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(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(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         trans->last_restarted_ip = 0;
 980
 981         trans_for_each_path(trans, path)
 982                 path->should_be_locked = false;
 983
 984         btree_trans_sort_paths(trans);
 985
 986         bch2_trans_unlock(trans);
 987         cond_resched();
 988
 989         if (unlikely(trans->memory_allocation_failure)) {
 990                 struct closure cl;
 991
 992                 closure_init_stack(&cl);
 993
 994                 do {
 995                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 996                         closure_sync(&cl);
 997                 } while (ret);
 998         }
 999
1000         /* Now, redo traversals in correct order: */
1001         i = 0;
1002         while (i < trans->nr_sorted) {
1003                 path = trans->paths + trans->sorted[i];
1004
1005                 /*
1006                  * Traversing a path can cause another path to be added at about
1007                  * the same position:
1008                  */
1009                 if (path->uptodate) {
1010                         __btree_path_get(path, false);
1011                         ret = bch2_btree_path_traverse_one(trans, path, 0, _THIS_IP_);
1012                         __btree_path_put(path, false);
1013
1014                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1015                             ret == -ENOMEM)
1016                                 goto retry_all;
1017                         if (ret)
1018                                 goto err;
1019                 } else {
1020                         i++;
1021                 }
1022         }
1023
1024         /*
1025          * We used to assert that all paths had been traversed here
1026          * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1027          * path->Should_be_locked is not set yet, we we might have unlocked and
1028          * then failed to relock a path - that's fine.
1029          */
1030 err:
1031         bch2_btree_cache_cannibalize_unlock(c);
1032
1033         trans->in_traverse_all = false;
1034
1035         trace_and_count(c, trans_traverse_all, trans, trace_ip);
1036         return ret;
1037 }
1038
1039 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1040                                                 unsigned l, int check_pos)
1041 {
1042         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1043                 return false;
1044         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1045                 return false;
1046         return true;
1047 }
1048
1049 static inline bool btree_path_good_node(struct btree_trans *trans,
1050                                         struct btree_path *path,
1051                                         unsigned l, int check_pos)
1052 {
1053         return is_btree_node(path, l) &&
1054                 bch2_btree_node_relock(trans, path, l) &&
1055                 btree_path_check_pos_in_node(path, l, check_pos);
1056 }
1057
1058 static void btree_path_set_level_down(struct btree_trans *trans,
1059                                       struct btree_path *path,
1060                                       unsigned new_level)
1061 {
1062         unsigned l;
1063
1064         path->level = new_level;
1065
1066         for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1067                 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1068                         btree_node_unlock(trans, path, l);
1069
1070         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1071         bch2_btree_path_verify(trans, path);
1072 }
1073
1074 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1075                                                          struct btree_path *path,
1076                                                          int check_pos)
1077 {
1078         unsigned i, l = path->level;
1079 again:
1080         while (btree_path_node(path, l) &&
1081                !btree_path_good_node(trans, path, l, check_pos))
1082                 __btree_path_set_level_up(trans, path, l++);
1083
1084         /* If we need intent locks, take them too: */
1085         for (i = l + 1;
1086              i < path->locks_want && btree_path_node(path, i);
1087              i++)
1088                 if (!bch2_btree_node_relock(trans, path, i)) {
1089                         while (l <= i)
1090                                 __btree_path_set_level_up(trans, path, l++);
1091                         goto again;
1092                 }
1093
1094         return l;
1095 }
1096
1097 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1098                                                      struct btree_path *path,
1099                                                      int check_pos)
1100 {
1101         return likely(btree_node_locked(path, path->level) &&
1102                       btree_path_check_pos_in_node(path, path->level, check_pos))
1103                 ? path->level
1104                 : __btree_path_up_until_good_node(trans, path, check_pos);
1105 }
1106
1107 /*
1108  * This is the main state machine for walking down the btree - walks down to a
1109  * specified depth
1110  *
1111  * Returns 0 on success, -EIO on error (error reading in a btree node).
1112  *
1113  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1114  * stashed in the iterator and returned from bch2_trans_exit().
1115  */
1116 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1117                                  struct btree_path *path,
1118                                  unsigned flags,
1119                                  unsigned long trace_ip)
1120 {
1121         unsigned depth_want = path->level;
1122         int ret = -((int) trans->restarted);
1123
1124         if (unlikely(ret))
1125                 goto out;
1126
1127         /*
1128          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1129          * and re-traverse the path without a transaction restart:
1130          */
1131         if (path->should_be_locked) {
1132                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1133                 goto out;
1134         }
1135
1136         if (path->cached) {
1137                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1138                 goto out;
1139         }
1140
1141         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1142                 goto out;
1143
1144         path->level = btree_path_up_until_good_node(trans, path, 0);
1145
1146         EBUG_ON(btree_path_node(path, path->level) &&
1147                 !btree_node_locked(path, path->level));
1148
1149         /*
1150          * Note: path->nodes[path->level] may be temporarily NULL here - that
1151          * would indicate to other code that we got to the end of the btree,
1152          * here it indicates that relocking the root failed - it's critical that
1153          * btree_path_lock_root() comes next and that it can't fail
1154          */
1155         while (path->level > depth_want) {
1156                 ret = btree_path_node(path, path->level)
1157                         ? btree_path_down(trans, path, flags, trace_ip)
1158                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1159                 if (unlikely(ret)) {
1160                         if (ret == 1) {
1161                                 /*
1162                                  * No nodes at this level - got to the end of
1163                                  * the btree:
1164                                  */
1165                                 ret = 0;
1166                                 goto out;
1167                         }
1168
1169                         __bch2_btree_path_unlock(trans, path);
1170                         path->level = depth_want;
1171                         path->l[path->level].b = ERR_PTR(ret);
1172                         goto out;
1173                 }
1174         }
1175
1176         path->uptodate = BTREE_ITER_UPTODATE;
1177 out:
1178         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1179                 panic("ret %s (%i) trans->restarted %s (%i)\n",
1180                       bch2_err_str(ret), ret,
1181                       bch2_err_str(trans->restarted), trans->restarted);
1182         bch2_btree_path_verify(trans, path);
1183         return ret;
1184 }
1185
1186 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1187                             struct btree_path *src)
1188 {
1189         unsigned i, offset = offsetof(struct btree_path, pos);
1190
1191         memcpy((void *) dst + offset,
1192                (void *) src + offset,
1193                sizeof(struct btree_path) - offset);
1194
1195         for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1196                 unsigned t = btree_node_locked_type(dst, i);
1197
1198                 if (t != BTREE_NODE_UNLOCKED)
1199                         six_lock_increment(&dst->l[i].b->c.lock, t);
1200         }
1201 }
1202
1203 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1204                                            bool intent)
1205 {
1206         struct btree_path *new = btree_path_alloc(trans, src);
1207
1208         btree_path_copy(trans, new, src);
1209         __btree_path_get(new, intent);
1210         return new;
1211 }
1212
1213 __flatten
1214 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *trans,
1215                          struct btree_path *path, bool intent,
1216                          unsigned long ip)
1217 {
1218         __btree_path_put(path, intent);
1219         path = btree_path_clone(trans, path, intent);
1220         path->preserve = false;
1221         return path;
1222 }
1223
1224 struct btree_path * __must_check
1225 __bch2_btree_path_set_pos(struct btree_trans *trans,
1226                    struct btree_path *path, struct bpos new_pos,
1227                    bool intent, unsigned long ip, int cmp)
1228 {
1229         unsigned level = path->level;
1230
1231         bch2_trans_verify_not_in_restart(trans);
1232         EBUG_ON(!path->ref);
1233
1234         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1235
1236         path->pos               = new_pos;
1237         trans->paths_sorted     = false;
1238
1239         if (unlikely(path->cached)) {
1240                 btree_node_unlock(trans, path, 0);
1241                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1242                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1243                 goto out;
1244         }
1245
1246         level = btree_path_up_until_good_node(trans, path, cmp);
1247
1248         if (btree_path_node(path, level)) {
1249                 struct btree_path_level *l = &path->l[level];
1250
1251                 BUG_ON(!btree_node_locked(path, level));
1252                 /*
1253                  * We might have to skip over many keys, or just a few: try
1254                  * advancing the node iterator, and if we have to skip over too
1255                  * many keys just reinit it (or if we're rewinding, since that
1256                  * is expensive).
1257                  */
1258                 if (cmp < 0 ||
1259                     !btree_path_advance_to_pos(path, l, 8))
1260                         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1261
1262                 /*
1263                  * Iterators to interior nodes should always be pointed at the first non
1264                  * whiteout:
1265                  */
1266                 if (unlikely(level))
1267                         bch2_btree_node_iter_peek(&l->iter, l->b);
1268         }
1269
1270         if (unlikely(level != path->level)) {
1271                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1272                 __bch2_btree_path_unlock(trans, path);
1273         }
1274 out:
1275         bch2_btree_path_verify(trans, path);
1276         return path;
1277 }
1278
1279 /* Btree path: main interface: */
1280
1281 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1282 {
1283         struct btree_path *sib;
1284
1285         sib = prev_btree_path(trans, path);
1286         if (sib && !btree_path_cmp(sib, path))
1287                 return sib;
1288
1289         sib = next_btree_path(trans, path);
1290         if (sib && !btree_path_cmp(sib, path))
1291                 return sib;
1292
1293         return NULL;
1294 }
1295
1296 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1297 {
1298         struct btree_path *sib;
1299
1300         sib = prev_btree_path(trans, path);
1301         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1302                 return sib;
1303
1304         sib = next_btree_path(trans, path);
1305         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1306                 return sib;
1307
1308         return NULL;
1309 }
1310
1311 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1312 {
1313         __bch2_btree_path_unlock(trans, path);
1314         btree_path_list_remove(trans, path);
1315         trans->paths_allocated &= ~(1ULL << path->idx);
1316 }
1317
1318 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1319 {
1320         struct btree_path *dup;
1321
1322         EBUG_ON(trans->paths + path->idx != path);
1323         EBUG_ON(!path->ref);
1324
1325         if (!__btree_path_put(path, intent))
1326                 return;
1327
1328         dup = path->preserve
1329                 ? have_path_at_pos(trans, path)
1330                 : have_node_at_pos(trans, path);
1331
1332         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1333                 return;
1334
1335         if (path->should_be_locked &&
1336             !trans->restarted &&
1337             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1338                 return;
1339
1340         if (dup) {
1341                 dup->preserve           |= path->preserve;
1342                 dup->should_be_locked   |= path->should_be_locked;
1343         }
1344
1345         __bch2_path_free(trans, path);
1346 }
1347
1348 static void bch2_path_put_nokeep(struct btree_trans *trans, struct btree_path *path,
1349                                  bool intent)
1350 {
1351         EBUG_ON(trans->paths + path->idx != path);
1352         EBUG_ON(!path->ref);
1353
1354         if (!__btree_path_put(path, intent))
1355                 return;
1356
1357         __bch2_path_free(trans, path);
1358 }
1359
1360 void bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1361 {
1362         panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1363               trans->restart_count, restart_count,
1364               (void *) trans->last_begin_ip);
1365 }
1366
1367 void bch2_trans_in_restart_error(struct btree_trans *trans)
1368 {
1369         panic("in transaction restart: %s, last restarted by %pS\n",
1370               bch2_err_str(trans->restarted),
1371               (void *) trans->last_restarted_ip);
1372 }
1373
1374 noinline __cold
1375 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1376 {
1377         struct btree_insert_entry *i;
1378         struct btree_write_buffered_key *wb;
1379
1380         prt_printf(buf, "transaction updates for %s journal seq %llu",
1381                trans->fn, trans->journal_res.seq);
1382         prt_newline(buf);
1383         printbuf_indent_add(buf, 2);
1384
1385         trans_for_each_update(trans, i) {
1386                 struct bkey_s_c old = { &i->old_k, i->old_v };
1387
1388                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1389                        bch2_btree_ids[i->btree_id],
1390                        i->cached,
1391                        (void *) i->ip_allocated);
1392                 prt_newline(buf);
1393
1394                 prt_printf(buf, "  old ");
1395                 bch2_bkey_val_to_text(buf, trans->c, old);
1396                 prt_newline(buf);
1397
1398                 prt_printf(buf, "  new ");
1399                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1400                 prt_newline(buf);
1401         }
1402
1403         trans_for_each_wb_update(trans, wb) {
1404                 prt_printf(buf, "update: btree=%s wb=1 %pS",
1405                        bch2_btree_ids[wb->btree],
1406                        (void *) i->ip_allocated);
1407                 prt_newline(buf);
1408
1409                 prt_printf(buf, "  new ");
1410                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(&wb->k));
1411                 prt_newline(buf);
1412         }
1413
1414         printbuf_indent_sub(buf, 2);
1415 }
1416
1417 noinline __cold
1418 void bch2_dump_trans_updates(struct btree_trans *trans)
1419 {
1420         struct printbuf buf = PRINTBUF;
1421
1422         bch2_trans_updates_to_text(&buf, trans);
1423         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1424         printbuf_exit(&buf);
1425 }
1426
1427 noinline __cold
1428 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1429 {
1430         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1431                    path->idx, path->ref, path->intent_ref,
1432                    path->preserve ? 'P' : ' ',
1433                    path->should_be_locked ? 'S' : ' ',
1434                    bch2_btree_ids[path->btree_id],
1435                    path->level);
1436         bch2_bpos_to_text(out, path->pos);
1437
1438         prt_printf(out, " locks %u", path->nodes_locked);
1439 #ifdef TRACK_PATH_ALLOCATED
1440         prt_printf(out, " %pS", (void *) path->ip_allocated);
1441 #endif
1442         prt_newline(out);
1443 }
1444
1445 noinline __cold
1446 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1447                                 bool nosort)
1448 {
1449         struct btree_path *path;
1450         unsigned idx;
1451
1452         if (!nosort)
1453                 btree_trans_sort_paths(trans);
1454
1455         trans_for_each_path_inorder(trans, path, idx)
1456                 bch2_btree_path_to_text(out, path);
1457 }
1458
1459 noinline __cold
1460 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1461 {
1462         __bch2_trans_paths_to_text(out, trans, false);
1463 }
1464
1465 noinline __cold
1466 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1467 {
1468         struct printbuf buf = PRINTBUF;
1469
1470         __bch2_trans_paths_to_text(&buf, trans, nosort);
1471         bch2_trans_updates_to_text(&buf, trans);
1472
1473         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1474         printbuf_exit(&buf);
1475 }
1476
1477 noinline __cold
1478 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1479 {
1480         __bch2_dump_trans_paths_updates(trans, false);
1481 }
1482
1483 noinline __cold
1484 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1485 {
1486         struct btree_transaction_stats *s = btree_trans_stats(trans);
1487         struct printbuf buf = PRINTBUF;
1488
1489         if (!s)
1490                 return;
1491
1492         bch2_trans_paths_to_text(&buf, trans);
1493
1494         if (!buf.allocation_failure) {
1495                 mutex_lock(&s->lock);
1496                 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1497                         s->nr_max_paths = trans->nr_max_paths =
1498                                 hweight64(trans->paths_allocated);
1499                         swap(s->max_paths_text, buf.buf);
1500                 }
1501                 mutex_unlock(&s->lock);
1502         }
1503
1504         printbuf_exit(&buf);
1505
1506         trans->nr_max_paths = hweight64(trans->paths_allocated);
1507 }
1508
1509 static noinline void btree_path_overflow(struct btree_trans *trans)
1510 {
1511         bch2_dump_trans_paths_updates(trans);
1512         panic("trans path oveflow\n");
1513 }
1514
1515 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1516                                                   struct btree_path *pos)
1517 {
1518         struct btree_path *path;
1519         unsigned idx;
1520
1521         if (unlikely(trans->paths_allocated ==
1522                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
1523                 btree_path_overflow(trans);
1524
1525         idx = __ffs64(~trans->paths_allocated);
1526
1527         /*
1528          * Do this before marking the new path as allocated, since it won't be
1529          * initialized yet:
1530          */
1531         if (unlikely(idx > trans->nr_max_paths))
1532                 bch2_trans_update_max_paths(trans);
1533
1534         trans->paths_allocated |= 1ULL << idx;
1535
1536         path = &trans->paths[idx];
1537         path->idx               = idx;
1538         path->ref               = 0;
1539         path->intent_ref        = 0;
1540         path->nodes_locked      = 0;
1541
1542         btree_path_list_add(trans, pos, path);
1543         trans->paths_sorted = false;
1544         return path;
1545 }
1546
1547 struct btree_path *bch2_path_get(struct btree_trans *trans,
1548                                  enum btree_id btree_id, struct bpos pos,
1549                                  unsigned locks_want, unsigned level,
1550                                  unsigned flags, unsigned long ip)
1551 {
1552         struct btree_path *path, *path_pos = NULL;
1553         bool cached = flags & BTREE_ITER_CACHED;
1554         bool intent = flags & BTREE_ITER_INTENT;
1555         int i;
1556
1557         bch2_trans_verify_not_in_restart(trans);
1558         bch2_trans_verify_locks(trans);
1559
1560         btree_trans_sort_paths(trans);
1561
1562         trans_for_each_path_inorder(trans, path, i) {
1563                 if (__btree_path_cmp(path,
1564                                      btree_id,
1565                                      cached,
1566                                      pos,
1567                                      level) > 0)
1568                         break;
1569
1570                 path_pos = path;
1571         }
1572
1573         if (path_pos &&
1574             path_pos->cached    == cached &&
1575             path_pos->btree_id  == btree_id &&
1576             path_pos->level     == level) {
1577                 __btree_path_get(path_pos, intent);
1578                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1579         } else {
1580                 path = btree_path_alloc(trans, path_pos);
1581                 path_pos = NULL;
1582
1583                 __btree_path_get(path, intent);
1584                 path->pos                       = pos;
1585                 path->btree_id                  = btree_id;
1586                 path->cached                    = cached;
1587                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1588                 path->should_be_locked          = false;
1589                 path->level                     = level;
1590                 path->locks_want                = locks_want;
1591                 path->nodes_locked              = 0;
1592                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1593                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1594 #ifdef TRACK_PATH_ALLOCATED
1595                 path->ip_allocated              = ip;
1596 #endif
1597                 trans->paths_sorted             = false;
1598         }
1599
1600         if (!(flags & BTREE_ITER_NOPRESERVE))
1601                 path->preserve = true;
1602
1603         if (path->intent_ref)
1604                 locks_want = max(locks_want, level + 1);
1605
1606         /*
1607          * If the path has locks_want greater than requested, we don't downgrade
1608          * it here - on transaction restart because btree node split needs to
1609          * upgrade locks, we might be putting/getting the iterator again.
1610          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1611          * a successful transaction commit.
1612          */
1613
1614         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1615         if (locks_want > path->locks_want)
1616                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want);
1617
1618         return path;
1619 }
1620
1621 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1622 {
1623
1624         struct btree_path_level *l = path_l(path);
1625         struct bkey_packed *_k;
1626         struct bkey_s_c k;
1627
1628         if (unlikely(!l->b))
1629                 return bkey_s_c_null;
1630
1631         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1632         EBUG_ON(!btree_node_locked(path, path->level));
1633
1634         if (!path->cached) {
1635                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1636                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1637
1638                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1639
1640                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1641                         goto hole;
1642         } else {
1643                 struct bkey_cached *ck = (void *) path->l[0].b;
1644
1645                 EBUG_ON(ck &&
1646                         (path->btree_id != ck->key.btree_id ||
1647                          !bkey_eq(path->pos, ck->key.pos)));
1648                 if (!ck || !ck->valid)
1649                         return bkey_s_c_null;
1650
1651                 *u = ck->k->k;
1652                 k = bkey_i_to_s_c(ck->k);
1653         }
1654
1655         return k;
1656 hole:
1657         bkey_init(u);
1658         u->p = path->pos;
1659         return (struct bkey_s_c) { u, NULL };
1660 }
1661
1662 /* Btree iterators: */
1663
1664 int __must_check
1665 __bch2_btree_iter_traverse(struct btree_iter *iter)
1666 {
1667         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1668 }
1669
1670 int __must_check
1671 bch2_btree_iter_traverse(struct btree_iter *iter)
1672 {
1673         int ret;
1674
1675         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1676                                         btree_iter_search_key(iter),
1677                                         iter->flags & BTREE_ITER_INTENT,
1678                                         btree_iter_ip_allocated(iter));
1679
1680         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1681         if (ret)
1682                 return ret;
1683
1684         btree_path_set_should_be_locked(iter->path);
1685         return 0;
1686 }
1687
1688 /* Iterate across nodes (leaf and interior nodes) */
1689
1690 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1691 {
1692         struct btree_trans *trans = iter->trans;
1693         struct btree *b = NULL;
1694         int ret;
1695
1696         EBUG_ON(iter->path->cached);
1697         bch2_btree_iter_verify(iter);
1698
1699         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1700         if (ret)
1701                 goto err;
1702
1703         b = btree_path_node(iter->path, iter->path->level);
1704         if (!b)
1705                 goto out;
1706
1707         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1708
1709         bkey_init(&iter->k);
1710         iter->k.p = iter->pos = b->key.k.p;
1711
1712         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1713                                         iter->flags & BTREE_ITER_INTENT,
1714                                         btree_iter_ip_allocated(iter));
1715         btree_path_set_should_be_locked(iter->path);
1716 out:
1717         bch2_btree_iter_verify_entry_exit(iter);
1718         bch2_btree_iter_verify(iter);
1719
1720         return b;
1721 err:
1722         b = ERR_PTR(ret);
1723         goto out;
1724 }
1725
1726 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1727 {
1728         struct btree_trans *trans = iter->trans;
1729         struct btree_path *path = iter->path;
1730         struct btree *b = NULL;
1731         int ret;
1732
1733         bch2_trans_verify_not_in_restart(trans);
1734         EBUG_ON(iter->path->cached);
1735         bch2_btree_iter_verify(iter);
1736
1737         /* already at end? */
1738         if (!btree_path_node(path, path->level))
1739                 return NULL;
1740
1741         /* got to end? */
1742         if (!btree_path_node(path, path->level + 1)) {
1743                 btree_path_set_level_up(trans, path);
1744                 return NULL;
1745         }
1746
1747         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1748                 __bch2_btree_path_unlock(trans, path);
1749                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1750                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1751                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1752                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1753                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1754                 goto err;
1755         }
1756
1757         b = btree_path_node(path, path->level + 1);
1758
1759         if (bpos_eq(iter->pos, b->key.k.p)) {
1760                 __btree_path_set_level_up(trans, path, path->level++);
1761         } else {
1762                 /*
1763                  * Haven't gotten to the end of the parent node: go back down to
1764                  * the next child node
1765                  */
1766                 path = iter->path =
1767                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1768                                            iter->flags & BTREE_ITER_INTENT,
1769                                            btree_iter_ip_allocated(iter));
1770
1771                 btree_path_set_level_down(trans, path, iter->min_depth);
1772
1773                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1774                 if (ret)
1775                         goto err;
1776
1777                 b = path->l[path->level].b;
1778         }
1779
1780         bkey_init(&iter->k);
1781         iter->k.p = iter->pos = b->key.k.p;
1782
1783         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1784                                         iter->flags & BTREE_ITER_INTENT,
1785                                         btree_iter_ip_allocated(iter));
1786         btree_path_set_should_be_locked(iter->path);
1787         BUG_ON(iter->path->uptodate);
1788 out:
1789         bch2_btree_iter_verify_entry_exit(iter);
1790         bch2_btree_iter_verify(iter);
1791
1792         return b;
1793 err:
1794         b = ERR_PTR(ret);
1795         goto out;
1796 }
1797
1798 /* Iterate across keys (in leaf nodes only) */
1799
1800 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1801 {
1802         if (likely(!(iter->flags & BTREE_ITER_ALL_LEVELS))) {
1803                 struct bpos pos = iter->k.p;
1804                 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1805                              ? bpos_eq(pos, SPOS_MAX)
1806                              : bkey_eq(pos, SPOS_MAX));
1807
1808                 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1809                         pos = bkey_successor(iter, pos);
1810                 bch2_btree_iter_set_pos(iter, pos);
1811                 return ret;
1812         } else {
1813                 if (!btree_path_node(iter->path, iter->path->level))
1814                         return true;
1815
1816                 iter->advanced = true;
1817                 return false;
1818         }
1819 }
1820
1821 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1822 {
1823         struct bpos pos = bkey_start_pos(&iter->k);
1824         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1825                      ? bpos_eq(pos, POS_MIN)
1826                      : bkey_eq(pos, POS_MIN));
1827
1828         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1829                 pos = bkey_predecessor(iter, pos);
1830         bch2_btree_iter_set_pos(iter, pos);
1831         return ret;
1832 }
1833
1834 static noinline
1835 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1836 {
1837         struct btree_insert_entry *i;
1838         struct bkey_i *ret = NULL;
1839
1840         trans_for_each_update(iter->trans, i) {
1841                 if (i->btree_id < iter->btree_id)
1842                         continue;
1843                 if (i->btree_id > iter->btree_id)
1844                         break;
1845                 if (bpos_lt(i->k->k.p, iter->path->pos))
1846                         continue;
1847                 if (i->key_cache_already_flushed)
1848                         continue;
1849                 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1850                         ret = i->k;
1851         }
1852
1853         return ret;
1854 }
1855
1856 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1857 {
1858         return iter->flags & BTREE_ITER_WITH_UPDATES
1859                 ? __bch2_btree_trans_peek_updates(iter)
1860                 : NULL;
1861 }
1862
1863 struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1864                                        struct btree_iter *iter,
1865                                        struct bpos end_pos)
1866 {
1867         struct bkey_i *k;
1868
1869         if (bpos_lt(iter->path->pos, iter->journal_pos))
1870                 iter->journal_idx = 0;
1871
1872         k = bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1873                                         iter->path->level,
1874                                         iter->path->pos,
1875                                         end_pos,
1876                                         &iter->journal_idx);
1877
1878         iter->journal_pos = k ? k->k.p : end_pos;
1879         return k;
1880 }
1881
1882 static noinline
1883 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1884                                               struct btree_iter *iter)
1885 {
1886         struct bkey_i *k = bch2_btree_journal_peek(trans, iter, iter->path->pos);
1887
1888         if (k) {
1889                 iter->k = k->k;
1890                 return bkey_i_to_s_c(k);
1891         } else {
1892                 return bkey_s_c_null;
1893         }
1894 }
1895
1896 static noinline
1897 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1898                                          struct btree_iter *iter,
1899                                          struct bkey_s_c k)
1900 {
1901         struct bkey_i *next_journal =
1902                 bch2_btree_journal_peek(trans, iter,
1903                                 k.k ? k.k->p : path_l(iter->path)->b->key.k.p);
1904
1905         if (next_journal) {
1906                 iter->k = next_journal->k;
1907                 k = bkey_i_to_s_c(next_journal);
1908         }
1909
1910         return k;
1911 }
1912
1913 /*
1914  * Checks btree key cache for key at iter->pos and returns it if present, or
1915  * bkey_s_c_null:
1916  */
1917 static noinline
1918 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1919 {
1920         struct btree_trans *trans = iter->trans;
1921         struct bch_fs *c = trans->c;
1922         struct bkey u;
1923         struct bkey_s_c k;
1924         int ret;
1925
1926         if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1927             bpos_eq(iter->pos, pos))
1928                 return bkey_s_c_null;
1929
1930         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1931                 return bkey_s_c_null;
1932
1933         if (!iter->key_cache_path)
1934                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1935                                                      iter->flags & BTREE_ITER_INTENT, 0,
1936                                                      iter->flags|BTREE_ITER_CACHED|
1937                                                      BTREE_ITER_CACHED_NOFILL,
1938                                                      _THIS_IP_);
1939
1940         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1941                                         iter->flags & BTREE_ITER_INTENT,
1942                                         btree_iter_ip_allocated(iter));
1943
1944         ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
1945                                          iter->flags|BTREE_ITER_CACHED) ?:
1946                 bch2_btree_path_relock(trans, iter->path, _THIS_IP_);
1947         if (unlikely(ret))
1948                 return bkey_s_c_err(ret);
1949
1950         btree_path_set_should_be_locked(iter->key_cache_path);
1951
1952         k = bch2_btree_path_peek_slot(iter->key_cache_path, &u);
1953         if (k.k && !bkey_err(k)) {
1954                 iter->k = u;
1955                 k.k = &iter->k;
1956         }
1957         return k;
1958 }
1959
1960 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1961 {
1962         struct btree_trans *trans = iter->trans;
1963         struct bkey_i *next_update;
1964         struct bkey_s_c k, k2;
1965         int ret;
1966
1967         EBUG_ON(iter->path->cached);
1968         bch2_btree_iter_verify(iter);
1969
1970         while (1) {
1971                 struct btree_path_level *l;
1972
1973                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
1974                                         iter->flags & BTREE_ITER_INTENT,
1975                                         btree_iter_ip_allocated(iter));
1976
1977                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1978                 if (unlikely(ret)) {
1979                         /* ensure that iter->k is consistent with iter->pos: */
1980                         bch2_btree_iter_set_pos(iter, iter->pos);
1981                         k = bkey_s_c_err(ret);
1982                         goto out;
1983                 }
1984
1985                 l = path_l(iter->path);
1986
1987                 if (unlikely(!l->b)) {
1988                         /* No btree nodes at requested level: */
1989                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
1990                         k = bkey_s_c_null;
1991                         goto out;
1992                 }
1993
1994                 btree_path_set_should_be_locked(iter->path);
1995
1996                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
1997
1998                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
1999                     k.k &&
2000                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2001                         k = k2;
2002                         ret = bkey_err(k);
2003                         if (ret) {
2004                                 bch2_btree_iter_set_pos(iter, iter->pos);
2005                                 goto out;
2006                         }
2007                 }
2008
2009                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2010                         k = btree_trans_peek_journal(trans, iter, k);
2011
2012                 next_update = btree_trans_peek_updates(iter);
2013
2014                 if (next_update &&
2015                     bpos_le(next_update->k.p,
2016                             k.k ? k.k->p : l->b->key.k.p)) {
2017                         iter->k = next_update->k;
2018                         k = bkey_i_to_s_c(next_update);
2019                 }
2020
2021                 if (k.k && bkey_deleted(k.k)) {
2022                         /*
2023                          * If we've got a whiteout, and it's after the search
2024                          * key, advance the search key to the whiteout instead
2025                          * of just after the whiteout - it might be a btree
2026                          * whiteout, with a real key at the same position, since
2027                          * in the btree deleted keys sort before non deleted.
2028                          */
2029                         search_key = !bpos_eq(search_key, k.k->p)
2030                                 ? k.k->p
2031                                 : bpos_successor(k.k->p);
2032                         continue;
2033                 }
2034
2035                 if (likely(k.k)) {
2036                         break;
2037                 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2038                         /* Advance to next leaf node: */
2039                         search_key = bpos_successor(l->b->key.k.p);
2040                 } else {
2041                         /* End of btree: */
2042                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2043                         k = bkey_s_c_null;
2044                         goto out;
2045                 }
2046         }
2047 out:
2048         bch2_btree_iter_verify(iter);
2049
2050         return k;
2051 }
2052
2053 /**
2054  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2055  * current position
2056  */
2057 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2058 {
2059         struct btree_trans *trans = iter->trans;
2060         struct bpos search_key = btree_iter_search_key(iter);
2061         struct bkey_s_c k;
2062         struct bpos iter_pos;
2063         int ret;
2064
2065         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2066         EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2067
2068         if (iter->update_path) {
2069                 bch2_path_put_nokeep(trans, iter->update_path,
2070                                      iter->flags & BTREE_ITER_INTENT);
2071                 iter->update_path = NULL;
2072         }
2073
2074         bch2_btree_iter_verify_entry_exit(iter);
2075
2076         while (1) {
2077                 k = __bch2_btree_iter_peek(iter, search_key);
2078                 if (unlikely(!k.k))
2079                         goto end;
2080                 if (unlikely(bkey_err(k)))
2081                         goto out_no_locked;
2082
2083                 /*
2084                  * iter->pos should be mononotically increasing, and always be
2085                  * equal to the key we just returned - except extents can
2086                  * straddle iter->pos:
2087                  */
2088                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2089                         iter_pos = k.k->p;
2090                 else
2091                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2092
2093                 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2094                              ? bkey_gt(iter_pos, end)
2095                              : bkey_ge(iter_pos, end)))
2096                         goto end;
2097
2098                 if (iter->update_path &&
2099                     !bkey_eq(iter->update_path->pos, k.k->p)) {
2100                         bch2_path_put_nokeep(trans, iter->update_path,
2101                                              iter->flags & BTREE_ITER_INTENT);
2102                         iter->update_path = NULL;
2103                 }
2104
2105                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2106                     (iter->flags & BTREE_ITER_INTENT) &&
2107                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2108                     !iter->update_path) {
2109                         struct bpos pos = k.k->p;
2110
2111                         if (pos.snapshot < iter->snapshot) {
2112                                 search_key = bpos_successor(k.k->p);
2113                                 continue;
2114                         }
2115
2116                         pos.snapshot = iter->snapshot;
2117
2118                         /*
2119                          * advance, same as on exit for iter->path, but only up
2120                          * to snapshot
2121                          */
2122                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2123                         iter->update_path = iter->path;
2124
2125                         iter->update_path = bch2_btree_path_set_pos(trans,
2126                                                 iter->update_path, pos,
2127                                                 iter->flags & BTREE_ITER_INTENT,
2128                                                 _THIS_IP_);
2129                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2130                         if (unlikely(ret)) {
2131                                 k = bkey_s_c_err(ret);
2132                                 goto out_no_locked;
2133                         }
2134                 }
2135
2136                 /*
2137                  * We can never have a key in a leaf node at POS_MAX, so
2138                  * we don't have to check these successor() calls:
2139                  */
2140                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2141                     !bch2_snapshot_is_ancestor(trans->c,
2142                                                iter->snapshot,
2143                                                k.k->p.snapshot)) {
2144                         search_key = bpos_successor(k.k->p);
2145                         continue;
2146                 }
2147
2148                 if (bkey_whiteout(k.k) &&
2149                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2150                         search_key = bkey_successor(iter, k.k->p);
2151                         continue;
2152                 }
2153
2154                 break;
2155         }
2156
2157         iter->pos = iter_pos;
2158
2159         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2160                                 iter->flags & BTREE_ITER_INTENT,
2161                                 btree_iter_ip_allocated(iter));
2162
2163         btree_path_set_should_be_locked(iter->path);
2164 out_no_locked:
2165         if (iter->update_path) {
2166                 ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_);
2167                 if (unlikely(ret))
2168                         k = bkey_s_c_err(ret);
2169                 else
2170                         btree_path_set_should_be_locked(iter->update_path);
2171         }
2172
2173         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2174                 iter->pos.snapshot = iter->snapshot;
2175
2176         ret = bch2_btree_iter_verify_ret(iter, k);
2177         if (unlikely(ret)) {
2178                 bch2_btree_iter_set_pos(iter, iter->pos);
2179                 k = bkey_s_c_err(ret);
2180         }
2181
2182         bch2_btree_iter_verify_entry_exit(iter);
2183
2184         return k;
2185 end:
2186         bch2_btree_iter_set_pos(iter, end);
2187         k = bkey_s_c_null;
2188         goto out_no_locked;
2189 }
2190
2191 /**
2192  * bch2_btree_iter_peek_all_levels: returns the first key greater than or equal
2193  * to iterator's current position, returning keys from every level of the btree.
2194  * For keys at different levels of the btree that compare equal, the key from
2195  * the lower level (leaf) is returned first.
2196  */
2197 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *iter)
2198 {
2199         struct btree_trans *trans = iter->trans;
2200         struct bkey_s_c k;
2201         int ret;
2202
2203         EBUG_ON(iter->path->cached);
2204         bch2_btree_iter_verify(iter);
2205         BUG_ON(iter->path->level < iter->min_depth);
2206         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
2207         EBUG_ON(!(iter->flags & BTREE_ITER_ALL_LEVELS));
2208
2209         while (1) {
2210                 iter->path = bch2_btree_path_set_pos(trans, iter->path, iter->pos,
2211                                         iter->flags & BTREE_ITER_INTENT,
2212                                         btree_iter_ip_allocated(iter));
2213
2214                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2215                 if (unlikely(ret)) {
2216                         /* ensure that iter->k is consistent with iter->pos: */
2217                         bch2_btree_iter_set_pos(iter, iter->pos);
2218                         k = bkey_s_c_err(ret);
2219                         goto out_no_locked;
2220                 }
2221
2222                 /* Already at end? */
2223                 if (!btree_path_node(iter->path, iter->path->level)) {
2224                         k = bkey_s_c_null;
2225                         goto out_no_locked;
2226                 }
2227
2228                 k = btree_path_level_peek_all(trans->c,
2229                                 &iter->path->l[iter->path->level], &iter->k);
2230
2231                 /* Check if we should go up to the parent node: */
2232                 if (!k.k ||
2233                     (iter->advanced &&
2234                      bpos_eq(path_l(iter->path)->b->key.k.p, iter->pos))) {
2235                         iter->pos = path_l(iter->path)->b->key.k.p;
2236                         btree_path_set_level_up(trans, iter->path);
2237                         iter->advanced = false;
2238                         continue;
2239                 }
2240
2241                 /*
2242                  * Check if we should go back down to a leaf:
2243                  * If we're not in a leaf node, we only return the current key
2244                  * if it exactly matches iter->pos - otherwise we first have to
2245                  * go back to the leaf:
2246                  */
2247                 if (iter->path->level != iter->min_depth &&
2248                     (iter->advanced ||
2249                      !k.k ||
2250                      !bpos_eq(iter->pos, k.k->p))) {
2251                         btree_path_set_level_down(trans, iter->path, iter->min_depth);
2252                         iter->pos = bpos_successor(iter->pos);
2253                         iter->advanced = false;
2254                         continue;
2255                 }
2256
2257                 /* Check if we should go to the next key: */
2258                 if (iter->path->level == iter->min_depth &&
2259                     iter->advanced &&
2260                     k.k &&
2261                     bpos_eq(iter->pos, k.k->p)) {
2262                         iter->pos = bpos_successor(iter->pos);
2263                         iter->advanced = false;
2264                         continue;
2265                 }
2266
2267                 if (iter->advanced &&
2268                     iter->path->level == iter->min_depth &&
2269                     !bpos_eq(k.k->p, iter->pos))
2270                         iter->advanced = false;
2271
2272                 BUG_ON(iter->advanced);
2273                 BUG_ON(!k.k);
2274                 break;
2275         }
2276
2277         iter->pos = k.k->p;
2278         btree_path_set_should_be_locked(iter->path);
2279 out_no_locked:
2280         bch2_btree_iter_verify(iter);
2281
2282         return k;
2283 }
2284
2285 /**
2286  * bch2_btree_iter_next: returns first key greater than iterator's current
2287  * position
2288  */
2289 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2290 {
2291         if (!bch2_btree_iter_advance(iter))
2292                 return bkey_s_c_null;
2293
2294         return bch2_btree_iter_peek(iter);
2295 }
2296
2297 /**
2298  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2299  * iterator's current position
2300  */
2301 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2302 {
2303         struct btree_trans *trans = iter->trans;
2304         struct bpos search_key = iter->pos;
2305         struct btree_path *saved_path = NULL;
2306         struct bkey_s_c k;
2307         struct bkey saved_k;
2308         const struct bch_val *saved_v;
2309         int ret;
2310
2311         EBUG_ON(iter->path->cached || iter->path->level);
2312         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2313
2314         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2315                 return bkey_s_c_err(-EIO);
2316
2317         bch2_btree_iter_verify(iter);
2318         bch2_btree_iter_verify_entry_exit(iter);
2319
2320         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2321                 search_key.snapshot = U32_MAX;
2322
2323         while (1) {
2324                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2325                                                 iter->flags & BTREE_ITER_INTENT,
2326                                                 btree_iter_ip_allocated(iter));
2327
2328                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2329                 if (unlikely(ret)) {
2330                         /* ensure that iter->k is consistent with iter->pos: */
2331                         bch2_btree_iter_set_pos(iter, iter->pos);
2332                         k = bkey_s_c_err(ret);
2333                         goto out_no_locked;
2334                 }
2335
2336                 k = btree_path_level_peek(trans, iter->path,
2337                                           &iter->path->l[0], &iter->k);
2338                 if (!k.k ||
2339                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2340                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2341                      : bpos_gt(k.k->p, search_key)))
2342                         k = btree_path_level_prev(trans, iter->path,
2343                                                   &iter->path->l[0], &iter->k);
2344
2345                 if (likely(k.k)) {
2346                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2347                                 if (k.k->p.snapshot == iter->snapshot)
2348                                         goto got_key;
2349
2350                                 /*
2351                                  * If we have a saved candidate, and we're no
2352                                  * longer at the same _key_ (not pos), return
2353                                  * that candidate
2354                                  */
2355                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2356                                         bch2_path_put_nokeep(trans, iter->path,
2357                                                       iter->flags & BTREE_ITER_INTENT);
2358                                         iter->path = saved_path;
2359                                         saved_path = NULL;
2360                                         iter->k = saved_k;
2361                                         k.v     = saved_v;
2362                                         goto got_key;
2363                                 }
2364
2365                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2366                                                               iter->snapshot,
2367                                                               k.k->p.snapshot)) {
2368                                         if (saved_path)
2369                                                 bch2_path_put_nokeep(trans, saved_path,
2370                                                       iter->flags & BTREE_ITER_INTENT);
2371                                         saved_path = btree_path_clone(trans, iter->path,
2372                                                                 iter->flags & BTREE_ITER_INTENT);
2373                                         saved_k = *k.k;
2374                                         saved_v = k.v;
2375                                 }
2376
2377                                 search_key = bpos_predecessor(k.k->p);
2378                                 continue;
2379                         }
2380 got_key:
2381                         if (bkey_whiteout(k.k) &&
2382                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2383                                 search_key = bkey_predecessor(iter, k.k->p);
2384                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2385                                         search_key.snapshot = U32_MAX;
2386                                 continue;
2387                         }
2388
2389                         break;
2390                 } else if (likely(!bpos_eq(iter->path->l[0].b->data->min_key, POS_MIN))) {
2391                         /* Advance to previous leaf node: */
2392                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2393                 } else {
2394                         /* Start of btree: */
2395                         bch2_btree_iter_set_pos(iter, POS_MIN);
2396                         k = bkey_s_c_null;
2397                         goto out_no_locked;
2398                 }
2399         }
2400
2401         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2402
2403         /* Extents can straddle iter->pos: */
2404         if (bkey_lt(k.k->p, iter->pos))
2405                 iter->pos = k.k->p;
2406
2407         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2408                 iter->pos.snapshot = iter->snapshot;
2409
2410         btree_path_set_should_be_locked(iter->path);
2411 out_no_locked:
2412         if (saved_path)
2413                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2414
2415         bch2_btree_iter_verify_entry_exit(iter);
2416         bch2_btree_iter_verify(iter);
2417
2418         return k;
2419 }
2420
2421 /**
2422  * bch2_btree_iter_prev: returns first key less than iterator's current
2423  * position
2424  */
2425 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2426 {
2427         if (!bch2_btree_iter_rewind(iter))
2428                 return bkey_s_c_null;
2429
2430         return bch2_btree_iter_peek_prev(iter);
2431 }
2432
2433 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2434 {
2435         struct btree_trans *trans = iter->trans;
2436         struct bpos search_key;
2437         struct bkey_s_c k;
2438         int ret;
2439
2440         bch2_btree_iter_verify(iter);
2441         bch2_btree_iter_verify_entry_exit(iter);
2442         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2443         EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2444
2445         /* extents can't span inode numbers: */
2446         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2447             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2448                 if (iter->pos.inode == KEY_INODE_MAX)
2449                         return bkey_s_c_null;
2450
2451                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2452         }
2453
2454         search_key = btree_iter_search_key(iter);
2455         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2456                                         iter->flags & BTREE_ITER_INTENT,
2457                                         btree_iter_ip_allocated(iter));
2458
2459         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2460         if (unlikely(ret)) {
2461                 k = bkey_s_c_err(ret);
2462                 goto out_no_locked;
2463         }
2464
2465         if ((iter->flags & BTREE_ITER_CACHED) ||
2466             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2467                 struct bkey_i *next_update;
2468
2469                 if ((next_update = btree_trans_peek_updates(iter)) &&
2470                     bpos_eq(next_update->k.p, iter->pos)) {
2471                         iter->k = next_update->k;
2472                         k = bkey_i_to_s_c(next_update);
2473                         goto out;
2474                 }
2475
2476                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2477                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2478                         goto out;
2479
2480                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2481                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2482                         if (!bkey_err(k))
2483                                 iter->k = *k.k;
2484                         /* We're not returning a key from iter->path: */
2485                         goto out_no_locked;
2486                 }
2487
2488                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2489                 if (unlikely(!k.k))
2490                         goto out_no_locked;
2491         } else {
2492                 struct bpos next;
2493                 struct bpos end = iter->pos;
2494
2495                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2496                         end.offset = U64_MAX;
2497
2498                 EBUG_ON(iter->path->level);
2499
2500                 if (iter->flags & BTREE_ITER_INTENT) {
2501                         struct btree_iter iter2;
2502
2503                         bch2_trans_copy_iter(&iter2, iter);
2504                         k = bch2_btree_iter_peek_upto(&iter2, end);
2505
2506                         if (k.k && !bkey_err(k)) {
2507                                 iter->k = iter2.k;
2508                                 k.k = &iter->k;
2509                         }
2510                         bch2_trans_iter_exit(trans, &iter2);
2511                 } else {
2512                         struct bpos pos = iter->pos;
2513
2514                         k = bch2_btree_iter_peek_upto(iter, end);
2515                         if (unlikely(bkey_err(k)))
2516                                 bch2_btree_iter_set_pos(iter, pos);
2517                         else
2518                                 iter->pos = pos;
2519                 }
2520
2521                 if (unlikely(bkey_err(k)))
2522                         goto out_no_locked;
2523
2524                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2525
2526                 if (bkey_lt(iter->pos, next)) {
2527                         bkey_init(&iter->k);
2528                         iter->k.p = iter->pos;
2529
2530                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2531                                 bch2_key_resize(&iter->k,
2532                                                 min_t(u64, KEY_SIZE_MAX,
2533                                                       (next.inode == iter->pos.inode
2534                                                        ? next.offset
2535                                                        : KEY_OFFSET_MAX) -
2536                                                       iter->pos.offset));
2537                                 EBUG_ON(!iter->k.size);
2538                         }
2539
2540                         k = (struct bkey_s_c) { &iter->k, NULL };
2541                 }
2542         }
2543 out:
2544         btree_path_set_should_be_locked(iter->path);
2545 out_no_locked:
2546         bch2_btree_iter_verify_entry_exit(iter);
2547         bch2_btree_iter_verify(iter);
2548         ret = bch2_btree_iter_verify_ret(iter, k);
2549         if (unlikely(ret))
2550                 return bkey_s_c_err(ret);
2551
2552         return k;
2553 }
2554
2555 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2556 {
2557         if (!bch2_btree_iter_advance(iter))
2558                 return bkey_s_c_null;
2559
2560         return bch2_btree_iter_peek_slot(iter);
2561 }
2562
2563 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2564 {
2565         if (!bch2_btree_iter_rewind(iter))
2566                 return bkey_s_c_null;
2567
2568         return bch2_btree_iter_peek_slot(iter);
2569 }
2570
2571 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2572 {
2573         struct bkey_s_c k;
2574
2575         while (btree_trans_too_many_iters(iter->trans) ||
2576                (k = bch2_btree_iter_peek_type(iter, iter->flags),
2577                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2578                 bch2_trans_begin(iter->trans);
2579
2580         return k;
2581 }
2582
2583 /* new transactional stuff: */
2584
2585 #ifdef CONFIG_BCACHEFS_DEBUG
2586 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2587 {
2588         struct btree_path *path;
2589         unsigned i;
2590
2591         BUG_ON(trans->nr_sorted != hweight64(trans->paths_allocated));
2592
2593         trans_for_each_path(trans, path) {
2594                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2595                 BUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2596         }
2597
2598         for (i = 0; i < trans->nr_sorted; i++) {
2599                 unsigned idx = trans->sorted[i];
2600
2601                 EBUG_ON(!(trans->paths_allocated & (1ULL << idx)));
2602                 BUG_ON(trans->paths[idx].sorted_idx != i);
2603         }
2604 }
2605
2606 static void btree_trans_verify_sorted(struct btree_trans *trans)
2607 {
2608         struct btree_path *path, *prev = NULL;
2609         unsigned i;
2610
2611         if (!bch2_debug_check_iterators)
2612                 return;
2613
2614         trans_for_each_path_inorder(trans, path, i) {
2615                 if (prev && btree_path_cmp(prev, path) > 0) {
2616                         __bch2_dump_trans_paths_updates(trans, true);
2617                         panic("trans paths out of order!\n");
2618                 }
2619                 prev = path;
2620         }
2621 }
2622 #else
2623 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2624 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2625 #endif
2626
2627 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2628 {
2629         int i, l = 0, r = trans->nr_sorted, inc = 1;
2630         bool swapped;
2631
2632         btree_trans_verify_sorted_refs(trans);
2633
2634         if (trans->paths_sorted)
2635                 goto out;
2636
2637         /*
2638          * Cocktail shaker sort: this is efficient because iterators will be
2639          * mostly sorted.
2640          */
2641         do {
2642                 swapped = false;
2643
2644                 for (i = inc > 0 ? l : r - 2;
2645                      i + 1 < r && i >= l;
2646                      i += inc) {
2647                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2648                                            trans->paths + trans->sorted[i + 1]) > 0) {
2649                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2650                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2651                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2652                                 swapped = true;
2653                         }
2654                 }
2655
2656                 if (inc > 0)
2657                         --r;
2658                 else
2659                         l++;
2660                 inc = -inc;
2661         } while (swapped);
2662
2663         trans->paths_sorted = true;
2664 out:
2665         btree_trans_verify_sorted(trans);
2666 }
2667
2668 static inline void btree_path_list_remove(struct btree_trans *trans,
2669                                           struct btree_path *path)
2670 {
2671         unsigned i;
2672
2673         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2674 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2675         trans->nr_sorted--;
2676         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2677                                 trans->sorted + path->sorted_idx + 1,
2678                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2679 #else
2680         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2681 #endif
2682         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2683                 trans->paths[trans->sorted[i]].sorted_idx = i;
2684
2685         path->sorted_idx = U8_MAX;
2686 }
2687
2688 static inline void btree_path_list_add(struct btree_trans *trans,
2689                                        struct btree_path *pos,
2690                                        struct btree_path *path)
2691 {
2692         unsigned i;
2693
2694         path->sorted_idx = pos ? pos->sorted_idx + 1 : trans->nr_sorted;
2695
2696 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2697         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2698                               trans->sorted + path->sorted_idx,
2699                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2700         trans->nr_sorted++;
2701         trans->sorted[path->sorted_idx] = path->idx;
2702 #else
2703         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2704 #endif
2705
2706         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2707                 trans->paths[trans->sorted[i]].sorted_idx = i;
2708
2709         btree_trans_verify_sorted_refs(trans);
2710 }
2711
2712 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2713 {
2714         if (iter->path)
2715                 bch2_path_put(trans, iter->path,
2716                               iter->flags & BTREE_ITER_INTENT);
2717         if (iter->update_path)
2718                 bch2_path_put_nokeep(trans, iter->update_path,
2719                               iter->flags & BTREE_ITER_INTENT);
2720         if (iter->key_cache_path)
2721                 bch2_path_put(trans, iter->key_cache_path,
2722                               iter->flags & BTREE_ITER_INTENT);
2723         iter->path = NULL;
2724         iter->update_path = NULL;
2725         iter->key_cache_path = NULL;
2726 }
2727
2728 static inline void bch2_trans_iter_init_inlined(struct btree_trans *trans,
2729                           struct btree_iter *iter,
2730                           unsigned btree_id, struct bpos pos,
2731                           unsigned flags)
2732 {
2733         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2734                                bch2_btree_iter_flags(trans, btree_id, flags),
2735                                _RET_IP_);
2736 }
2737
2738 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2739                           struct btree_iter *iter,
2740                           unsigned btree_id, struct bpos pos,
2741                           unsigned flags)
2742 {
2743         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2744                                bch2_btree_iter_flags(trans, btree_id, flags),
2745                                _RET_IP_);
2746 }
2747
2748 void bch2_trans_node_iter_init(struct btree_trans *trans,
2749                                struct btree_iter *iter,
2750                                enum btree_id btree_id,
2751                                struct bpos pos,
2752                                unsigned locks_want,
2753                                unsigned depth,
2754                                unsigned flags)
2755 {
2756        flags |= BTREE_ITER_NOT_EXTENTS;
2757        flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2758        flags |= BTREE_ITER_ALL_SNAPSHOTS;
2759
2760         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2761                                __bch2_btree_iter_flags(trans, btree_id, flags),
2762                                _RET_IP_);
2763
2764         iter->min_depth = depth;
2765
2766         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2767         BUG_ON(iter->path->level        != depth);
2768         BUG_ON(iter->min_depth          != depth);
2769 }
2770
2771 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2772 {
2773         *dst = *src;
2774         if (src->path)
2775                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2776         if (src->update_path)
2777                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2778         dst->key_cache_path = NULL;
2779 }
2780
2781 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2782 {
2783         unsigned new_top = trans->mem_top + size;
2784         size_t old_bytes = trans->mem_bytes;
2785         size_t new_bytes = roundup_pow_of_two(new_top);
2786         void *new_mem;
2787         void *p;
2788
2789         trans->mem_max = max(trans->mem_max, new_top);
2790
2791         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2792
2793         new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2794         if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2795                 new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2796                 new_bytes = BTREE_TRANS_MEM_MAX;
2797                 kfree(trans->mem);
2798         }
2799
2800         if (!new_mem)
2801                 return ERR_PTR(-ENOMEM);
2802
2803         trans->mem = new_mem;
2804         trans->mem_bytes = new_bytes;
2805
2806         if (old_bytes) {
2807                 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2808                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2809         }
2810
2811         p = trans->mem + trans->mem_top;
2812         trans->mem_top += size;
2813         memset(p, 0, size);
2814         return p;
2815 }
2816
2817 static noinline void bch2_trans_reset_srcu_lock(struct btree_trans *trans)
2818 {
2819         struct bch_fs *c = trans->c;
2820         struct btree_path *path;
2821
2822         trans_for_each_path(trans, path)
2823                 if (path->cached && !btree_node_locked(path, 0))
2824                         path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2825
2826         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2827         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2828         trans->srcu_lock_time   = jiffies;
2829 }
2830
2831 /**
2832  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2833  * @trans: transaction to reset
2834  *
2835  * While iterating over nodes or updating nodes a attempt to lock a btree node
2836  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2837  * occurs bch2_trans_begin() should be called and the transaction retried.
2838  */
2839 u32 bch2_trans_begin(struct btree_trans *trans)
2840 {
2841         struct btree_path *path;
2842
2843         bch2_trans_reset_updates(trans);
2844
2845         trans->restart_count++;
2846         trans->mem_top                  = 0;
2847
2848         trans_for_each_path(trans, path) {
2849                 path->should_be_locked = false;
2850
2851                 /*
2852                  * If the transaction wasn't restarted, we're presuming to be
2853                  * doing something new: dont keep iterators excpt the ones that
2854                  * are in use - except for the subvolumes btree:
2855                  */
2856                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2857                         path->preserve = false;
2858
2859                 /*
2860                  * XXX: we probably shouldn't be doing this if the transaction
2861                  * was restarted, but currently we still overflow transaction
2862                  * iterators if we do that
2863                  */
2864                 if (!path->ref && !path->preserve)
2865                         __bch2_path_free(trans, path);
2866                 else
2867                         path->preserve = false;
2868         }
2869
2870         if (!trans->restarted &&
2871             (need_resched() ||
2872              local_clock() - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2873                 bch2_trans_unlock(trans);
2874                 cond_resched();
2875                 bch2_trans_relock(trans);
2876         }
2877
2878         if (unlikely(time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2879                 bch2_trans_reset_srcu_lock(trans);
2880
2881         trans->last_begin_ip = _RET_IP_;
2882         if (trans->restarted) {
2883                 bch2_btree_path_traverse_all(trans);
2884                 trans->notrace_relock_fail = false;
2885         }
2886
2887         trans->last_begin_time = local_clock();
2888         return trans->restart_count;
2889 }
2890
2891 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
2892 {
2893         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
2894         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2895         void *p = NULL;
2896
2897         BUG_ON(trans->used_mempool);
2898
2899 #ifdef __KERNEL__
2900         p = this_cpu_xchg(c->btree_paths_bufs->path, NULL);
2901 #endif
2902         if (!p)
2903                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
2904
2905         trans->paths            = p; p += paths_bytes;
2906         trans->updates          = p; p += updates_bytes;
2907 }
2908
2909 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2910
2911 unsigned bch2_trans_get_fn_idx(const char *fn)
2912 {
2913         unsigned i;
2914
2915         for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2916                 if (!bch2_btree_transaction_fns[i] ||
2917                     bch2_btree_transaction_fns[i] == fn) {
2918                         bch2_btree_transaction_fns[i] = fn;
2919                         return i;
2920                 }
2921
2922         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2923         return i;
2924 }
2925
2926 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, unsigned fn_idx)
2927         __acquires(&c->btree_trans_barrier)
2928 {
2929         struct btree_transaction_stats *s;
2930
2931         bch2_assert_btree_nodes_not_locked();
2932
2933         memset(trans, 0, sizeof(*trans));
2934         trans->c                = c;
2935         trans->fn               = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2936                 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2937         trans->last_begin_time  = local_clock();
2938         trans->fn_idx           = fn_idx;
2939         trans->locking_wait.task = current;
2940         trans->journal_replay_not_finished =
2941                 !test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
2942         closure_init_stack(&trans->ref);
2943
2944         bch2_trans_alloc_paths(trans, c);
2945
2946         s = btree_trans_stats(trans);
2947         if (s && s->max_mem) {
2948                 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2949
2950                 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2951
2952                 if (!unlikely(trans->mem)) {
2953                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2954                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2955                 } else {
2956                         trans->mem_bytes = expected_mem_bytes;
2957                 }
2958         }
2959
2960         if (s) {
2961                 trans->nr_max_paths = s->nr_max_paths;
2962                 trans->wb_updates_size = s->wb_updates_size;
2963         }
2964
2965         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2966         trans->srcu_lock_time   = jiffies;
2967
2968         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2969                 struct btree_trans *pos;
2970
2971                 mutex_lock(&c->btree_trans_lock);
2972                 list_for_each_entry(pos, &c->btree_trans_list, list) {
2973                         /*
2974                          * We'd much prefer to be stricter here and completely
2975                          * disallow multiple btree_trans in the same thread -
2976                          * but the data move path calls bch2_write when we
2977                          * already have a btree_trans initialized.
2978                          */
2979                         BUG_ON(trans->locking_wait.task->pid == pos->locking_wait.task->pid &&
2980                                bch2_trans_locked(pos));
2981
2982                         if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
2983                                 list_add_tail(&trans->list, &pos->list);
2984                                 goto list_add_done;
2985                         }
2986                 }
2987                 list_add_tail(&trans->list, &c->btree_trans_list);
2988 list_add_done:
2989                 mutex_unlock(&c->btree_trans_lock);
2990         }
2991 }
2992
2993 static void check_btree_paths_leaked(struct btree_trans *trans)
2994 {
2995 #ifdef CONFIG_BCACHEFS_DEBUG
2996         struct bch_fs *c = trans->c;
2997         struct btree_path *path;
2998
2999         trans_for_each_path(trans, path)
3000                 if (path->ref)
3001                         goto leaked;
3002         return;
3003 leaked:
3004         bch_err(c, "btree paths leaked from %s!", trans->fn);
3005         trans_for_each_path(trans, path)
3006                 if (path->ref)
3007                         printk(KERN_ERR "  btree %s %pS\n",
3008                                bch2_btree_ids[path->btree_id],
3009                                (void *) path->ip_allocated);
3010         /* Be noisy about this: */
3011         bch2_fatal_error(c);
3012 #endif
3013 }
3014
3015 void bch2_trans_exit(struct btree_trans *trans)
3016         __releases(&c->btree_trans_barrier)
3017 {
3018         struct btree_insert_entry *i;
3019         struct bch_fs *c = trans->c;
3020         struct btree_transaction_stats *s = btree_trans_stats(trans);
3021
3022         bch2_trans_unlock(trans);
3023
3024         closure_sync(&trans->ref);
3025
3026         if (s)
3027                 s->max_mem = max(s->max_mem, trans->mem_max);
3028
3029         trans_for_each_update(trans, i)
3030                 __btree_path_put(i->path, true);
3031         trans->nr_updates               = 0;
3032
3033         check_btree_paths_leaked(trans);
3034
3035         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
3036                 mutex_lock(&c->btree_trans_lock);
3037                 list_del(&trans->list);
3038                 mutex_unlock(&c->btree_trans_lock);
3039         }
3040
3041         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3042
3043         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3044
3045         kfree(trans->extra_journal_entries.data);
3046
3047         if (trans->fs_usage_deltas) {
3048                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3049                     REPLICAS_DELTA_LIST_MAX)
3050                         mempool_free(trans->fs_usage_deltas,
3051                                      &c->replicas_delta_pool);
3052                 else
3053                         kfree(trans->fs_usage_deltas);
3054         }
3055
3056         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3057                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3058         else
3059                 kfree(trans->mem);
3060
3061 #ifdef __KERNEL__
3062         /*
3063          * Userspace doesn't have a real percpu implementation:
3064          */
3065         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3066 #endif
3067
3068         if (trans->paths)
3069                 mempool_free(trans->paths, &c->btree_paths_pool);
3070
3071         trans->mem      = (void *) 0x1;
3072         trans->paths    = (void *) 0x1;
3073 }
3074
3075 static void __maybe_unused
3076 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3077                                       struct btree_bkey_cached_common *b)
3078 {
3079         struct six_lock_count c = six_lock_counts(&b->lock);
3080         struct task_struct *owner;
3081         pid_t pid;
3082
3083         rcu_read_lock();
3084         owner = READ_ONCE(b->lock.owner);
3085         pid = owner ? owner->pid : 0;
3086         rcu_read_unlock();
3087
3088         prt_tab(out);
3089         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3090                    b->level, bch2_btree_ids[b->btree_id]);
3091         bch2_bpos_to_text(out, btree_node_pos(b));
3092
3093         prt_tab(out);
3094         prt_printf(out, " locks %u:%u:%u held by pid %u",
3095                    c.n[0], c.n[1], c.n[2], pid);
3096 }
3097
3098 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3099 {
3100         struct btree_path *path;
3101         struct btree_bkey_cached_common *b;
3102         static char lock_types[] = { 'r', 'i', 'w' };
3103         unsigned l;
3104
3105         if (!out->nr_tabstops) {
3106                 printbuf_tabstop_push(out, 16);
3107                 printbuf_tabstop_push(out, 32);
3108         }
3109
3110         prt_printf(out, "%i %s\n", trans->locking_wait.task->pid, trans->fn);
3111
3112         trans_for_each_path(trans, path) {
3113                 if (!path->nodes_locked)
3114                         continue;
3115
3116                 prt_printf(out, "  path %u %c l=%u %s:",
3117                        path->idx,
3118                        path->cached ? 'c' : 'b',
3119                        path->level,
3120                        bch2_btree_ids[path->btree_id]);
3121                 bch2_bpos_to_text(out, path->pos);
3122                 prt_newline(out);
3123
3124                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3125                         if (btree_node_locked(path, l) &&
3126                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3127                                 prt_printf(out, "    %c l=%u ",
3128                                            lock_types[btree_node_locked_type(path, l)], l);
3129                                 bch2_btree_bkey_cached_common_to_text(out, b);
3130                                 prt_newline(out);
3131                         }
3132                 }
3133         }
3134
3135         b = READ_ONCE(trans->locking);
3136         if (b) {
3137                 prt_printf(out, "  blocked for %lluus on",
3138                            div_u64(local_clock() - trans->locking_wait.start_time,
3139                                    1000));
3140                 prt_newline(out);
3141                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3142                 bch2_btree_bkey_cached_common_to_text(out, b);
3143                 prt_newline(out);
3144         }
3145 }
3146
3147 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3148 {
3149         struct btree_transaction_stats *s;
3150
3151         for (s = c->btree_transaction_stats;
3152              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3153              s++) {
3154                 kfree(s->max_paths_text);
3155                 bch2_time_stats_exit(&s->lock_hold_times);
3156         }
3157
3158         if (c->btree_trans_barrier_initialized)
3159                 cleanup_srcu_struct(&c->btree_trans_barrier);
3160         mempool_exit(&c->btree_trans_mem_pool);
3161         mempool_exit(&c->btree_paths_pool);
3162 }
3163
3164 int bch2_fs_btree_iter_init(struct bch_fs *c)
3165 {
3166         struct btree_transaction_stats *s;
3167         unsigned nr = BTREE_ITER_MAX;
3168         int ret;
3169
3170         for (s = c->btree_transaction_stats;
3171              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3172              s++) {
3173                 bch2_time_stats_init(&s->lock_hold_times);
3174                 mutex_init(&s->lock);
3175         }
3176
3177         INIT_LIST_HEAD(&c->btree_trans_list);
3178         mutex_init(&c->btree_trans_lock);
3179
3180         ret   = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3181                         sizeof(struct btree_path) * nr +
3182                         sizeof(struct btree_insert_entry) * nr) ?:
3183                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3184                                           BTREE_TRANS_MEM_MAX) ?:
3185                 init_srcu_struct(&c->btree_trans_barrier);
3186         if (!ret)
3187                 c->btree_trans_barrier_initialized = true;
3188         return ret;
3189 }