  98 }
  99
 100 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
 101 {
 102         struct bpos pos = iter->pos;
 103
 104         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 105             !bkey_eq(pos, POS_MAX))
 106                 pos = bkey_successor(iter, pos);
 107         return pos;
 108 }
 109
 110 static inline bool btree_path_pos_before_node(struct btree_path *path,
 111                                               struct btree *b)
 112 {
 113         return bpos_lt(path->pos, b->data->min_key);
 114 }
 115
 116 static inline bool btree_path_pos_after_node(struct btree_path *path,
 117                                              struct btree *b)
 118 {
 119         return bpos_gt(path->pos, b->key.k.p);
 120 }
 121
 122 static inline bool btree_path_pos_in_node(struct btree_path *path,
 123                                           struct btree *b)
 124 {
 125         return path->btree_id == b->c.btree_id &&
 126                 !btree_path_pos_before_node(path, b) &&
 127                 !btree_path_pos_after_node(path, b);
 128 }
 129
 130 /* Btree iterator: */
 131
 132 #ifdef CONFIG_BCACHEFS_DEBUG
 133
 134 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 135                                           struct btree_path *path)
 136 {
 137         struct bkey_cached *ck;
 138         bool locked = btree_node_locked(path, 0);
 139
 140         if (!bch2_btree_node_relock(trans, path, 0))
 141                 return;
 142
 143         ck = (void *) path->l[0].b;
 144         BUG_ON(ck->key.btree_id != path->btree_id ||
 145                !bkey_eq(ck->key.pos, path->pos));
 146
 147         if (!locked)
 148                 btree_node_unlock(trans, path, 0);
 149 }
 150
 151 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 152                                 struct btree_path *path, unsigned level)
 153 {
 154         struct btree_path_level *l;
 155         struct btree_node_iter tmp;
 156         bool locked;
 157         struct bkey_packed *p, *k;
 158         struct printbuf buf1 = PRINTBUF;
 159         struct printbuf buf2 = PRINTBUF;
 160         struct printbuf buf3 = PRINTBUF;
 161         const char *msg;
 162
 163         if (!bch2_debug_check_iterators)
 164                 return;
 165
 166         l       = &path->l[level];
 167         tmp     = l->iter;
 168         locked  = btree_node_locked(path, level);
 169
 170         if (path->cached) {
 171                 if (!level)
 172                         bch2_btree_path_verify_cached(trans, path);
 173                 return;
 174         }
 175
 176         if (!btree_path_node(path, level))
 177                 return;
 178
 179         if (!bch2_btree_node_relock_notrace(trans, path, level))
 180                 return;
 181
 182         BUG_ON(!btree_path_pos_in_node(path, l->b));
 183
 184         bch2_btree_node_iter_verify(&l->iter, l->b);
 185
 186         /*
 187          * For interior nodes, the iterator will have skipped past deleted keys:
 188          */
 189         p = level
 190                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 191                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 192         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 193
 194         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 195                 msg = "before";
 196                 goto err;
 197         }
 198
 199         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 200                 msg = "after";
 201                 goto err;
 202         }
 203
 204         if (!locked)
 205                 btree_node_unlock(trans, path, level);
 206         return;
 207 err:
 208         bch2_bpos_to_text(&buf1, path->pos);
 209
 210         if (p) {
 211                 struct bkey uk = bkey_unpack_key(l->b, p);
 212
 213                 bch2_bkey_to_text(&buf2, &uk);
 214         } else {
 215                 prt_printf(&buf2, "(none)");
 216         }
 217
 218         if (k) {
 219                 struct bkey uk = bkey_unpack_key(l->b, k);
 220
 221                 bch2_bkey_to_text(&buf3, &uk);
 222         } else {
 223                 prt_printf(&buf3, "(none)");
 224         }
 225
 226         panic("path should be %s key at level %u:\n"
 227               "path pos %s\n"
 228               "prev key %s\n"
 229               "cur  key %s\n",
 230               msg, level, buf1.buf, buf2.buf, buf3.buf);
 231 }
 232
 233 static void bch2_btree_path_verify(struct btree_trans *trans,
 234                                    struct btree_path *path)
 235 {
 236         struct bch_fs *c = trans->c;
 237         unsigned i;
 238
 239         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 240
 241         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 242                 if (!path->l[i].b) {
 243                         BUG_ON(!path->cached &&
 244                                c->btree_roots[path->btree_id].b->c.level > i);
 245                         break;
 246                 }
 247
 248                 bch2_btree_path_verify_level(trans, path, i);
 249         }
 250
 251         bch2_btree_path_verify_locks(path);
 252 }
 253
 254 void bch2_trans_verify_paths(struct btree_trans *trans)
 255 {
 256         struct btree_path *path;
 257
 258         trans_for_each_path(trans, path)
 259                 bch2_btree_path_verify(trans, path);
 260 }
 261
 262 static void bch2_btree_iter_verify(struct btree_iter *iter)
 263 {
 264         struct btree_trans *trans = iter->trans;
 265
 266         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 267
 268         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
 269
 270         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 271                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 272
 273         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 274                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 275                !btree_type_has_snapshots(iter->btree_id));
 276
 277         if (iter->update_path)
 278                 bch2_btree_path_verify(trans, iter->update_path);
 279         bch2_btree_path_verify(trans, iter->path);
 280 }
 281
 282 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 283 {
 284         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 285                !iter->pos.snapshot);
 286
 287         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 288                iter->pos.snapshot != iter->snapshot);
 289
 290         BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
 291                bkey_gt(iter->pos, iter->k.p));
 292 }
 293
 294 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 295 {
 296         struct btree_trans *trans = iter->trans;
 297         struct btree_iter copy;
 298         struct bkey_s_c prev;
 299         int ret = 0;
 300
 301         if (!bch2_debug_check_iterators)
 302                 return 0;
 303
 304         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 305                 return 0;
 306
 307         if (bkey_err(k) || !k.k)
 308                 return 0;
 309
 310         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 311                                           iter->snapshot,
 312                                           k.k->p.snapshot));
 313
 314         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 315                              BTREE_ITER_NOPRESERVE|
 316                              BTREE_ITER_ALL_SNAPSHOTS);
 317         prev = bch2_btree_iter_prev(&copy);
 318         if (!prev.k)
 319                 goto out;
 320
 321         ret = bkey_err(prev);
 322         if (ret)
 323                 goto out;
 324
 325         if (bkey_eq(prev.k->p, k.k->p) &&
 326             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 327                                       prev.k->p.snapshot) > 0) {
 328                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 329
 330                 bch2_bkey_to_text(&buf1, k.k);
 331                 bch2_bkey_to_text(&buf2, prev.k);
 332
 333                 panic("iter snap %u\n"
 334                       "k    %s\n"
 335                       "prev %s\n",
 336                       iter->snapshot,
 337                       buf1.buf, buf2.buf);
 338         }
 339 out:
 340         bch2_trans_iter_exit(trans, &copy);
 341         return ret;
 342 }
 343
 344 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 345                             struct bpos pos, bool key_cache)
 346 {
 347         struct btree_path *path;
 348         unsigned idx;
 349         struct printbuf buf = PRINTBUF;
 350
 351         btree_trans_sort_paths(trans);
 352
 353         trans_for_each_path_inorder(trans, path, idx) {
 354                 int cmp = cmp_int(path->btree_id, id) ?:
 355                         cmp_int(path->cached, key_cache);
 356
 357                 if (cmp > 0)
 358                         break;
 359                 if (cmp < 0)
 360                         continue;
 361
 362                 if (!btree_node_locked(path, 0) ||
 363                     !path->should_be_locked)
 364                         continue;
 365
 366                 if (!key_cache) {
 367                         if (bkey_ge(pos, path->l[0].b->data->min_key) &&
 368                             bkey_le(pos, path->l[0].b->key.k.p))
 369                                 return;
 370                 } else {
 371                         if (bkey_eq(pos, path->pos))
 372                                 return;
 373                 }
 374         }
 375
 376         bch2_dump_trans_paths_updates(trans);
 377         bch2_bpos_to_text(&buf, pos);
 378
 379         panic("not locked: %s %s%s\n",
 380               bch2_btree_ids[id], buf.buf,
 381               key_cache ? " cached" : "");
 382 }
 383
 384 #else
 385
 386 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 387                                                 struct btree_path *path, unsigned l) {}
 388 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 389                                           struct btree_path *path) {}
 390 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 391 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 392 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 393
 394 #endif
 395
 396 /* Btree path: fixups after btree updates */
 397
 398 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 399                                         struct btree *b,
 400                                         struct bset_tree *t,
 401                                         struct bkey_packed *k)
 402 {
 403         struct btree_node_iter_set *set;
 404
 405         btree_node_iter_for_each(iter, set)
 406                 if (set->end == t->end_offset) {
 407                         set->k = __btree_node_key_to_offset(b, k);
 408                         bch2_btree_node_iter_sort(iter, b);
 409                         return;
 410                 }
 411
 412         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 413 }
 414
 415 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 416                                                struct btree *b,
 417                                                struct bkey_packed *where)
 418 {
 419         struct btree_path_level *l = &path->l[b->c.level];
 420
 421         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 422                 return;
 423
 424         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 425                 bch2_btree_node_iter_advance(&l->iter, l->b);
 426 }
 427
 428 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 429                                       struct btree *b,
 430                                       struct bkey_packed *where)
 431 {
 432         struct btree_path *path;
 433
 434         trans_for_each_path_with_node(trans, b, path) {
 435                 __bch2_btree_path_fix_key_modified(path, b, where);
 436                 bch2_btree_path_verify_level(trans, path, b->c.level);
 437         }
 438 }
 439
 440 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 441                                        struct btree *b,
 442                                        struct btree_node_iter *node_iter,
 443                                        struct bset_tree *t,
 444                                        struct bkey_packed *where,
 445                                        unsigned clobber_u64s,
 446                                        unsigned new_u64s)
 447 {
 448         const struct bkey_packed *end = btree_bkey_last(b, t);
 449         struct btree_node_iter_set *set;
 450         unsigned offset = __btree_node_key_to_offset(b, where);
 451         int shift = new_u64s - clobber_u64s;
 452         unsigned old_end = t->end_offset - shift;
 453         unsigned orig_iter_pos = node_iter->data[0].k;
 454         bool iter_current_key_modified =
 455                 orig_iter_pos >= offset &&
 456                 orig_iter_pos <= offset + clobber_u64s;
 457
 458         btree_node_iter_for_each(node_iter, set)
 459                 if (set->end == old_end)
 460                         goto found;
 461
 462         /* didn't find the bset in the iterator - might have to readd it: */
 463         if (new_u64s &&
 464             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 465                 bch2_btree_node_iter_push(node_iter, b, where, end);
 466                 goto fixup_done;
 467         } else {
 468                 /* Iterator is after key that changed */
 469                 return;
 470         }
 471 found:
 472         set->end = t->end_offset;
 473
 474         /* Iterator hasn't gotten to the key that changed yet: */
 475         if (set->k < offset)
 476                 return;
 477
 478         if (new_u64s &&
 479             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 480                 set->k = offset;
 481         } else if (set->k < offset + clobber_u64s) {
 482                 set->k = offset + new_u64s;
 483                 if (set->k == set->end)
 484                         bch2_btree_node_iter_set_drop(node_iter, set);
 485         } else {
 486                 /* Iterator is after key that changed */
 487                 set->k = (int) set->k + shift;
 488                 return;
 489         }
 490
 491         bch2_btree_node_iter_sort(node_iter, b);
 492 fixup_done:
 493         if (node_iter->data[0].k != orig_iter_pos)
 494                 iter_current_key_modified = true;
 495
 496         /*
 497          * When a new key is added, and the node iterator now points to that
 498          * key, the iterator might have skipped past deleted keys that should
 499          * come after the key the iterator now points to. We have to rewind to
 500          * before those deleted keys - otherwise
 501          * bch2_btree_node_iter_prev_all() breaks:
 502          */
 503         if (!bch2_btree_node_iter_end(node_iter) &&
 504             iter_current_key_modified &&
 505             b->c.level) {
 506                 struct bset_tree *t;
 507                 struct bkey_packed *k, *k2, *p;
 508
 509                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 510
 511                 for_each_bset(b, t) {
 512                         bool set_pos = false;
 513
 514                         if (node_iter->data[0].end == t->end_offset)
 515                                 continue;
 516
 517                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 518
 519                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 520                                bkey_iter_cmp(b, k, p) < 0) {
 521                                 k2 = p;
 522                                 set_pos = true;
 523                         }
 524
 525                         if (set_pos)
 526                                 btree_node_iter_set_set_pos(node_iter,
 527                                                             b, t, k2);
 528                 }
 529         }
 530 }
 531
 532 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 533                               struct btree_path *path,
 534                               struct btree *b,
 535                               struct btree_node_iter *node_iter,
 536                               struct bkey_packed *where,
 537                               unsigned clobber_u64s,
 538                               unsigned new_u64s)
 539 {
 540         struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
 541         struct btree_path *linked;
 542
 543         if (node_iter != &path->l[b->c.level].iter) {
 544                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 545                                            where, clobber_u64s, new_u64s);
 546
 547                 if (bch2_debug_check_iterators)
 548                         bch2_btree_node_iter_verify(node_iter, b);
 549         }
 550
 551         trans_for_each_path_with_node(trans, b, linked) {
 552                 __bch2_btree_node_iter_fix(linked, b,
 553                                            &linked->l[b->c.level].iter, t,
 554                                            where, clobber_u64s, new_u64s);
 555                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 556         }
 557 }
 558
 559 /* Btree path level: pointer to a particular btree node and node iter */
 560
 561 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
 562                                                   struct btree_path_level *l,
 563                                                   struct bkey *u,
 564                                                   struct bkey_packed *k)
 565 {
 566         if (unlikely(!k)) {
 567                 /*
 568                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 569                  * a hole
 570                  */
 571                 u->type = KEY_TYPE_deleted;
 572                 return bkey_s_c_null;
 573         }
 574
 575         return bkey_disassemble(l->b, k, u);
 576 }
 577
 578 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
 579                                                         struct btree_path_level *l,
 580                                                         struct bkey *u)
 581 {
 582         return __btree_iter_unpack(c, l, u,
 583                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 584 }
 585
 586 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
 587                                                     struct btree_path *path,
 588                                                     struct btree_path_level *l,
 589                                                     struct bkey *u)
 590 {
 591         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 592                         bch2_btree_node_iter_peek(&l->iter, l->b));
 593
 594         path->pos = k.k ? k.k->p : l->b->key.k.p;
 595         trans->paths_sorted = false;
 596         bch2_btree_path_verify_level(trans, path, l - path->l);
 597         return k;
 598 }
 599
 600 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
 601                                                     struct btree_path *path,
 602                                                     struct btree_path_level *l,
 603                                                     struct bkey *u)
 604 {
 605         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 606                         bch2_btree_node_iter_prev(&l->iter, l->b));
 607
 608         path->pos = k.k ? k.k->p : l->b->data->min_key;
 609         trans->paths_sorted = false;
 610         bch2_btree_path_verify_level(trans, path, l - path->l);
 611         return k;
 612 }
 613
 614 static inline bool btree_path_advance_to_pos(struct btree_path *path,
 615                                              struct btree_path_level *l,
 616                                              int max_advance)
 617 {
 618         struct bkey_packed *k;
 619         int nr_advanced = 0;
 620
 621         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 622                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 623                 if (max_advance > 0 && nr_advanced >= max_advance)
 624                         return false;
 625
 626                 bch2_btree_node_iter_advance(&l->iter, l->b);
 627                 nr_advanced++;
 628         }
 629
 630         return true;
 631 }
 632
 633 static inline void __btree_path_level_init(struct btree_path *path,
 634                                            unsigned level)
 635 {
 636         struct btree_path_level *l = &path->l[level];
 637
 638         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
 639
 640         /*
 641          * Iterators to interior nodes should always be pointed at the first non
 642          * whiteout:
 643          */
 644         if (level)
 645                 bch2_btree_node_iter_peek(&l->iter, l->b);
 646 }
 647
 648 void bch2_btree_path_level_init(struct btree_trans *trans,
 649                                 struct btree_path *path,
 650                                 struct btree *b)
 651 {
 652         BUG_ON(path->cached);
 653
 654         EBUG_ON(!btree_path_pos_in_node(path, b));
 655         EBUG_ON(b->c.lock.state.seq & 1);
 656
 657         path->l[b->c.level].lock_seq = b->c.lock.state.seq;
 658         path->l[b->c.level].b = b;
 659         __btree_path_level_init(path, b->c.level);
 660 }
 661
 662 /* Btree path: fixups after btree node updates: */
 663
 664 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
 665 {
 666         struct bch_fs *c = trans->c;
 667         struct btree_insert_entry *i;
 668
 669         trans_for_each_update(trans, i)
 670                 if (!i->cached &&
 671                     i->level    == b->c.level &&
 672                     i->btree_id == b->c.btree_id &&
 673                     bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
 674                     bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
 675                         i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 676
 677                         if (unlikely(trans->journal_replay_not_finished)) {
 678                                 struct bkey_i *j_k =
 679                                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 680                                                                     i->k->k.p);
 681
 682                                 if (j_k) {
 683                                         i->old_k = j_k->k;
 684                                         i->old_v = &j_k->v;
 685                                 }
 686                         }
 687                 }
 688 }
 689
 690 /*
 691  * A btree node is being replaced - update the iterator to point to the new
 692  * node:
 693  */
 694 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
 695 {
 696         struct btree_path *path;
 697
 698         trans_for_each_path(trans, path)
 699                 if (path->uptodate == BTREE_ITER_UPTODATE &&
 700                     !path->cached &&
 701                     btree_path_pos_in_node(path, b)) {
 702                         enum btree_node_locked_type t =
 703                                 btree_lock_want(path, b->c.level);
 704
 705                         if (t != BTREE_NODE_UNLOCKED) {
 706                                 btree_node_unlock(trans, path, b->c.level);
 707                                 six_lock_increment(&b->c.lock, t);
 708                                 mark_btree_node_locked(trans, path, b->c.level, t);
 709                         }
 710
 711                         bch2_btree_path_level_init(trans, path, b);
 712                 }
 713
 714         bch2_trans_revalidate_updates_in_node(trans, b);
 715 }
 716
 717 /*
 718  * A btree node has been modified in such a way as to invalidate iterators - fix
 719  * them:
 720  */
 721 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
 722 {
 723         struct btree_path *path;
 724
 725         trans_for_each_path_with_node(trans, b, path)
 726                 __btree_path_level_init(path, b->c.level);
 727
 728         bch2_trans_revalidate_updates_in_node(trans, b);
 729 }
 730
 731 /* Btree path: traverse, set_pos: */
 732
 733 static inline int btree_path_lock_root(struct btree_trans *trans,
 734                                        struct btree_path *path,
 735                                        unsigned depth_want,
 736                                        unsigned long trace_ip)
 737 {
 738         struct bch_fs *c = trans->c;
 739         struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
 740         enum six_lock_type lock_type;
 741         unsigned i;
 742         int ret;
 743
 744         EBUG_ON(path->nodes_locked);
 745
 746         while (1) {
 747                 b = READ_ONCE(*rootp);
 748                 path->level = READ_ONCE(b->c.level);
 749
 750                 if (unlikely(path->level < depth_want)) {
 751                         /*
 752                          * the root is at a lower depth than the depth we want:
 753                          * got to the end of the btree, or we're walking nodes
 754                          * greater than some depth and there are no nodes >=
 755                          * that depth
 756                          */
 757                         path->level = depth_want;
 758                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
 759                                 path->l[i].b = NULL;
 760                         return 1;
 761                 }
 762
 763                 lock_type = __btree_lock_want(path, path->level);
 764                 ret = btree_node_lock(trans, path, &b->c,
 765                                       path->level, lock_type, trace_ip);
 766                 if (unlikely(ret)) {
 767                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
 768                                 continue;
 769                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 770                                 return ret;
 771                         BUG();
 772                 }
 773
 774                 if (likely(b == READ_ONCE(*rootp) &&
 775                            b->c.level == path->level &&
 776                            !race_fault())) {
 777                         for (i = 0; i < path->level; i++)
 778                                 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
 779                         path->l[path->level].b = b;
 780                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
 781                                 path->l[i].b = NULL;
 782
 783                         mark_btree_node_locked(trans, path, path->level, lock_type);
 784                         bch2_btree_path_level_init(trans, path, b);
 785                         return 0;
 786                 }
 787
 788                 six_unlock_type(&b->c.lock, lock_type);
 789         }
 790 }
 791
 792 noinline
 793 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
 794 {
 795         struct bch_fs *c = trans->c;
 796         struct btree_path_level *l = path_l(path);
 797         struct btree_node_iter node_iter = l->iter;
 798         struct bkey_packed *k;
 799         struct bkey_buf tmp;
 800         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 801                 ? (path->level > 1 ? 0 :  2)
 802                 : (path->level > 1 ? 1 : 16);
 803         bool was_locked = btree_node_locked(path, path->level);
 804         int ret = 0;
 805
 806         bch2_bkey_buf_init(&tmp);
 807
 808         while (nr-- && !ret) {
 809                 if (!bch2_btree_node_relock(trans, path, path->level))
 810                         break;
 811
 812                 bch2_btree_node_iter_advance(&node_iter, l->b);
 813                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
 814                 if (!k)
 815                         break;
 816
 817                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
 818                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
 819                                                path->level - 1);
 820         }
 821
 822         if (!was_locked)
 823                 btree_node_unlock(trans, path, path->level);
 824
 825         bch2_bkey_buf_exit(&tmp, c);
 826         return ret;
 827 }
 828
 829 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
 830                                  struct btree_and_journal_iter *jiter)
 831 {
 832         struct bch_fs *c = trans->c;
 833         struct bkey_s_c k;
 834         struct bkey_buf tmp;
 835         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 836                 ? (path->level > 1 ? 0 :  2)
 837                 : (path->level > 1 ? 1 : 16);
 838         bool was_locked = btree_node_locked(path, path->level);
 839         int ret = 0;
 840
 841         bch2_bkey_buf_init(&tmp);
 842
 843         while (nr-- && !ret) {
 844                 if (!bch2_btree_node_relock(trans, path, path->level))
 845                         break;
 846
 847                 bch2_btree_and_journal_iter_advance(jiter);
 848                 k = bch2_btree_and_journal_iter_peek(jiter);
 849                 if (!k.k)
 850                         break;
 851
 852                 bch2_bkey_buf_reassemble(&tmp, c, k);
 853                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
 854                                                path->level - 1);
 855         }
 856
 857         if (!was_locked)
 858                 btree_node_unlock(trans, path, path->level);
 859
 860         bch2_bkey_buf_exit(&tmp, c);
 861         return ret;
 862 }
 863
 864 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
 865                                             struct btree_path *path,
 866                                             unsigned plevel, struct btree *b)
 867 {
 868         struct btree_path_level *l = &path->l[plevel];
 869         bool locked = btree_node_locked(path, plevel);
 870         struct bkey_packed *k;
 871         struct bch_btree_ptr_v2 *bp;
 872
 873         if (!bch2_btree_node_relock(trans, path, plevel))
 874                 return;
 875
 876         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 877         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
 878
 879         bp = (void *) bkeyp_val(&l->b->format, k);
 880         bp->mem_ptr = (unsigned long)b;
 881
 882         if (!locked)
 883                 btree_node_unlock(trans, path, plevel);
 884 }
 885
 886 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
 887                                                      struct btree_path *path,
 888                                                      unsigned flags,
 889                                                      struct bkey_buf *out)
 890 {
 891         struct bch_fs *c = trans->c;
 892         struct btree_path_level *l = path_l(path);
 893         struct btree_and_journal_iter jiter;
 894         struct bkey_s_c k;
 895         int ret = 0;
 896
 897         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
 898
 899         k = bch2_btree_and_journal_iter_peek(&jiter);
 900
 901         bch2_bkey_buf_reassemble(out, c, k);
 902
 903         if (flags & BTREE_ITER_PREFETCH)
 904                 ret = btree_path_prefetch_j(trans, path, &jiter);
 905
 906         bch2_btree_and_journal_iter_exit(&jiter);
 907         return ret;
 908 }
 909
 910 static __always_inline int btree_path_down(struct btree_trans *trans,
 911                                            struct btree_path *path,
 912                                            unsigned flags,
 913                                            unsigned long trace_ip)
 914 {
 915         struct bch_fs *c = trans->c;
 916         struct btree_path_level *l = path_l(path);
 917         struct btree *b;
 918         unsigned level = path->level - 1;
 919         enum six_lock_type lock_type = __btree_lock_want(path, level);
 920         struct bkey_buf tmp;
 921         int ret;
 922
 923         EBUG_ON(!btree_node_locked(path, path->level));
 924
 925         bch2_bkey_buf_init(&tmp);
 926
 927         if (unlikely(trans->journal_replay_not_finished)) {
 928                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
 929                 if (ret)
 930                         goto err;
 931         } else {
 932                 bch2_bkey_buf_unpack(&tmp, c, l->b,
 933                                  bch2_btree_node_iter_peek(&l->iter, l->b));
 934
 935                 if (flags & BTREE_ITER_PREFETCH) {
 936                         ret = btree_path_prefetch(trans, path);
 937                         if (ret)
 938                                 goto err;
 939                 }
 940         }
 941
 942         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
 943         ret = PTR_ERR_OR_ZERO(b);
 944         if (unlikely(ret))
 945                 goto err;
 946
 947         if (likely(!trans->journal_replay_not_finished &&
 948                    tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
 949             unlikely(b != btree_node_mem_ptr(tmp.k)))
 950                 btree_node_mem_ptr_set(trans, path, level + 1, b);
 951
 952         if (btree_node_read_locked(path, level + 1))
 953                 btree_node_unlock(trans, path, level + 1);
 954
 955         mark_btree_node_locked(trans, path, level, lock_type);
 956         path->level = level;
 957         bch2_btree_path_level_init(trans, path, b);
 958
 959         bch2_btree_path_verify_locks(path);
 960 err:
 961         bch2_bkey_buf_exit(&tmp, c);
 962         return ret;
 963 }
 964
 965
 966 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
 967 {
 968         struct bch_fs *c = trans->c;
 969         struct btree_path *path;
 970         unsigned long trace_ip = _RET_IP_;
 971         int i, ret = 0;
 972
 973         if (trans->in_traverse_all)
 974                 return -BCH_ERR_transaction_restart_in_traverse_all;
 975
 976         trans->in_traverse_all = true;
 977 retry_all:
 978         trans->restarted = 0;
 979
 980         trans_for_each_path(trans, path)
 981                 path->should_be_locked = false;
 982
 983         btree_trans_sort_paths(trans);
 984
 985         bch2_trans_unlock(trans);
 986         cond_resched();
 987
 988         if (unlikely(trans->memory_allocation_failure)) {
 989                 struct closure cl;
 990
 991                 closure_init_stack(&cl);
 992
 993                 do {
 994                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 995                         closure_sync(&cl);
 996                 } while (ret);
 997         }
 998
 999         /* Now, redo traversals in correct order: */
1000         i = 0;
1001         while (i < trans->nr_sorted) {
1002                 path = trans->paths + trans->sorted[i];
1003
1004                 /*
1005                  * Traversing a path can cause another path to be added at about
1006                  * the same position:
1007                  */
1008                 if (path->uptodate) {
1009                         __btree_path_get(path, false);
1010                         ret = bch2_btree_path_traverse_one(trans, path, 0, _THIS_IP_);
1011                         __btree_path_put(path, false);
1012
1013                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1014                             ret == -ENOMEM)
1015                                 goto retry_all;
1016                         if (ret)
1017                                 goto err;
1018                 } else {
1019                         i++;
1020                 }
1021         }
1022
1023         /*
1024          * We used to assert that all paths had been traversed here
1025          * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1026          * path->Should_be_locked is not set yet, we we might have unlocked and
1027          * then failed to relock a path - that's fine.
1028          */
1029 err:
1030         bch2_btree_cache_cannibalize_unlock(c);
1031
1032         trans->in_traverse_all = false;
1033
1034         trace_and_count(c, trans_traverse_all, trans, trace_ip);
1035         return ret;
1036 }
1037
1038 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1039                                                 unsigned l, int check_pos)
1040 {
1041         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1042                 return false;
1043         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1044                 return false;
1045         return true;
1046 }
1047
1048 static inline bool btree_path_good_node(struct btree_trans *trans,
1049                                         struct btree_path *path,
1050                                         unsigned l, int check_pos)
1051 {
1052         return is_btree_node(path, l) &&
1053                 bch2_btree_node_relock(trans, path, l) &&
1054                 btree_path_check_pos_in_node(path, l, check_pos);
1055 }
1056
1057 static void btree_path_set_level_down(struct btree_trans *trans,
1058                                       struct btree_path *path,
1059                                       unsigned new_level)
1060 {
1061         unsigned l;
1062
1063         path->level = new_level;
1064
1065         for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1066                 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1067                         btree_node_unlock(trans, path, l);
1068
1069         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1070         bch2_btree_path_verify(trans, path);
1071 }
1072
1073 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1074                                                          struct btree_path *path,
1075                                                          int check_pos)
1076 {
1077         unsigned i, l = path->level;
1078 again:
1079         while (btree_path_node(path, l) &&
1080                !btree_path_good_node(trans, path, l, check_pos))
1081                 __btree_path_set_level_up(trans, path, l++);
1082
1083         /* If we need intent locks, take them too: */
1084         for (i = l + 1;
1085              i < path->locks_want && btree_path_node(path, i);
1086              i++)
1087                 if (!bch2_btree_node_relock(trans, path, i)) {
1088                         while (l <= i)
1089                                 __btree_path_set_level_up(trans, path, l++);
1090                         goto again;
1091                 }
1092
1093         return l;
1094 }
1095
1096 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1097                                                      struct btree_path *path,
1098                                                      int check_pos)
1099 {
1100         return likely(btree_node_locked(path, path->level) &&
1101                       btree_path_check_pos_in_node(path, path->level, check_pos))
1102                 ? path->level
1103                 : __btree_path_up_until_good_node(trans, path, check_pos);
1104 }
1105
1106 /*
1107  * This is the main state machine for walking down the btree - walks down to a
1108  * specified depth
1109  *
1110  * Returns 0 on success, -EIO on error (error reading in a btree node).
1111  *
1112  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1113  * stashed in the iterator and returned from bch2_trans_exit().
1114  */
1115 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1116                                  struct btree_path *path,
1117                                  unsigned flags,
1118                                  unsigned long trace_ip)
1119 {
1120         unsigned depth_want = path->level;
1121         int ret = -((int) trans->restarted);
1122
1123         if (unlikely(ret))
1124                 goto out;
1125
1126         /*
1127          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1128          * and re-traverse the path without a transaction restart:
1129          */
1130         if (path->should_be_locked) {
1131                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1132                 goto out;
1133         }
1134
1135         if (path->cached) {
1136                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1137                 goto out;
1138         }
1139
1140         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1141                 goto out;
1142
1143         path->level = btree_path_up_until_good_node(trans, path, 0);
1144
1145         EBUG_ON(btree_path_node(path, path->level) &&
1146                 !btree_node_locked(path, path->level));
1147
1148         /*
1149          * Note: path->nodes[path->level] may be temporarily NULL here - that
1150          * would indicate to other code that we got to the end of the btree,
1151          * here it indicates that relocking the root failed - it's critical that
1152          * btree_path_lock_root() comes next and that it can't fail
1153          */
1154         while (path->level > depth_want) {
1155                 ret = btree_path_node(path, path->level)
1156                         ? btree_path_down(trans, path, flags, trace_ip)
1157                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1158                 if (unlikely(ret)) {
1159                         if (ret == 1) {
1160                                 /*
1161                                  * No nodes at this level - got to the end of
1162                                  * the btree:
1163                                  */
1164                                 ret = 0;
1165                                 goto out;
1166                         }
1167
1168                         __bch2_btree_path_unlock(trans, path);
1169                         path->level = depth_want;
1170                         path->l[path->level].b = ERR_PTR(ret);
1171                         goto out;
1172                 }
1173         }
1174
1175         path->uptodate = BTREE_ITER_UPTODATE;
1176 out:
1177         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1178                 panic("ret %s (%i) trans->restarted %s (%i)\n",
1179                       bch2_err_str(ret), ret,
1180                       bch2_err_str(trans->restarted), trans->restarted);
1181         bch2_btree_path_verify(trans, path);
1182         return ret;
1183 }
1184
1185 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1186                             struct btree_path *src)
1187 {
1188         unsigned i, offset = offsetof(struct btree_path, pos);
1189
1190         memcpy((void *) dst + offset,
1191                (void *) src + offset,
1192                sizeof(struct btree_path) - offset);
1193
1194         for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1195                 unsigned t = btree_node_locked_type(dst, i);
1196
1197                 if (t != BTREE_NODE_UNLOCKED)
1198                         six_lock_increment(&dst->l[i].b->c.lock, t);
1199         }
1200 }
1201
1202 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1203                                            bool intent)
1204 {
1205         struct btree_path *new = btree_path_alloc(trans, src);
1206
1207         btree_path_copy(trans, new, src);
1208         __btree_path_get(new, intent);
1209         return new;
1210 }
1211
1212 __flatten
1213 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *trans,
1214                          struct btree_path *path, bool intent,
1215                          unsigned long ip)
1216 {
1217         __btree_path_put(path, intent);
1218         path = btree_path_clone(trans, path, intent);
1219         path->preserve = false;
1220         return path;
1221 }
1222
1223 struct btree_path * __must_check
1224 __bch2_btree_path_set_pos(struct btree_trans *trans,
1225                    struct btree_path *path, struct bpos new_pos,
1226                    bool intent, unsigned long ip, int cmp)
1227 {
1228         unsigned level = path->level;
1229
1230         bch2_trans_verify_not_in_restart(trans);
1231         EBUG_ON(!path->ref);
1232
1233         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1234
1235         path->pos               = new_pos;
1236         trans->paths_sorted     = false;
1237
1238         if (unlikely(path->cached)) {
1239                 btree_node_unlock(trans, path, 0);
1240                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1241                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1242                 goto out;
1243         }
1244
1245         level = btree_path_up_until_good_node(trans, path, cmp);
1246
1247         if (btree_path_node(path, level)) {
1248                 struct btree_path_level *l = &path->l[level];
1249
1250                 BUG_ON(!btree_node_locked(path, level));
1251                 /*
1252                  * We might have to skip over many keys, or just a few: try
1253                  * advancing the node iterator, and if we have to skip over too
1254                  * many keys just reinit it (or if we're rewinding, since that
1255                  * is expensive).
1256                  */
1257                 if (cmp < 0 ||
1258                     !btree_path_advance_to_pos(path, l, 8))
1259                         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1260
1261                 /*
1262                  * Iterators to interior nodes should always be pointed at the first non
1263                  * whiteout:
1264                  */
1265                 if (unlikely(level))
1266                         bch2_btree_node_iter_peek(&l->iter, l->b);
1267         }
1268
1269         if (unlikely(level != path->level)) {
1270                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1271                 __bch2_btree_path_unlock(trans, path);
1272         }
1273 out:
1274         bch2_btree_path_verify(trans, path);
1275         return path;
1276 }
1277
1278 /* Btree path: main interface: */
1279
1280 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1281 {
1282         struct btree_path *sib;
1283
1284         sib = prev_btree_path(trans, path);
1285         if (sib && !btree_path_cmp(sib, path))
1286                 return sib;
1287
1288         sib = next_btree_path(trans, path);
1289         if (sib && !btree_path_cmp(sib, path))
1290                 return sib;
1291
1292         return NULL;
1293 }
1294
1295 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1296 {
1297         struct btree_path *sib;
1298
1299         sib = prev_btree_path(trans, path);
1300         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1301                 return sib;
1302
1303         sib = next_btree_path(trans, path);
1304         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1305                 return sib;
1306
1307         return NULL;
1308 }
1309
1310 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1311 {
1312         __bch2_btree_path_unlock(trans, path);
1313         btree_path_list_remove(trans, path);
1314         trans->paths_allocated &= ~(1ULL << path->idx);
1315 }
1316
1317 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1318 {
1319         struct btree_path *dup;
1320
1321         EBUG_ON(trans->paths + path->idx != path);
1322         EBUG_ON(!path->ref);
1323
1324         if (!__btree_path_put(path, intent))
1325                 return;
1326
1327         dup = path->preserve
1328                 ? have_path_at_pos(trans, path)
1329                 : have_node_at_pos(trans, path);
1330
1331         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1332                 return;
1333
1334         if (path->should_be_locked &&
1335             !trans->restarted &&
1336             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1337                 return;
1338
1339         if (dup) {
1340                 dup->preserve           |= path->preserve;
1341                 dup->should_be_locked   |= path->should_be_locked;
1342         }
1343
1344         __bch2_path_free(trans, path);
1345 }
1346
1347 static void bch2_path_put_nokeep(struct btree_trans *trans, struct btree_path *path,
1348                                  bool intent)
1349 {
1350         EBUG_ON(trans->paths + path->idx != path);
1351         EBUG_ON(!path->ref);
1352
1353         if (!__btree_path_put(path, intent))
1354                 return;
1355
1356         __bch2_path_free(trans, path);
1357 }
1358
1359 void bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1360 {
1361         panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1362               trans->restart_count, restart_count,
1363               (void *) trans->last_restarted_ip);
1364 }
1365
1366 void bch2_trans_in_restart_error(struct btree_trans *trans)
1367 {
1368         panic("in transaction restart: %s, last restarted by %pS\n",
1369               bch2_err_str(trans->restarted),
1370               (void *) trans->last_restarted_ip);
1371 }
1372
1373 noinline __cold
1374 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1375 {
1376         struct btree_insert_entry *i;
1377         struct btree_write_buffered_key *wb;
1378
1379         prt_printf(buf, "transaction updates for %s journal seq %llu",
1380                trans->fn, trans->journal_res.seq);
1381         prt_newline(buf);
1382         printbuf_indent_add(buf, 2);
1383
1384         trans_for_each_update(trans, i) {
1385                 struct bkey_s_c old = { &i->old_k, i->old_v };
1386
1387                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1388                        bch2_btree_ids[i->btree_id],
1389                        i->cached,
1390                        (void *) i->ip_allocated);
1391                 prt_newline(buf);
1392
1393                 prt_printf(buf, "  old ");
1394                 bch2_bkey_val_to_text(buf, trans->c, old);
1395                 prt_newline(buf);
1396
1397                 prt_printf(buf, "  new ");
1398                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1399                 prt_newline(buf);
1400         }
1401
1402         trans_for_each_wb_update(trans, wb) {
1403                 prt_printf(buf, "update: btree=%s wb=1 %pS",
1404                        bch2_btree_ids[wb->btree],
1405                        (void *) i->ip_allocated);
1406                 prt_newline(buf);
1407
1408                 prt_printf(buf, "  new ");
1409                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(&wb->k));
1410                 prt_newline(buf);
1411         }
1412
1413         printbuf_indent_sub(buf, 2);
1414 }
1415
1416 noinline __cold
1417 void bch2_dump_trans_updates(struct btree_trans *trans)
1418 {
1419         struct printbuf buf = PRINTBUF;
1420
1421         bch2_trans_updates_to_text(&buf, trans);
1422         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1423         printbuf_exit(&buf);
1424 }
1425
1426 noinline __cold
1427 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1428 {
1429         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1430                    path->idx, path->ref, path->intent_ref,
1431                    path->preserve ? 'P' : ' ',
1432                    path->should_be_locked ? 'S' : ' ',
1433                    bch2_btree_ids[path->btree_id],
1434                    path->level);
1435         bch2_bpos_to_text(out, path->pos);
1436
1437         prt_printf(out, " locks %u", path->nodes_locked);
1438 #ifdef TRACK_PATH_ALLOCATED
1439         prt_printf(out, " %pS", (void *) path->ip_allocated);
1440 #endif
1441         prt_newline(out);
1442 }
1443
1444 noinline __cold
1445 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1446                                 bool nosort)
1447 {
1448         struct btree_path *path;
1449         unsigned idx;
1450
1451         if (!nosort)
1452                 btree_trans_sort_paths(trans);
1453
1454         trans_for_each_path_inorder(trans, path, idx)
1455                 bch2_btree_path_to_text(out, path);
1456 }
1457
1458 noinline __cold
1459 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1460 {
1461         __bch2_trans_paths_to_text(out, trans, false);
1462 }
1463
1464 noinline __cold
1465 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1466 {
1467         struct printbuf buf = PRINTBUF;
1468
1469         __bch2_trans_paths_to_text(&buf, trans, nosort);
1470         bch2_trans_updates_to_text(&buf, trans);
1471
1472         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1473         printbuf_exit(&buf);
1474 }
1475
1476 noinline __cold
1477 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1478 {
1479         __bch2_dump_trans_paths_updates(trans, false);
1480 }
1481
1482 noinline __cold
1483 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1484 {
1485         struct btree_transaction_stats *s = btree_trans_stats(trans);
1486         struct printbuf buf = PRINTBUF;
1487
1488         if (!s)
1489                 return;
1490
1491         bch2_trans_paths_to_text(&buf, trans);
1492
1493         if (!buf.allocation_failure) {
1494                 mutex_lock(&s->lock);
1495                 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1496                         s->nr_max_paths = trans->nr_max_paths =
1497                                 hweight64(trans->paths_allocated);
1498                         swap(s->max_paths_text, buf.buf);
1499                 }
1500                 mutex_unlock(&s->lock);
1501         }
1502
1503         printbuf_exit(&buf);
1504
1505         trans->nr_max_paths = hweight64(trans->paths_allocated);
1506 }
1507
1508 static noinline void btree_path_overflow(struct btree_trans *trans)
1509 {
1510         bch2_dump_trans_paths_updates(trans);
1511         panic("trans path oveflow\n");
1512 }
1513
1514 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1515                                                   struct btree_path *pos)
1516 {
1517         struct btree_path *path;
1518         unsigned idx;
1519
1520         if (unlikely(trans->paths_allocated ==
1521                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
1522                 btree_path_overflow(trans);
1523
1524         idx = __ffs64(~trans->paths_allocated);
1525
1526         /*
1527          * Do this before marking the new path as allocated, since it won't be
1528          * initialized yet:
1529          */
1530         if (unlikely(idx > trans->nr_max_paths))
1531                 bch2_trans_update_max_paths(trans);
1532
1533         trans->paths_allocated |= 1ULL << idx;
1534
1535         path = &trans->paths[idx];
1536         path->idx               = idx;
1537         path->ref               = 0;
1538         path->intent_ref        = 0;
1539         path->nodes_locked      = 0;
1540
1541         btree_path_list_add(trans, pos, path);
1542         trans->paths_sorted = false;
1543         return path;
1544 }
1545
1546 struct btree_path *bch2_path_get(struct btree_trans *trans,
1547                                  enum btree_id btree_id, struct bpos pos,
1548                                  unsigned locks_want, unsigned level,
1549                                  unsigned flags, unsigned long ip)
1550 {
1551         struct btree_path *path, *path_pos = NULL;
1552         bool cached = flags & BTREE_ITER_CACHED;
1553         bool intent = flags & BTREE_ITER_INTENT;
1554         int i;
1555
1556         bch2_trans_verify_not_in_restart(trans);
1557         bch2_trans_verify_locks(trans);
1558
1559         btree_trans_sort_paths(trans);
1560
1561         trans_for_each_path_inorder(trans, path, i) {
1562                 if (__btree_path_cmp(path,
1563                                      btree_id,
1564                                      cached,
1565                                      pos,
1566                                      level) > 0)
1567                         break;
1568
1569                 path_pos = path;
1570         }
1571
1572         if (path_pos &&
1573             path_pos->cached    == cached &&
1574             path_pos->btree_id  == btree_id &&
1575             path_pos->level     == level) {
1576                 __btree_path_get(path_pos, intent);
1577                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1578         } else {
1579                 path = btree_path_alloc(trans, path_pos);
1580                 path_pos = NULL;
1581
1582                 __btree_path_get(path, intent);
1583                 path->pos                       = pos;
1584                 path->btree_id                  = btree_id;
1585                 path->cached                    = cached;
1586                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1587                 path->should_be_locked          = false;
1588                 path->level                     = level;
1589                 path->locks_want                = locks_want;
1590                 path->nodes_locked              = 0;
1591                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1592                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1593 #ifdef TRACK_PATH_ALLOCATED
1594                 path->ip_allocated              = ip;
1595 #endif
1596                 trans->paths_sorted             = false;
1597         }
1598
1599         if (!(flags & BTREE_ITER_NOPRESERVE))
1600                 path->preserve = true;
1601
1602         if (path->intent_ref)
1603                 locks_want = max(locks_want, level + 1);
1604
1605         /*
1606          * If the path has locks_want greater than requested, we don't downgrade
1607          * it here - on transaction restart because btree node split needs to
1608          * upgrade locks, we might be putting/getting the iterator again.
1609          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1610          * a successful transaction commit.
1611          */
1612
1613         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1614         if (locks_want > path->locks_want)
1615                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want);
1616
1617         return path;
1618 }
1619
1620 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1621 {
1622
1623         struct btree_path_level *l = path_l(path);
1624         struct bkey_packed *_k;
1625         struct bkey_s_c k;
1626
1627         if (unlikely(!l->b))
1628                 return bkey_s_c_null;
1629
1630         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1631         EBUG_ON(!btree_node_locked(path, path->level));
1632
1633         if (!path->cached) {
1634                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1635                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1636
1637                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1638
1639                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1640                         goto hole;
1641         } else {
1642                 struct bkey_cached *ck = (void *) path->l[0].b;
1643
1644                 EBUG_ON(ck &&
1645                         (path->btree_id != ck->key.btree_id ||
1646                          !bkey_eq(path->pos, ck->key.pos)));
1647                 if (!ck || !ck->valid)
1648                         return bkey_s_c_null;
1649
1650                 *u = ck->k->k;
1651                 k = bkey_i_to_s_c(ck->k);
1652         }
1653
1654         return k;
1655 hole:
1656         bkey_init(u);
1657         u->p = path->pos;
1658         return (struct bkey_s_c) { u, NULL };
1659 }
1660
1661 /* Btree iterators: */
1662
1663 int __must_check
1664 __bch2_btree_iter_traverse(struct btree_iter *iter)
1665 {
1666         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1667 }
1668
1669 int __must_check
1670 bch2_btree_iter_traverse(struct btree_iter *iter)
1671 {
1672         int ret;
1673
1674         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1675                                         btree_iter_search_key(iter),
1676                                         iter->flags & BTREE_ITER_INTENT,
1677                                         btree_iter_ip_allocated(iter));
1678
1679         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1680         if (ret)
1681                 return ret;
1682
1683         btree_path_set_should_be_locked(iter->path);
1684         return 0;
1685 }
1686
1687 /* Iterate across nodes (leaf and interior nodes) */
1688
1689 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1690 {
1691         struct btree_trans *trans = iter->trans;
1692         struct btree *b = NULL;
1693         int ret;
1694
1695         EBUG_ON(iter->path->cached);
1696         bch2_btree_iter_verify(iter);
1697
1698         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1699         if (ret)
1700                 goto err;
1701
1702         b = btree_path_node(iter->path, iter->path->level);
1703         if (!b)
1704                 goto out;
1705
1706         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1707
1708         bkey_init(&iter->k);
1709         iter->k.p = iter->pos = b->key.k.p;
1710
1711         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1712                                         iter->flags & BTREE_ITER_INTENT,
1713                                         btree_iter_ip_allocated(iter));
1714         btree_path_set_should_be_locked(iter->path);
1715 out:
1716         bch2_btree_iter_verify_entry_exit(iter);
1717         bch2_btree_iter_verify(iter);
1718
1719         return b;
1720 err:
1721         b = ERR_PTR(ret);
1722         goto out;
1723 }
1724
1725 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1726 {
1727         struct btree_trans *trans = iter->trans;
1728         struct btree_path *path = iter->path;
1729         struct btree *b = NULL;
1730         int ret;
1731
1732         bch2_trans_verify_not_in_restart(trans);
1733         EBUG_ON(iter->path->cached);
1734         bch2_btree_iter_verify(iter);
1735
1736         /* already at end? */
1737         if (!btree_path_node(path, path->level))
1738                 return NULL;
1739
1740         /* got to end? */
1741         if (!btree_path_node(path, path->level + 1)) {
1742                 btree_path_set_level_up(trans, path);
1743                 return NULL;
1744         }
1745
1746         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1747                 __bch2_btree_path_unlock(trans, path);
1748                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1749                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1750                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1751                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1752                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1753                 goto err;
1754         }
1755
1756         b = btree_path_node(path, path->level + 1);
1757
1758         if (bpos_eq(iter->pos, b->key.k.p)) {
1759                 __btree_path_set_level_up(trans, path, path->level++);
1760         } else {
1761                 /*
1762                  * Haven't gotten to the end of the parent node: go back down to
1763                  * the next child node
1764                  */
1765                 path = iter->path =
1766                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1767                                            iter->flags & BTREE_ITER_INTENT,
1768                                            btree_iter_ip_allocated(iter));
1769
1770                 btree_path_set_level_down(trans, path, iter->min_depth);
1771
1772                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1773                 if (ret)
1774                         goto err;
1775
1776                 b = path->l[path->level].b;
1777         }
1778
1779         bkey_init(&iter->k);
1780         iter->k.p = iter->pos = b->key.k.p;
1781
1782         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1783                                         iter->flags & BTREE_ITER_INTENT,
1784                                         btree_iter_ip_allocated(iter));
1785         btree_path_set_should_be_locked(iter->path);
1786         BUG_ON(iter->path->uptodate);
1787 out:
1788         bch2_btree_iter_verify_entry_exit(iter);
1789         bch2_btree_iter_verify(iter);
1790
1791         return b;
1792 err:
1793         b = ERR_PTR(ret);
1794         goto out;
1795 }
1796
1797 /* Iterate across keys (in leaf nodes only) */
1798
1799 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1800 {
1801         if (likely(!(iter->flags & BTREE_ITER_ALL_LEVELS))) {
1802                 struct bpos pos = iter->k.p;
1803                 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1804                              ? bpos_eq(pos, SPOS_MAX)
1805                              : bkey_eq(pos, SPOS_MAX));
1806
1807                 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1808                         pos = bkey_successor(iter, pos);
1809                 bch2_btree_iter_set_pos(iter, pos);
1810                 return ret;
1811         } else {
1812                 if (!btree_path_node(iter->path, iter->path->level))
1813                         return true;
1814
1815                 iter->advanced = true;
1816                 return false;
1817         }
1818 }
1819
1820 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1821 {
1822         struct bpos pos = bkey_start_pos(&iter->k);
1823         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1824                      ? bpos_eq(pos, POS_MIN)
1825                      : bkey_eq(pos, POS_MIN));
1826
1827         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1828                 pos = bkey_predecessor(iter, pos);
1829         bch2_btree_iter_set_pos(iter, pos);
1830         return ret;
1831 }
1832
1833 static noinline
1834 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1835 {
1836         struct btree_insert_entry *i;
1837         struct bkey_i *ret = NULL;
1838
1839         trans_for_each_update(iter->trans, i) {
1840                 if (i->btree_id < iter->btree_id)
1841                         continue;
1842                 if (i->btree_id > iter->btree_id)
1843                         break;
1844                 if (bpos_lt(i->k->k.p, iter->path->pos))
1845                         continue;
1846                 if (i->key_cache_already_flushed)
1847                         continue;
1848                 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1849                         ret = i->k;
1850         }
1851
1852         return ret;
1853 }
1854
1855 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1856 {
1857         return iter->flags & BTREE_ITER_WITH_UPDATES
1858                 ? __bch2_btree_trans_peek_updates(iter)
1859                 : NULL;
1860 }
1861
1862 struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1863                                        struct btree_iter *iter,
1864                                        struct bpos end_pos)
1865 {
1866         struct bkey_i *k;
1867
1868         if (bpos_lt(iter->path->pos, iter->journal_pos))
1869                 iter->journal_idx = 0;
1870
1871         k = bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1872                                         iter->path->level,
1873                                         iter->path->pos,
1874                                         end_pos,
1875                                         &iter->journal_idx);
1876
1877         iter->journal_pos = k ? k->k.p : end_pos;
1878         return k;
1879 }
1880
1881 static noinline
1882 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1883                                               struct btree_iter *iter)
1884 {
1885         struct bkey_i *k = bch2_btree_journal_peek(trans, iter, iter->path->pos);
1886
1887         if (k) {
1888                 iter->k = k->k;
1889                 return bkey_i_to_s_c(k);
1890         } else {
1891                 return bkey_s_c_null;
1892         }
1893 }
1894
1895 static noinline
1896 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1897                                          struct btree_iter *iter,
1898                                          struct bkey_s_c k)
1899 {
1900         struct bkey_i *next_journal =
1901                 bch2_btree_journal_peek(trans, iter,
1902                                 k.k ? k.k->p : path_l(iter->path)->b->key.k.p);
1903
1904         if (next_journal) {
1905                 iter->k = next_journal->k;
1906                 k = bkey_i_to_s_c(next_journal);
1907         }
1908
1909         return k;
1910 }
1911
1912 /*
1913  * Checks btree key cache for key at iter->pos and returns it if present, or
1914  * bkey_s_c_null:
1915  */
1916 static noinline
1917 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1918 {
1919         struct btree_trans *trans = iter->trans;
1920         struct bch_fs *c = trans->c;
1921         struct bkey u;
1922         struct bkey_s_c k;
1923         int ret;
1924
1925         if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1926             bpos_eq(iter->pos, pos))
1927                 return bkey_s_c_null;
1928
1929         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1930                 return bkey_s_c_null;
1931
1932         if (!iter->key_cache_path)
1933                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1934                                                      iter->flags & BTREE_ITER_INTENT, 0,
1935                                                      iter->flags|BTREE_ITER_CACHED|
1936                                                      BTREE_ITER_CACHED_NOFILL,
1937                                                      _THIS_IP_);
1938
1939         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1940                                         iter->flags & BTREE_ITER_INTENT,
1941                                         btree_iter_ip_allocated(iter));
1942
1943         ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
1944                                          iter->flags|BTREE_ITER_CACHED) ?:
1945                 bch2_btree_path_relock(trans, iter->path, _THIS_IP_);
1946         if (unlikely(ret))
1947                 return bkey_s_c_err(ret);
1948
1949         btree_path_set_should_be_locked(iter->key_cache_path);
1950
1951         k = bch2_btree_path_peek_slot(iter->key_cache_path, &u);
1952         if (k.k && !bkey_err(k)) {
1953                 iter->k = u;
1954                 k.k = &iter->k;
1955         }
1956         return k;
1957 }
1958
1959 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1960 {
1961         struct btree_trans *trans = iter->trans;
1962         struct bkey_i *next_update;
1963         struct bkey_s_c k, k2;
1964         int ret;
1965
1966         EBUG_ON(iter->path->cached);
1967         bch2_btree_iter_verify(iter);
1968
1969         while (1) {
1970                 struct btree_path_level *l;
1971
1972                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
1973                                         iter->flags & BTREE_ITER_INTENT,
1974                                         btree_iter_ip_allocated(iter));
1975
1976                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1977                 if (unlikely(ret)) {
1978                         /* ensure that iter->k is consistent with iter->pos: */
1979                         bch2_btree_iter_set_pos(iter, iter->pos);
1980                         k = bkey_s_c_err(ret);
1981                         goto out;
1982                 }
1983
1984                 l = path_l(iter->path);
1985
1986                 if (unlikely(!l->b)) {
1987                         /* No btree nodes at requested level: */
1988                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
1989                         k = bkey_s_c_null;
1990                         goto out;
1991                 }
1992
1993                 btree_path_set_should_be_locked(iter->path);
1994
1995                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
1996
1997                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
1998                     k.k &&
1999                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2000                         k = k2;
2001                         ret = bkey_err(k);
2002                         if (ret) {
2003                                 bch2_btree_iter_set_pos(iter, iter->pos);
2004                                 goto out;
2005                         }
2006                 }
2007
2008                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2009                         k = btree_trans_peek_journal(trans, iter, k);
2010
2011                 next_update = btree_trans_peek_updates(iter);
2012
2013                 if (next_update &&
2014                     bpos_le(next_update->k.p,
2015                             k.k ? k.k->p : l->b->key.k.p)) {
2016                         iter->k = next_update->k;
2017                         k = bkey_i_to_s_c(next_update);
2018                 }
2019
2020                 if (k.k && bkey_deleted(k.k)) {
2021                         /*
2022                          * If we've got a whiteout, and it's after the search
2023                          * key, advance the search key to the whiteout instead
2024                          * of just after the whiteout - it might be a btree
2025                          * whiteout, with a real key at the same position, since
2026                          * in the btree deleted keys sort before non deleted.
2027                          */
2028                         search_key = !bpos_eq(search_key, k.k->p)
2029                                 ? k.k->p
2030                                 : bpos_successor(k.k->p);
2031                         continue;
2032                 }
2033
2034                 if (likely(k.k)) {
2035                         break;
2036                 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2037                         /* Advance to next leaf node: */
2038                         search_key = bpos_successor(l->b->key.k.p);
2039                 } else {
2040                         /* End of btree: */
2041                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2042                         k = bkey_s_c_null;
2043                         goto out;
2044                 }
2045         }
2046 out:
2047         bch2_btree_iter_verify(iter);
2048
2049         return k;
2050 }
2051
2052 /**
2053  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2054  * current position
2055  */
2056 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2057 {
2058         struct btree_trans *trans = iter->trans;
2059         struct bpos search_key = btree_iter_search_key(iter);
2060         struct bkey_s_c k;
2061         struct bpos iter_pos;
2062         int ret;
2063
2064         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2065         EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2066
2067         if (iter->update_path) {
2068                 bch2_path_put_nokeep(trans, iter->update_path,
2069                                      iter->flags & BTREE_ITER_INTENT);
2070                 iter->update_path = NULL;
2071         }
2072
2073         bch2_btree_iter_verify_entry_exit(iter);
2074
2075         while (1) {
2076                 k = __bch2_btree_iter_peek(iter, search_key);
2077                 if (unlikely(!k.k))
2078                         goto end;
2079                 if (unlikely(bkey_err(k)))
2080                         goto out_no_locked;
2081
2082                 /*
2083                  * iter->pos should be mononotically increasing, and always be
2084                  * equal to the key we just returned - except extents can
2085                  * straddle iter->pos:
2086                  */
2087                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2088                         iter_pos = k.k->p;
2089                 else
2090                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2091
2092                 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2093                              ? bkey_gt(iter_pos, end)
2094                              : bkey_ge(iter_pos, end)))
2095                         goto end;
2096
2097                 if (iter->update_path &&
2098                     !bkey_eq(iter->update_path->pos, k.k->p)) {
2099                         bch2_path_put_nokeep(trans, iter->update_path,
2100                                              iter->flags & BTREE_ITER_INTENT);
2101                         iter->update_path = NULL;
2102                 }
2103
2104                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2105                     (iter->flags & BTREE_ITER_INTENT) &&
2106                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2107                     !iter->update_path) {
2108                         struct bpos pos = k.k->p;
2109
2110                         if (pos.snapshot < iter->snapshot) {
2111                                 search_key = bpos_successor(k.k->p);
2112                                 continue;
2113                         }
2114
2115                         pos.snapshot = iter->snapshot;
2116
2117                         /*
2118                          * advance, same as on exit for iter->path, but only up
2119                          * to snapshot
2120                          */
2121                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2122                         iter->update_path = iter->path;
2123
2124                         iter->update_path = bch2_btree_path_set_pos(trans,
2125                                                 iter->update_path, pos,
2126                                                 iter->flags & BTREE_ITER_INTENT,
2127                                                 _THIS_IP_);
2128                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2129                         if (unlikely(ret)) {
2130                                 k = bkey_s_c_err(ret);
2131                                 goto out_no_locked;
2132                         }
2133                 }
2134
2135                 /*
2136                  * We can never have a key in a leaf node at POS_MAX, so
2137                  * we don't have to check these successor() calls:
2138                  */
2139                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2140                     !bch2_snapshot_is_ancestor(trans->c,
2141                                                iter->snapshot,
2142                                                k.k->p.snapshot)) {
2143                         search_key = bpos_successor(k.k->p);
2144                         continue;
2145                 }
2146
2147                 if (bkey_whiteout(k.k) &&
2148                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2149                         search_key = bkey_successor(iter, k.k->p);
2150                         continue;
2151                 }
2152
2153                 break;
2154         }
2155
2156         iter->pos = iter_pos;
2157
2158         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2159                                 iter->flags & BTREE_ITER_INTENT,
2160                                 btree_iter_ip_allocated(iter));
2161
2162         btree_path_set_should_be_locked(iter->path);
2163 out_no_locked:
2164         if (iter->update_path) {
2165                 ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_);
2166                 if (unlikely(ret))
2167                         k = bkey_s_c_err(ret);
2168                 else
2169                         btree_path_set_should_be_locked(iter->update_path);
2170         }
2171
2172         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2173                 iter->pos.snapshot = iter->snapshot;
2174
2175         ret = bch2_btree_iter_verify_ret(iter, k);
2176         if (unlikely(ret)) {
2177                 bch2_btree_iter_set_pos(iter, iter->pos);
2178                 k = bkey_s_c_err(ret);
2179         }
2180
2181         bch2_btree_iter_verify_entry_exit(iter);
2182
2183         return k;
2184 end:
2185         bch2_btree_iter_set_pos(iter, end);
2186         k = bkey_s_c_null;
2187         goto out_no_locked;
2188 }
2189
2190 /**
2191  * bch2_btree_iter_peek_all_levels: returns the first key greater than or equal
2192  * to iterator's current position, returning keys from every level of the btree.
2193  * For keys at different levels of the btree that compare equal, the key from
2194  * the lower level (leaf) is returned first.
2195  */
2196 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *iter)
2197 {
2198         struct btree_trans *trans = iter->trans;
2199         struct bkey_s_c k;
2200         int ret;
2201
2202         EBUG_ON(iter->path->cached);
2203         bch2_btree_iter_verify(iter);
2204         BUG_ON(iter->path->level < iter->min_depth);
2205         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
2206         EBUG_ON(!(iter->flags & BTREE_ITER_ALL_LEVELS));
2207
2208         while (1) {
2209                 iter->path = bch2_btree_path_set_pos(trans, iter->path, iter->pos,
2210                                         iter->flags & BTREE_ITER_INTENT,
2211                                         btree_iter_ip_allocated(iter));
2212
2213                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2214                 if (unlikely(ret)) {
2215                         /* ensure that iter->k is consistent with iter->pos: */
2216                         bch2_btree_iter_set_pos(iter, iter->pos);
2217                         k = bkey_s_c_err(ret);
2218                         goto out_no_locked;
2219                 }
2220
2221                 /* Already at end? */
2222                 if (!btree_path_node(iter->path, iter->path->level)) {
2223                         k = bkey_s_c_null;
2224                         goto out_no_locked;
2225                 }
2226
2227                 k = btree_path_level_peek_all(trans->c,
2228                                 &iter->path->l[iter->path->level], &iter->k);
2229
2230                 /* Check if we should go up to the parent node: */
2231                 if (!k.k ||
2232                     (iter->advanced &&
2233                      bpos_eq(path_l(iter->path)->b->key.k.p, iter->pos))) {
2234                         iter->pos = path_l(iter->path)->b->key.k.p;
2235                         btree_path_set_level_up(trans, iter->path);
2236                         iter->advanced = false;
2237                         continue;
2238                 }
2239
2240                 /*
2241                  * Check if we should go back down to a leaf:
2242                  * If we're not in a leaf node, we only return the current key
2243                  * if it exactly matches iter->pos - otherwise we first have to
2244                  * go back to the leaf:
2245                  */
2246                 if (iter->path->level != iter->min_depth &&
2247                     (iter->advanced ||
2248                      !k.k ||
2249                      !bpos_eq(iter->pos, k.k->p))) {
2250                         btree_path_set_level_down(trans, iter->path, iter->min_depth);
2251                         iter->pos = bpos_successor(iter->pos);
2252                         iter->advanced = false;
2253                         continue;
2254                 }
2255
2256                 /* Check if we should go to the next key: */
2257                 if (iter->path->level == iter->min_depth &&
2258                     iter->advanced &&
2259                     k.k &&
2260                     bpos_eq(iter->pos, k.k->p)) {
2261                         iter->pos = bpos_successor(iter->pos);
2262                         iter->advanced = false;
2263                         continue;
2264                 }
2265
2266                 if (iter->advanced &&
2267                     iter->path->level == iter->min_depth &&
2268                     !bpos_eq(k.k->p, iter->pos))
2269                         iter->advanced = false;
2270
2271                 BUG_ON(iter->advanced);
2272                 BUG_ON(!k.k);
2273                 break;
2274         }
2275
2276         iter->pos = k.k->p;
2277         btree_path_set_should_be_locked(iter->path);
2278 out_no_locked:
2279         bch2_btree_iter_verify(iter);
2280
2281         return k;
2282 }
2283
2284 /**
2285  * bch2_btree_iter_next: returns first key greater than iterator's current
2286  * position
2287  */
2288 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2289 {
2290         if (!bch2_btree_iter_advance(iter))
2291                 return bkey_s_c_null;
2292
2293         return bch2_btree_iter_peek(iter);
2294 }
2295
2296 /**
2297  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2298  * iterator's current position
2299  */
2300 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2301 {
2302         struct btree_trans *trans = iter->trans;
2303         struct bpos search_key = iter->pos;
2304         struct btree_path *saved_path = NULL;
2305         struct bkey_s_c k;
2306         struct bkey saved_k;
2307         const struct bch_val *saved_v;
2308         int ret;
2309
2310         EBUG_ON(iter->path->cached || iter->path->level);
2311         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2312
2313         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2314                 return bkey_s_c_err(-EIO);
2315
2316         bch2_btree_iter_verify(iter);
2317         bch2_btree_iter_verify_entry_exit(iter);
2318
2319         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2320                 search_key.snapshot = U32_MAX;
2321
2322         while (1) {
2323                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2324                                                 iter->flags & BTREE_ITER_INTENT,
2325                                                 btree_iter_ip_allocated(iter));
2326
2327                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2328                 if (unlikely(ret)) {
2329                         /* ensure that iter->k is consistent with iter->pos: */
2330                         bch2_btree_iter_set_pos(iter, iter->pos);
2331                         k = bkey_s_c_err(ret);
2332                         goto out_no_locked;
2333                 }
2334
2335                 k = btree_path_level_peek(trans, iter->path,
2336                                           &iter->path->l[0], &iter->k);
2337                 if (!k.k ||
2338                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2339                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2340                      : bpos_gt(k.k->p, search_key)))
2341                         k = btree_path_level_prev(trans, iter->path,
2342                                                   &iter->path->l[0], &iter->k);
2343
2344                 if (likely(k.k)) {
2345                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2346                                 if (k.k->p.snapshot == iter->snapshot)
2347                                         goto got_key;
2348
2349                                 /*
2350                                  * If we have a saved candidate, and we're no
2351                                  * longer at the same _key_ (not pos), return
2352                                  * that candidate
2353                                  */
2354                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2355                                         bch2_path_put_nokeep(trans, iter->path,
2356                                                       iter->flags & BTREE_ITER_INTENT);
2357                                         iter->path = saved_path;
2358                                         saved_path = NULL;
2359                                         iter->k = saved_k;
2360                                         k.v     = saved_v;
2361                                         goto got_key;
2362                                 }
2363
2364                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2365                                                               iter->snapshot,
2366                                                               k.k->p.snapshot)) {
2367                                         if (saved_path)
2368                                                 bch2_path_put_nokeep(trans, saved_path,
2369                                                       iter->flags & BTREE_ITER_INTENT);
2370                                         saved_path = btree_path_clone(trans, iter->path,
2371                                                                 iter->flags & BTREE_ITER_INTENT);
2372                                         saved_k = *k.k;
2373                                         saved_v = k.v;
2374                                 }
2375
2376                                 search_key = bpos_predecessor(k.k->p);
2377                                 continue;
2378                         }
2379 got_key:
2380                         if (bkey_whiteout(k.k) &&
2381                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2382                                 search_key = bkey_predecessor(iter, k.k->p);
2383                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2384                                         search_key.snapshot = U32_MAX;
2385                                 continue;
2386                         }
2387
2388                         break;
2389                 } else if (likely(!bpos_eq(iter->path->l[0].b->data->min_key, POS_MIN))) {
2390                         /* Advance to previous leaf node: */
2391                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2392                 } else {
2393                         /* Start of btree: */
2394                         bch2_btree_iter_set_pos(iter, POS_MIN);
2395                         k = bkey_s_c_null;
2396                         goto out_no_locked;
2397                 }
2398         }
2399
2400         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2401
2402         /* Extents can straddle iter->pos: */
2403         if (bkey_lt(k.k->p, iter->pos))
2404                 iter->pos = k.k->p;
2405
2406         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2407                 iter->pos.snapshot = iter->snapshot;
2408
2409         btree_path_set_should_be_locked(iter->path);
2410 out_no_locked:
2411         if (saved_path)
2412                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2413
2414         bch2_btree_iter_verify_entry_exit(iter);
2415         bch2_btree_iter_verify(iter);
2416
2417         return k;
2418 }
2419
2420 /**
2421  * bch2_btree_iter_prev: returns first key less than iterator's current
2422  * position
2423  */
2424 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2425 {
2426         if (!bch2_btree_iter_rewind(iter))
2427                 return bkey_s_c_null;
2428
2429         return bch2_btree_iter_peek_prev(iter);
2430 }
2431
2432 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2433 {
2434         struct btree_trans *trans = iter->trans;
2435         struct bpos search_key;
2436         struct bkey_s_c k;
2437         int ret;
2438
2439         bch2_btree_iter_verify(iter);
2440         bch2_btree_iter_verify_entry_exit(iter);
2441         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2442         EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2443
2444         /* extents can't span inode numbers: */
2445         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2446             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2447                 if (iter->pos.inode == KEY_INODE_MAX)
2448                         return bkey_s_c_null;
2449
2450                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2451         }
2452
2453         search_key = btree_iter_search_key(iter);
2454         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2455                                         iter->flags & BTREE_ITER_INTENT,
2456                                         btree_iter_ip_allocated(iter));
2457
2458         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2459         if (unlikely(ret)) {
2460                 k = bkey_s_c_err(ret);
2461                 goto out_no_locked;
2462         }
2463
2464         if ((iter->flags & BTREE_ITER_CACHED) ||
2465             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2466                 struct bkey_i *next_update;
2467
2468                 if ((next_update = btree_trans_peek_updates(iter)) &&
2469                     bpos_eq(next_update->k.p, iter->pos)) {
2470                         iter->k = next_update->k;
2471                         k = bkey_i_to_s_c(next_update);
2472                         goto out;
2473                 }
2474
2475                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2476                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2477                         goto out;
2478
2479                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2480                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2481                         if (!bkey_err(k))
2482                                 iter->k = *k.k;
2483                         /* We're not returning a key from iter->path: */
2484                         goto out_no_locked;
2485                 }
2486
2487                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2488                 if (unlikely(!k.k))
2489                         goto out_no_locked;
2490         } else {
2491                 struct bpos next;
2492                 struct bpos end = iter->pos;
2493
2494                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2495                         end.offset = U64_MAX;
2496
2497                 EBUG_ON(iter->path->level);
2498
2499                 if (iter->flags & BTREE_ITER_INTENT) {
2500                         struct btree_iter iter2;
2501
2502                         bch2_trans_copy_iter(&iter2, iter);
2503                         k = bch2_btree_iter_peek_upto(&iter2, end);
2504
2505                         if (k.k && !bkey_err(k)) {
2506                                 iter->k = iter2.k;
2507                                 k.k = &iter->k;
2508                         }
2509                         bch2_trans_iter_exit(trans, &iter2);
2510                 } else {
2511                         struct bpos pos = iter->pos;
2512
2513                         k = bch2_btree_iter_peek_upto(iter, end);
2514                         if (unlikely(bkey_err(k)))
2515                                 bch2_btree_iter_set_pos(iter, pos);
2516                         else
2517                                 iter->pos = pos;
2518                 }
2519
2520                 if (unlikely(bkey_err(k)))
2521                         goto out_no_locked;
2522
2523                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2524
2525                 if (bkey_lt(iter->pos, next)) {
2526                         bkey_init(&iter->k);
2527                         iter->k.p = iter->pos;
2528
2529                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2530                                 bch2_key_resize(&iter->k,
2531                                                 min_t(u64, KEY_SIZE_MAX,
2532                                                       (next.inode == iter->pos.inode
2533                                                        ? next.offset
2534                                                        : KEY_OFFSET_MAX) -
2535                                                       iter->pos.offset));
2536                                 EBUG_ON(!iter->k.size);
2537                         }
2538
2539                         k = (struct bkey_s_c) { &iter->k, NULL };
2540                 }
2541         }
2542 out:
2543         btree_path_set_should_be_locked(iter->path);
2544 out_no_locked:
2545         bch2_btree_iter_verify_entry_exit(iter);
2546         bch2_btree_iter_verify(iter);
2547         ret = bch2_btree_iter_verify_ret(iter, k);
2548         if (unlikely(ret))
2549                 return bkey_s_c_err(ret);
2550
2551         return k;
2552 }
2553
2554 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2555 {
2556         if (!bch2_btree_iter_advance(iter))
2557                 return bkey_s_c_null;
2558
2559         return bch2_btree_iter_peek_slot(iter);
2560 }
2561
2562 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2563 {
2564         if (!bch2_btree_iter_rewind(iter))
2565                 return bkey_s_c_null;
2566
2567         return bch2_btree_iter_peek_slot(iter);
2568 }
2569
2570 /* new transactional stuff: */
2571
2572 #ifdef CONFIG_BCACHEFS_DEBUG
2573 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2574 {
2575         struct btree_path *path;
2576         unsigned i;
2577
2578         BUG_ON(trans->nr_sorted != hweight64(trans->paths_allocated));
2579
2580         trans_for_each_path(trans, path) {
2581                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2582                 BUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2583         }
2584
2585         for (i = 0; i < trans->nr_sorted; i++) {
2586                 unsigned idx = trans->sorted[i];
2587
2588                 EBUG_ON(!(trans->paths_allocated & (1ULL << idx)));
2589                 BUG_ON(trans->paths[idx].sorted_idx != i);
2590         }
2591 }
2592
2593 static void btree_trans_verify_sorted(struct btree_trans *trans)
2594 {
2595         struct btree_path *path, *prev = NULL;
2596         unsigned i;
2597
2598         if (!bch2_debug_check_iterators)
2599                 return;
2600
2601         trans_for_each_path_inorder(trans, path, i) {
2602                 if (prev && btree_path_cmp(prev, path) > 0) {
2603                         __bch2_dump_trans_paths_updates(trans, true);
2604                         panic("trans paths out of order!\n");
2605                 }
2606                 prev = path;
2607         }
2608 }
2609 #else
2610 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2611 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2612 #endif
2613
2614 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2615 {
2616         int i, l = 0, r = trans->nr_sorted, inc = 1;
2617         bool swapped;
2618
2619         btree_trans_verify_sorted_refs(trans);
2620
2621         if (trans->paths_sorted)
2622                 goto out;
2623
2624         /*
2625          * Cocktail shaker sort: this is efficient because iterators will be
2626          * mostly sorted.
2627          */
2628         do {
2629                 swapped = false;
2630
2631                 for (i = inc > 0 ? l : r - 2;
2632                      i + 1 < r && i >= l;
2633                      i += inc) {
2634                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2635                                            trans->paths + trans->sorted[i + 1]) > 0) {
2636                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2637                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2638                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2639                                 swapped = true;
2640                         }
2641                 }
2642
2643                 if (inc > 0)
2644                         --r;
2645                 else
2646                         l++;
2647                 inc = -inc;
2648         } while (swapped);
2649
2650         trans->paths_sorted = true;
2651 out:
2652         btree_trans_verify_sorted(trans);
2653 }
2654
2655 static inline void btree_path_list_remove(struct btree_trans *trans,
2656                                           struct btree_path *path)
2657 {
2658         unsigned i;
2659
2660         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2661 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2662         trans->nr_sorted--;
2663         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2664                                 trans->sorted + path->sorted_idx + 1,
2665                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2666 #else
2667         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2668 #endif
2669         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2670                 trans->paths[trans->sorted[i]].sorted_idx = i;
2671
2672         path->sorted_idx = U8_MAX;
2673 }
2674
2675 static inline void btree_path_list_add(struct btree_trans *trans,
2676                                        struct btree_path *pos,
2677                                        struct btree_path *path)
2678 {
2679         unsigned i;
2680
2681         path->sorted_idx = pos ? pos->sorted_idx + 1 : trans->nr_sorted;
2682
2683 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2684         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2685                               trans->sorted + path->sorted_idx,
2686                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2687         trans->nr_sorted++;
2688         trans->sorted[path->sorted_idx] = path->idx;
2689 #else
2690         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2691 #endif
2692
2693         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2694                 trans->paths[trans->sorted[i]].sorted_idx = i;
2695
2696         btree_trans_verify_sorted_refs(trans);
2697 }
2698
2699 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2700 {
2701         if (iter->path)
2702                 bch2_path_put(trans, iter->path,
2703                               iter->flags & BTREE_ITER_INTENT);
2704         if (iter->update_path)
2705                 bch2_path_put_nokeep(trans, iter->update_path,
2706                               iter->flags & BTREE_ITER_INTENT);
2707         if (iter->key_cache_path)
2708                 bch2_path_put(trans, iter->key_cache_path,
2709                               iter->flags & BTREE_ITER_INTENT);
2710         iter->path = NULL;
2711         iter->update_path = NULL;
2712         iter->key_cache_path = NULL;
2713 }
2714
2715 static inline void bch2_trans_iter_init_inlined(struct btree_trans *trans,
2716                           struct btree_iter *iter,
2717                           unsigned btree_id, struct bpos pos,
2718                           unsigned flags)
2719 {
2720         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2721                                bch2_btree_iter_flags(trans, btree_id, flags),
2722                                _RET_IP_);
2723 }
2724
2725 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2726                           struct btree_iter *iter,
2727                           unsigned btree_id, struct bpos pos,
2728                           unsigned flags)
2729 {
2730         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2731                                bch2_btree_iter_flags(trans, btree_id, flags),
2732                                _RET_IP_);
2733 }
2734
2735 void bch2_trans_node_iter_init(struct btree_trans *trans,
2736                                struct btree_iter *iter,
2737                                enum btree_id btree_id,
2738                                struct bpos pos,
2739                                unsigned locks_want,
2740                                unsigned depth,
2741                                unsigned flags)
2742 {
2743        flags |= BTREE_ITER_NOT_EXTENTS;
2744        flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2745        flags |= BTREE_ITER_ALL_SNAPSHOTS;
2746
2747         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2748                                __bch2_btree_iter_flags(trans, btree_id, flags),
2749                                _RET_IP_);
2750
2751         iter->min_depth = depth;
2752
2753         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2754         BUG_ON(iter->path->level        != depth);
2755         BUG_ON(iter->min_depth          != depth);
2756 }
2757
2758 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2759 {
2760         *dst = *src;
2761         if (src->path)
2762                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2763         if (src->update_path)
2764                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2765         dst->key_cache_path = NULL;
2766 }
2767
2768 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2769 {
2770         unsigned new_top = trans->mem_top + size;
2771         size_t old_bytes = trans->mem_bytes;
2772         size_t new_bytes = roundup_pow_of_two(new_top);
2773         void *new_mem;
2774         void *p;
2775
2776         trans->mem_max = max(trans->mem_max, new_top);
2777
2778         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2779
2780         new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2781         if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2782                 new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2783                 new_bytes = BTREE_TRANS_MEM_MAX;
2784                 kfree(trans->mem);
2785         }
2786
2787         if (!new_mem)
2788                 return ERR_PTR(-ENOMEM);
2789
2790         trans->mem = new_mem;
2791         trans->mem_bytes = new_bytes;
2792
2793         if (old_bytes) {
2794                 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2795                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2796         }
2797
2798         p = trans->mem + trans->mem_top;
2799         trans->mem_top += size;
2800         memset(p, 0, size);
2801         return p;
2802 }
2803
2804 static noinline void bch2_trans_reset_srcu_lock(struct btree_trans *trans)
2805 {
2806         struct bch_fs *c = trans->c;
2807         struct btree_path *path;
2808
2809         trans_for_each_path(trans, path)
2810                 if (path->cached && !btree_node_locked(path, 0))
2811                         path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2812
2813         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2814         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2815         trans->srcu_lock_time   = jiffies;
2816 }
2817
2818 /**
2819  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2820  * @trans: transaction to reset
2821  *
2822  * While iterating over nodes or updating nodes a attempt to lock a btree node
2823  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2824  * occurs bch2_trans_begin() should be called and the transaction retried.
2825  */
2826 u32 bch2_trans_begin(struct btree_trans *trans)
2827 {
2828         struct btree_path *path;
2829
2830         bch2_trans_reset_updates(trans);
2831
2832         trans->restart_count++;
2833         trans->mem_top                  = 0;
2834
2835         trans_for_each_path(trans, path) {
2836                 path->should_be_locked = false;
2837
2838                 /*
2839                  * If the transaction wasn't restarted, we're presuming to be
2840                  * doing something new: dont keep iterators excpt the ones that
2841                  * are in use - except for the subvolumes btree:
2842                  */
2843                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2844                         path->preserve = false;
2845
2846                 /*
2847                  * XXX: we probably shouldn't be doing this if the transaction
2848                  * was restarted, but currently we still overflow transaction
2849                  * iterators if we do that
2850                  */
2851                 if (!path->ref && !path->preserve)
2852                         __bch2_path_free(trans, path);
2853                 else
2854                         path->preserve = false;
2855         }
2856
2857         if (!trans->restarted &&
2858             (need_resched() ||
2859              local_clock() - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2860                 bch2_trans_unlock(trans);
2861                 cond_resched();
2862                 bch2_trans_relock(trans);
2863         }
2864
2865         if (unlikely(time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2866                 bch2_trans_reset_srcu_lock(trans);
2867
2868         trans->last_restarted_ip = _RET_IP_;
2869         if (trans->restarted) {
2870                 bch2_btree_path_traverse_all(trans);
2871                 trans->notrace_relock_fail = false;
2872         }
2873
2874         trans->last_begin_time = local_clock();
2875         return trans->restart_count;
2876 }
2877
2878 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
2879 {
2880         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
2881         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2882         void *p = NULL;
2883
2884         BUG_ON(trans->used_mempool);
2885
2886 #ifdef __KERNEL__
2887         p = this_cpu_xchg(c->btree_paths_bufs->path, NULL);
2888 #endif
2889         if (!p)
2890                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
2891
2892         trans->paths            = p; p += paths_bytes;
2893         trans->updates          = p; p += updates_bytes;
2894 }
2895
2896 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2897
2898 unsigned bch2_trans_get_fn_idx(const char *fn)
2899 {
2900         unsigned i;
2901
2902         for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2903                 if (!bch2_btree_transaction_fns[i] ||
2904                     bch2_btree_transaction_fns[i] == fn) {
2905                         bch2_btree_transaction_fns[i] = fn;
2906                         return i;
2907                 }
2908
2909         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2910         return i;
2911 }
2912
2913 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, unsigned fn_idx)
2914         __acquires(&c->btree_trans_barrier)
2915 {
2916         struct btree_transaction_stats *s;
2917
2918         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
2919
2920         memset(trans, 0, sizeof(*trans));
2921         trans->c                = c;
2922         trans->fn               = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2923                 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2924         trans->last_begin_time  = local_clock();
2925         trans->fn_idx           = fn_idx;
2926         trans->locking_wait.task = current;
2927         trans->journal_replay_not_finished =
2928                 !test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
2929         closure_init_stack(&trans->ref);
2930
2931         bch2_trans_alloc_paths(trans, c);
2932
2933         s = btree_trans_stats(trans);
2934         if (s && s->max_mem) {
2935                 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2936
2937                 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2938
2939                 if (!unlikely(trans->mem)) {
2940                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2941                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2942                 } else {
2943                         trans->mem_bytes = expected_mem_bytes;
2944                 }
2945         }
2946
2947         if (s) {
2948                 trans->nr_max_paths = s->nr_max_paths;
2949                 trans->wb_updates_size = s->wb_updates_size;
2950         }
2951
2952         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2953         trans->srcu_lock_time   = jiffies;
2954
2955         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2956                 struct btree_trans *pos;
2957
2958                 mutex_lock(&c->btree_trans_lock);
2959                 list_for_each_entry(pos, &c->btree_trans_list, list) {
2960                         if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
2961                                 list_add_tail(&trans->list, &pos->list);
2962                                 goto list_add_done;
2963                         }
2964                 }
2965                 list_add_tail(&trans->list, &c->btree_trans_list);
2966 list_add_done:
2967                 mutex_unlock(&c->btree_trans_lock);
2968         }
2969 }
2970
2971 static void check_btree_paths_leaked(struct btree_trans *trans)
2972 {
2973 #ifdef CONFIG_BCACHEFS_DEBUG
2974         struct bch_fs *c = trans->c;
2975         struct btree_path *path;
2976
2977         trans_for_each_path(trans, path)
2978                 if (path->ref)
2979                         goto leaked;
2980         return;
2981 leaked:
2982         bch_err(c, "btree paths leaked from %s!", trans->fn);
2983         trans_for_each_path(trans, path)
2984                 if (path->ref)
2985                         printk(KERN_ERR "  btree %s %pS\n",
2986                                bch2_btree_ids[path->btree_id],
2987                                (void *) path->ip_allocated);
2988         /* Be noisy about this: */
2989         bch2_fatal_error(c);
2990 #endif
2991 }
2992
2993 void bch2_trans_exit(struct btree_trans *trans)
2994         __releases(&c->btree_trans_barrier)
2995 {
2996         struct btree_insert_entry *i;
2997         struct bch_fs *c = trans->c;
2998         struct btree_transaction_stats *s = btree_trans_stats(trans);
2999
3000         bch2_trans_unlock(trans);
3001
3002         closure_sync(&trans->ref);
3003
3004         if (s)
3005                 s->max_mem = max(s->max_mem, trans->mem_max);
3006
3007         trans_for_each_update(trans, i)
3008                 __btree_path_put(i->path, true);
3009         trans->nr_updates               = 0;
3010
3011         check_btree_paths_leaked(trans);
3012
3013         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
3014                 mutex_lock(&c->btree_trans_lock);
3015                 list_del(&trans->list);
3016                 mutex_unlock(&c->btree_trans_lock);
3017         }
3018
3019         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3020
3021         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3022
3023         kfree(trans->extra_journal_entries.data);
3024
3025         if (trans->fs_usage_deltas) {
3026                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3027                     REPLICAS_DELTA_LIST_MAX)
3028                         mempool_free(trans->fs_usage_deltas,
3029                                      &c->replicas_delta_pool);
3030                 else
3031                         kfree(trans->fs_usage_deltas);
3032         }
3033
3034         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3035                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3036         else
3037                 kfree(trans->mem);
3038
3039 #ifdef __KERNEL__
3040         /*
3041          * Userspace doesn't have a real percpu implementation:
3042          */
3043         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3044 #endif
3045
3046         if (trans->paths)
3047                 mempool_free(trans->paths, &c->btree_paths_pool);
3048
3049         trans->mem      = (void *) 0x1;
3050         trans->paths    = (void *) 0x1;
3051 }
3052
3053 static void __maybe_unused
3054 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3055                                       struct btree_bkey_cached_common *b)
3056 {
3057         struct six_lock_count c = six_lock_counts(&b->lock);
3058         struct task_struct *owner;
3059         pid_t pid;
3060
3061         rcu_read_lock();
3062         owner = READ_ONCE(b->lock.owner);
3063         pid = owner ? owner->pid : 0;
3064         rcu_read_unlock();
3065
3066         prt_tab(out);
3067         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3068                    b->level, bch2_btree_ids[b->btree_id]);
3069         bch2_bpos_to_text(out, btree_node_pos(b));
3070
3071         prt_tab(out);
3072         prt_printf(out, " locks %u:%u:%u held by pid %u",
3073                    c.n[0], c.n[1], c.n[2], pid);
3074 }
3075
3076 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3077 {
3078         struct btree_path *path;
3079         struct btree_bkey_cached_common *b;
3080         static char lock_types[] = { 'r', 'i', 'w' };
3081         unsigned l;
3082
3083         if (!out->nr_tabstops) {
3084                 printbuf_tabstop_push(out, 16);
3085                 printbuf_tabstop_push(out, 32);
3086         }
3087
3088         prt_printf(out, "%i %s\n", trans->locking_wait.task->pid, trans->fn);
3089
3090         trans_for_each_path(trans, path) {
3091                 if (!path->nodes_locked)
3092                         continue;
3093
3094                 prt_printf(out, "  path %u %c l=%u %s:",
3095                        path->idx,
3096                        path->cached ? 'c' : 'b',
3097                        path->level,
3098                        bch2_btree_ids[path->btree_id]);
3099                 bch2_bpos_to_text(out, path->pos);
3100                 prt_newline(out);
3101
3102                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3103                         if (btree_node_locked(path, l) &&
3104                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3105                                 prt_printf(out, "    %c l=%u ",
3106                                            lock_types[btree_node_locked_type(path, l)], l);
3107                                 bch2_btree_bkey_cached_common_to_text(out, b);
3108                                 prt_newline(out);
3109                         }
3110                 }
3111         }
3112
3113         b = READ_ONCE(trans->locking);
3114         if (b) {
3115                 prt_printf(out, "  blocked for %lluus on",
3116                            div_u64(local_clock() - trans->locking_wait.start_time,
3117                                    1000));
3118                 prt_newline(out);
3119                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3120                 bch2_btree_bkey_cached_common_to_text(out, b);
3121                 prt_newline(out);
3122         }
3123 }
3124
3125 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3126 {
3127         struct btree_transaction_stats *s;
3128
3129         for (s = c->btree_transaction_stats;
3130              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3131              s++) {
3132                 kfree(s->max_paths_text);
3133                 bch2_time_stats_exit(&s->lock_hold_times);
3134         }
3135
3136         if (c->btree_trans_barrier_initialized)
3137                 cleanup_srcu_struct(&c->btree_trans_barrier);
3138         mempool_exit(&c->btree_trans_mem_pool);
3139         mempool_exit(&c->btree_paths_pool);
3140 }
3141
3142 int bch2_fs_btree_iter_init(struct bch_fs *c)
3143 {
3144         struct btree_transaction_stats *s;
3145         unsigned nr = BTREE_ITER_MAX;
3146         int ret;
3147
3148         for (s = c->btree_transaction_stats;
3149              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3150              s++) {
3151                 bch2_time_stats_init(&s->lock_hold_times);
3152                 mutex_init(&s->lock);
3153         }
3154
3155         INIT_LIST_HEAD(&c->btree_trans_list);
3156         mutex_init(&c->btree_trans_lock);
3157
3158         ret   = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3159                         sizeof(struct btree_path) * nr +
3160                         sizeof(struct btree_insert_entry) * nr) ?:
3161                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3162                                           BTREE_TRANS_MEM_MAX) ?:
3163                 init_srcu_struct(&c->btree_trans_barrier);
3164         if (!ret)
3165                 c->btree_trans_barrier_initialized = true;
3166         return ret;
3167 }