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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_buf.h"
   6 #include "btree_cache.h"
   7 #include "btree_iter.h"
   8 #include "btree_journal_iter.h"
   9 #include "btree_key_cache.h"
  10 #include "btree_locking.h"
  11 #include "btree_update.h"
  12 #include "debug.h"
  13 #include "error.h"
  14 #include "extents.h"
  15 #include "journal.h"
  16 #include "replicas.h"
  17 #include "subvolume.h"
  18 #include "trace.h"
  19
  20 #include <linux/random.h>
  21 #include <linux/prefetch.h>
  22
  23 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
  24 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
  25                                        struct btree_path *);
  26
  27 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
  28 {
  29 #ifdef TRACK_PATH_ALLOCATED
  30         return iter->ip_allocated;
  31 #else
  32         return 0;
  33 #endif
  34 }
  35
  36 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
  37
  38 static inline int __btree_path_cmp(const struct btree_path *l,
  39                                    enum btree_id        r_btree_id,
  40                                    bool                 r_cached,
  41                                    struct bpos          r_pos,
  42                                    unsigned             r_level)
  43 {
  44         /*
  45          * Must match lock ordering as defined by __bch2_btree_node_lock:
  46          */
  47         return   cmp_int(l->btree_id,   r_btree_id) ?:
  48                  cmp_int((int) l->cached,       (int) r_cached) ?:
  49                  bpos_cmp(l->pos,       r_pos) ?:
  50                 -cmp_int(l->level,      r_level);
  51 }
  52
  53 static inline int btree_path_cmp(const struct btree_path *l,
  54                                  const struct btree_path *r)
  55 {
  56         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  57 }
  58
  59 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  60 {
  61         /* Are we iterating over keys in all snapshots? */
  62         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  63                 p = bpos_successor(p);
  64         } else {
  65                 p = bpos_nosnap_successor(p);
  66                 p.snapshot = iter->snapshot;
  67         }
  68
  69         return p;
  70 }
  71
  72 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  73 {
  74         /* Are we iterating over keys in all snapshots? */
  75         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  76                 p = bpos_predecessor(p);
  77         } else {
  78                 p = bpos_nosnap_predecessor(p);
  79                 p.snapshot = iter->snapshot;
  80         }
  81
  82         return p;
  83 }
  84
  85 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
  86 {
  87         struct bpos pos = iter->pos;
  88
  89         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
  90             !bkey_eq(pos, POS_MAX))
  91                 pos = bkey_successor(iter, pos);
  92         return pos;
  93 }
  94
  95 static inline bool btree_path_pos_before_node(struct btree_path *path,
  96                                               struct btree *b)
  97 {
  98         return bpos_lt(path->pos, b->data->min_key);
  99 }
 100
 101 static inline bool btree_path_pos_after_node(struct btree_path *path,
 102                                              struct btree *b)
 103 {
 104         return bpos_gt(path->pos, b->key.k.p);
 105 }
 106
 107 static inline bool btree_path_pos_in_node(struct btree_path *path,
 108                                           struct btree *b)
 109 {
 110         return path->btree_id == b->c.btree_id &&
 111                 !btree_path_pos_before_node(path, b) &&
 112                 !btree_path_pos_after_node(path, b);
 113 }
 114
 115 /* Btree iterator: */
 116
 117 #ifdef CONFIG_BCACHEFS_DEBUG
 118
 119 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 120                                           struct btree_path *path)
 121 {
 122         struct bkey_cached *ck;
 123         bool locked = btree_node_locked(path, 0);
 124
 125         if (!bch2_btree_node_relock(trans, path, 0))
 126                 return;
 127
 128         ck = (void *) path->l[0].b;
 129         BUG_ON(ck->key.btree_id != path->btree_id ||
 130                !bkey_eq(ck->key.pos, path->pos));
 131
 132         if (!locked)
 133                 btree_node_unlock(trans, path, 0);
 134 }
 135
 136 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 137                                 struct btree_path *path, unsigned level)
 138 {
 139         struct btree_path_level *l;
 140         struct btree_node_iter tmp;
 141         bool locked;
 142         struct bkey_packed *p, *k;
 143         struct printbuf buf1 = PRINTBUF;
 144         struct printbuf buf2 = PRINTBUF;
 145         struct printbuf buf3 = PRINTBUF;
 146         const char *msg;
 147
 148         if (!bch2_debug_check_iterators)
 149                 return;
 150
 151         l       = &path->l[level];
 152         tmp     = l->iter;
 153         locked  = btree_node_locked(path, level);
 154
 155         if (path->cached) {
 156                 if (!level)
 157                         bch2_btree_path_verify_cached(trans, path);
 158                 return;
 159         }
 160
 161         if (!btree_path_node(path, level))
 162                 return;
 163
 164         if (!bch2_btree_node_relock_notrace(trans, path, level))
 165                 return;
 166
 167         BUG_ON(!btree_path_pos_in_node(path, l->b));
 168
 169         bch2_btree_node_iter_verify(&l->iter, l->b);
 170
 171         /*
 172          * For interior nodes, the iterator will have skipped past deleted keys:
 173          */
 174         p = level
 175                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 176                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 177         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 178
 179         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 180                 msg = "before";
 181                 goto err;
 182         }
 183
 184         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 185                 msg = "after";
 186                 goto err;
 187         }
 188
 189         if (!locked)
 190                 btree_node_unlock(trans, path, level);
 191         return;
 192 err:
 193         bch2_bpos_to_text(&buf1, path->pos);
 194
 195         if (p) {
 196                 struct bkey uk = bkey_unpack_key(l->b, p);
 197
 198                 bch2_bkey_to_text(&buf2, &uk);
 199         } else {
 200                 prt_printf(&buf2, "(none)");
 201         }
 202
 203         if (k) {
 204                 struct bkey uk = bkey_unpack_key(l->b, k);
 205
 206                 bch2_bkey_to_text(&buf3, &uk);
 207         } else {
 208                 prt_printf(&buf3, "(none)");
 209         }
 210
 211         panic("path should be %s key at level %u:\n"
 212               "path pos %s\n"
 213               "prev key %s\n"
 214               "cur  key %s\n",
 215               msg, level, buf1.buf, buf2.buf, buf3.buf);
 216 }
 217
 218 static void bch2_btree_path_verify(struct btree_trans *trans,
 219                                    struct btree_path *path)
 220 {
 221         struct bch_fs *c = trans->c;
 222         unsigned i;
 223
 224         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 225
 226         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 227                 if (!path->l[i].b) {
 228                         BUG_ON(!path->cached &&
 229                                bch2_btree_id_root(c, path->btree_id)->b->c.level > i);
 230                         break;
 231                 }
 232
 233                 bch2_btree_path_verify_level(trans, path, i);
 234         }
 235
 236         bch2_btree_path_verify_locks(path);
 237 }
 238
 239 void bch2_trans_verify_paths(struct btree_trans *trans)
 240 {
 241         struct btree_path *path;
 242
 243         trans_for_each_path(trans, path)
 244                 bch2_btree_path_verify(trans, path);
 245 }
 246
 247 static void bch2_btree_iter_verify(struct btree_iter *iter)
 248 {
 249         struct btree_trans *trans = iter->trans;
 250
 251         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 252
 253         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
 254
 255         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 256                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 257
 258         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 259                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 260                !btree_type_has_snapshots(iter->btree_id));
 261
 262         if (iter->update_path)
 263                 bch2_btree_path_verify(trans, iter->update_path);
 264         bch2_btree_path_verify(trans, iter->path);
 265 }
 266
 267 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 268 {
 269         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 270                !iter->pos.snapshot);
 271
 272         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 273                iter->pos.snapshot != iter->snapshot);
 274
 275         BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
 276                bkey_gt(iter->pos, iter->k.p));
 277 }
 278
 279 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 280 {
 281         struct btree_trans *trans = iter->trans;
 282         struct btree_iter copy;
 283         struct bkey_s_c prev;
 284         int ret = 0;
 285
 286         if (!bch2_debug_check_iterators)
 287                 return 0;
 288
 289         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 290                 return 0;
 291
 292         if (bkey_err(k) || !k.k)
 293                 return 0;
 294
 295         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 296                                           iter->snapshot,
 297                                           k.k->p.snapshot));
 298
 299         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 300                              BTREE_ITER_NOPRESERVE|
 301                              BTREE_ITER_ALL_SNAPSHOTS);
 302         prev = bch2_btree_iter_prev(&copy);
 303         if (!prev.k)
 304                 goto out;
 305
 306         ret = bkey_err(prev);
 307         if (ret)
 308                 goto out;
 309
 310         if (bkey_eq(prev.k->p, k.k->p) &&
 311             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 312                                       prev.k->p.snapshot) > 0) {
 313                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 314
 315                 bch2_bkey_to_text(&buf1, k.k);
 316                 bch2_bkey_to_text(&buf2, prev.k);
 317
 318                 panic("iter snap %u\n"
 319                       "k    %s\n"
 320                       "prev %s\n",
 321                       iter->snapshot,
 322                       buf1.buf, buf2.buf);
 323         }
 324 out:
 325         bch2_trans_iter_exit(trans, &copy);
 326         return ret;
 327 }
 328
 329 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 330                             struct bpos pos, bool key_cache)
 331 {
 332         struct btree_path *path;
 333         unsigned idx;
 334         struct printbuf buf = PRINTBUF;
 335
 336         btree_trans_sort_paths(trans);
 337
 338         trans_for_each_path_inorder(trans, path, idx) {
 339                 int cmp = cmp_int(path->btree_id, id) ?:
 340                         cmp_int(path->cached, key_cache);
 341
 342                 if (cmp > 0)
 343                         break;
 344                 if (cmp < 0)
 345                         continue;
 346
 347                 if (!btree_node_locked(path, 0) ||
 348                     !path->should_be_locked)
 349                         continue;
 350
 351                 if (!key_cache) {
 352                         if (bkey_ge(pos, path->l[0].b->data->min_key) &&
 353                             bkey_le(pos, path->l[0].b->key.k.p))
 354                                 return;
 355                 } else {
 356                         if (bkey_eq(pos, path->pos))
 357                                 return;
 358                 }
 359         }
 360
 361         bch2_dump_trans_paths_updates(trans);
 362         bch2_bpos_to_text(&buf, pos);
 363
 364         panic("not locked: %s %s%s\n",
 365               bch2_btree_ids[id], buf.buf,
 366               key_cache ? " cached" : "");
 367 }
 368
 369 #else
 370
 371 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 372                                                 struct btree_path *path, unsigned l) {}
 373 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 374                                           struct btree_path *path) {}
 375 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 376 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 377 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 378
 379 #endif
 380
 381 /* Btree path: fixups after btree updates */
 382
 383 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 384                                         struct btree *b,
 385                                         struct bset_tree *t,
 386                                         struct bkey_packed *k)
 387 {
 388         struct btree_node_iter_set *set;
 389
 390         btree_node_iter_for_each(iter, set)
 391                 if (set->end == t->end_offset) {
 392                         set->k = __btree_node_key_to_offset(b, k);
 393                         bch2_btree_node_iter_sort(iter, b);
 394                         return;
 395                 }
 396
 397         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 398 }
 399
 400 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 401                                                struct btree *b,
 402                                                struct bkey_packed *where)
 403 {
 404         struct btree_path_level *l = &path->l[b->c.level];
 405
 406         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 407                 return;
 408
 409         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 410                 bch2_btree_node_iter_advance(&l->iter, l->b);
 411 }
 412
 413 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 414                                       struct btree *b,
 415                                       struct bkey_packed *where)
 416 {
 417         struct btree_path *path;
 418
 419         trans_for_each_path_with_node(trans, b, path) {
 420                 __bch2_btree_path_fix_key_modified(path, b, where);
 421                 bch2_btree_path_verify_level(trans, path, b->c.level);
 422         }
 423 }
 424
 425 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 426                                        struct btree *b,
 427                                        struct btree_node_iter *node_iter,
 428                                        struct bset_tree *t,
 429                                        struct bkey_packed *where,
 430                                        unsigned clobber_u64s,
 431                                        unsigned new_u64s)
 432 {
 433         const struct bkey_packed *end = btree_bkey_last(b, t);
 434         struct btree_node_iter_set *set;
 435         unsigned offset = __btree_node_key_to_offset(b, where);
 436         int shift = new_u64s - clobber_u64s;
 437         unsigned old_end = t->end_offset - shift;
 438         unsigned orig_iter_pos = node_iter->data[0].k;
 439         bool iter_current_key_modified =
 440                 orig_iter_pos >= offset &&
 441                 orig_iter_pos <= offset + clobber_u64s;
 442
 443         btree_node_iter_for_each(node_iter, set)
 444                 if (set->end == old_end)
 445                         goto found;
 446
 447         /* didn't find the bset in the iterator - might have to readd it: */
 448         if (new_u64s &&
 449             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 450                 bch2_btree_node_iter_push(node_iter, b, where, end);
 451                 goto fixup_done;
 452         } else {
 453                 /* Iterator is after key that changed */
 454                 return;
 455         }
 456 found:
 457         set->end = t->end_offset;
 458
 459         /* Iterator hasn't gotten to the key that changed yet: */
 460         if (set->k < offset)
 461                 return;
 462
 463         if (new_u64s &&
 464             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 465                 set->k = offset;
 466         } else if (set->k < offset + clobber_u64s) {
 467                 set->k = offset + new_u64s;
 468                 if (set->k == set->end)
 469                         bch2_btree_node_iter_set_drop(node_iter, set);
 470         } else {
 471                 /* Iterator is after key that changed */
 472                 set->k = (int) set->k + shift;
 473                 return;
 474         }
 475
 476         bch2_btree_node_iter_sort(node_iter, b);
 477 fixup_done:
 478         if (node_iter->data[0].k != orig_iter_pos)
 479                 iter_current_key_modified = true;
 480
 481         /*
 482          * When a new key is added, and the node iterator now points to that
 483          * key, the iterator might have skipped past deleted keys that should
 484          * come after the key the iterator now points to. We have to rewind to
 485          * before those deleted keys - otherwise
 486          * bch2_btree_node_iter_prev_all() breaks:
 487          */
 488         if (!bch2_btree_node_iter_end(node_iter) &&
 489             iter_current_key_modified &&
 490             b->c.level) {
 491                 struct bset_tree *t;
 492                 struct bkey_packed *k, *k2, *p;
 493
 494                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 495
 496                 for_each_bset(b, t) {
 497                         bool set_pos = false;
 498
 499                         if (node_iter->data[0].end == t->end_offset)
 500                                 continue;
 501
 502                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 503
 504                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 505                                bkey_iter_cmp(b, k, p) < 0) {
 506                                 k2 = p;
 507                                 set_pos = true;
 508                         }
 509
 510                         if (set_pos)
 511                                 btree_node_iter_set_set_pos(node_iter,
 512                                                             b, t, k2);
 513                 }
 514         }
 515 }
 516
 517 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 518                               struct btree_path *path,
 519                               struct btree *b,
 520                               struct btree_node_iter *node_iter,
 521                               struct bkey_packed *where,
 522                               unsigned clobber_u64s,
 523                               unsigned new_u64s)
 524 {
 525         struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
 526         struct btree_path *linked;
 527
 528         if (node_iter != &path->l[b->c.level].iter) {
 529                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 530                                            where, clobber_u64s, new_u64s);
 531
 532                 if (bch2_debug_check_iterators)
 533                         bch2_btree_node_iter_verify(node_iter, b);
 534         }
 535
 536         trans_for_each_path_with_node(trans, b, linked) {
 537                 __bch2_btree_node_iter_fix(linked, b,
 538                                            &linked->l[b->c.level].iter, t,
 539                                            where, clobber_u64s, new_u64s);
 540                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 541         }
 542 }
 543
 544 /* Btree path level: pointer to a particular btree node and node iter */
 545
 546 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
 547                                                   struct btree_path_level *l,
 548                                                   struct bkey *u,
 549                                                   struct bkey_packed *k)
 550 {
 551         if (unlikely(!k)) {
 552                 /*
 553                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 554                  * a hole
 555                  */
 556                 u->type = KEY_TYPE_deleted;
 557                 return bkey_s_c_null;
 558         }
 559
 560         return bkey_disassemble(l->b, k, u);
 561 }
 562
 563 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
 564                                                         struct btree_path_level *l,
 565                                                         struct bkey *u)
 566 {
 567         return __btree_iter_unpack(c, l, u,
 568                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 569 }
 570
 571 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
 572                                                     struct btree_path *path,
 573                                                     struct btree_path_level *l,
 574                                                     struct bkey *u)
 575 {
 576         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 577                         bch2_btree_node_iter_peek(&l->iter, l->b));
 578
 579         path->pos = k.k ? k.k->p : l->b->key.k.p;
 580         trans->paths_sorted = false;
 581         bch2_btree_path_verify_level(trans, path, l - path->l);
 582         return k;
 583 }
 584
 585 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
 586                                                     struct btree_path *path,
 587                                                     struct btree_path_level *l,
 588                                                     struct bkey *u)
 589 {
 590         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 591                         bch2_btree_node_iter_prev(&l->iter, l->b));
 592
 593         path->pos = k.k ? k.k->p : l->b->data->min_key;
 594         trans->paths_sorted = false;
 595         bch2_btree_path_verify_level(trans, path, l - path->l);
 596         return k;
 597 }
 598
 599 static inline bool btree_path_advance_to_pos(struct btree_path *path,
 600                                              struct btree_path_level *l,
 601                                              int max_advance)
 602 {
 603         struct bkey_packed *k;
 604         int nr_advanced = 0;
 605
 606         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 607                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 608                 if (max_advance > 0 && nr_advanced >= max_advance)
 609                         return false;
 610
 611                 bch2_btree_node_iter_advance(&l->iter, l->b);
 612                 nr_advanced++;
 613         }
 614
 615         return true;
 616 }
 617
 618 static inline void __btree_path_level_init(struct btree_path *path,
 619                                            unsigned level)
 620 {
 621         struct btree_path_level *l = &path->l[level];
 622
 623         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
 624
 625         /*
 626          * Iterators to interior nodes should always be pointed at the first non
 627          * whiteout:
 628          */
 629         if (level)
 630                 bch2_btree_node_iter_peek(&l->iter, l->b);
 631 }
 632
 633 void bch2_btree_path_level_init(struct btree_trans *trans,
 634                                 struct btree_path *path,
 635                                 struct btree *b)
 636 {
 637         BUG_ON(path->cached);
 638
 639         EBUG_ON(!btree_path_pos_in_node(path, b));
 640
 641         path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 642         path->l[b->c.level].b = b;
 643         __btree_path_level_init(path, b->c.level);
 644 }
 645
 646 /* Btree path: fixups after btree node updates: */
 647
 648 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
 649 {
 650         struct bch_fs *c = trans->c;
 651         struct btree_insert_entry *i;
 652
 653         trans_for_each_update(trans, i)
 654                 if (!i->cached &&
 655                     i->level    == b->c.level &&
 656                     i->btree_id == b->c.btree_id &&
 657                     bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
 658                     bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
 659                         i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 660
 661                         if (unlikely(trans->journal_replay_not_finished)) {
 662                                 struct bkey_i *j_k =
 663                                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 664                                                                     i->k->k.p);
 665
 666                                 if (j_k) {
 667                                         i->old_k = j_k->k;
 668                                         i->old_v = &j_k->v;
 669                                 }
 670                         }
 671                 }
 672 }
 673
 674 /*
 675  * A btree node is being replaced - update the iterator to point to the new
 676  * node:
 677  */
 678 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
 679 {
 680         struct btree_path *path;
 681
 682         trans_for_each_path(trans, path)
 683                 if (path->uptodate == BTREE_ITER_UPTODATE &&
 684                     !path->cached &&
 685                     btree_path_pos_in_node(path, b)) {
 686                         enum btree_node_locked_type t =
 687                                 btree_lock_want(path, b->c.level);
 688
 689                         if (t != BTREE_NODE_UNLOCKED) {
 690                                 btree_node_unlock(trans, path, b->c.level);
 691                                 six_lock_increment(&b->c.lock, (enum six_lock_type) t);
 692                                 mark_btree_node_locked(trans, path, b->c.level, (enum six_lock_type) t);
 693                         }
 694
 695                         bch2_btree_path_level_init(trans, path, b);
 696                 }
 697
 698         bch2_trans_revalidate_updates_in_node(trans, b);
 699 }
 700
 701 /*
 702  * A btree node has been modified in such a way as to invalidate iterators - fix
 703  * them:
 704  */
 705 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
 706 {
 707         struct btree_path *path;
 708
 709         trans_for_each_path_with_node(trans, b, path)
 710                 __btree_path_level_init(path, b->c.level);
 711
 712         bch2_trans_revalidate_updates_in_node(trans, b);
 713 }
 714
 715 /* Btree path: traverse, set_pos: */
 716
 717 static inline int btree_path_lock_root(struct btree_trans *trans,
 718                                        struct btree_path *path,
 719                                        unsigned depth_want,
 720                                        unsigned long trace_ip)
 721 {
 722         struct bch_fs *c = trans->c;
 723         struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b;
 724         enum six_lock_type lock_type;
 725         unsigned i;
 726         int ret;
 727
 728         EBUG_ON(path->nodes_locked);
 729
 730         while (1) {
 731                 b = READ_ONCE(*rootp);
 732                 path->level = READ_ONCE(b->c.level);
 733
 734                 if (unlikely(path->level < depth_want)) {
 735                         /*
 736                          * the root is at a lower depth than the depth we want:
 737                          * got to the end of the btree, or we're walking nodes
 738                          * greater than some depth and there are no nodes >=
 739                          * that depth
 740                          */
 741                         path->level = depth_want;
 742                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
 743                                 path->l[i].b = NULL;
 744                         return 1;
 745                 }
 746
 747                 lock_type = __btree_lock_want(path, path->level);
 748                 ret = btree_node_lock(trans, path, &b->c,
 749                                       path->level, lock_type, trace_ip);
 750                 if (unlikely(ret)) {
 751                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
 752                                 continue;
 753                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 754                                 return ret;
 755                         BUG();
 756                 }
 757
 758                 if (likely(b == READ_ONCE(*rootp) &&
 759                            b->c.level == path->level &&
 760                            !race_fault())) {
 761                         for (i = 0; i < path->level; i++)
 762                                 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
 763                         path->l[path->level].b = b;
 764                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
 765                                 path->l[i].b = NULL;
 766
 767                         mark_btree_node_locked(trans, path, path->level, lock_type);
 768                         bch2_btree_path_level_init(trans, path, b);
 769                         return 0;
 770                 }
 771
 772                 six_unlock_type(&b->c.lock, lock_type);
 773         }
 774 }
 775
 776 noinline
 777 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
 778 {
 779         struct bch_fs *c = trans->c;
 780         struct btree_path_level *l = path_l(path);
 781         struct btree_node_iter node_iter = l->iter;
 782         struct bkey_packed *k;
 783         struct bkey_buf tmp;
 784         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 785                 ? (path->level > 1 ? 0 :  2)
 786                 : (path->level > 1 ? 1 : 16);
 787         bool was_locked = btree_node_locked(path, path->level);
 788         int ret = 0;
 789
 790         bch2_bkey_buf_init(&tmp);
 791
 792         while (nr-- && !ret) {
 793                 if (!bch2_btree_node_relock(trans, path, path->level))
 794                         break;
 795
 796                 bch2_btree_node_iter_advance(&node_iter, l->b);
 797                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
 798                 if (!k)
 799                         break;
 800
 801                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
 802                 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
 803                                                path->level - 1);
 804         }
 805
 806         if (!was_locked)
 807                 btree_node_unlock(trans, path, path->level);
 808
 809         bch2_bkey_buf_exit(&tmp, c);
 810         return ret;
 811 }
 812
 813 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
 814                                  struct btree_and_journal_iter *jiter)
 815 {
 816         struct bch_fs *c = trans->c;
 817         struct bkey_s_c k;
 818         struct bkey_buf tmp;
 819         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 820                 ? (path->level > 1 ? 0 :  2)
 821                 : (path->level > 1 ? 1 : 16);
 822         bool was_locked = btree_node_locked(path, path->level);
 823         int ret = 0;
 824
 825         bch2_bkey_buf_init(&tmp);
 826
 827         while (nr-- && !ret) {
 828                 if (!bch2_btree_node_relock(trans, path, path->level))
 829                         break;
 830
 831                 bch2_btree_and_journal_iter_advance(jiter);
 832                 k = bch2_btree_and_journal_iter_peek(jiter);
 833                 if (!k.k)
 834                         break;
 835
 836                 bch2_bkey_buf_reassemble(&tmp, c, k);
 837                 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
 838                                                path->level - 1);
 839         }
 840
 841         if (!was_locked)
 842                 btree_node_unlock(trans, path, path->level);
 843
 844         bch2_bkey_buf_exit(&tmp, c);
 845         return ret;
 846 }
 847
 848 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
 849                                             struct btree_path *path,
 850                                             unsigned plevel, struct btree *b)
 851 {
 852         struct btree_path_level *l = &path->l[plevel];
 853         bool locked = btree_node_locked(path, plevel);
 854         struct bkey_packed *k;
 855         struct bch_btree_ptr_v2 *bp;
 856
 857         if (!bch2_btree_node_relock(trans, path, plevel))
 858                 return;
 859
 860         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 861         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
 862
 863         bp = (void *) bkeyp_val(&l->b->format, k);
 864         bp->mem_ptr = (unsigned long)b;
 865
 866         if (!locked)
 867                 btree_node_unlock(trans, path, plevel);
 868 }
 869
 870 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
 871                                                      struct btree_path *path,
 872                                                      unsigned flags,
 873                                                      struct bkey_buf *out)
 874 {
 875         struct bch_fs *c = trans->c;
 876         struct btree_path_level *l = path_l(path);
 877         struct btree_and_journal_iter jiter;
 878         struct bkey_s_c k;
 879         int ret = 0;
 880
 881         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
 882
 883         k = bch2_btree_and_journal_iter_peek(&jiter);
 884
 885         bch2_bkey_buf_reassemble(out, c, k);
 886
 887         if (flags & BTREE_ITER_PREFETCH)
 888                 ret = btree_path_prefetch_j(trans, path, &jiter);
 889
 890         bch2_btree_and_journal_iter_exit(&jiter);
 891         return ret;
 892 }
 893
 894 static __always_inline int btree_path_down(struct btree_trans *trans,
 895                                            struct btree_path *path,
 896                                            unsigned flags,
 897                                            unsigned long trace_ip)
 898 {
 899         struct bch_fs *c = trans->c;
 900         struct btree_path_level *l = path_l(path);
 901         struct btree *b;
 902         unsigned level = path->level - 1;
 903         enum six_lock_type lock_type = __btree_lock_want(path, level);
 904         struct bkey_buf tmp;
 905         int ret;
 906
 907         EBUG_ON(!btree_node_locked(path, path->level));
 908
 909         bch2_bkey_buf_init(&tmp);
 910
 911         if (unlikely(trans->journal_replay_not_finished)) {
 912                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
 913                 if (ret)
 914                         goto err;
 915         } else {
 916                 bch2_bkey_buf_unpack(&tmp, c, l->b,
 917                                  bch2_btree_node_iter_peek(&l->iter, l->b));
 918
 919                 if (flags & BTREE_ITER_PREFETCH) {
 920                         ret = btree_path_prefetch(trans, path);
 921                         if (ret)
 922                                 goto err;
 923                 }
 924         }
 925
 926         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
 927         ret = PTR_ERR_OR_ZERO(b);
 928         if (unlikely(ret))
 929                 goto err;
 930
 931         if (likely(!trans->journal_replay_not_finished &&
 932                    tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
 933             unlikely(b != btree_node_mem_ptr(tmp.k)))
 934                 btree_node_mem_ptr_set(trans, path, level + 1, b);
 935
 936         if (btree_node_read_locked(path, level + 1))
 937                 btree_node_unlock(trans, path, level + 1);
 938
 939         mark_btree_node_locked(trans, path, level, lock_type);
 940         path->level = level;
 941         bch2_btree_path_level_init(trans, path, b);
 942
 943         bch2_btree_path_verify_locks(path);
 944 err:
 945         bch2_bkey_buf_exit(&tmp, c);
 946         return ret;
 947 }
 948
 949
 950 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
 951 {
 952         struct bch_fs *c = trans->c;
 953         struct btree_path *path;
 954         unsigned long trace_ip = _RET_IP_;
 955         int i, ret = 0;
 956
 957         if (trans->in_traverse_all)
 958                 return -BCH_ERR_transaction_restart_in_traverse_all;
 959
 960         trans->in_traverse_all = true;
 961 retry_all:
 962         trans->restarted = 0;
 963         trans->last_restarted_ip = 0;
 964
 965         trans_for_each_path(trans, path)
 966                 path->should_be_locked = false;
 967
 968         btree_trans_sort_paths(trans);
 969
 970         bch2_trans_unlock(trans);
 971         cond_resched();
 972
 973         if (unlikely(trans->memory_allocation_failure)) {
 974                 struct closure cl;
 975
 976                 closure_init_stack(&cl);
 977
 978                 do {
 979                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 980                         closure_sync(&cl);
 981                 } while (ret);
 982         }
 983
 984         /* Now, redo traversals in correct order: */
 985         i = 0;
 986         while (i < trans->nr_sorted) {
 987                 path = trans->paths + trans->sorted[i];
 988
 989                 /*
 990                  * Traversing a path can cause another path to be added at about
 991                  * the same position:
 992                  */
 993                 if (path->uptodate) {
 994                         __btree_path_get(path, false);
 995                         ret = bch2_btree_path_traverse_one(trans, path, 0, _THIS_IP_);
 996                         __btree_path_put(path, false);
 997
 998                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
 999                             bch2_err_matches(ret, ENOMEM))
1000                                 goto retry_all;
1001                         if (ret)
1002                                 goto err;
1003                 } else {
1004                         i++;
1005                 }
1006         }
1007
1008         /*
1009          * We used to assert that all paths had been traversed here
1010          * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1011          * path->Should_be_locked is not set yet, we we might have unlocked and
1012          * then failed to relock a path - that's fine.
1013          */
1014 err:
1015         bch2_btree_cache_cannibalize_unlock(c);
1016
1017         trans->in_traverse_all = false;
1018
1019         trace_and_count(c, trans_traverse_all, trans, trace_ip);
1020         return ret;
1021 }
1022
1023 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1024                                                 unsigned l, int check_pos)
1025 {
1026         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1027                 return false;
1028         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1029                 return false;
1030         return true;
1031 }
1032
1033 static inline bool btree_path_good_node(struct btree_trans *trans,
1034                                         struct btree_path *path,
1035                                         unsigned l, int check_pos)
1036 {
1037         return is_btree_node(path, l) &&
1038                 bch2_btree_node_relock(trans, path, l) &&
1039                 btree_path_check_pos_in_node(path, l, check_pos);
1040 }
1041
1042 static void btree_path_set_level_down(struct btree_trans *trans,
1043                                       struct btree_path *path,
1044                                       unsigned new_level)
1045 {
1046         unsigned l;
1047
1048         path->level = new_level;
1049
1050         for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1051                 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1052                         btree_node_unlock(trans, path, l);
1053
1054         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1055         bch2_btree_path_verify(trans, path);
1056 }
1057
1058 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1059                                                          struct btree_path *path,
1060                                                          int check_pos)
1061 {
1062         unsigned i, l = path->level;
1063 again:
1064         while (btree_path_node(path, l) &&
1065                !btree_path_good_node(trans, path, l, check_pos))
1066                 __btree_path_set_level_up(trans, path, l++);
1067
1068         /* If we need intent locks, take them too: */
1069         for (i = l + 1;
1070              i < path->locks_want && btree_path_node(path, i);
1071              i++)
1072                 if (!bch2_btree_node_relock(trans, path, i)) {
1073                         while (l <= i)
1074                                 __btree_path_set_level_up(trans, path, l++);
1075                         goto again;
1076                 }
1077
1078         return l;
1079 }
1080
1081 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1082                                                      struct btree_path *path,
1083                                                      int check_pos)
1084 {
1085         return likely(btree_node_locked(path, path->level) &&
1086                       btree_path_check_pos_in_node(path, path->level, check_pos))
1087                 ? path->level
1088                 : __btree_path_up_until_good_node(trans, path, check_pos);
1089 }
1090
1091 /*
1092  * This is the main state machine for walking down the btree - walks down to a
1093  * specified depth
1094  *
1095  * Returns 0 on success, -EIO on error (error reading in a btree node).
1096  *
1097  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1098  * stashed in the iterator and returned from bch2_trans_exit().
1099  */
1100 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1101                                  struct btree_path *path,
1102                                  unsigned flags,
1103                                  unsigned long trace_ip)
1104 {
1105         unsigned depth_want = path->level;
1106         int ret = -((int) trans->restarted);
1107
1108         if (unlikely(ret))
1109                 goto out;
1110
1111         /*
1112          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1113          * and re-traverse the path without a transaction restart:
1114          */
1115         if (path->should_be_locked) {
1116                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1117                 goto out;
1118         }
1119
1120         if (path->cached) {
1121                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1122                 goto out;
1123         }
1124
1125         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1126                 goto out;
1127
1128         path->level = btree_path_up_until_good_node(trans, path, 0);
1129
1130         EBUG_ON(btree_path_node(path, path->level) &&
1131                 !btree_node_locked(path, path->level));
1132
1133         /*
1134          * Note: path->nodes[path->level] may be temporarily NULL here - that
1135          * would indicate to other code that we got to the end of the btree,
1136          * here it indicates that relocking the root failed - it's critical that
1137          * btree_path_lock_root() comes next and that it can't fail
1138          */
1139         while (path->level > depth_want) {
1140                 ret = btree_path_node(path, path->level)
1141                         ? btree_path_down(trans, path, flags, trace_ip)
1142                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1143                 if (unlikely(ret)) {
1144                         if (ret == 1) {
1145                                 /*
1146                                  * No nodes at this level - got to the end of
1147                                  * the btree:
1148                                  */
1149                                 ret = 0;
1150                                 goto out;
1151                         }
1152
1153                         __bch2_btree_path_unlock(trans, path);
1154                         path->level = depth_want;
1155                         path->l[path->level].b = ERR_PTR(ret);
1156                         goto out;
1157                 }
1158         }
1159
1160         path->uptodate = BTREE_ITER_UPTODATE;
1161 out:
1162         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1163                 panic("ret %s (%i) trans->restarted %s (%i)\n",
1164                       bch2_err_str(ret), ret,
1165                       bch2_err_str(trans->restarted), trans->restarted);
1166         bch2_btree_path_verify(trans, path);
1167         return ret;
1168 }
1169
1170 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1171                             struct btree_path *src)
1172 {
1173         unsigned i, offset = offsetof(struct btree_path, pos);
1174
1175         memcpy((void *) dst + offset,
1176                (void *) src + offset,
1177                sizeof(struct btree_path) - offset);
1178
1179         for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1180                 unsigned t = btree_node_locked_type(dst, i);
1181
1182                 if (t != BTREE_NODE_UNLOCKED)
1183                         six_lock_increment(&dst->l[i].b->c.lock, t);
1184         }
1185 }
1186
1187 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1188                                            bool intent)
1189 {
1190         struct btree_path *new = btree_path_alloc(trans, src);
1191
1192         btree_path_copy(trans, new, src);
1193         __btree_path_get(new, intent);
1194         return new;
1195 }
1196
1197 __flatten
1198 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *trans,
1199                          struct btree_path *path, bool intent,
1200                          unsigned long ip)
1201 {
1202         __btree_path_put(path, intent);
1203         path = btree_path_clone(trans, path, intent);
1204         path->preserve = false;
1205         return path;
1206 }
1207
1208 struct btree_path * __must_check
1209 __bch2_btree_path_set_pos(struct btree_trans *trans,
1210                    struct btree_path *path, struct bpos new_pos,
1211                    bool intent, unsigned long ip, int cmp)
1212 {
1213         unsigned level = path->level;
1214
1215         bch2_trans_verify_not_in_restart(trans);
1216         EBUG_ON(!path->ref);
1217
1218         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1219
1220         path->pos               = new_pos;
1221         trans->paths_sorted     = false;
1222
1223         if (unlikely(path->cached)) {
1224                 btree_node_unlock(trans, path, 0);
1225                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1226                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1227                 goto out;
1228         }
1229
1230         level = btree_path_up_until_good_node(trans, path, cmp);
1231
1232         if (btree_path_node(path, level)) {
1233                 struct btree_path_level *l = &path->l[level];
1234
1235                 BUG_ON(!btree_node_locked(path, level));
1236                 /*
1237                  * We might have to skip over many keys, or just a few: try
1238                  * advancing the node iterator, and if we have to skip over too
1239                  * many keys just reinit it (or if we're rewinding, since that
1240                  * is expensive).
1241                  */
1242                 if (cmp < 0 ||
1243                     !btree_path_advance_to_pos(path, l, 8))
1244                         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1245
1246                 /*
1247                  * Iterators to interior nodes should always be pointed at the first non
1248                  * whiteout:
1249                  */
1250                 if (unlikely(level))
1251                         bch2_btree_node_iter_peek(&l->iter, l->b);
1252         }
1253
1254         if (unlikely(level != path->level)) {
1255                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1256                 __bch2_btree_path_unlock(trans, path);
1257         }
1258 out:
1259         bch2_btree_path_verify(trans, path);
1260         return path;
1261 }
1262
1263 /* Btree path: main interface: */
1264
1265 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1266 {
1267         struct btree_path *sib;
1268
1269         sib = prev_btree_path(trans, path);
1270         if (sib && !btree_path_cmp(sib, path))
1271                 return sib;
1272
1273         sib = next_btree_path(trans, path);
1274         if (sib && !btree_path_cmp(sib, path))
1275                 return sib;
1276
1277         return NULL;
1278 }
1279
1280 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1281 {
1282         struct btree_path *sib;
1283
1284         sib = prev_btree_path(trans, path);
1285         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1286                 return sib;
1287
1288         sib = next_btree_path(trans, path);
1289         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1290                 return sib;
1291
1292         return NULL;
1293 }
1294
1295 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1296 {
1297         __bch2_btree_path_unlock(trans, path);
1298         btree_path_list_remove(trans, path);
1299         trans->paths_allocated &= ~(1ULL << path->idx);
1300 }
1301
1302 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1303 {
1304         struct btree_path *dup;
1305
1306         EBUG_ON(trans->paths + path->idx != path);
1307         EBUG_ON(!path->ref);
1308
1309         if (!__btree_path_put(path, intent))
1310                 return;
1311
1312         dup = path->preserve
1313                 ? have_path_at_pos(trans, path)
1314                 : have_node_at_pos(trans, path);
1315
1316         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1317                 return;
1318
1319         if (path->should_be_locked &&
1320             !trans->restarted &&
1321             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1322                 return;
1323
1324         if (dup) {
1325                 dup->preserve           |= path->preserve;
1326                 dup->should_be_locked   |= path->should_be_locked;
1327         }
1328
1329         __bch2_path_free(trans, path);
1330 }
1331
1332 static void bch2_path_put_nokeep(struct btree_trans *trans, struct btree_path *path,
1333                                  bool intent)
1334 {
1335         EBUG_ON(trans->paths + path->idx != path);
1336         EBUG_ON(!path->ref);
1337
1338         if (!__btree_path_put(path, intent))
1339                 return;
1340
1341         __bch2_path_free(trans, path);
1342 }
1343
1344 void bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1345 {
1346         panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1347               trans->restart_count, restart_count,
1348               (void *) trans->last_begin_ip);
1349 }
1350
1351 void bch2_trans_in_restart_error(struct btree_trans *trans)
1352 {
1353         panic("in transaction restart: %s, last restarted by %pS\n",
1354               bch2_err_str(trans->restarted),
1355               (void *) trans->last_restarted_ip);
1356 }
1357
1358 noinline __cold
1359 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1360 {
1361         struct btree_insert_entry *i;
1362         struct btree_write_buffered_key *wb;
1363
1364         prt_printf(buf, "transaction updates for %s journal seq %llu",
1365                trans->fn, trans->journal_res.seq);
1366         prt_newline(buf);
1367         printbuf_indent_add(buf, 2);
1368
1369         trans_for_each_update(trans, i) {
1370                 struct bkey_s_c old = { &i->old_k, i->old_v };
1371
1372                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1373                        bch2_btree_ids[i->btree_id],
1374                        i->cached,
1375                        (void *) i->ip_allocated);
1376                 prt_newline(buf);
1377
1378                 prt_printf(buf, "  old ");
1379                 bch2_bkey_val_to_text(buf, trans->c, old);
1380                 prt_newline(buf);
1381
1382                 prt_printf(buf, "  new ");
1383                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1384                 prt_newline(buf);
1385         }
1386
1387         trans_for_each_wb_update(trans, wb) {
1388                 prt_printf(buf, "update: btree=%s wb=1 %pS",
1389                        bch2_btree_ids[wb->btree],
1390                        (void *) i->ip_allocated);
1391                 prt_newline(buf);
1392
1393                 prt_printf(buf, "  new ");
1394                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(&wb->k));
1395                 prt_newline(buf);
1396         }
1397
1398         printbuf_indent_sub(buf, 2);
1399 }
1400
1401 noinline __cold
1402 void bch2_dump_trans_updates(struct btree_trans *trans)
1403 {
1404         struct printbuf buf = PRINTBUF;
1405
1406         bch2_trans_updates_to_text(&buf, trans);
1407         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1408         printbuf_exit(&buf);
1409 }
1410
1411 noinline __cold
1412 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1413 {
1414         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1415                    path->idx, path->ref, path->intent_ref,
1416                    path->preserve ? 'P' : ' ',
1417                    path->should_be_locked ? 'S' : ' ',
1418                    bch2_btree_ids[path->btree_id],
1419                    path->level);
1420         bch2_bpos_to_text(out, path->pos);
1421
1422         prt_printf(out, " locks %u", path->nodes_locked);
1423 #ifdef TRACK_PATH_ALLOCATED
1424         prt_printf(out, " %pS", (void *) path->ip_allocated);
1425 #endif
1426         prt_newline(out);
1427 }
1428
1429 static noinline __cold
1430 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1431                                 bool nosort)
1432 {
1433         struct btree_path *path;
1434         unsigned idx;
1435
1436         if (!nosort)
1437                 btree_trans_sort_paths(trans);
1438
1439         trans_for_each_path_inorder(trans, path, idx)
1440                 bch2_btree_path_to_text(out, path);
1441 }
1442
1443 noinline __cold
1444 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1445 {
1446         __bch2_trans_paths_to_text(out, trans, false);
1447 }
1448
1449 static noinline __cold
1450 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1451 {
1452         struct printbuf buf = PRINTBUF;
1453
1454         __bch2_trans_paths_to_text(&buf, trans, nosort);
1455         bch2_trans_updates_to_text(&buf, trans);
1456
1457         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1458         printbuf_exit(&buf);
1459 }
1460
1461 noinline __cold
1462 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1463 {
1464         __bch2_dump_trans_paths_updates(trans, false);
1465 }
1466
1467 noinline __cold
1468 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1469 {
1470         struct btree_transaction_stats *s = btree_trans_stats(trans);
1471         struct printbuf buf = PRINTBUF;
1472
1473         if (!s)
1474                 return;
1475
1476         bch2_trans_paths_to_text(&buf, trans);
1477
1478         if (!buf.allocation_failure) {
1479                 mutex_lock(&s->lock);
1480                 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1481                         s->nr_max_paths = trans->nr_max_paths =
1482                                 hweight64(trans->paths_allocated);
1483                         swap(s->max_paths_text, buf.buf);
1484                 }
1485                 mutex_unlock(&s->lock);
1486         }
1487
1488         printbuf_exit(&buf);
1489
1490         trans->nr_max_paths = hweight64(trans->paths_allocated);
1491 }
1492
1493 static noinline void btree_path_overflow(struct btree_trans *trans)
1494 {
1495         bch2_dump_trans_paths_updates(trans);
1496         panic("trans path oveflow\n");
1497 }
1498
1499 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1500                                                   struct btree_path *pos)
1501 {
1502         struct btree_path *path;
1503         unsigned idx;
1504
1505         if (unlikely(trans->paths_allocated ==
1506                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
1507                 btree_path_overflow(trans);
1508
1509         idx = __ffs64(~trans->paths_allocated);
1510
1511         /*
1512          * Do this before marking the new path as allocated, since it won't be
1513          * initialized yet:
1514          */
1515         if (unlikely(idx > trans->nr_max_paths))
1516                 bch2_trans_update_max_paths(trans);
1517
1518         trans->paths_allocated |= 1ULL << idx;
1519
1520         path = &trans->paths[idx];
1521         path->idx               = idx;
1522         path->ref               = 0;
1523         path->intent_ref        = 0;
1524         path->nodes_locked      = 0;
1525
1526         btree_path_list_add(trans, pos, path);
1527         trans->paths_sorted = false;
1528         return path;
1529 }
1530
1531 struct btree_path *bch2_path_get(struct btree_trans *trans,
1532                                  enum btree_id btree_id, struct bpos pos,
1533                                  unsigned locks_want, unsigned level,
1534                                  unsigned flags, unsigned long ip)
1535 {
1536         struct btree_path *path, *path_pos = NULL;
1537         bool cached = flags & BTREE_ITER_CACHED;
1538         bool intent = flags & BTREE_ITER_INTENT;
1539         int i;
1540
1541         bch2_trans_verify_not_in_restart(trans);
1542         bch2_trans_verify_locks(trans);
1543
1544         btree_trans_sort_paths(trans);
1545
1546         trans_for_each_path_inorder(trans, path, i) {
1547                 if (__btree_path_cmp(path,
1548                                      btree_id,
1549                                      cached,
1550                                      pos,
1551                                      level) > 0)
1552                         break;
1553
1554                 path_pos = path;
1555         }
1556
1557         if (path_pos &&
1558             path_pos->cached    == cached &&
1559             path_pos->btree_id  == btree_id &&
1560             path_pos->level     == level) {
1561                 __btree_path_get(path_pos, intent);
1562                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1563         } else {
1564                 path = btree_path_alloc(trans, path_pos);
1565                 path_pos = NULL;
1566
1567                 __btree_path_get(path, intent);
1568                 path->pos                       = pos;
1569                 path->btree_id                  = btree_id;
1570                 path->cached                    = cached;
1571                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1572                 path->should_be_locked          = false;
1573                 path->level                     = level;
1574                 path->locks_want                = locks_want;
1575                 path->nodes_locked              = 0;
1576                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1577                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1578 #ifdef TRACK_PATH_ALLOCATED
1579                 path->ip_allocated              = ip;
1580 #endif
1581                 trans->paths_sorted             = false;
1582         }
1583
1584         if (!(flags & BTREE_ITER_NOPRESERVE))
1585                 path->preserve = true;
1586
1587         if (path->intent_ref)
1588                 locks_want = max(locks_want, level + 1);
1589
1590         /*
1591          * If the path has locks_want greater than requested, we don't downgrade
1592          * it here - on transaction restart because btree node split needs to
1593          * upgrade locks, we might be putting/getting the iterator again.
1594          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1595          * a successful transaction commit.
1596          */
1597
1598         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1599         if (locks_want > path->locks_want)
1600                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want);
1601
1602         return path;
1603 }
1604
1605 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1606 {
1607
1608         struct btree_path_level *l = path_l(path);
1609         struct bkey_packed *_k;
1610         struct bkey_s_c k;
1611
1612         if (unlikely(!l->b))
1613                 return bkey_s_c_null;
1614
1615         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1616         EBUG_ON(!btree_node_locked(path, path->level));
1617
1618         if (!path->cached) {
1619                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1620                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1621
1622                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1623
1624                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1625                         goto hole;
1626         } else {
1627                 struct bkey_cached *ck = (void *) path->l[0].b;
1628
1629                 EBUG_ON(ck &&
1630                         (path->btree_id != ck->key.btree_id ||
1631                          !bkey_eq(path->pos, ck->key.pos)));
1632                 if (!ck || !ck->valid)
1633                         return bkey_s_c_null;
1634
1635                 *u = ck->k->k;
1636                 k = bkey_i_to_s_c(ck->k);
1637         }
1638
1639         return k;
1640 hole:
1641         bkey_init(u);
1642         u->p = path->pos;
1643         return (struct bkey_s_c) { u, NULL };
1644 }
1645
1646 /* Btree iterators: */
1647
1648 int __must_check
1649 __bch2_btree_iter_traverse(struct btree_iter *iter)
1650 {
1651         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1652 }
1653
1654 int __must_check
1655 bch2_btree_iter_traverse(struct btree_iter *iter)
1656 {
1657         int ret;
1658
1659         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1660                                         btree_iter_search_key(iter),
1661                                         iter->flags & BTREE_ITER_INTENT,
1662                                         btree_iter_ip_allocated(iter));
1663
1664         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1665         if (ret)
1666                 return ret;
1667
1668         btree_path_set_should_be_locked(iter->path);
1669         return 0;
1670 }
1671
1672 /* Iterate across nodes (leaf and interior nodes) */
1673
1674 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1675 {
1676         struct btree_trans *trans = iter->trans;
1677         struct btree *b = NULL;
1678         int ret;
1679
1680         EBUG_ON(iter->path->cached);
1681         bch2_btree_iter_verify(iter);
1682
1683         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1684         if (ret)
1685                 goto err;
1686
1687         b = btree_path_node(iter->path, iter->path->level);
1688         if (!b)
1689                 goto out;
1690
1691         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1692
1693         bkey_init(&iter->k);
1694         iter->k.p = iter->pos = b->key.k.p;
1695
1696         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1697                                         iter->flags & BTREE_ITER_INTENT,
1698                                         btree_iter_ip_allocated(iter));
1699         btree_path_set_should_be_locked(iter->path);
1700 out:
1701         bch2_btree_iter_verify_entry_exit(iter);
1702         bch2_btree_iter_verify(iter);
1703
1704         return b;
1705 err:
1706         b = ERR_PTR(ret);
1707         goto out;
1708 }
1709
1710 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1711 {
1712         struct btree *b;
1713
1714         while (b = bch2_btree_iter_peek_node(iter),
1715                bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1716                 bch2_trans_begin(iter->trans);
1717
1718         return b;
1719 }
1720
1721 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1722 {
1723         struct btree_trans *trans = iter->trans;
1724         struct btree_path *path = iter->path;
1725         struct btree *b = NULL;
1726         int ret;
1727
1728         bch2_trans_verify_not_in_restart(trans);
1729         EBUG_ON(iter->path->cached);
1730         bch2_btree_iter_verify(iter);
1731
1732         /* already at end? */
1733         if (!btree_path_node(path, path->level))
1734                 return NULL;
1735
1736         /* got to end? */
1737         if (!btree_path_node(path, path->level + 1)) {
1738                 btree_path_set_level_up(trans, path);
1739                 return NULL;
1740         }
1741
1742         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1743                 __bch2_btree_path_unlock(trans, path);
1744                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1745                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1746                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1747                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1748                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1749                 goto err;
1750         }
1751
1752         b = btree_path_node(path, path->level + 1);
1753
1754         if (bpos_eq(iter->pos, b->key.k.p)) {
1755                 __btree_path_set_level_up(trans, path, path->level++);
1756         } else {
1757                 /*
1758                  * Haven't gotten to the end of the parent node: go back down to
1759                  * the next child node
1760                  */
1761                 path = iter->path =
1762                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1763                                            iter->flags & BTREE_ITER_INTENT,
1764                                            btree_iter_ip_allocated(iter));
1765
1766                 btree_path_set_level_down(trans, path, iter->min_depth);
1767
1768                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1769                 if (ret)
1770                         goto err;
1771
1772                 b = path->l[path->level].b;
1773         }
1774
1775         bkey_init(&iter->k);
1776         iter->k.p = iter->pos = b->key.k.p;
1777
1778         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1779                                         iter->flags & BTREE_ITER_INTENT,
1780                                         btree_iter_ip_allocated(iter));
1781         btree_path_set_should_be_locked(iter->path);
1782         BUG_ON(iter->path->uptodate);
1783 out:
1784         bch2_btree_iter_verify_entry_exit(iter);
1785         bch2_btree_iter_verify(iter);
1786
1787         return b;
1788 err:
1789         b = ERR_PTR(ret);
1790         goto out;
1791 }
1792
1793 /* Iterate across keys (in leaf nodes only) */
1794
1795 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1796 {
1797         if (likely(!(iter->flags & BTREE_ITER_ALL_LEVELS))) {
1798                 struct bpos pos = iter->k.p;
1799                 bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1800                              ? bpos_eq(pos, SPOS_MAX)
1801                              : bkey_eq(pos, SPOS_MAX));
1802
1803                 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1804                         pos = bkey_successor(iter, pos);
1805                 bch2_btree_iter_set_pos(iter, pos);
1806                 return ret;
1807         } else {
1808                 if (!btree_path_node(iter->path, iter->path->level))
1809                         return true;
1810
1811                 iter->advanced = true;
1812                 return false;
1813         }
1814 }
1815
1816 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1817 {
1818         struct bpos pos = bkey_start_pos(&iter->k);
1819         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1820                      ? bpos_eq(pos, POS_MIN)
1821                      : bkey_eq(pos, POS_MIN));
1822
1823         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1824                 pos = bkey_predecessor(iter, pos);
1825         bch2_btree_iter_set_pos(iter, pos);
1826         return ret;
1827 }
1828
1829 static noinline
1830 struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
1831 {
1832         struct btree_insert_entry *i;
1833         struct bkey_i *ret = NULL;
1834
1835         trans_for_each_update(iter->trans, i) {
1836                 if (i->btree_id < iter->btree_id)
1837                         continue;
1838                 if (i->btree_id > iter->btree_id)
1839                         break;
1840                 if (bpos_lt(i->k->k.p, iter->path->pos))
1841                         continue;
1842                 if (i->key_cache_already_flushed)
1843                         continue;
1844                 if (!ret || bpos_lt(i->k->k.p, ret->k.p))
1845                         ret = i->k;
1846         }
1847
1848         return ret;
1849 }
1850
1851 static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter)
1852 {
1853         return iter->flags & BTREE_ITER_WITH_UPDATES
1854                 ? __bch2_btree_trans_peek_updates(iter)
1855                 : NULL;
1856 }
1857
1858 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1859                                               struct btree_iter *iter,
1860                                               struct bpos end_pos)
1861 {
1862         struct bkey_i *k;
1863
1864         if (bpos_lt(iter->path->pos, iter->journal_pos))
1865                 iter->journal_idx = 0;
1866
1867         k = bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1868                                         iter->path->level,
1869                                         iter->path->pos,
1870                                         end_pos,
1871                                         &iter->journal_idx);
1872
1873         iter->journal_pos = k ? k->k.p : end_pos;
1874         return k;
1875 }
1876
1877 static noinline
1878 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1879                                               struct btree_iter *iter)
1880 {
1881         struct bkey_i *k = bch2_btree_journal_peek(trans, iter, iter->path->pos);
1882
1883         if (k) {
1884                 iter->k = k->k;
1885                 return bkey_i_to_s_c(k);
1886         } else {
1887                 return bkey_s_c_null;
1888         }
1889 }
1890
1891 static noinline
1892 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1893                                          struct btree_iter *iter,
1894                                          struct bkey_s_c k)
1895 {
1896         struct bkey_i *next_journal =
1897                 bch2_btree_journal_peek(trans, iter,
1898                                 k.k ? k.k->p : path_l(iter->path)->b->key.k.p);
1899
1900         if (next_journal) {
1901                 iter->k = next_journal->k;
1902                 k = bkey_i_to_s_c(next_journal);
1903         }
1904
1905         return k;
1906 }
1907
1908 /*
1909  * Checks btree key cache for key at iter->pos and returns it if present, or
1910  * bkey_s_c_null:
1911  */
1912 static noinline
1913 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1914 {
1915         struct btree_trans *trans = iter->trans;
1916         struct bch_fs *c = trans->c;
1917         struct bkey u;
1918         struct bkey_s_c k;
1919         int ret;
1920
1921         if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1922             bpos_eq(iter->pos, pos))
1923                 return bkey_s_c_null;
1924
1925         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1926                 return bkey_s_c_null;
1927
1928         if (!iter->key_cache_path)
1929                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1930                                                      iter->flags & BTREE_ITER_INTENT, 0,
1931                                                      iter->flags|BTREE_ITER_CACHED|
1932                                                      BTREE_ITER_CACHED_NOFILL,
1933                                                      _THIS_IP_);
1934
1935         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1936                                         iter->flags & BTREE_ITER_INTENT,
1937                                         btree_iter_ip_allocated(iter));
1938
1939         ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
1940                                          iter->flags|BTREE_ITER_CACHED) ?:
1941                 bch2_btree_path_relock(trans, iter->path, _THIS_IP_);
1942         if (unlikely(ret))
1943                 return bkey_s_c_err(ret);
1944
1945         btree_path_set_should_be_locked(iter->key_cache_path);
1946
1947         k = bch2_btree_path_peek_slot(iter->key_cache_path, &u);
1948         if (k.k && !bkey_err(k)) {
1949                 iter->k = u;
1950                 k.k = &iter->k;
1951         }
1952         return k;
1953 }
1954
1955 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1956 {
1957         struct btree_trans *trans = iter->trans;
1958         struct bkey_i *next_update;
1959         struct bkey_s_c k, k2;
1960         int ret;
1961
1962         EBUG_ON(iter->path->cached);
1963         bch2_btree_iter_verify(iter);
1964
1965         while (1) {
1966                 struct btree_path_level *l;
1967
1968                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
1969                                         iter->flags & BTREE_ITER_INTENT,
1970                                         btree_iter_ip_allocated(iter));
1971
1972                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1973                 if (unlikely(ret)) {
1974                         /* ensure that iter->k is consistent with iter->pos: */
1975                         bch2_btree_iter_set_pos(iter, iter->pos);
1976                         k = bkey_s_c_err(ret);
1977                         goto out;
1978                 }
1979
1980                 l = path_l(iter->path);
1981
1982                 if (unlikely(!l->b)) {
1983                         /* No btree nodes at requested level: */
1984                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
1985                         k = bkey_s_c_null;
1986                         goto out;
1987                 }
1988
1989                 btree_path_set_should_be_locked(iter->path);
1990
1991                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
1992
1993                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
1994                     k.k &&
1995                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
1996                         k = k2;
1997                         ret = bkey_err(k);
1998                         if (ret) {
1999                                 bch2_btree_iter_set_pos(iter, iter->pos);
2000                                 goto out;
2001                         }
2002                 }
2003
2004                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2005                         k = btree_trans_peek_journal(trans, iter, k);
2006
2007                 next_update = btree_trans_peek_updates(iter);
2008
2009                 if (next_update &&
2010                     bpos_le(next_update->k.p,
2011                             k.k ? k.k->p : l->b->key.k.p)) {
2012                         iter->k = next_update->k;
2013                         k = bkey_i_to_s_c(next_update);
2014                 }
2015
2016                 if (k.k && bkey_deleted(k.k)) {
2017                         /*
2018                          * If we've got a whiteout, and it's after the search
2019                          * key, advance the search key to the whiteout instead
2020                          * of just after the whiteout - it might be a btree
2021                          * whiteout, with a real key at the same position, since
2022                          * in the btree deleted keys sort before non deleted.
2023                          */
2024                         search_key = !bpos_eq(search_key, k.k->p)
2025                                 ? k.k->p
2026                                 : bpos_successor(k.k->p);
2027                         continue;
2028                 }
2029
2030                 if (likely(k.k)) {
2031                         break;
2032                 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2033                         /* Advance to next leaf node: */
2034                         search_key = bpos_successor(l->b->key.k.p);
2035                 } else {
2036                         /* End of btree: */
2037                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2038                         k = bkey_s_c_null;
2039                         goto out;
2040                 }
2041         }
2042 out:
2043         bch2_btree_iter_verify(iter);
2044
2045         return k;
2046 }
2047
2048 /**
2049  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
2050  * current position
2051  */
2052 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2053 {
2054         struct btree_trans *trans = iter->trans;
2055         struct bpos search_key = btree_iter_search_key(iter);
2056         struct bkey_s_c k;
2057         struct bpos iter_pos;
2058         int ret;
2059
2060         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2061         EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2062
2063         if (iter->update_path) {
2064                 bch2_path_put_nokeep(trans, iter->update_path,
2065                                      iter->flags & BTREE_ITER_INTENT);
2066                 iter->update_path = NULL;
2067         }
2068
2069         bch2_btree_iter_verify_entry_exit(iter);
2070
2071         while (1) {
2072                 k = __bch2_btree_iter_peek(iter, search_key);
2073                 if (unlikely(!k.k))
2074                         goto end;
2075                 if (unlikely(bkey_err(k)))
2076                         goto out_no_locked;
2077
2078                 /*
2079                  * iter->pos should be mononotically increasing, and always be
2080                  * equal to the key we just returned - except extents can
2081                  * straddle iter->pos:
2082                  */
2083                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2084                         iter_pos = k.k->p;
2085                 else
2086                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2087
2088                 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2089                              ? bkey_gt(iter_pos, end)
2090                              : bkey_ge(iter_pos, end)))
2091                         goto end;
2092
2093                 if (iter->update_path &&
2094                     !bkey_eq(iter->update_path->pos, k.k->p)) {
2095                         bch2_path_put_nokeep(trans, iter->update_path,
2096                                              iter->flags & BTREE_ITER_INTENT);
2097                         iter->update_path = NULL;
2098                 }
2099
2100                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2101                     (iter->flags & BTREE_ITER_INTENT) &&
2102                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2103                     !iter->update_path) {
2104                         struct bpos pos = k.k->p;
2105
2106                         if (pos.snapshot < iter->snapshot) {
2107                                 search_key = bpos_successor(k.k->p);
2108                                 continue;
2109                         }
2110
2111                         pos.snapshot = iter->snapshot;
2112
2113                         /*
2114                          * advance, same as on exit for iter->path, but only up
2115                          * to snapshot
2116                          */
2117                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2118                         iter->update_path = iter->path;
2119
2120                         iter->update_path = bch2_btree_path_set_pos(trans,
2121                                                 iter->update_path, pos,
2122                                                 iter->flags & BTREE_ITER_INTENT,
2123                                                 _THIS_IP_);
2124                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2125                         if (unlikely(ret)) {
2126                                 k = bkey_s_c_err(ret);
2127                                 goto out_no_locked;
2128                         }
2129                 }
2130
2131                 /*
2132                  * We can never have a key in a leaf node at POS_MAX, so
2133                  * we don't have to check these successor() calls:
2134                  */
2135                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2136                     !bch2_snapshot_is_ancestor(trans->c,
2137                                                iter->snapshot,
2138                                                k.k->p.snapshot)) {
2139                         search_key = bpos_successor(k.k->p);
2140                         continue;
2141                 }
2142
2143                 if (bkey_whiteout(k.k) &&
2144                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2145                         search_key = bkey_successor(iter, k.k->p);
2146                         continue;
2147                 }
2148
2149                 break;
2150         }
2151
2152         iter->pos = iter_pos;
2153
2154         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2155                                 iter->flags & BTREE_ITER_INTENT,
2156                                 btree_iter_ip_allocated(iter));
2157
2158         btree_path_set_should_be_locked(iter->path);
2159 out_no_locked:
2160         if (iter->update_path) {
2161                 ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_);
2162                 if (unlikely(ret))
2163                         k = bkey_s_c_err(ret);
2164                 else
2165                         btree_path_set_should_be_locked(iter->update_path);
2166         }
2167
2168         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2169                 iter->pos.snapshot = iter->snapshot;
2170
2171         ret = bch2_btree_iter_verify_ret(iter, k);
2172         if (unlikely(ret)) {
2173                 bch2_btree_iter_set_pos(iter, iter->pos);
2174                 k = bkey_s_c_err(ret);
2175         }
2176
2177         bch2_btree_iter_verify_entry_exit(iter);
2178
2179         return k;
2180 end:
2181         bch2_btree_iter_set_pos(iter, end);
2182         k = bkey_s_c_null;
2183         goto out_no_locked;
2184 }
2185
2186 /**
2187  * bch2_btree_iter_peek_all_levels: returns the first key greater than or equal
2188  * to iterator's current position, returning keys from every level of the btree.
2189  * For keys at different levels of the btree that compare equal, the key from
2190  * the lower level (leaf) is returned first.
2191  */
2192 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *iter)
2193 {
2194         struct btree_trans *trans = iter->trans;
2195         struct bkey_s_c k;
2196         int ret;
2197
2198         EBUG_ON(iter->path->cached);
2199         bch2_btree_iter_verify(iter);
2200         BUG_ON(iter->path->level < iter->min_depth);
2201         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
2202         EBUG_ON(!(iter->flags & BTREE_ITER_ALL_LEVELS));
2203
2204         while (1) {
2205                 iter->path = bch2_btree_path_set_pos(trans, iter->path, iter->pos,
2206                                         iter->flags & BTREE_ITER_INTENT,
2207                                         btree_iter_ip_allocated(iter));
2208
2209                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2210                 if (unlikely(ret)) {
2211                         /* ensure that iter->k is consistent with iter->pos: */
2212                         bch2_btree_iter_set_pos(iter, iter->pos);
2213                         k = bkey_s_c_err(ret);
2214                         goto out_no_locked;
2215                 }
2216
2217                 /* Already at end? */
2218                 if (!btree_path_node(iter->path, iter->path->level)) {
2219                         k = bkey_s_c_null;
2220                         goto out_no_locked;
2221                 }
2222
2223                 k = btree_path_level_peek_all(trans->c,
2224                                 &iter->path->l[iter->path->level], &iter->k);
2225
2226                 /* Check if we should go up to the parent node: */
2227                 if (!k.k ||
2228                     (iter->advanced &&
2229                      bpos_eq(path_l(iter->path)->b->key.k.p, iter->pos))) {
2230                         iter->pos = path_l(iter->path)->b->key.k.p;
2231                         btree_path_set_level_up(trans, iter->path);
2232                         iter->advanced = false;
2233                         continue;
2234                 }
2235
2236                 /*
2237                  * Check if we should go back down to a leaf:
2238                  * If we're not in a leaf node, we only return the current key
2239                  * if it exactly matches iter->pos - otherwise we first have to
2240                  * go back to the leaf:
2241                  */
2242                 if (iter->path->level != iter->min_depth &&
2243                     (iter->advanced ||
2244                      !k.k ||
2245                      !bpos_eq(iter->pos, k.k->p))) {
2246                         btree_path_set_level_down(trans, iter->path, iter->min_depth);
2247                         iter->pos = bpos_successor(iter->pos);
2248                         iter->advanced = false;
2249                         continue;
2250                 }
2251
2252                 /* Check if we should go to the next key: */
2253                 if (iter->path->level == iter->min_depth &&
2254                     iter->advanced &&
2255                     k.k &&
2256                     bpos_eq(iter->pos, k.k->p)) {
2257                         iter->pos = bpos_successor(iter->pos);
2258                         iter->advanced = false;
2259                         continue;
2260                 }
2261
2262                 if (iter->advanced &&
2263                     iter->path->level == iter->min_depth &&
2264                     !bpos_eq(k.k->p, iter->pos))
2265                         iter->advanced = false;
2266
2267                 BUG_ON(iter->advanced);
2268                 BUG_ON(!k.k);
2269                 break;
2270         }
2271
2272         iter->pos = k.k->p;
2273         btree_path_set_should_be_locked(iter->path);
2274 out_no_locked:
2275         bch2_btree_iter_verify(iter);
2276
2277         return k;
2278 }
2279
2280 /**
2281  * bch2_btree_iter_next: returns first key greater than iterator's current
2282  * position
2283  */
2284 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2285 {
2286         if (!bch2_btree_iter_advance(iter))
2287                 return bkey_s_c_null;
2288
2289         return bch2_btree_iter_peek(iter);
2290 }
2291
2292 /**
2293  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2294  * iterator's current position
2295  */
2296 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2297 {
2298         struct btree_trans *trans = iter->trans;
2299         struct bpos search_key = iter->pos;
2300         struct btree_path *saved_path = NULL;
2301         struct bkey_s_c k;
2302         struct bkey saved_k;
2303         const struct bch_val *saved_v;
2304         int ret;
2305
2306         EBUG_ON(iter->path->cached || iter->path->level);
2307         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2308
2309         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2310                 return bkey_s_c_err(-EIO);
2311
2312         bch2_btree_iter_verify(iter);
2313         bch2_btree_iter_verify_entry_exit(iter);
2314
2315         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2316                 search_key.snapshot = U32_MAX;
2317
2318         while (1) {
2319                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2320                                                 iter->flags & BTREE_ITER_INTENT,
2321                                                 btree_iter_ip_allocated(iter));
2322
2323                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2324                 if (unlikely(ret)) {
2325                         /* ensure that iter->k is consistent with iter->pos: */
2326                         bch2_btree_iter_set_pos(iter, iter->pos);
2327                         k = bkey_s_c_err(ret);
2328                         goto out_no_locked;
2329                 }
2330
2331                 k = btree_path_level_peek(trans, iter->path,
2332                                           &iter->path->l[0], &iter->k);
2333                 if (!k.k ||
2334                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2335                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2336                      : bpos_gt(k.k->p, search_key)))
2337                         k = btree_path_level_prev(trans, iter->path,
2338                                                   &iter->path->l[0], &iter->k);
2339
2340                 if (likely(k.k)) {
2341                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2342                                 if (k.k->p.snapshot == iter->snapshot)
2343                                         goto got_key;
2344
2345                                 /*
2346                                  * If we have a saved candidate, and we're no
2347                                  * longer at the same _key_ (not pos), return
2348                                  * that candidate
2349                                  */
2350                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2351                                         bch2_path_put_nokeep(trans, iter->path,
2352                                                       iter->flags & BTREE_ITER_INTENT);
2353                                         iter->path = saved_path;
2354                                         saved_path = NULL;
2355                                         iter->k = saved_k;
2356                                         k.v     = saved_v;
2357                                         goto got_key;
2358                                 }
2359
2360                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2361                                                               iter->snapshot,
2362                                                               k.k->p.snapshot)) {
2363                                         if (saved_path)
2364                                                 bch2_path_put_nokeep(trans, saved_path,
2365                                                       iter->flags & BTREE_ITER_INTENT);
2366                                         saved_path = btree_path_clone(trans, iter->path,
2367                                                                 iter->flags & BTREE_ITER_INTENT);
2368                                         saved_k = *k.k;
2369                                         saved_v = k.v;
2370                                 }
2371
2372                                 search_key = bpos_predecessor(k.k->p);
2373                                 continue;
2374                         }
2375 got_key:
2376                         if (bkey_whiteout(k.k) &&
2377                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2378                                 search_key = bkey_predecessor(iter, k.k->p);
2379                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2380                                         search_key.snapshot = U32_MAX;
2381                                 continue;
2382                         }
2383
2384                         break;
2385                 } else if (likely(!bpos_eq(iter->path->l[0].b->data->min_key, POS_MIN))) {
2386                         /* Advance to previous leaf node: */
2387                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2388                 } else {
2389                         /* Start of btree: */
2390                         bch2_btree_iter_set_pos(iter, POS_MIN);
2391                         k = bkey_s_c_null;
2392                         goto out_no_locked;
2393                 }
2394         }
2395
2396         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2397
2398         /* Extents can straddle iter->pos: */
2399         if (bkey_lt(k.k->p, iter->pos))
2400                 iter->pos = k.k->p;
2401
2402         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2403                 iter->pos.snapshot = iter->snapshot;
2404
2405         btree_path_set_should_be_locked(iter->path);
2406 out_no_locked:
2407         if (saved_path)
2408                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2409
2410         bch2_btree_iter_verify_entry_exit(iter);
2411         bch2_btree_iter_verify(iter);
2412
2413         return k;
2414 }
2415
2416 /**
2417  * bch2_btree_iter_prev: returns first key less than iterator's current
2418  * position
2419  */
2420 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2421 {
2422         if (!bch2_btree_iter_rewind(iter))
2423                 return bkey_s_c_null;
2424
2425         return bch2_btree_iter_peek_prev(iter);
2426 }
2427
2428 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2429 {
2430         struct btree_trans *trans = iter->trans;
2431         struct bpos search_key;
2432         struct bkey_s_c k;
2433         int ret;
2434
2435         bch2_btree_iter_verify(iter);
2436         bch2_btree_iter_verify_entry_exit(iter);
2437         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2438         EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2439
2440         /* extents can't span inode numbers: */
2441         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2442             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2443                 if (iter->pos.inode == KEY_INODE_MAX)
2444                         return bkey_s_c_null;
2445
2446                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2447         }
2448
2449         search_key = btree_iter_search_key(iter);
2450         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2451                                         iter->flags & BTREE_ITER_INTENT,
2452                                         btree_iter_ip_allocated(iter));
2453
2454         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2455         if (unlikely(ret)) {
2456                 k = bkey_s_c_err(ret);
2457                 goto out_no_locked;
2458         }
2459
2460         if ((iter->flags & BTREE_ITER_CACHED) ||
2461             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2462                 struct bkey_i *next_update;
2463
2464                 if ((next_update = btree_trans_peek_updates(iter)) &&
2465                     bpos_eq(next_update->k.p, iter->pos)) {
2466                         iter->k = next_update->k;
2467                         k = bkey_i_to_s_c(next_update);
2468                         goto out;
2469                 }
2470
2471                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2472                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2473                         goto out;
2474
2475                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2476                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2477                         if (!bkey_err(k))
2478                                 iter->k = *k.k;
2479                         /* We're not returning a key from iter->path: */
2480                         goto out_no_locked;
2481                 }
2482
2483                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2484                 if (unlikely(!k.k))
2485                         goto out_no_locked;
2486         } else {
2487                 struct bpos next;
2488                 struct bpos end = iter->pos;
2489
2490                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2491                         end.offset = U64_MAX;
2492
2493                 EBUG_ON(iter->path->level);
2494
2495                 if (iter->flags & BTREE_ITER_INTENT) {
2496                         struct btree_iter iter2;
2497
2498                         bch2_trans_copy_iter(&iter2, iter);
2499                         k = bch2_btree_iter_peek_upto(&iter2, end);
2500
2501                         if (k.k && !bkey_err(k)) {
2502                                 iter->k = iter2.k;
2503                                 k.k = &iter->k;
2504                         }
2505                         bch2_trans_iter_exit(trans, &iter2);
2506                 } else {
2507                         struct bpos pos = iter->pos;
2508
2509                         k = bch2_btree_iter_peek_upto(iter, end);
2510                         if (unlikely(bkey_err(k)))
2511                                 bch2_btree_iter_set_pos(iter, pos);
2512                         else
2513                                 iter->pos = pos;
2514                 }
2515
2516                 if (unlikely(bkey_err(k)))
2517                         goto out_no_locked;
2518
2519                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2520
2521                 if (bkey_lt(iter->pos, next)) {
2522                         bkey_init(&iter->k);
2523                         iter->k.p = iter->pos;
2524
2525                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2526                                 bch2_key_resize(&iter->k,
2527                                                 min_t(u64, KEY_SIZE_MAX,
2528                                                       (next.inode == iter->pos.inode
2529                                                        ? next.offset
2530                                                        : KEY_OFFSET_MAX) -
2531                                                       iter->pos.offset));
2532                                 EBUG_ON(!iter->k.size);
2533                         }
2534
2535                         k = (struct bkey_s_c) { &iter->k, NULL };
2536                 }
2537         }
2538 out:
2539         btree_path_set_should_be_locked(iter->path);
2540 out_no_locked:
2541         bch2_btree_iter_verify_entry_exit(iter);
2542         bch2_btree_iter_verify(iter);
2543         ret = bch2_btree_iter_verify_ret(iter, k);
2544         if (unlikely(ret))
2545                 return bkey_s_c_err(ret);
2546
2547         return k;
2548 }
2549
2550 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2551 {
2552         if (!bch2_btree_iter_advance(iter))
2553                 return bkey_s_c_null;
2554
2555         return bch2_btree_iter_peek_slot(iter);
2556 }
2557
2558 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2559 {
2560         if (!bch2_btree_iter_rewind(iter))
2561                 return bkey_s_c_null;
2562
2563         return bch2_btree_iter_peek_slot(iter);
2564 }
2565
2566 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2567 {
2568         struct bkey_s_c k;
2569
2570         while (btree_trans_too_many_iters(iter->trans) ||
2571                (k = bch2_btree_iter_peek_type(iter, iter->flags),
2572                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2573                 bch2_trans_begin(iter->trans);
2574
2575         return k;
2576 }
2577
2578 /* new transactional stuff: */
2579
2580 #ifdef CONFIG_BCACHEFS_DEBUG
2581 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2582 {
2583         struct btree_path *path;
2584         unsigned i;
2585
2586         BUG_ON(trans->nr_sorted != hweight64(trans->paths_allocated));
2587
2588         trans_for_each_path(trans, path) {
2589                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2590                 BUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2591         }
2592
2593         for (i = 0; i < trans->nr_sorted; i++) {
2594                 unsigned idx = trans->sorted[i];
2595
2596                 EBUG_ON(!(trans->paths_allocated & (1ULL << idx)));
2597                 BUG_ON(trans->paths[idx].sorted_idx != i);
2598         }
2599 }
2600
2601 static void btree_trans_verify_sorted(struct btree_trans *trans)
2602 {
2603         struct btree_path *path, *prev = NULL;
2604         unsigned i;
2605
2606         if (!bch2_debug_check_iterators)
2607                 return;
2608
2609         trans_for_each_path_inorder(trans, path, i) {
2610                 if (prev && btree_path_cmp(prev, path) > 0) {
2611                         __bch2_dump_trans_paths_updates(trans, true);
2612                         panic("trans paths out of order!\n");
2613                 }
2614                 prev = path;
2615         }
2616 }
2617 #else
2618 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2619 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2620 #endif
2621
2622 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2623 {
2624         int i, l = 0, r = trans->nr_sorted, inc = 1;
2625         bool swapped;
2626
2627         btree_trans_verify_sorted_refs(trans);
2628
2629         if (trans->paths_sorted)
2630                 goto out;
2631
2632         /*
2633          * Cocktail shaker sort: this is efficient because iterators will be
2634          * mostly sorted.
2635          */
2636         do {
2637                 swapped = false;
2638
2639                 for (i = inc > 0 ? l : r - 2;
2640                      i + 1 < r && i >= l;
2641                      i += inc) {
2642                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2643                                            trans->paths + trans->sorted[i + 1]) > 0) {
2644                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2645                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2646                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2647                                 swapped = true;
2648                         }
2649                 }
2650
2651                 if (inc > 0)
2652                         --r;
2653                 else
2654                         l++;
2655                 inc = -inc;
2656         } while (swapped);
2657
2658         trans->paths_sorted = true;
2659 out:
2660         btree_trans_verify_sorted(trans);
2661 }
2662
2663 static inline void btree_path_list_remove(struct btree_trans *trans,
2664                                           struct btree_path *path)
2665 {
2666         unsigned i;
2667
2668         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2669 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2670         trans->nr_sorted--;
2671         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2672                                 trans->sorted + path->sorted_idx + 1,
2673                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2674 #else
2675         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2676 #endif
2677         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2678                 trans->paths[trans->sorted[i]].sorted_idx = i;
2679
2680         path->sorted_idx = U8_MAX;
2681 }
2682
2683 static inline void btree_path_list_add(struct btree_trans *trans,
2684                                        struct btree_path *pos,
2685                                        struct btree_path *path)
2686 {
2687         unsigned i;
2688
2689         path->sorted_idx = pos ? pos->sorted_idx + 1 : trans->nr_sorted;
2690
2691 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2692         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2693                               trans->sorted + path->sorted_idx,
2694                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2695         trans->nr_sorted++;
2696         trans->sorted[path->sorted_idx] = path->idx;
2697 #else
2698         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2699 #endif
2700
2701         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2702                 trans->paths[trans->sorted[i]].sorted_idx = i;
2703
2704         btree_trans_verify_sorted_refs(trans);
2705 }
2706
2707 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2708 {
2709         if (iter->update_path)
2710                 bch2_path_put_nokeep(trans, iter->update_path,
2711                               iter->flags & BTREE_ITER_INTENT);
2712         if (iter->path)
2713                 bch2_path_put(trans, iter->path,
2714                               iter->flags & BTREE_ITER_INTENT);
2715         if (iter->key_cache_path)
2716                 bch2_path_put(trans, iter->key_cache_path,
2717                               iter->flags & BTREE_ITER_INTENT);
2718         iter->path = NULL;
2719         iter->update_path = NULL;
2720         iter->key_cache_path = NULL;
2721 }
2722
2723 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2724                           struct btree_iter *iter,
2725                           unsigned btree_id, struct bpos pos,
2726                           unsigned flags)
2727 {
2728         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2729                                bch2_btree_iter_flags(trans, btree_id, flags),
2730                                _RET_IP_);
2731 }
2732
2733 void bch2_trans_node_iter_init(struct btree_trans *trans,
2734                                struct btree_iter *iter,
2735                                enum btree_id btree_id,
2736                                struct bpos pos,
2737                                unsigned locks_want,
2738                                unsigned depth,
2739                                unsigned flags)
2740 {
2741        flags |= BTREE_ITER_NOT_EXTENTS;
2742        flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2743        flags |= BTREE_ITER_ALL_SNAPSHOTS;
2744
2745         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2746                                __bch2_btree_iter_flags(trans, btree_id, flags),
2747                                _RET_IP_);
2748
2749         iter->min_depth = depth;
2750
2751         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2752         BUG_ON(iter->path->level        != depth);
2753         BUG_ON(iter->min_depth          != depth);
2754 }
2755
2756 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2757 {
2758         *dst = *src;
2759         if (src->path)
2760                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2761         if (src->update_path)
2762                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2763         dst->key_cache_path = NULL;
2764 }
2765
2766 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2767 {
2768         unsigned new_top = trans->mem_top + size;
2769         size_t old_bytes = trans->mem_bytes;
2770         size_t new_bytes = roundup_pow_of_two(new_top);
2771         int ret;
2772         void *new_mem;
2773         void *p;
2774
2775         trans->mem_max = max(trans->mem_max, new_top);
2776
2777         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2778
2779         new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2780         if (unlikely(!new_mem)) {
2781                 bch2_trans_unlock(trans);
2782
2783                 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2784                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2785                         new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2786                         new_bytes = BTREE_TRANS_MEM_MAX;
2787                         kfree(trans->mem);
2788                 }
2789
2790                 if (!new_mem)
2791                         return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2792
2793                 trans->mem = new_mem;
2794                 trans->mem_bytes = new_bytes;
2795
2796                 ret = bch2_trans_relock(trans);
2797                 if (ret)
2798                         return ERR_PTR(ret);
2799         }
2800
2801         trans->mem = new_mem;
2802         trans->mem_bytes = new_bytes;
2803
2804         if (old_bytes) {
2805                 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2806                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2807         }
2808
2809         p = trans->mem + trans->mem_top;
2810         trans->mem_top += size;
2811         memset(p, 0, size);
2812         return p;
2813 }
2814
2815 static noinline void bch2_trans_reset_srcu_lock(struct btree_trans *trans)
2816 {
2817         struct bch_fs *c = trans->c;
2818         struct btree_path *path;
2819
2820         trans_for_each_path(trans, path)
2821                 if (path->cached && !btree_node_locked(path, 0))
2822                         path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2823
2824         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2825         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2826         trans->srcu_lock_time   = jiffies;
2827 }
2828
2829 /**
2830  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2831  * @trans: transaction to reset
2832  *
2833  * While iterating over nodes or updating nodes a attempt to lock a btree node
2834  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2835  * occurs bch2_trans_begin() should be called and the transaction retried.
2836  */
2837 u32 bch2_trans_begin(struct btree_trans *trans)
2838 {
2839         struct btree_path *path;
2840         u64 now;
2841
2842         bch2_trans_reset_updates(trans);
2843
2844         trans->restart_count++;
2845         trans->mem_top                  = 0;
2846
2847         trans_for_each_path(trans, path) {
2848                 path->should_be_locked = false;
2849
2850                 /*
2851                  * If the transaction wasn't restarted, we're presuming to be
2852                  * doing something new: dont keep iterators excpt the ones that
2853                  * are in use - except for the subvolumes btree:
2854                  */
2855                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2856                         path->preserve = false;
2857
2858                 /*
2859                  * XXX: we probably shouldn't be doing this if the transaction
2860                  * was restarted, but currently we still overflow transaction
2861                  * iterators if we do that
2862                  */
2863                 if (!path->ref && !path->preserve)
2864                         __bch2_path_free(trans, path);
2865                 else
2866                         path->preserve = false;
2867         }
2868
2869         now = local_clock();
2870         if (!trans->restarted &&
2871             (need_resched() ||
2872              now - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2873                 drop_locks_do(trans, (cond_resched(), 0));
2874                 now = local_clock();
2875         }
2876         trans->last_begin_time = now;
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         return trans->restart_count;
2888 }
2889
2890 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
2891 {
2892         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
2893         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2894         void *p = NULL;
2895
2896         BUG_ON(trans->used_mempool);
2897
2898 #ifdef __KERNEL__
2899         p = this_cpu_xchg(c->btree_paths_bufs->path, NULL);
2900 #endif
2901         if (!p)
2902                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
2903         /*
2904          * paths need to be zeroed, bch2_check_for_deadlock looks at paths in
2905          * other threads
2906          */
2907
2908         trans->paths            = p; p += paths_bytes;
2909         trans->updates          = p; p += updates_bytes;
2910 }
2911
2912 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2913
2914 unsigned bch2_trans_get_fn_idx(const char *fn)
2915 {
2916         unsigned i;
2917
2918         for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2919                 if (!bch2_btree_transaction_fns[i] ||
2920                     bch2_btree_transaction_fns[i] == fn) {
2921                         bch2_btree_transaction_fns[i] = fn;
2922                         return i;
2923                 }
2924
2925         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2926         return i;
2927 }
2928
2929 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, unsigned fn_idx)
2930         __acquires(&c->btree_trans_barrier)
2931 {
2932         struct btree_transaction_stats *s;
2933
2934         bch2_assert_btree_nodes_not_locked();
2935
2936         memset(trans, 0, sizeof(*trans));
2937         trans->c                = c;
2938         trans->fn               = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2939                 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2940         trans->last_begin_time  = local_clock();
2941         trans->fn_idx           = fn_idx;
2942         trans->locking_wait.task = current;
2943         trans->journal_replay_not_finished =
2944                 !test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
2945         closure_init_stack(&trans->ref);
2946
2947         bch2_trans_alloc_paths(trans, c);
2948
2949         s = btree_trans_stats(trans);
2950         if (s && s->max_mem) {
2951                 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2952
2953                 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2954
2955                 if (!unlikely(trans->mem)) {
2956                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2957                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2958                 } else {
2959                         trans->mem_bytes = expected_mem_bytes;
2960                 }
2961         }
2962
2963         if (s) {
2964                 trans->nr_max_paths = s->nr_max_paths;
2965                 trans->wb_updates_size = s->wb_updates_size;
2966         }
2967
2968         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2969         trans->srcu_lock_time   = jiffies;
2970
2971         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
2972                 struct btree_trans *pos;
2973
2974                 seqmutex_lock(&c->btree_trans_lock);
2975                 list_for_each_entry(pos, &c->btree_trans_list, list) {
2976                         /*
2977                          * We'd much prefer to be stricter here and completely
2978                          * disallow multiple btree_trans in the same thread -
2979                          * but the data move path calls bch2_write when we
2980                          * already have a btree_trans initialized.
2981                          */
2982                         BUG_ON(trans->locking_wait.task->pid == pos->locking_wait.task->pid &&
2983                                bch2_trans_locked(pos));
2984
2985                         if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
2986                                 list_add_tail(&trans->list, &pos->list);
2987                                 goto list_add_done;
2988                         }
2989                 }
2990                 list_add_tail(&trans->list, &c->btree_trans_list);
2991 list_add_done:
2992                 seqmutex_unlock(&c->btree_trans_lock);
2993         }
2994 }
2995
2996 static void check_btree_paths_leaked(struct btree_trans *trans)
2997 {
2998 #ifdef CONFIG_BCACHEFS_DEBUG
2999         struct bch_fs *c = trans->c;
3000         struct btree_path *path;
3001
3002         trans_for_each_path(trans, path)
3003                 if (path->ref)
3004                         goto leaked;
3005         return;
3006 leaked:
3007         bch_err(c, "btree paths leaked from %s!", trans->fn);
3008         trans_for_each_path(trans, path)
3009                 if (path->ref)
3010                         printk(KERN_ERR "  btree %s %pS\n",
3011                                bch2_btree_ids[path->btree_id],
3012                                (void *) path->ip_allocated);
3013         /* Be noisy about this: */
3014         bch2_fatal_error(c);
3015 #endif
3016 }
3017
3018 void bch2_trans_exit(struct btree_trans *trans)
3019         __releases(&c->btree_trans_barrier)
3020 {
3021         struct btree_insert_entry *i;
3022         struct bch_fs *c = trans->c;
3023         struct btree_transaction_stats *s = btree_trans_stats(trans);
3024
3025         bch2_trans_unlock(trans);
3026
3027         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG_TRANSACTIONS)) {
3028                 seqmutex_lock(&c->btree_trans_lock);
3029                 list_del(&trans->list);
3030                 seqmutex_unlock(&c->btree_trans_lock);
3031         }
3032
3033         closure_sync(&trans->ref);
3034
3035         if (s)
3036                 s->max_mem = max(s->max_mem, trans->mem_max);
3037
3038         trans_for_each_update(trans, i)
3039                 __btree_path_put(i->path, true);
3040         trans->nr_updates               = 0;
3041
3042         check_btree_paths_leaked(trans);
3043
3044         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3045
3046         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
3047
3048         kfree(trans->extra_journal_entries.data);
3049
3050         if (trans->fs_usage_deltas) {
3051                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3052                     REPLICAS_DELTA_LIST_MAX)
3053                         mempool_free(trans->fs_usage_deltas,
3054                                      &c->replicas_delta_pool);
3055                 else
3056                         kfree(trans->fs_usage_deltas);
3057         }
3058
3059         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3060                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3061         else
3062                 kfree(trans->mem);
3063
3064 #ifdef __KERNEL__
3065         /*
3066          * Userspace doesn't have a real percpu implementation:
3067          */
3068         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3069 #endif
3070
3071         if (trans->paths)
3072                 mempool_free(trans->paths, &c->btree_paths_pool);
3073
3074         trans->mem      = (void *) 0x1;
3075         trans->paths    = (void *) 0x1;
3076 }
3077
3078 static void __maybe_unused
3079 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3080                                       struct btree_bkey_cached_common *b)
3081 {
3082         struct six_lock_count c = six_lock_counts(&b->lock);
3083         struct task_struct *owner;
3084         pid_t pid;
3085
3086         rcu_read_lock();
3087         owner = READ_ONCE(b->lock.owner);
3088         pid = owner ? owner->pid : 0;
3089         rcu_read_unlock();
3090
3091         prt_tab(out);
3092         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3093                    b->level, bch2_btree_ids[b->btree_id]);
3094         bch2_bpos_to_text(out, btree_node_pos(b));
3095
3096         prt_tab(out);
3097         prt_printf(out, " locks %u:%u:%u held by pid %u",
3098                    c.n[0], c.n[1], c.n[2], pid);
3099 }
3100
3101 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3102 {
3103         struct btree_path *path;
3104         struct btree_bkey_cached_common *b;
3105         static char lock_types[] = { 'r', 'i', 'w' };
3106         unsigned l, idx;
3107
3108         if (!out->nr_tabstops) {
3109                 printbuf_tabstop_push(out, 16);
3110                 printbuf_tabstop_push(out, 32);
3111         }
3112
3113         prt_printf(out, "%i %s\n", trans->locking_wait.task->pid, trans->fn);
3114
3115         trans_for_each_path_safe(trans, path, idx) {
3116                 if (!path->nodes_locked)
3117                         continue;
3118
3119                 prt_printf(out, "  path %u %c l=%u %s:",
3120                        path->idx,
3121                        path->cached ? 'c' : 'b',
3122                        path->level,
3123                        bch2_btree_ids[path->btree_id]);
3124                 bch2_bpos_to_text(out, path->pos);
3125                 prt_newline(out);
3126
3127                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3128                         if (btree_node_locked(path, l) &&
3129                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3130                                 prt_printf(out, "    %c l=%u ",
3131                                            lock_types[btree_node_locked_type(path, l)], l);
3132                                 bch2_btree_bkey_cached_common_to_text(out, b);
3133                                 prt_newline(out);
3134                         }
3135                 }
3136         }
3137
3138         b = READ_ONCE(trans->locking);
3139         if (b) {
3140                 prt_printf(out, "  blocked for %lluus on",
3141                            div_u64(local_clock() - trans->locking_wait.start_time,
3142                                    1000));
3143                 prt_newline(out);
3144                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3145                 bch2_btree_bkey_cached_common_to_text(out, b);
3146                 prt_newline(out);
3147         }
3148 }
3149
3150 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3151 {
3152         struct btree_transaction_stats *s;
3153
3154         for (s = c->btree_transaction_stats;
3155              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3156              s++) {
3157                 kfree(s->max_paths_text);
3158                 bch2_time_stats_exit(&s->lock_hold_times);
3159         }
3160
3161         if (c->btree_trans_barrier_initialized)
3162                 cleanup_srcu_struct(&c->btree_trans_barrier);
3163         mempool_exit(&c->btree_trans_mem_pool);
3164         mempool_exit(&c->btree_paths_pool);
3165 }
3166
3167 int bch2_fs_btree_iter_init(struct bch_fs *c)
3168 {
3169         struct btree_transaction_stats *s;
3170         unsigned nr = BTREE_ITER_MAX;
3171         int ret;
3172
3173         for (s = c->btree_transaction_stats;
3174              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3175              s++) {
3176                 bch2_time_stats_init(&s->lock_hold_times);
3177                 mutex_init(&s->lock);
3178         }
3179
3180         INIT_LIST_HEAD(&c->btree_trans_list);
3181         seqmutex_init(&c->btree_trans_lock);
3182
3183         ret   = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3184                         sizeof(struct btree_path) * nr +
3185                         sizeof(struct btree_insert_entry) * nr) ?:
3186                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3187                                           BTREE_TRANS_MEM_MAX) ?:
3188                 init_srcu_struct(&c->btree_trans_barrier);
3189         if (!ret)
3190                 c->btree_trans_barrier_initialized = true;
3191         return ret;
3192 }