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 "journal_io.h"
  17 #include "replicas.h"
  18 #include "snapshot.h"
  19 #include "trace.h"
  20
  21 #include <linux/random.h>
  22 #include <linux/prefetch.h>
  23
  24 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
  25 static inline void btree_path_list_add(struct btree_trans *,
  26                         btree_path_idx_t, btree_path_idx_t);
  27
  28 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
  29 {
  30 #ifdef TRACK_PATH_ALLOCATED
  31         return iter->ip_allocated;
  32 #else
  33         return 0;
  34 #endif
  35 }
  36
  37 static btree_path_idx_t btree_path_alloc(struct btree_trans *, btree_path_idx_t);
  38 static void bch2_trans_srcu_lock(struct btree_trans *);
  39
  40 static inline int __btree_path_cmp(const struct btree_path *l,
  41                                    enum btree_id        r_btree_id,
  42                                    bool                 r_cached,
  43                                    struct bpos          r_pos,
  44                                    unsigned             r_level)
  45 {
  46         /*
  47          * Must match lock ordering as defined by __bch2_btree_node_lock:
  48          */
  49         return   cmp_int(l->btree_id,   r_btree_id) ?:
  50                  cmp_int((int) l->cached,       (int) r_cached) ?:
  51                  bpos_cmp(l->pos,       r_pos) ?:
  52                 -cmp_int(l->level,      r_level);
  53 }
  54
  55 static inline int btree_path_cmp(const struct btree_path *l,
  56                                  const struct btree_path *r)
  57 {
  58         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  59 }
  60
  61 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  62 {
  63         /* Are we iterating over keys in all snapshots? */
  64         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  65                 p = bpos_successor(p);
  66         } else {
  67                 p = bpos_nosnap_successor(p);
  68                 p.snapshot = iter->snapshot;
  69         }
  70
  71         return p;
  72 }
  73
  74 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  75 {
  76         /* Are we iterating over keys in all snapshots? */
  77         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  78                 p = bpos_predecessor(p);
  79         } else {
  80                 p = bpos_nosnap_predecessor(p);
  81                 p.snapshot = iter->snapshot;
  82         }
  83
  84         return p;
  85 }
  86
  87 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
  88 {
  89         struct bpos pos = iter->pos;
  90
  91         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
  92             !bkey_eq(pos, POS_MAX))
  93                 pos = bkey_successor(iter, pos);
  94         return pos;
  95 }
  96
  97 static inline bool btree_path_pos_before_node(struct btree_path *path,
  98                                               struct btree *b)
  99 {
 100         return bpos_lt(path->pos, b->data->min_key);
 101 }
 102
 103 static inline bool btree_path_pos_after_node(struct btree_path *path,
 104                                              struct btree *b)
 105 {
 106         return bpos_gt(path->pos, b->key.k.p);
 107 }
 108
 109 static inline bool btree_path_pos_in_node(struct btree_path *path,
 110                                           struct btree *b)
 111 {
 112         return path->btree_id == b->c.btree_id &&
 113                 !btree_path_pos_before_node(path, b) &&
 114                 !btree_path_pos_after_node(path, b);
 115 }
 116
 117 /* Btree iterator: */
 118
 119 #ifdef CONFIG_BCACHEFS_DEBUG
 120
 121 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 122                                           struct btree_path *path)
 123 {
 124         struct bkey_cached *ck;
 125         bool locked = btree_node_locked(path, 0);
 126
 127         if (!bch2_btree_node_relock(trans, path, 0))
 128                 return;
 129
 130         ck = (void *) path->l[0].b;
 131         BUG_ON(ck->key.btree_id != path->btree_id ||
 132                !bkey_eq(ck->key.pos, path->pos));
 133
 134         if (!locked)
 135                 btree_node_unlock(trans, path, 0);
 136 }
 137
 138 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 139                                 struct btree_path *path, unsigned level)
 140 {
 141         struct btree_path_level *l;
 142         struct btree_node_iter tmp;
 143         bool locked;
 144         struct bkey_packed *p, *k;
 145         struct printbuf buf1 = PRINTBUF;
 146         struct printbuf buf2 = PRINTBUF;
 147         struct printbuf buf3 = PRINTBUF;
 148         const char *msg;
 149
 150         if (!bch2_debug_check_iterators)
 151                 return;
 152
 153         l       = &path->l[level];
 154         tmp     = l->iter;
 155         locked  = btree_node_locked(path, level);
 156
 157         if (path->cached) {
 158                 if (!level)
 159                         bch2_btree_path_verify_cached(trans, path);
 160                 return;
 161         }
 162
 163         if (!btree_path_node(path, level))
 164                 return;
 165
 166         if (!bch2_btree_node_relock_notrace(trans, path, level))
 167                 return;
 168
 169         BUG_ON(!btree_path_pos_in_node(path, l->b));
 170
 171         bch2_btree_node_iter_verify(&l->iter, l->b);
 172
 173         /*
 174          * For interior nodes, the iterator will have skipped past deleted keys:
 175          */
 176         p = level
 177                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 178                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 179         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 180
 181         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 182                 msg = "before";
 183                 goto err;
 184         }
 185
 186         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 187                 msg = "after";
 188                 goto err;
 189         }
 190
 191         if (!locked)
 192                 btree_node_unlock(trans, path, level);
 193         return;
 194 err:
 195         bch2_bpos_to_text(&buf1, path->pos);
 196
 197         if (p) {
 198                 struct bkey uk = bkey_unpack_key(l->b, p);
 199
 200                 bch2_bkey_to_text(&buf2, &uk);
 201         } else {
 202                 prt_printf(&buf2, "(none)");
 203         }
 204
 205         if (k) {
 206                 struct bkey uk = bkey_unpack_key(l->b, k);
 207
 208                 bch2_bkey_to_text(&buf3, &uk);
 209         } else {
 210                 prt_printf(&buf3, "(none)");
 211         }
 212
 213         panic("path should be %s key at level %u:\n"
 214               "path pos %s\n"
 215               "prev key %s\n"
 216               "cur  key %s\n",
 217               msg, level, buf1.buf, buf2.buf, buf3.buf);
 218 }
 219
 220 static void bch2_btree_path_verify(struct btree_trans *trans,
 221                                    struct btree_path *path)
 222 {
 223         struct bch_fs *c = trans->c;
 224         unsigned i;
 225
 226         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 227
 228         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 229                 if (!path->l[i].b) {
 230                         BUG_ON(!path->cached &&
 231                                bch2_btree_id_root(c, path->btree_id)->b->c.level > i);
 232                         break;
 233                 }
 234
 235                 bch2_btree_path_verify_level(trans, path, i);
 236         }
 237
 238         bch2_btree_path_verify_locks(path);
 239 }
 240
 241 void bch2_trans_verify_paths(struct btree_trans *trans)
 242 {
 243         struct btree_path *path;
 244         unsigned iter;
 245
 246         trans_for_each_path(trans, path, iter)
 247                 bch2_btree_path_verify(trans, path);
 248 }
 249
 250 static void bch2_btree_iter_verify(struct btree_iter *iter)
 251 {
 252         struct btree_trans *trans = iter->trans;
 253
 254         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 255
 256         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != btree_iter_path(trans, iter)->cached);
 257
 258         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 259                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 260
 261         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 262                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 263                !btree_type_has_snapshot_field(iter->btree_id));
 264
 265         if (iter->update_path)
 266                 bch2_btree_path_verify(trans, &trans->paths[iter->update_path]);
 267         bch2_btree_path_verify(trans, btree_iter_path(trans, iter));
 268 }
 269
 270 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 271 {
 272         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 273                !iter->pos.snapshot);
 274
 275         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 276                iter->pos.snapshot != iter->snapshot);
 277
 278         BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
 279                bkey_gt(iter->pos, iter->k.p));
 280 }
 281
 282 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 283 {
 284         struct btree_trans *trans = iter->trans;
 285         struct btree_iter copy;
 286         struct bkey_s_c prev;
 287         int ret = 0;
 288
 289         if (!bch2_debug_check_iterators)
 290                 return 0;
 291
 292         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 293                 return 0;
 294
 295         if (bkey_err(k) || !k.k)
 296                 return 0;
 297
 298         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 299                                           iter->snapshot,
 300                                           k.k->p.snapshot));
 301
 302         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 303                              BTREE_ITER_NOPRESERVE|
 304                              BTREE_ITER_ALL_SNAPSHOTS);
 305         prev = bch2_btree_iter_prev(&copy);
 306         if (!prev.k)
 307                 goto out;
 308
 309         ret = bkey_err(prev);
 310         if (ret)
 311                 goto out;
 312
 313         if (bkey_eq(prev.k->p, k.k->p) &&
 314             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 315                                       prev.k->p.snapshot) > 0) {
 316                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 317
 318                 bch2_bkey_to_text(&buf1, k.k);
 319                 bch2_bkey_to_text(&buf2, prev.k);
 320
 321                 panic("iter snap %u\n"
 322                       "k    %s\n"
 323                       "prev %s\n",
 324                       iter->snapshot,
 325                       buf1.buf, buf2.buf);
 326         }
 327 out:
 328         bch2_trans_iter_exit(trans, &copy);
 329         return ret;
 330 }
 331
 332 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 333                             struct bpos pos, bool key_cache)
 334 {
 335         struct btree_path *path;
 336         struct trans_for_each_path_inorder_iter iter;
 337         struct printbuf buf = PRINTBUF;
 338
 339         btree_trans_sort_paths(trans);
 340
 341         trans_for_each_path_inorder(trans, path, iter) {
 342                 int cmp = cmp_int(path->btree_id, id) ?:
 343                         cmp_int(path->cached, key_cache);
 344
 345                 if (cmp > 0)
 346                         break;
 347                 if (cmp < 0)
 348                         continue;
 349
 350                 if (!btree_node_locked(path, 0) ||
 351                     !path->should_be_locked)
 352                         continue;
 353
 354                 if (!key_cache) {
 355                         if (bkey_ge(pos, path->l[0].b->data->min_key) &&
 356                             bkey_le(pos, path->l[0].b->key.k.p))
 357                                 return;
 358                 } else {
 359                         if (bkey_eq(pos, path->pos))
 360                                 return;
 361                 }
 362         }
 363
 364         bch2_dump_trans_paths_updates(trans);
 365         bch2_bpos_to_text(&buf, pos);
 366
 367         panic("not locked: %s %s%s\n",
 368               bch2_btree_id_str(id), buf.buf,
 369               key_cache ? " cached" : "");
 370 }
 371
 372 #else
 373
 374 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 375                                                 struct btree_path *path, unsigned l) {}
 376 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 377                                           struct btree_path *path) {}
 378 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 379 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 380 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 381
 382 #endif
 383
 384 /* Btree path: fixups after btree updates */
 385
 386 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 387                                         struct btree *b,
 388                                         struct bset_tree *t,
 389                                         struct bkey_packed *k)
 390 {
 391         struct btree_node_iter_set *set;
 392
 393         btree_node_iter_for_each(iter, set)
 394                 if (set->end == t->end_offset) {
 395                         set->k = __btree_node_key_to_offset(b, k);
 396                         bch2_btree_node_iter_sort(iter, b);
 397                         return;
 398                 }
 399
 400         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 401 }
 402
 403 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 404                                                struct btree *b,
 405                                                struct bkey_packed *where)
 406 {
 407         struct btree_path_level *l = &path->l[b->c.level];
 408
 409         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 410                 return;
 411
 412         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 413                 bch2_btree_node_iter_advance(&l->iter, l->b);
 414 }
 415
 416 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 417                                       struct btree *b,
 418                                       struct bkey_packed *where)
 419 {
 420         struct btree_path *path;
 421         unsigned i;
 422
 423         trans_for_each_path_with_node(trans, b, path, i) {
 424                 __bch2_btree_path_fix_key_modified(path, b, where);
 425                 bch2_btree_path_verify_level(trans, path, b->c.level);
 426         }
 427 }
 428
 429 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 430                                        struct btree *b,
 431                                        struct btree_node_iter *node_iter,
 432                                        struct bset_tree *t,
 433                                        struct bkey_packed *where,
 434                                        unsigned clobber_u64s,
 435                                        unsigned new_u64s)
 436 {
 437         const struct bkey_packed *end = btree_bkey_last(b, t);
 438         struct btree_node_iter_set *set;
 439         unsigned offset = __btree_node_key_to_offset(b, where);
 440         int shift = new_u64s - clobber_u64s;
 441         unsigned old_end = t->end_offset - shift;
 442         unsigned orig_iter_pos = node_iter->data[0].k;
 443         bool iter_current_key_modified =
 444                 orig_iter_pos >= offset &&
 445                 orig_iter_pos <= offset + clobber_u64s;
 446
 447         btree_node_iter_for_each(node_iter, set)
 448                 if (set->end == old_end)
 449                         goto found;
 450
 451         /* didn't find the bset in the iterator - might have to readd it: */
 452         if (new_u64s &&
 453             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 454                 bch2_btree_node_iter_push(node_iter, b, where, end);
 455                 goto fixup_done;
 456         } else {
 457                 /* Iterator is after key that changed */
 458                 return;
 459         }
 460 found:
 461         set->end = t->end_offset;
 462
 463         /* Iterator hasn't gotten to the key that changed yet: */
 464         if (set->k < offset)
 465                 return;
 466
 467         if (new_u64s &&
 468             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 469                 set->k = offset;
 470         } else if (set->k < offset + clobber_u64s) {
 471                 set->k = offset + new_u64s;
 472                 if (set->k == set->end)
 473                         bch2_btree_node_iter_set_drop(node_iter, set);
 474         } else {
 475                 /* Iterator is after key that changed */
 476                 set->k = (int) set->k + shift;
 477                 return;
 478         }
 479
 480         bch2_btree_node_iter_sort(node_iter, b);
 481 fixup_done:
 482         if (node_iter->data[0].k != orig_iter_pos)
 483                 iter_current_key_modified = true;
 484
 485         /*
 486          * When a new key is added, and the node iterator now points to that
 487          * key, the iterator might have skipped past deleted keys that should
 488          * come after the key the iterator now points to. We have to rewind to
 489          * before those deleted keys - otherwise
 490          * bch2_btree_node_iter_prev_all() breaks:
 491          */
 492         if (!bch2_btree_node_iter_end(node_iter) &&
 493             iter_current_key_modified &&
 494             b->c.level) {
 495                 struct bkey_packed *k, *k2, *p;
 496
 497                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 498
 499                 for_each_bset(b, t) {
 500                         bool set_pos = false;
 501
 502                         if (node_iter->data[0].end == t->end_offset)
 503                                 continue;
 504
 505                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 506
 507                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 508                                bkey_iter_cmp(b, k, p) < 0) {
 509                                 k2 = p;
 510                                 set_pos = true;
 511                         }
 512
 513                         if (set_pos)
 514                                 btree_node_iter_set_set_pos(node_iter,
 515                                                             b, t, k2);
 516                 }
 517         }
 518 }
 519
 520 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 521                               struct btree_path *path,
 522                               struct btree *b,
 523                               struct btree_node_iter *node_iter,
 524                               struct bkey_packed *where,
 525                               unsigned clobber_u64s,
 526                               unsigned new_u64s)
 527 {
 528         struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where);
 529         struct btree_path *linked;
 530         unsigned i;
 531
 532         if (node_iter != &path->l[b->c.level].iter) {
 533                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 534                                            where, clobber_u64s, new_u64s);
 535
 536                 if (bch2_debug_check_iterators)
 537                         bch2_btree_node_iter_verify(node_iter, b);
 538         }
 539
 540         trans_for_each_path_with_node(trans, b, linked, i) {
 541                 __bch2_btree_node_iter_fix(linked, b,
 542                                            &linked->l[b->c.level].iter, t,
 543                                            where, clobber_u64s, new_u64s);
 544                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 545         }
 546 }
 547
 548 /* Btree path level: pointer to a particular btree node and node iter */
 549
 550 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
 551                                                   struct btree_path_level *l,
 552                                                   struct bkey *u,
 553                                                   struct bkey_packed *k)
 554 {
 555         if (unlikely(!k)) {
 556                 /*
 557                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 558                  * a hole
 559                  */
 560                 u->type = KEY_TYPE_deleted;
 561                 return bkey_s_c_null;
 562         }
 563
 564         return bkey_disassemble(l->b, k, u);
 565 }
 566
 567 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
 568                                                         struct btree_path_level *l,
 569                                                         struct bkey *u)
 570 {
 571         return __btree_iter_unpack(c, l, u,
 572                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 573 }
 574
 575 static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans,
 576                                                     struct btree_path *path,
 577                                                     struct btree_path_level *l,
 578                                                     struct bkey *u)
 579 {
 580         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 581                         bch2_btree_node_iter_peek(&l->iter, l->b));
 582
 583         path->pos = k.k ? k.k->p : l->b->key.k.p;
 584         trans->paths_sorted = false;
 585         bch2_btree_path_verify_level(trans, path, l - path->l);
 586         return k;
 587 }
 588
 589 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
 590                                                     struct btree_path *path,
 591                                                     struct btree_path_level *l,
 592                                                     struct bkey *u)
 593 {
 594         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 595                         bch2_btree_node_iter_prev(&l->iter, l->b));
 596
 597         path->pos = k.k ? k.k->p : l->b->data->min_key;
 598         trans->paths_sorted = false;
 599         bch2_btree_path_verify_level(trans, path, l - path->l);
 600         return k;
 601 }
 602
 603 static inline bool btree_path_advance_to_pos(struct btree_path *path,
 604                                              struct btree_path_level *l,
 605                                              int max_advance)
 606 {
 607         struct bkey_packed *k;
 608         int nr_advanced = 0;
 609
 610         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 611                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 612                 if (max_advance > 0 && nr_advanced >= max_advance)
 613                         return false;
 614
 615                 bch2_btree_node_iter_advance(&l->iter, l->b);
 616                 nr_advanced++;
 617         }
 618
 619         return true;
 620 }
 621
 622 static inline void __btree_path_level_init(struct btree_path *path,
 623                                            unsigned level)
 624 {
 625         struct btree_path_level *l = &path->l[level];
 626
 627         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
 628
 629         /*
 630          * Iterators to interior nodes should always be pointed at the first non
 631          * whiteout:
 632          */
 633         if (level)
 634                 bch2_btree_node_iter_peek(&l->iter, l->b);
 635 }
 636
 637 void bch2_btree_path_level_init(struct btree_trans *trans,
 638                                 struct btree_path *path,
 639                                 struct btree *b)
 640 {
 641         BUG_ON(path->cached);
 642
 643         EBUG_ON(!btree_path_pos_in_node(path, b));
 644
 645         path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 646         path->l[b->c.level].b = b;
 647         __btree_path_level_init(path, b->c.level);
 648 }
 649
 650 /* Btree path: fixups after btree node updates: */
 651
 652 static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b)
 653 {
 654         struct bch_fs *c = trans->c;
 655
 656         trans_for_each_update(trans, i)
 657                 if (!i->cached &&
 658                     i->level    == b->c.level &&
 659                     i->btree_id == b->c.btree_id &&
 660                     bpos_cmp(i->k->k.p, b->data->min_key) >= 0 &&
 661                     bpos_cmp(i->k->k.p, b->data->max_key) <= 0) {
 662                         i->old_v = bch2_btree_path_peek_slot(trans->paths + i->path, &i->old_k).v;
 663
 664                         if (unlikely(trans->journal_replay_not_finished)) {
 665                                 struct bkey_i *j_k =
 666                                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 667                                                                     i->k->k.p);
 668
 669                                 if (j_k) {
 670                                         i->old_k = j_k->k;
 671                                         i->old_v = &j_k->v;
 672                                 }
 673                         }
 674                 }
 675 }
 676
 677 /*
 678  * A btree node is being replaced - update the iterator to point to the new
 679  * node:
 680  */
 681 void bch2_trans_node_add(struct btree_trans *trans,
 682                          struct btree_path *path,
 683                          struct btree *b)
 684 {
 685         struct btree_path *prev;
 686
 687         BUG_ON(!btree_path_pos_in_node(path, b));
 688
 689         while ((prev = prev_btree_path(trans, path)) &&
 690                btree_path_pos_in_node(prev, b))
 691                 path = prev;
 692
 693         for (;
 694              path && btree_path_pos_in_node(path, b);
 695              path = next_btree_path(trans, path))
 696                 if (path->uptodate == BTREE_ITER_UPTODATE && !path->cached) {
 697                         enum btree_node_locked_type t =
 698                                 btree_lock_want(path, b->c.level);
 699
 700                         if (t != BTREE_NODE_UNLOCKED) {
 701                                 btree_node_unlock(trans, path, b->c.level);
 702                                 six_lock_increment(&b->c.lock, (enum six_lock_type) t);
 703                                 mark_btree_node_locked(trans, path, b->c.level, t);
 704                         }
 705
 706                         bch2_btree_path_level_init(trans, path, b);
 707                 }
 708
 709         bch2_trans_revalidate_updates_in_node(trans, b);
 710 }
 711
 712 /*
 713  * A btree node has been modified in such a way as to invalidate iterators - fix
 714  * them:
 715  */
 716 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
 717 {
 718         struct btree_path *path;
 719         unsigned i;
 720
 721         trans_for_each_path_with_node(trans, b, path, i)
 722                 __btree_path_level_init(path, b->c.level);
 723
 724         bch2_trans_revalidate_updates_in_node(trans, b);
 725 }
 726
 727 /* Btree path: traverse, set_pos: */
 728
 729 static inline int btree_path_lock_root(struct btree_trans *trans,
 730                                        struct btree_path *path,
 731                                        unsigned depth_want,
 732                                        unsigned long trace_ip)
 733 {
 734         struct bch_fs *c = trans->c;
 735         struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b;
 736         enum six_lock_type lock_type;
 737         unsigned i;
 738         int ret;
 739
 740         EBUG_ON(path->nodes_locked);
 741
 742         while (1) {
 743                 b = READ_ONCE(*rootp);
 744                 path->level = READ_ONCE(b->c.level);
 745
 746                 if (unlikely(path->level < depth_want)) {
 747                         /*
 748                          * the root is at a lower depth than the depth we want:
 749                          * got to the end of the btree, or we're walking nodes
 750                          * greater than some depth and there are no nodes >=
 751                          * that depth
 752                          */
 753                         path->level = depth_want;
 754                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
 755                                 path->l[i].b = NULL;
 756                         return 1;
 757                 }
 758
 759                 lock_type = __btree_lock_want(path, path->level);
 760                 ret = btree_node_lock(trans, path, &b->c,
 761                                       path->level, lock_type, trace_ip);
 762                 if (unlikely(ret)) {
 763                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
 764                                 continue;
 765                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 766                                 return ret;
 767                         BUG();
 768                 }
 769
 770                 if (likely(b == READ_ONCE(*rootp) &&
 771                            b->c.level == path->level &&
 772                            !race_fault())) {
 773                         for (i = 0; i < path->level; i++)
 774                                 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
 775                         path->l[path->level].b = b;
 776                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
 777                                 path->l[i].b = NULL;
 778
 779                         mark_btree_node_locked(trans, path, path->level,
 780                                                (enum btree_node_locked_type) lock_type);
 781                         bch2_btree_path_level_init(trans, path, b);
 782                         return 0;
 783                 }
 784
 785                 six_unlock_type(&b->c.lock, lock_type);
 786         }
 787 }
 788
 789 noinline
 790 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
 791 {
 792         struct bch_fs *c = trans->c;
 793         struct btree_path_level *l = path_l(path);
 794         struct btree_node_iter node_iter = l->iter;
 795         struct bkey_packed *k;
 796         struct bkey_buf tmp;
 797         unsigned nr = test_bit(BCH_FS_started, &c->flags)
 798                 ? (path->level > 1 ? 0 :  2)
 799                 : (path->level > 1 ? 1 : 16);
 800         bool was_locked = btree_node_locked(path, path->level);
 801         int ret = 0;
 802
 803         bch2_bkey_buf_init(&tmp);
 804
 805         while (nr-- && !ret) {
 806                 if (!bch2_btree_node_relock(trans, path, path->level))
 807                         break;
 808
 809                 bch2_btree_node_iter_advance(&node_iter, l->b);
 810                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
 811                 if (!k)
 812                         break;
 813
 814                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
 815                 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
 816                                                path->level - 1);
 817         }
 818
 819         if (!was_locked)
 820                 btree_node_unlock(trans, path, path->level);
 821
 822         bch2_bkey_buf_exit(&tmp, c);
 823         return ret;
 824 }
 825
 826 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
 827                                  struct btree_and_journal_iter *jiter)
 828 {
 829         struct bch_fs *c = trans->c;
 830         struct bkey_s_c k;
 831         struct bkey_buf tmp;
 832         unsigned nr = test_bit(BCH_FS_started, &c->flags)
 833                 ? (path->level > 1 ? 0 :  2)
 834                 : (path->level > 1 ? 1 : 16);
 835         bool was_locked = btree_node_locked(path, path->level);
 836         int ret = 0;
 837
 838         bch2_bkey_buf_init(&tmp);
 839
 840         while (nr-- && !ret) {
 841                 if (!bch2_btree_node_relock(trans, path, path->level))
 842                         break;
 843
 844                 bch2_btree_and_journal_iter_advance(jiter);
 845                 k = bch2_btree_and_journal_iter_peek(jiter);
 846                 if (!k.k)
 847                         break;
 848
 849                 bch2_bkey_buf_reassemble(&tmp, c, k);
 850                 ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id,
 851                                                path->level - 1);
 852         }
 853
 854         if (!was_locked)
 855                 btree_node_unlock(trans, path, path->level);
 856
 857         bch2_bkey_buf_exit(&tmp, c);
 858         return ret;
 859 }
 860
 861 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
 862                                             struct btree_path *path,
 863                                             unsigned plevel, struct btree *b)
 864 {
 865         struct btree_path_level *l = &path->l[plevel];
 866         bool locked = btree_node_locked(path, plevel);
 867         struct bkey_packed *k;
 868         struct bch_btree_ptr_v2 *bp;
 869
 870         if (!bch2_btree_node_relock(trans, path, plevel))
 871                 return;
 872
 873         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 874         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
 875
 876         bp = (void *) bkeyp_val(&l->b->format, k);
 877         bp->mem_ptr = (unsigned long)b;
 878
 879         if (!locked)
 880                 btree_node_unlock(trans, path, plevel);
 881 }
 882
 883 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
 884                                                      struct btree_path *path,
 885                                                      unsigned flags,
 886                                                      struct bkey_buf *out)
 887 {
 888         struct bch_fs *c = trans->c;
 889         struct btree_path_level *l = path_l(path);
 890         struct btree_and_journal_iter jiter;
 891         struct bkey_s_c k;
 892         int ret = 0;
 893
 894         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
 895
 896         k = bch2_btree_and_journal_iter_peek(&jiter);
 897
 898         bch2_bkey_buf_reassemble(out, c, k);
 899
 900         if (flags & BTREE_ITER_PREFETCH)
 901                 ret = btree_path_prefetch_j(trans, path, &jiter);
 902
 903         bch2_btree_and_journal_iter_exit(&jiter);
 904         return ret;
 905 }
 906
 907 static __always_inline int btree_path_down(struct btree_trans *trans,
 908                                            struct btree_path *path,
 909                                            unsigned flags,
 910                                            unsigned long trace_ip)
 911 {
 912         struct bch_fs *c = trans->c;
 913         struct btree_path_level *l = path_l(path);
 914         struct btree *b;
 915         unsigned level = path->level - 1;
 916         enum six_lock_type lock_type = __btree_lock_want(path, level);
 917         struct bkey_buf tmp;
 918         int ret;
 919
 920         EBUG_ON(!btree_node_locked(path, path->level));
 921
 922         bch2_bkey_buf_init(&tmp);
 923
 924         if (unlikely(trans->journal_replay_not_finished)) {
 925                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
 926                 if (ret)
 927                         goto err;
 928         } else {
 929                 bch2_bkey_buf_unpack(&tmp, c, l->b,
 930                                  bch2_btree_node_iter_peek(&l->iter, l->b));
 931
 932                 if (flags & BTREE_ITER_PREFETCH) {
 933                         ret = btree_path_prefetch(trans, path);
 934                         if (ret)
 935                                 goto err;
 936                 }
 937         }
 938
 939         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
 940         ret = PTR_ERR_OR_ZERO(b);
 941         if (unlikely(ret))
 942                 goto err;
 943
 944         if (likely(!trans->journal_replay_not_finished &&
 945                    tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
 946             unlikely(b != btree_node_mem_ptr(tmp.k)))
 947                 btree_node_mem_ptr_set(trans, path, level + 1, b);
 948
 949         if (btree_node_read_locked(path, level + 1))
 950                 btree_node_unlock(trans, path, level + 1);
 951
 952         mark_btree_node_locked(trans, path, level,
 953                                (enum btree_node_locked_type) lock_type);
 954         path->level = level;
 955         bch2_btree_path_level_init(trans, path, b);
 956
 957         bch2_btree_path_verify_locks(path);
 958 err:
 959         bch2_bkey_buf_exit(&tmp, c);
 960         return ret;
 961 }
 962
 963
 964 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
 965 {
 966         struct bch_fs *c = trans->c;
 967         struct btree_path *path;
 968         unsigned long trace_ip = _RET_IP_;
 969         unsigned i;
 970         int ret = 0;
 971
 972         if (trans->in_traverse_all)
 973                 return -BCH_ERR_transaction_restart_in_traverse_all;
 974
 975         trans->in_traverse_all = true;
 976 retry_all:
 977         trans->restarted = 0;
 978         trans->last_restarted_ip = 0;
 979
 980         trans_for_each_path(trans, path, i)
 981                 path->should_be_locked = false;
 982
 983         btree_trans_sort_paths(trans);
 984
 985         bch2_trans_unlock(trans);
 986         cond_resched();
 987
 988         if (unlikely(trans->memory_allocation_failure)) {
 989                 struct closure cl;
 990
 991                 closure_init_stack(&cl);
 992
 993                 do {
 994                         ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
 995                         closure_sync(&cl);
 996                 } while (ret);
 997         }
 998
 999         /* Now, redo traversals in correct order: */
1000         i = 0;
1001         while (i < trans->nr_sorted) {
1002                 btree_path_idx_t idx = trans->sorted[i];
1003
1004                 /*
1005                  * Traversing a path can cause another path to be added at about
1006                  * the same position:
1007                  */
1008                 if (trans->paths[idx].uptodate) {
1009                         __btree_path_get(&trans->paths[idx], false);
1010                         ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_);
1011                         __btree_path_put(&trans->paths[idx], false);
1012
1013                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1014                             bch2_err_matches(ret, ENOMEM))
1015                                 goto retry_all;
1016                         if (ret)
1017                                 goto err;
1018                 } else {
1019                         i++;
1020                 }
1021         }
1022
1023         /*
1024          * We used to assert that all paths had been traversed here
1025          * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1026          * path->should_be_locked is not set yet, we might have unlocked and
1027          * then failed to relock a path - that's fine.
1028          */
1029 err:
1030         bch2_btree_cache_cannibalize_unlock(trans);
1031
1032         trans->in_traverse_all = false;
1033
1034         trace_and_count(c, trans_traverse_all, trans, trace_ip);
1035         return ret;
1036 }
1037
1038 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1039                                                 unsigned l, int check_pos)
1040 {
1041         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1042                 return false;
1043         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1044                 return false;
1045         return true;
1046 }
1047
1048 static inline bool btree_path_good_node(struct btree_trans *trans,
1049                                         struct btree_path *path,
1050                                         unsigned l, int check_pos)
1051 {
1052         return is_btree_node(path, l) &&
1053                 bch2_btree_node_relock(trans, path, l) &&
1054                 btree_path_check_pos_in_node(path, l, check_pos);
1055 }
1056
1057 static void btree_path_set_level_down(struct btree_trans *trans,
1058                                       struct btree_path *path,
1059                                       unsigned new_level)
1060 {
1061         unsigned l;
1062
1063         path->level = new_level;
1064
1065         for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1066                 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1067                         btree_node_unlock(trans, path, l);
1068
1069         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1070         bch2_btree_path_verify(trans, path);
1071 }
1072
1073 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1074                                                          struct btree_path *path,
1075                                                          int check_pos)
1076 {
1077         unsigned i, l = path->level;
1078 again:
1079         while (btree_path_node(path, l) &&
1080                !btree_path_good_node(trans, path, l, check_pos))
1081                 __btree_path_set_level_up(trans, path, l++);
1082
1083         /* If we need intent locks, take them too: */
1084         for (i = l + 1;
1085              i < path->locks_want && btree_path_node(path, i);
1086              i++)
1087                 if (!bch2_btree_node_relock(trans, path, i)) {
1088                         while (l <= i)
1089                                 __btree_path_set_level_up(trans, path, l++);
1090                         goto again;
1091                 }
1092
1093         return l;
1094 }
1095
1096 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1097                                                      struct btree_path *path,
1098                                                      int check_pos)
1099 {
1100         return likely(btree_node_locked(path, path->level) &&
1101                       btree_path_check_pos_in_node(path, path->level, check_pos))
1102                 ? path->level
1103                 : __btree_path_up_until_good_node(trans, path, check_pos);
1104 }
1105
1106 /*
1107  * This is the main state machine for walking down the btree - walks down to a
1108  * specified depth
1109  *
1110  * Returns 0 on success, -EIO on error (error reading in a btree node).
1111  *
1112  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1113  * stashed in the iterator and returned from bch2_trans_exit().
1114  */
1115 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1116                                  btree_path_idx_t path_idx,
1117                                  unsigned flags,
1118                                  unsigned long trace_ip)
1119 {
1120         struct btree_path *path = &trans->paths[path_idx];
1121         unsigned depth_want = path->level;
1122         int ret = -((int) trans->restarted);
1123
1124         if (unlikely(ret))
1125                 goto out;
1126
1127         if (unlikely(!trans->srcu_held))
1128                 bch2_trans_srcu_lock(trans);
1129
1130         /*
1131          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1132          * and re-traverse the path without a transaction restart:
1133          */
1134         if (path->should_be_locked) {
1135                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1136                 goto out;
1137         }
1138
1139         if (path->cached) {
1140                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1141                 goto out;
1142         }
1143
1144         path = &trans->paths[path_idx];
1145
1146         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1147                 goto out;
1148
1149         path->level = btree_path_up_until_good_node(trans, path, 0);
1150
1151         EBUG_ON(btree_path_node(path, path->level) &&
1152                 !btree_node_locked(path, path->level));
1153
1154         /*
1155          * Note: path->nodes[path->level] may be temporarily NULL here - that
1156          * would indicate to other code that we got to the end of the btree,
1157          * here it indicates that relocking the root failed - it's critical that
1158          * btree_path_lock_root() comes next and that it can't fail
1159          */
1160         while (path->level > depth_want) {
1161                 ret = btree_path_node(path, path->level)
1162                         ? btree_path_down(trans, path, flags, trace_ip)
1163                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1164                 if (unlikely(ret)) {
1165                         if (ret == 1) {
1166                                 /*
1167                                  * No nodes at this level - got to the end of
1168                                  * the btree:
1169                                  */
1170                                 ret = 0;
1171                                 goto out;
1172                         }
1173
1174                         __bch2_btree_path_unlock(trans, path);
1175                         path->level = depth_want;
1176                         path->l[path->level].b = ERR_PTR(ret);
1177                         goto out;
1178                 }
1179         }
1180
1181         path->uptodate = BTREE_ITER_UPTODATE;
1182 out:
1183         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1184                 panic("ret %s (%i) trans->restarted %s (%i)\n",
1185                       bch2_err_str(ret), ret,
1186                       bch2_err_str(trans->restarted), trans->restarted);
1187         bch2_btree_path_verify(trans, path);
1188         return ret;
1189 }
1190
1191 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1192                             struct btree_path *src)
1193 {
1194         unsigned i, offset = offsetof(struct btree_path, pos);
1195
1196         memcpy((void *) dst + offset,
1197                (void *) src + offset,
1198                sizeof(struct btree_path) - offset);
1199
1200         for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1201                 unsigned t = btree_node_locked_type(dst, i);
1202
1203                 if (t != BTREE_NODE_UNLOCKED)
1204                         six_lock_increment(&dst->l[i].b->c.lock, t);
1205         }
1206 }
1207
1208 static btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src,
1209                                          bool intent)
1210 {
1211         btree_path_idx_t new = btree_path_alloc(trans, src);
1212
1213         btree_path_copy(trans, trans->paths + new, trans->paths + src);
1214         __btree_path_get(trans->paths + new, intent);
1215         return new;
1216 }
1217
1218 __flatten
1219 btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans,
1220                         btree_path_idx_t path, bool intent, unsigned long ip)
1221 {
1222         __btree_path_put(trans->paths + path, intent);
1223         path = btree_path_clone(trans, path, intent);
1224         trans->paths[path].preserve = false;
1225         return path;
1226 }
1227
1228 btree_path_idx_t __must_check
1229 __bch2_btree_path_set_pos(struct btree_trans *trans,
1230                           btree_path_idx_t path_idx, struct bpos new_pos,
1231                           bool intent, unsigned long ip)
1232 {
1233         int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos);
1234
1235         bch2_trans_verify_not_in_restart(trans);
1236         EBUG_ON(!trans->paths[path_idx].ref);
1237
1238         path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip);
1239
1240         struct btree_path *path = trans->paths + path_idx;
1241         path->pos               = new_pos;
1242         trans->paths_sorted     = false;
1243
1244         if (unlikely(path->cached)) {
1245                 btree_node_unlock(trans, path, 0);
1246                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1247                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1248                 goto out;
1249         }
1250
1251         unsigned level = btree_path_up_until_good_node(trans, path, cmp);
1252
1253         if (btree_path_node(path, level)) {
1254                 struct btree_path_level *l = &path->l[level];
1255
1256                 BUG_ON(!btree_node_locked(path, level));
1257                 /*
1258                  * We might have to skip over many keys, or just a few: try
1259                  * advancing the node iterator, and if we have to skip over too
1260                  * many keys just reinit it (or if we're rewinding, since that
1261                  * is expensive).
1262                  */
1263                 if (cmp < 0 ||
1264                     !btree_path_advance_to_pos(path, l, 8))
1265                         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1266
1267                 /*
1268                  * Iterators to interior nodes should always be pointed at the first non
1269                  * whiteout:
1270                  */
1271                 if (unlikely(level))
1272                         bch2_btree_node_iter_peek(&l->iter, l->b);
1273         }
1274
1275         if (unlikely(level != path->level)) {
1276                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1277                 __bch2_btree_path_unlock(trans, path);
1278         }
1279 out:
1280         bch2_btree_path_verify(trans, path);
1281         return path_idx;
1282 }
1283
1284 /* Btree path: main interface: */
1285
1286 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1287 {
1288         struct btree_path *sib;
1289
1290         sib = prev_btree_path(trans, path);
1291         if (sib && !btree_path_cmp(sib, path))
1292                 return sib;
1293
1294         sib = next_btree_path(trans, path);
1295         if (sib && !btree_path_cmp(sib, path))
1296                 return sib;
1297
1298         return NULL;
1299 }
1300
1301 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1302 {
1303         struct btree_path *sib;
1304
1305         sib = prev_btree_path(trans, path);
1306         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1307                 return sib;
1308
1309         sib = next_btree_path(trans, path);
1310         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1311                 return sib;
1312
1313         return NULL;
1314 }
1315
1316 static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path)
1317 {
1318         __bch2_btree_path_unlock(trans, trans->paths + path);
1319         btree_path_list_remove(trans, trans->paths + path);
1320         __clear_bit(path, trans->paths_allocated);
1321 }
1322
1323 void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent)
1324 {
1325         struct btree_path *path = trans->paths + path_idx, *dup;
1326
1327         if (!__btree_path_put(path, intent))
1328                 return;
1329
1330         dup = path->preserve
1331                 ? have_path_at_pos(trans, path)
1332                 : have_node_at_pos(trans, path);
1333
1334         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1335                 return;
1336
1337         if (path->should_be_locked &&
1338             !trans->restarted &&
1339             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1340                 return;
1341
1342         if (dup) {
1343                 dup->preserve           |= path->preserve;
1344                 dup->should_be_locked   |= path->should_be_locked;
1345         }
1346
1347         __bch2_path_free(trans, path_idx);
1348 }
1349
1350 static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path,
1351                                  bool intent)
1352 {
1353         if (!__btree_path_put(trans->paths + path, intent))
1354                 return;
1355
1356         __bch2_path_free(trans, path);
1357 }
1358
1359 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1360 {
1361         panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1362               trans->restart_count, restart_count,
1363               (void *) trans->last_begin_ip);
1364 }
1365
1366 void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1367 {
1368         panic("in transaction restart: %s, last restarted by %pS\n",
1369               bch2_err_str(trans->restarted),
1370               (void *) trans->last_restarted_ip);
1371 }
1372
1373 noinline __cold
1374 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1375 {
1376         prt_printf(buf, "transaction updates for %s journal seq %llu",
1377                trans->fn, trans->journal_res.seq);
1378         prt_newline(buf);
1379         printbuf_indent_add(buf, 2);
1380
1381         trans_for_each_update(trans, i) {
1382                 struct bkey_s_c old = { &i->old_k, i->old_v };
1383
1384                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1385                        bch2_btree_id_str(i->btree_id),
1386                        i->cached,
1387                        (void *) i->ip_allocated);
1388                 prt_newline(buf);
1389
1390                 prt_printf(buf, "  old ");
1391                 bch2_bkey_val_to_text(buf, trans->c, old);
1392                 prt_newline(buf);
1393
1394                 prt_printf(buf, "  new ");
1395                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1396                 prt_newline(buf);
1397         }
1398
1399         for (struct jset_entry *e = trans->journal_entries;
1400              e != btree_trans_journal_entries_top(trans);
1401              e = vstruct_next(e))
1402                 bch2_journal_entry_to_text(buf, trans->c, e);
1403
1404         printbuf_indent_sub(buf, 2);
1405 }
1406
1407 noinline __cold
1408 void bch2_dump_trans_updates(struct btree_trans *trans)
1409 {
1410         struct printbuf buf = PRINTBUF;
1411
1412         bch2_trans_updates_to_text(&buf, trans);
1413         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1414         printbuf_exit(&buf);
1415 }
1416
1417 static void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
1418 {
1419         struct btree_path *path = trans->paths + path_idx;
1420
1421         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1422                    path_idx, path->ref, path->intent_ref,
1423                    path->preserve ? 'P' : ' ',
1424                    path->should_be_locked ? 'S' : ' ',
1425                    bch2_btree_id_str(path->btree_id),
1426                    path->level);
1427         bch2_bpos_to_text(out, path->pos);
1428
1429         prt_printf(out, " locks %u", path->nodes_locked);
1430 #ifdef TRACK_PATH_ALLOCATED
1431         prt_printf(out, " %pS", (void *) path->ip_allocated);
1432 #endif
1433         prt_newline(out);
1434 }
1435
1436 static noinline __cold
1437 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1438                                 bool nosort)
1439 {
1440         struct trans_for_each_path_inorder_iter iter;
1441
1442         if (!nosort)
1443                 btree_trans_sort_paths(trans);
1444
1445         trans_for_each_path_idx_inorder(trans, iter)
1446                 bch2_btree_path_to_text(out, trans, iter.path_idx);
1447 }
1448
1449 noinline __cold
1450 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1451 {
1452         __bch2_trans_paths_to_text(out, trans, false);
1453 }
1454
1455 static noinline __cold
1456 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1457 {
1458         struct printbuf buf = PRINTBUF;
1459
1460         __bch2_trans_paths_to_text(&buf, trans, nosort);
1461         bch2_trans_updates_to_text(&buf, trans);
1462
1463         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1464         printbuf_exit(&buf);
1465 }
1466
1467 noinline __cold
1468 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1469 {
1470         __bch2_dump_trans_paths_updates(trans, false);
1471 }
1472
1473 noinline __cold
1474 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1475 {
1476         struct btree_transaction_stats *s = btree_trans_stats(trans);
1477         struct printbuf buf = PRINTBUF;
1478         size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths);
1479
1480         if (!s)
1481                 return;
1482
1483         bch2_trans_paths_to_text(&buf, trans);
1484
1485         if (!buf.allocation_failure) {
1486                 mutex_lock(&s->lock);
1487                 if (nr > s->nr_max_paths) {
1488                         s->nr_max_paths = nr;
1489                         swap(s->max_paths_text, buf.buf);
1490                 }
1491                 mutex_unlock(&s->lock);
1492         }
1493
1494         printbuf_exit(&buf);
1495
1496         trans->nr_paths_max = nr;
1497 }
1498
1499 noinline __cold
1500 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
1501 {
1502         if (trace_trans_restart_too_many_iters_enabled()) {
1503                 struct printbuf buf = PRINTBUF;
1504
1505                 bch2_trans_paths_to_text(&buf, trans);
1506                 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
1507                 printbuf_exit(&buf);
1508         }
1509
1510         count_event(trans->c, trans_restart_too_many_iters);
1511
1512         return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
1513 }
1514
1515 static noinline void btree_paths_realloc(struct btree_trans *trans)
1516 {
1517         unsigned nr = trans->nr_paths * 2;
1518
1519         void *p = kzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) +
1520                           nr + 8 +
1521                           sizeof(struct btree_trans_paths) +
1522                           nr * sizeof(struct btree_path) +
1523                           nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL);
1524
1525         unsigned long *paths_allocated = p;
1526         p += BITS_TO_LONGS(nr) * sizeof(unsigned long);
1527         struct btree_path *paths = p;
1528         p += nr * sizeof(struct btree_path);
1529         u8 *sorted = p;
1530         p += nr + 8;
1531         struct btree_insert_entry *updates = p;
1532
1533         *trans_paths_nr(paths) = nr;
1534
1535         memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long));
1536         memcpy(sorted, trans->sorted, trans->nr_sorted);
1537         memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path));
1538         memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_path));
1539
1540         unsigned long *old = trans->paths_allocated;
1541
1542         rcu_assign_pointer(trans->paths_allocated,      paths_allocated);
1543         rcu_assign_pointer(trans->sorted,               sorted);
1544         rcu_assign_pointer(trans->paths,                paths);
1545         rcu_assign_pointer(trans->updates,              updates);
1546
1547         trans->nr_paths         = nr;
1548
1549         if (old != trans->_paths_allocated)
1550                 kfree_rcu_mightsleep(trans->paths_allocated);
1551 }
1552
1553 static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans,
1554                                                 btree_path_idx_t pos)
1555 {
1556         btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths);
1557
1558         if (unlikely(idx == trans->nr_paths))
1559                 btree_paths_realloc(trans);
1560
1561         /*
1562          * Do this before marking the new path as allocated, since it won't be
1563          * initialized yet:
1564          */
1565         if (unlikely(idx > trans->nr_paths_max))
1566                 bch2_trans_update_max_paths(trans);
1567
1568         __set_bit(idx, trans->paths_allocated);
1569
1570         struct btree_path *path = &trans->paths[idx];
1571         path->ref               = 0;
1572         path->intent_ref        = 0;
1573         path->nodes_locked      = 0;
1574
1575         btree_path_list_add(trans, pos, idx);
1576         trans->paths_sorted = false;
1577         return idx;
1578 }
1579
1580 btree_path_idx_t bch2_path_get(struct btree_trans *trans,
1581                              enum btree_id btree_id, struct bpos pos,
1582                              unsigned locks_want, unsigned level,
1583                              unsigned flags, unsigned long ip)
1584 {
1585         struct btree_path *path;
1586         bool cached = flags & BTREE_ITER_CACHED;
1587         bool intent = flags & BTREE_ITER_INTENT;
1588         struct trans_for_each_path_inorder_iter iter;
1589         btree_path_idx_t path_pos = 0, path_idx;
1590
1591         bch2_trans_verify_not_in_restart(trans);
1592         bch2_trans_verify_locks(trans);
1593
1594         btree_trans_sort_paths(trans);
1595
1596         trans_for_each_path_inorder(trans, path, iter) {
1597                 if (__btree_path_cmp(path,
1598                                      btree_id,
1599                                      cached,
1600                                      pos,
1601                                      level) > 0)
1602                         break;
1603
1604                 path_pos = iter.path_idx;
1605         }
1606
1607         if (path_pos &&
1608             trans->paths[path_pos].cached       == cached &&
1609             trans->paths[path_pos].btree_id     == btree_id &&
1610             trans->paths[path_pos].level        == level) {
1611                 __btree_path_get(trans->paths + path_pos, intent);
1612                 path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1613                 path = trans->paths + path_idx;
1614         } else {
1615                 path_idx = btree_path_alloc(trans, path_pos);
1616                 path = trans->paths + path_idx;
1617
1618                 __btree_path_get(path, intent);
1619                 path->pos                       = pos;
1620                 path->btree_id                  = btree_id;
1621                 path->cached                    = cached;
1622                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1623                 path->should_be_locked          = false;
1624                 path->level                     = level;
1625                 path->locks_want                = locks_want;
1626                 path->nodes_locked              = 0;
1627                 for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++)
1628                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1629 #ifdef TRACK_PATH_ALLOCATED
1630                 path->ip_allocated              = ip;
1631 #endif
1632                 trans->paths_sorted             = false;
1633         }
1634
1635         if (!(flags & BTREE_ITER_NOPRESERVE))
1636                 path->preserve = true;
1637
1638         if (path->intent_ref)
1639                 locks_want = max(locks_want, level + 1);
1640
1641         /*
1642          * If the path has locks_want greater than requested, we don't downgrade
1643          * it here - on transaction restart because btree node split needs to
1644          * upgrade locks, we might be putting/getting the iterator again.
1645          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1646          * a successful transaction commit.
1647          */
1648
1649         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1650         if (locks_want > path->locks_want)
1651                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
1652
1653         return path_idx;
1654 }
1655
1656 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1657 {
1658
1659         struct btree_path_level *l = path_l(path);
1660         struct bkey_packed *_k;
1661         struct bkey_s_c k;
1662
1663         if (unlikely(!l->b))
1664                 return bkey_s_c_null;
1665
1666         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1667         EBUG_ON(!btree_node_locked(path, path->level));
1668
1669         if (!path->cached) {
1670                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1671                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1672
1673                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1674
1675                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1676                         goto hole;
1677         } else {
1678                 struct bkey_cached *ck = (void *) path->l[0].b;
1679
1680                 EBUG_ON(ck &&
1681                         (path->btree_id != ck->key.btree_id ||
1682                          !bkey_eq(path->pos, ck->key.pos)));
1683                 if (!ck || !ck->valid)
1684                         return bkey_s_c_null;
1685
1686                 *u = ck->k->k;
1687                 k = bkey_i_to_s_c(ck->k);
1688         }
1689
1690         return k;
1691 hole:
1692         bkey_init(u);
1693         u->p = path->pos;
1694         return (struct bkey_s_c) { u, NULL };
1695 }
1696
1697 /* Btree iterators: */
1698
1699 int __must_check
1700 __bch2_btree_iter_traverse(struct btree_iter *iter)
1701 {
1702         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1703 }
1704
1705 int __must_check
1706 bch2_btree_iter_traverse(struct btree_iter *iter)
1707 {
1708         struct btree_trans *trans = iter->trans;
1709         int ret;
1710
1711         iter->path = bch2_btree_path_set_pos(trans, iter->path,
1712                                         btree_iter_search_key(iter),
1713                                         iter->flags & BTREE_ITER_INTENT,
1714                                         btree_iter_ip_allocated(iter));
1715
1716         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1717         if (ret)
1718                 return ret;
1719
1720         btree_path_set_should_be_locked(trans->paths + iter->path);
1721         return 0;
1722 }
1723
1724 /* Iterate across nodes (leaf and interior nodes) */
1725
1726 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1727 {
1728         struct btree_trans *trans = iter->trans;
1729         struct btree *b = NULL;
1730         int ret;
1731
1732         EBUG_ON(trans->paths[iter->path].cached);
1733         bch2_btree_iter_verify(iter);
1734
1735         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1736         if (ret)
1737                 goto err;
1738
1739         struct btree_path *path = btree_iter_path(trans, iter);
1740         b = btree_path_node(path, path->level);
1741         if (!b)
1742                 goto out;
1743
1744         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1745
1746         bkey_init(&iter->k);
1747         iter->k.p = iter->pos = b->key.k.p;
1748
1749         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1750                                         iter->flags & BTREE_ITER_INTENT,
1751                                         btree_iter_ip_allocated(iter));
1752         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1753 out:
1754         bch2_btree_iter_verify_entry_exit(iter);
1755         bch2_btree_iter_verify(iter);
1756
1757         return b;
1758 err:
1759         b = ERR_PTR(ret);
1760         goto out;
1761 }
1762
1763 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1764 {
1765         struct btree *b;
1766
1767         while (b = bch2_btree_iter_peek_node(iter),
1768                bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1769                 bch2_trans_begin(iter->trans);
1770
1771         return b;
1772 }
1773
1774 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1775 {
1776         struct btree_trans *trans = iter->trans;
1777         struct btree *b = NULL;
1778         int ret;
1779
1780         EBUG_ON(trans->paths[iter->path].cached);
1781         bch2_trans_verify_not_in_restart(trans);
1782         bch2_btree_iter_verify(iter);
1783
1784         struct btree_path *path = btree_iter_path(trans, iter);
1785
1786         /* already at end? */
1787         if (!btree_path_node(path, path->level))
1788                 return NULL;
1789
1790         /* got to end? */
1791         if (!btree_path_node(path, path->level + 1)) {
1792                 btree_path_set_level_up(trans, path);
1793                 return NULL;
1794         }
1795
1796         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1797                 __bch2_btree_path_unlock(trans, path);
1798                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1799                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1800                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1801                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1802                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1803                 goto err;
1804         }
1805
1806         b = btree_path_node(path, path->level + 1);
1807
1808         if (bpos_eq(iter->pos, b->key.k.p)) {
1809                 __btree_path_set_level_up(trans, path, path->level++);
1810         } else {
1811                 /*
1812                  * Haven't gotten to the end of the parent node: go back down to
1813                  * the next child node
1814                  */
1815                 iter->path = bch2_btree_path_set_pos(trans, iter->path,
1816                                         bpos_successor(iter->pos),
1817                                         iter->flags & BTREE_ITER_INTENT,
1818                                         btree_iter_ip_allocated(iter));
1819
1820                 path = btree_iter_path(trans, iter);
1821                 btree_path_set_level_down(trans, path, iter->min_depth);
1822
1823                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1824                 if (ret)
1825                         goto err;
1826
1827                 path = btree_iter_path(trans, iter);
1828                 b = path->l[path->level].b;
1829         }
1830
1831         bkey_init(&iter->k);
1832         iter->k.p = iter->pos = b->key.k.p;
1833
1834         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1835                                         iter->flags & BTREE_ITER_INTENT,
1836                                         btree_iter_ip_allocated(iter));
1837         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1838         EBUG_ON(btree_iter_path(trans, iter)->uptodate);
1839 out:
1840         bch2_btree_iter_verify_entry_exit(iter);
1841         bch2_btree_iter_verify(iter);
1842
1843         return b;
1844 err:
1845         b = ERR_PTR(ret);
1846         goto out;
1847 }
1848
1849 /* Iterate across keys (in leaf nodes only) */
1850
1851 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1852 {
1853         struct bpos pos = iter->k.p;
1854         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1855                      ? bpos_eq(pos, SPOS_MAX)
1856                      : bkey_eq(pos, SPOS_MAX));
1857
1858         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1859                 pos = bkey_successor(iter, pos);
1860         bch2_btree_iter_set_pos(iter, pos);
1861         return ret;
1862 }
1863
1864 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1865 {
1866         struct bpos pos = bkey_start_pos(&iter->k);
1867         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1868                      ? bpos_eq(pos, POS_MIN)
1869                      : bkey_eq(pos, POS_MIN));
1870
1871         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1872                 pos = bkey_predecessor(iter, pos);
1873         bch2_btree_iter_set_pos(iter, pos);
1874         return ret;
1875 }
1876
1877 static noinline
1878 void bch2_btree_trans_peek_prev_updates(struct btree_trans *trans, struct btree_iter *iter,
1879                                         struct bkey_s_c *k)
1880 {
1881         struct bpos end = path_l(btree_iter_path(trans, iter))->b->data->min_key;
1882
1883         trans_for_each_update(trans, i)
1884                 if (!i->key_cache_already_flushed &&
1885                     i->btree_id == iter->btree_id &&
1886                     bpos_le(i->k->k.p, iter->pos) &&
1887                     bpos_ge(i->k->k.p, k->k ? k->k->p : end)) {
1888                         iter->k = i->k->k;
1889                         *k = bkey_i_to_s_c(i->k);
1890                 }
1891 }
1892
1893 static noinline
1894 void bch2_btree_trans_peek_updates(struct btree_trans *trans, struct btree_iter *iter,
1895                                    struct bkey_s_c *k)
1896 {
1897         struct btree_path *path = btree_iter_path(trans, iter);
1898         struct bpos end = path_l(path)->b->key.k.p;
1899
1900         trans_for_each_update(trans, i)
1901                 if (!i->key_cache_already_flushed &&
1902                     i->btree_id == iter->btree_id &&
1903                     bpos_ge(i->k->k.p, path->pos) &&
1904                     bpos_le(i->k->k.p, k->k ? k->k->p : end)) {
1905                         iter->k = i->k->k;
1906                         *k = bkey_i_to_s_c(i->k);
1907                 }
1908 }
1909
1910 static noinline
1911 void bch2_btree_trans_peek_slot_updates(struct btree_trans *trans, struct btree_iter *iter,
1912                                         struct bkey_s_c *k)
1913 {
1914         trans_for_each_update(trans, i)
1915                 if (!i->key_cache_already_flushed &&
1916                     i->btree_id == iter->btree_id &&
1917                     bpos_eq(i->k->k.p, iter->pos)) {
1918                         iter->k = i->k->k;
1919                         *k = bkey_i_to_s_c(i->k);
1920                 }
1921 }
1922
1923 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1924                                               struct btree_iter *iter,
1925                                               struct bpos end_pos)
1926 {
1927         struct btree_path *path = btree_iter_path(trans, iter);
1928
1929         return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1930                                            path->level,
1931                                            path->pos,
1932                                            end_pos,
1933                                            &iter->journal_idx);
1934 }
1935
1936 static noinline
1937 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1938                                               struct btree_iter *iter)
1939 {
1940         struct btree_path *path = btree_iter_path(trans, iter);
1941         struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos);
1942
1943         if (k) {
1944                 iter->k = k->k;
1945                 return bkey_i_to_s_c(k);
1946         } else {
1947                 return bkey_s_c_null;
1948         }
1949 }
1950
1951 static noinline
1952 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1953                                          struct btree_iter *iter,
1954                                          struct bkey_s_c k)
1955 {
1956         struct btree_path *path = btree_iter_path(trans, iter);
1957         struct bkey_i *next_journal =
1958                 bch2_btree_journal_peek(trans, iter,
1959                                 k.k ? k.k->p : path_l(path)->b->key.k.p);
1960
1961         if (next_journal) {
1962                 iter->k = next_journal->k;
1963                 k = bkey_i_to_s_c(next_journal);
1964         }
1965
1966         return k;
1967 }
1968
1969 /*
1970  * Checks btree key cache for key at iter->pos and returns it if present, or
1971  * bkey_s_c_null:
1972  */
1973 static noinline
1974 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1975 {
1976         struct btree_trans *trans = iter->trans;
1977         struct bch_fs *c = trans->c;
1978         struct bkey u;
1979         struct bkey_s_c k;
1980         int ret;
1981
1982         if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1983             bpos_eq(iter->pos, pos))
1984                 return bkey_s_c_null;
1985
1986         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1987                 return bkey_s_c_null;
1988
1989         if (!iter->key_cache_path)
1990                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1991                                                      iter->flags & BTREE_ITER_INTENT, 0,
1992                                                      iter->flags|BTREE_ITER_CACHED|
1993                                                      BTREE_ITER_CACHED_NOFILL,
1994                                                      _THIS_IP_);
1995
1996         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1997                                         iter->flags & BTREE_ITER_INTENT,
1998                                         btree_iter_ip_allocated(iter));
1999
2000         ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
2001                                          iter->flags|BTREE_ITER_CACHED) ?:
2002                 bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_);
2003         if (unlikely(ret))
2004                 return bkey_s_c_err(ret);
2005
2006         btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
2007
2008         k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
2009         if (k.k && !bkey_err(k)) {
2010                 iter->k = u;
2011                 k.k = &iter->k;
2012         }
2013         return k;
2014 }
2015
2016 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
2017 {
2018         struct btree_trans *trans = iter->trans;
2019         struct bkey_s_c k, k2;
2020         int ret;
2021
2022         EBUG_ON(btree_iter_path(trans, iter)->cached);
2023         bch2_btree_iter_verify(iter);
2024
2025         while (1) {
2026                 struct btree_path_level *l;
2027
2028                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2029                                         iter->flags & BTREE_ITER_INTENT,
2030                                         btree_iter_ip_allocated(iter));
2031
2032                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2033                 if (unlikely(ret)) {
2034                         /* ensure that iter->k is consistent with iter->pos: */
2035                         bch2_btree_iter_set_pos(iter, iter->pos);
2036                         k = bkey_s_c_err(ret);
2037                         goto out;
2038                 }
2039
2040                 struct btree_path *path = btree_iter_path(trans, iter);
2041                 l = path_l(path);
2042
2043                 if (unlikely(!l->b)) {
2044                         /* No btree nodes at requested level: */
2045                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2046                         k = bkey_s_c_null;
2047                         goto out;
2048                 }
2049
2050                 btree_path_set_should_be_locked(path);
2051
2052                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
2053
2054                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2055                     k.k &&
2056                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2057                         k = k2;
2058                         ret = bkey_err(k);
2059                         if (ret) {
2060                                 bch2_btree_iter_set_pos(iter, iter->pos);
2061                                 goto out;
2062                         }
2063                 }
2064
2065                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2066                         k = btree_trans_peek_journal(trans, iter, k);
2067
2068                 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2069                              trans->nr_updates))
2070                         bch2_btree_trans_peek_updates(trans, iter, &k);
2071
2072                 if (k.k && bkey_deleted(k.k)) {
2073                         /*
2074                          * If we've got a whiteout, and it's after the search
2075                          * key, advance the search key to the whiteout instead
2076                          * of just after the whiteout - it might be a btree
2077                          * whiteout, with a real key at the same position, since
2078                          * in the btree deleted keys sort before non deleted.
2079                          */
2080                         search_key = !bpos_eq(search_key, k.k->p)
2081                                 ? k.k->p
2082                                 : bpos_successor(k.k->p);
2083                         continue;
2084                 }
2085
2086                 if (likely(k.k)) {
2087                         break;
2088                 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2089                         /* Advance to next leaf node: */
2090                         search_key = bpos_successor(l->b->key.k.p);
2091                 } else {
2092                         /* End of btree: */
2093                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2094                         k = bkey_s_c_null;
2095                         goto out;
2096                 }
2097         }
2098 out:
2099         bch2_btree_iter_verify(iter);
2100
2101         return k;
2102 }
2103
2104 /**
2105  * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
2106  * iterator's current position
2107  * @iter:       iterator to peek from
2108  * @end:        search limit: returns keys less than or equal to @end
2109  *
2110  * Returns:     key if found, or an error extractable with bkey_err().
2111  */
2112 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2113 {
2114         struct btree_trans *trans = iter->trans;
2115         struct bpos search_key = btree_iter_search_key(iter);
2116         struct bkey_s_c k;
2117         struct bpos iter_pos;
2118         int ret;
2119
2120         EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2121
2122         if (iter->update_path) {
2123                 bch2_path_put_nokeep(trans, iter->update_path,
2124                                      iter->flags & BTREE_ITER_INTENT);
2125                 iter->update_path = 0;
2126         }
2127
2128         bch2_btree_iter_verify_entry_exit(iter);
2129
2130         while (1) {
2131                 k = __bch2_btree_iter_peek(iter, search_key);
2132                 if (unlikely(!k.k))
2133                         goto end;
2134                 if (unlikely(bkey_err(k)))
2135                         goto out_no_locked;
2136
2137                 /*
2138                  * iter->pos should be mononotically increasing, and always be
2139                  * equal to the key we just returned - except extents can
2140                  * straddle iter->pos:
2141                  */
2142                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2143                         iter_pos = k.k->p;
2144                 else
2145                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2146
2147                 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2148                              ? bkey_gt(iter_pos, end)
2149                              : bkey_ge(iter_pos, end)))
2150                         goto end;
2151
2152                 if (iter->update_path &&
2153                     !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
2154                         bch2_path_put_nokeep(trans, iter->update_path,
2155                                              iter->flags & BTREE_ITER_INTENT);
2156                         iter->update_path = 0;
2157                 }
2158
2159                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2160                     (iter->flags & BTREE_ITER_INTENT) &&
2161                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2162                     !iter->update_path) {
2163                         struct bpos pos = k.k->p;
2164
2165                         if (pos.snapshot < iter->snapshot) {
2166                                 search_key = bpos_successor(k.k->p);
2167                                 continue;
2168                         }
2169
2170                         pos.snapshot = iter->snapshot;
2171
2172                         /*
2173                          * advance, same as on exit for iter->path, but only up
2174                          * to snapshot
2175                          */
2176                         __btree_path_get(trans->paths + iter->path, iter->flags & BTREE_ITER_INTENT);
2177                         iter->update_path = iter->path;
2178
2179                         iter->update_path = bch2_btree_path_set_pos(trans,
2180                                                 iter->update_path, pos,
2181                                                 iter->flags & BTREE_ITER_INTENT,
2182                                                 _THIS_IP_);
2183                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2184                         if (unlikely(ret)) {
2185                                 k = bkey_s_c_err(ret);
2186                                 goto out_no_locked;
2187                         }
2188                 }
2189
2190                 /*
2191                  * We can never have a key in a leaf node at POS_MAX, so
2192                  * we don't have to check these successor() calls:
2193                  */
2194                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2195                     !bch2_snapshot_is_ancestor(trans->c,
2196                                                iter->snapshot,
2197                                                k.k->p.snapshot)) {
2198                         search_key = bpos_successor(k.k->p);
2199                         continue;
2200                 }
2201
2202                 if (bkey_whiteout(k.k) &&
2203                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2204                         search_key = bkey_successor(iter, k.k->p);
2205                         continue;
2206                 }
2207
2208                 break;
2209         }
2210
2211         iter->pos = iter_pos;
2212
2213         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2214                                 iter->flags & BTREE_ITER_INTENT,
2215                                 btree_iter_ip_allocated(iter));
2216
2217         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2218 out_no_locked:
2219         if (iter->update_path) {
2220                 ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_);
2221                 if (unlikely(ret))
2222                         k = bkey_s_c_err(ret);
2223                 else
2224                         btree_path_set_should_be_locked(trans->paths + iter->update_path);
2225         }
2226
2227         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2228                 iter->pos.snapshot = iter->snapshot;
2229
2230         ret = bch2_btree_iter_verify_ret(iter, k);
2231         if (unlikely(ret)) {
2232                 bch2_btree_iter_set_pos(iter, iter->pos);
2233                 k = bkey_s_c_err(ret);
2234         }
2235
2236         bch2_btree_iter_verify_entry_exit(iter);
2237
2238         return k;
2239 end:
2240         bch2_btree_iter_set_pos(iter, end);
2241         k = bkey_s_c_null;
2242         goto out_no_locked;
2243 }
2244
2245 /**
2246  * bch2_btree_iter_next() - returns first key greater than iterator's current
2247  * position
2248  * @iter:       iterator to peek from
2249  *
2250  * Returns:     key if found, or an error extractable with bkey_err().
2251  */
2252 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2253 {
2254         if (!bch2_btree_iter_advance(iter))
2255                 return bkey_s_c_null;
2256
2257         return bch2_btree_iter_peek(iter);
2258 }
2259
2260 /**
2261  * bch2_btree_iter_peek_prev() - returns first key less than or equal to
2262  * iterator's current position
2263  * @iter:       iterator to peek from
2264  *
2265  * Returns:     key if found, or an error extractable with bkey_err().
2266  */
2267 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2268 {
2269         struct btree_trans *trans = iter->trans;
2270         struct bpos search_key = iter->pos;
2271         struct bkey_s_c k;
2272         struct bkey saved_k;
2273         const struct bch_val *saved_v;
2274         btree_path_idx_t saved_path = 0;
2275         int ret;
2276
2277         EBUG_ON(btree_iter_path(trans, iter)->cached ||
2278                 btree_iter_path(trans, iter)->level);
2279
2280         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2281                 return bkey_s_c_err(-EIO);
2282
2283         bch2_btree_iter_verify(iter);
2284         bch2_btree_iter_verify_entry_exit(iter);
2285
2286         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2287                 search_key.snapshot = U32_MAX;
2288
2289         while (1) {
2290                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2291                                                 iter->flags & BTREE_ITER_INTENT,
2292                                                 btree_iter_ip_allocated(iter));
2293
2294                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2295                 if (unlikely(ret)) {
2296                         /* ensure that iter->k is consistent with iter->pos: */
2297                         bch2_btree_iter_set_pos(iter, iter->pos);
2298                         k = bkey_s_c_err(ret);
2299                         goto out_no_locked;
2300                 }
2301
2302                 struct btree_path *path = btree_iter_path(trans, iter);
2303
2304                 k = btree_path_level_peek(trans, path, &path->l[0], &iter->k);
2305                 if (!k.k ||
2306                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2307                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2308                      : bpos_gt(k.k->p, search_key)))
2309                         k = btree_path_level_prev(trans, path, &path->l[0], &iter->k);
2310
2311                 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2312                              trans->nr_updates))
2313                         bch2_btree_trans_peek_prev_updates(trans, iter, &k);
2314
2315                 if (likely(k.k)) {
2316                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2317                                 if (k.k->p.snapshot == iter->snapshot)
2318                                         goto got_key;
2319
2320                                 /*
2321                                  * If we have a saved candidate, and we're no
2322                                  * longer at the same _key_ (not pos), return
2323                                  * that candidate
2324                                  */
2325                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2326                                         bch2_path_put_nokeep(trans, iter->path,
2327                                                       iter->flags & BTREE_ITER_INTENT);
2328                                         iter->path = saved_path;
2329                                         saved_path = 0;
2330                                         iter->k = saved_k;
2331                                         k.v     = saved_v;
2332                                         goto got_key;
2333                                 }
2334
2335                                 if (bch2_snapshot_is_ancestor(trans->c,
2336                                                               iter->snapshot,
2337                                                               k.k->p.snapshot)) {
2338                                         if (saved_path)
2339                                                 bch2_path_put_nokeep(trans, saved_path,
2340                                                       iter->flags & BTREE_ITER_INTENT);
2341                                         saved_path = btree_path_clone(trans, iter->path,
2342                                                                 iter->flags & BTREE_ITER_INTENT);
2343                                         path = btree_iter_path(trans, iter);
2344                                         saved_k = *k.k;
2345                                         saved_v = k.v;
2346                                 }
2347
2348                                 search_key = bpos_predecessor(k.k->p);
2349                                 continue;
2350                         }
2351 got_key:
2352                         if (bkey_whiteout(k.k) &&
2353                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2354                                 search_key = bkey_predecessor(iter, k.k->p);
2355                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2356                                         search_key.snapshot = U32_MAX;
2357                                 continue;
2358                         }
2359
2360                         btree_path_set_should_be_locked(path);
2361                         break;
2362                 } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
2363                         /* Advance to previous leaf node: */
2364                         search_key = bpos_predecessor(path->l[0].b->data->min_key);
2365                 } else {
2366                         /* Start of btree: */
2367                         bch2_btree_iter_set_pos(iter, POS_MIN);
2368                         k = bkey_s_c_null;
2369                         goto out_no_locked;
2370                 }
2371         }
2372
2373         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2374
2375         /* Extents can straddle iter->pos: */
2376         if (bkey_lt(k.k->p, iter->pos))
2377                 iter->pos = k.k->p;
2378
2379         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2380                 iter->pos.snapshot = iter->snapshot;
2381 out_no_locked:
2382         if (saved_path)
2383                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2384
2385         bch2_btree_iter_verify_entry_exit(iter);
2386         bch2_btree_iter_verify(iter);
2387
2388         return k;
2389 }
2390
2391 /**
2392  * bch2_btree_iter_prev() - returns first key less than iterator's current
2393  * position
2394  * @iter:       iterator to peek from
2395  *
2396  * Returns:     key if found, or an error extractable with bkey_err().
2397  */
2398 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2399 {
2400         if (!bch2_btree_iter_rewind(iter))
2401                 return bkey_s_c_null;
2402
2403         return bch2_btree_iter_peek_prev(iter);
2404 }
2405
2406 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2407 {
2408         struct btree_trans *trans = iter->trans;
2409         struct bpos search_key;
2410         struct bkey_s_c k;
2411         int ret;
2412
2413         bch2_btree_iter_verify(iter);
2414         bch2_btree_iter_verify_entry_exit(iter);
2415         EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2416
2417         /* extents can't span inode numbers: */
2418         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2419             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2420                 if (iter->pos.inode == KEY_INODE_MAX)
2421                         return bkey_s_c_null;
2422
2423                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2424         }
2425
2426         search_key = btree_iter_search_key(iter);
2427         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2428                                         iter->flags & BTREE_ITER_INTENT,
2429                                         btree_iter_ip_allocated(iter));
2430
2431         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2432         if (unlikely(ret)) {
2433                 k = bkey_s_c_err(ret);
2434                 goto out_no_locked;
2435         }
2436
2437         if ((iter->flags & BTREE_ITER_CACHED) ||
2438             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2439                 k = bkey_s_c_null;
2440
2441                 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2442                              trans->nr_updates)) {
2443                         bch2_btree_trans_peek_slot_updates(trans, iter, &k);
2444                         if (k.k)
2445                                 goto out;
2446                 }
2447
2448                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2449                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2450                         goto out;
2451
2452                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2453                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2454                         if (!bkey_err(k))
2455                                 iter->k = *k.k;
2456                         /* We're not returning a key from iter->path: */
2457                         goto out_no_locked;
2458                 }
2459
2460                 k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k);
2461                 if (unlikely(!k.k))
2462                         goto out_no_locked;
2463         } else {
2464                 struct bpos next;
2465                 struct bpos end = iter->pos;
2466
2467                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2468                         end.offset = U64_MAX;
2469
2470                 EBUG_ON(btree_iter_path(trans, iter)->level);
2471
2472                 if (iter->flags & BTREE_ITER_INTENT) {
2473                         struct btree_iter iter2;
2474
2475                         bch2_trans_copy_iter(&iter2, iter);
2476                         k = bch2_btree_iter_peek_upto(&iter2, end);
2477
2478                         if (k.k && !bkey_err(k)) {
2479                                 iter->k = iter2.k;
2480                                 k.k = &iter->k;
2481                         }
2482                         bch2_trans_iter_exit(trans, &iter2);
2483                 } else {
2484                         struct bpos pos = iter->pos;
2485
2486                         k = bch2_btree_iter_peek_upto(iter, end);
2487                         if (unlikely(bkey_err(k)))
2488                                 bch2_btree_iter_set_pos(iter, pos);
2489                         else
2490                                 iter->pos = pos;
2491                 }
2492
2493                 if (unlikely(bkey_err(k)))
2494                         goto out_no_locked;
2495
2496                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2497
2498                 if (bkey_lt(iter->pos, next)) {
2499                         bkey_init(&iter->k);
2500                         iter->k.p = iter->pos;
2501
2502                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2503                                 bch2_key_resize(&iter->k,
2504                                                 min_t(u64, KEY_SIZE_MAX,
2505                                                       (next.inode == iter->pos.inode
2506                                                        ? next.offset
2507                                                        : KEY_OFFSET_MAX) -
2508                                                       iter->pos.offset));
2509                                 EBUG_ON(!iter->k.size);
2510                         }
2511
2512                         k = (struct bkey_s_c) { &iter->k, NULL };
2513                 }
2514         }
2515 out:
2516         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2517 out_no_locked:
2518         bch2_btree_iter_verify_entry_exit(iter);
2519         bch2_btree_iter_verify(iter);
2520         ret = bch2_btree_iter_verify_ret(iter, k);
2521         if (unlikely(ret))
2522                 return bkey_s_c_err(ret);
2523
2524         return k;
2525 }
2526
2527 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2528 {
2529         if (!bch2_btree_iter_advance(iter))
2530                 return bkey_s_c_null;
2531
2532         return bch2_btree_iter_peek_slot(iter);
2533 }
2534
2535 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2536 {
2537         if (!bch2_btree_iter_rewind(iter))
2538                 return bkey_s_c_null;
2539
2540         return bch2_btree_iter_peek_slot(iter);
2541 }
2542
2543 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2544 {
2545         struct bkey_s_c k;
2546
2547         while (btree_trans_too_many_iters(iter->trans) ||
2548                (k = bch2_btree_iter_peek_type(iter, iter->flags),
2549                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2550                 bch2_trans_begin(iter->trans);
2551
2552         return k;
2553 }
2554
2555 /* new transactional stuff: */
2556
2557 #ifdef CONFIG_BCACHEFS_DEBUG
2558 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2559 {
2560         struct btree_path *path;
2561         unsigned i;
2562
2563         BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1);
2564
2565         trans_for_each_path(trans, path, i) {
2566                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2567                 BUG_ON(trans->sorted[path->sorted_idx] != i);
2568         }
2569
2570         for (i = 0; i < trans->nr_sorted; i++) {
2571                 unsigned idx = trans->sorted[i];
2572
2573                 BUG_ON(!test_bit(idx, trans->paths_allocated));
2574                 BUG_ON(trans->paths[idx].sorted_idx != i);
2575         }
2576 }
2577
2578 static void btree_trans_verify_sorted(struct btree_trans *trans)
2579 {
2580         struct btree_path *path, *prev = NULL;
2581         struct trans_for_each_path_inorder_iter iter;
2582
2583         if (!bch2_debug_check_iterators)
2584                 return;
2585
2586         trans_for_each_path_inorder(trans, path, iter) {
2587                 if (prev && btree_path_cmp(prev, path) > 0) {
2588                         __bch2_dump_trans_paths_updates(trans, true);
2589                         panic("trans paths out of order!\n");
2590                 }
2591                 prev = path;
2592         }
2593 }
2594 #else
2595 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2596 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2597 #endif
2598
2599 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2600 {
2601         int i, l = 0, r = trans->nr_sorted, inc = 1;
2602         bool swapped;
2603
2604         btree_trans_verify_sorted_refs(trans);
2605
2606         if (trans->paths_sorted)
2607                 goto out;
2608
2609         /*
2610          * Cocktail shaker sort: this is efficient because iterators will be
2611          * mostly sorted.
2612          */
2613         do {
2614                 swapped = false;
2615
2616                 for (i = inc > 0 ? l : r - 2;
2617                      i + 1 < r && i >= l;
2618                      i += inc) {
2619                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2620                                            trans->paths + trans->sorted[i + 1]) > 0) {
2621                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2622                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2623                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2624                                 swapped = true;
2625                         }
2626                 }
2627
2628                 if (inc > 0)
2629                         --r;
2630                 else
2631                         l++;
2632                 inc = -inc;
2633         } while (swapped);
2634
2635         trans->paths_sorted = true;
2636 out:
2637         btree_trans_verify_sorted(trans);
2638 }
2639
2640 static inline void btree_path_list_remove(struct btree_trans *trans,
2641                                           struct btree_path *path)
2642 {
2643         unsigned i;
2644
2645         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2646 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2647         trans->nr_sorted--;
2648         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2649                                 trans->sorted + path->sorted_idx + 1,
2650                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2651 #else
2652         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2653 #endif
2654         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2655                 trans->paths[trans->sorted[i]].sorted_idx = i;
2656
2657         path->sorted_idx = U8_MAX;
2658 }
2659
2660 static inline void btree_path_list_add(struct btree_trans *trans,
2661                                        btree_path_idx_t pos,
2662                                        btree_path_idx_t path_idx)
2663 {
2664         struct btree_path *path = trans->paths + path_idx;
2665         unsigned i;
2666
2667         path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted;
2668
2669 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2670         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2671                               trans->sorted + path->sorted_idx,
2672                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8));
2673         trans->nr_sorted++;
2674         trans->sorted[path->sorted_idx] = path_idx;
2675 #else
2676         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx);
2677 #endif
2678
2679         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2680                 trans->paths[trans->sorted[i]].sorted_idx = i;
2681
2682         btree_trans_verify_sorted_refs(trans);
2683 }
2684
2685 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2686 {
2687         if (iter->update_path)
2688                 bch2_path_put_nokeep(trans, iter->update_path,
2689                               iter->flags & BTREE_ITER_INTENT);
2690         if (iter->path)
2691                 bch2_path_put(trans, iter->path,
2692                               iter->flags & BTREE_ITER_INTENT);
2693         if (iter->key_cache_path)
2694                 bch2_path_put(trans, iter->key_cache_path,
2695                               iter->flags & BTREE_ITER_INTENT);
2696         iter->path              = 0;
2697         iter->update_path       = 0;
2698         iter->key_cache_path    = 0;
2699         iter->trans             = NULL;
2700 }
2701
2702 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2703                           struct btree_iter *iter,
2704                           enum btree_id btree_id, struct bpos pos,
2705                           unsigned flags)
2706 {
2707         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2708                                bch2_btree_iter_flags(trans, btree_id, flags),
2709                                _RET_IP_);
2710 }
2711
2712 void bch2_trans_node_iter_init(struct btree_trans *trans,
2713                                struct btree_iter *iter,
2714                                enum btree_id btree_id,
2715                                struct bpos pos,
2716                                unsigned locks_want,
2717                                unsigned depth,
2718                                unsigned flags)
2719 {
2720         flags |= BTREE_ITER_NOT_EXTENTS;
2721         flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2722         flags |= BTREE_ITER_ALL_SNAPSHOTS;
2723
2724         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2725                                __bch2_btree_iter_flags(trans, btree_id, flags),
2726                                _RET_IP_);
2727
2728         iter->min_depth = depth;
2729
2730         struct btree_path *path = btree_iter_path(trans, iter);
2731         BUG_ON(path->locks_want  < min(locks_want, BTREE_MAX_DEPTH));
2732         BUG_ON(path->level      != depth);
2733         BUG_ON(iter->min_depth  != depth);
2734 }
2735
2736 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2737 {
2738         struct btree_trans *trans = src->trans;
2739
2740         *dst = *src;
2741         if (src->path)
2742                 __btree_path_get(trans->paths + src->path, src->flags & BTREE_ITER_INTENT);
2743         if (src->update_path)
2744                 __btree_path_get(trans->paths + src->update_path, src->flags & BTREE_ITER_INTENT);
2745         dst->key_cache_path = 0;
2746 }
2747
2748 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2749 {
2750         struct bch_fs *c = trans->c;
2751         unsigned new_top = trans->mem_top + size;
2752         unsigned old_bytes = trans->mem_bytes;
2753         unsigned new_bytes = roundup_pow_of_two(new_top);
2754         int ret;
2755         void *new_mem;
2756         void *p;
2757
2758         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2759
2760         struct btree_transaction_stats *s = btree_trans_stats(trans);
2761         if (s)
2762                 s->max_mem = max(s->max_mem, new_bytes);
2763
2764         new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2765         if (unlikely(!new_mem)) {
2766                 bch2_trans_unlock(trans);
2767
2768                 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2769                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2770                         new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2771                         new_bytes = BTREE_TRANS_MEM_MAX;
2772                         kfree(trans->mem);
2773                 }
2774
2775                 if (!new_mem)
2776                         return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2777
2778                 trans->mem = new_mem;
2779                 trans->mem_bytes = new_bytes;
2780
2781                 ret = bch2_trans_relock(trans);
2782                 if (ret)
2783                         return ERR_PTR(ret);
2784         }
2785
2786         trans->mem = new_mem;
2787         trans->mem_bytes = new_bytes;
2788
2789         if (old_bytes) {
2790                 trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2791                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2792         }
2793
2794         p = trans->mem + trans->mem_top;
2795         trans->mem_top += size;
2796         memset(p, 0, size);
2797         return p;
2798 }
2799
2800 static inline void check_srcu_held_too_long(struct btree_trans *trans)
2801 {
2802         WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
2803              "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
2804              (jiffies - trans->srcu_lock_time) / HZ);
2805 }
2806
2807 void bch2_trans_srcu_unlock(struct btree_trans *trans)
2808 {
2809         if (trans->srcu_held) {
2810                 struct bch_fs *c = trans->c;
2811                 struct btree_path *path;
2812                 unsigned i;
2813
2814                 trans_for_each_path(trans, path, i)
2815                         if (path->cached && !btree_node_locked(path, 0))
2816                                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2817
2818                 check_srcu_held_too_long(trans);
2819                 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2820                 trans->srcu_held = false;
2821         }
2822 }
2823
2824 static void bch2_trans_srcu_lock(struct btree_trans *trans)
2825 {
2826         if (!trans->srcu_held) {
2827                 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
2828                 trans->srcu_lock_time   = jiffies;
2829                 trans->srcu_held = true;
2830         }
2831 }
2832
2833 /**
2834  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2835  * @trans: transaction to reset
2836  *
2837  * Returns:     current restart counter, to be used with trans_was_restarted()
2838  *
2839  * While iterating over nodes or updating nodes a attempt to lock a btree node
2840  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2841  * occurs bch2_trans_begin() should be called and the transaction retried.
2842  */
2843 u32 bch2_trans_begin(struct btree_trans *trans)
2844 {
2845         struct btree_path *path;
2846         unsigned i;
2847         u64 now;
2848
2849         bch2_trans_reset_updates(trans);
2850
2851         trans->restart_count++;
2852         trans->mem_top                  = 0;
2853         trans->journal_entries          = NULL;
2854
2855         trans_for_each_path(trans, path, i) {
2856                 path->should_be_locked = false;
2857
2858                 /*
2859                  * If the transaction wasn't restarted, we're presuming to be
2860                  * doing something new: dont keep iterators excpt the ones that
2861                  * are in use - except for the subvolumes btree:
2862                  */
2863                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2864                         path->preserve = false;
2865
2866                 /*
2867                  * XXX: we probably shouldn't be doing this if the transaction
2868                  * was restarted, but currently we still overflow transaction
2869                  * iterators if we do that
2870                  */
2871                 if (!path->ref && !path->preserve)
2872                         __bch2_path_free(trans, i);
2873                 else
2874                         path->preserve = false;
2875         }
2876
2877         now = local_clock();
2878         if (!trans->restarted &&
2879             (need_resched() ||
2880              now - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2881                 drop_locks_do(trans, (cond_resched(), 0));
2882                 now = local_clock();
2883         }
2884         trans->last_begin_time = now;
2885
2886         if (unlikely(trans->srcu_held &&
2887                      time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2888                 bch2_trans_srcu_unlock(trans);
2889
2890         trans->last_begin_ip = _RET_IP_;
2891         if (trans->restarted) {
2892                 bch2_btree_path_traverse_all(trans);
2893                 trans->notrace_relock_fail = false;
2894         }
2895
2896         return trans->restart_count;
2897 }
2898
2899 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2900
2901 unsigned bch2_trans_get_fn_idx(const char *fn)
2902 {
2903         unsigned i;
2904
2905         for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2906                 if (!bch2_btree_transaction_fns[i] ||
2907                     bch2_btree_transaction_fns[i] == fn) {
2908                         bch2_btree_transaction_fns[i] = fn;
2909                         return i;
2910                 }
2911
2912         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2913         return i;
2914 }
2915
2916 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
2917         __acquires(&c->btree_trans_barrier)
2918 {
2919         struct btree_trans *trans;
2920
2921         if (IS_ENABLED(__KERNEL__)) {
2922                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
2923                 if (trans) {
2924                         memset(trans, 0, offsetof(struct btree_trans, list));
2925                         goto got_trans;
2926                 }
2927         }
2928
2929         trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
2930         memset(trans, 0, sizeof(*trans));
2931         closure_init_stack(&trans->ref);
2932
2933         seqmutex_lock(&c->btree_trans_lock);
2934         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
2935                 struct btree_trans *pos;
2936                 pid_t pid = current->pid;
2937
2938                 trans->locking_wait.task = current;
2939
2940                 list_for_each_entry(pos, &c->btree_trans_list, list) {
2941                         struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task);
2942                         /*
2943                          * We'd much prefer to be stricter here and completely
2944                          * disallow multiple btree_trans in the same thread -
2945                          * but the data move path calls bch2_write when we
2946                          * already have a btree_trans initialized.
2947                          */
2948                         BUG_ON(pos_task &&
2949                                pid == pos_task->pid &&
2950                                bch2_trans_locked(pos));
2951
2952                         if (pos_task && pid < pos_task->pid) {
2953                                 list_add_tail(&trans->list, &pos->list);
2954                                 goto list_add_done;
2955                         }
2956                 }
2957         }
2958         list_add_tail(&trans->list, &c->btree_trans_list);
2959 list_add_done:
2960         seqmutex_unlock(&c->btree_trans_lock);
2961 got_trans:
2962         trans->c                = c;
2963         trans->last_begin_time  = local_clock();
2964         trans->fn_idx           = fn_idx;
2965         trans->locking_wait.task = current;
2966         trans->journal_replay_not_finished =
2967                 unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) &&
2968                 atomic_inc_not_zero(&c->journal_keys.ref);
2969         trans->nr_paths         = ARRAY_SIZE(trans->_paths);
2970         trans->paths_allocated  = trans->_paths_allocated;
2971         trans->sorted           = trans->_sorted;
2972         trans->paths            = trans->_paths;
2973         trans->updates          = trans->_updates;
2974
2975         *trans_paths_nr(trans->paths) = BTREE_ITER_MAX;
2976
2977         trans->paths_allocated[0] = 1;
2978
2979         if (fn_idx < BCH_TRANSACTIONS_NR) {
2980                 trans->fn = bch2_btree_transaction_fns[fn_idx];
2981
2982                 struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx];
2983
2984                 if (s->max_mem) {
2985                         unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2986
2987                         trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2988                         if (likely(trans->mem))
2989                                 trans->mem_bytes = expected_mem_bytes;
2990                 }
2991
2992                 trans->nr_paths_max = s->nr_max_paths;
2993                 trans->journal_entries_size = s->journal_entries_size;
2994         }
2995
2996         trans->srcu_idx         = srcu_read_lock(&c->btree_trans_barrier);
2997         trans->srcu_lock_time   = jiffies;
2998         trans->srcu_held        = true;
2999         return trans;
3000 }
3001
3002 static void check_btree_paths_leaked(struct btree_trans *trans)
3003 {
3004 #ifdef CONFIG_BCACHEFS_DEBUG
3005         struct bch_fs *c = trans->c;
3006         struct btree_path *path;
3007         unsigned i;
3008
3009         trans_for_each_path(trans, path, i)
3010                 if (path->ref)
3011                         goto leaked;
3012         return;
3013 leaked:
3014         bch_err(c, "btree paths leaked from %s!", trans->fn);
3015         trans_for_each_path(trans, path, i)
3016                 if (path->ref)
3017                         printk(KERN_ERR "  btree %s %pS\n",
3018                                bch2_btree_id_str(path->btree_id),
3019                                (void *) path->ip_allocated);
3020         /* Be noisy about this: */
3021         bch2_fatal_error(c);
3022 #endif
3023 }
3024
3025 void bch2_trans_put(struct btree_trans *trans)
3026         __releases(&c->btree_trans_barrier)
3027 {
3028         struct bch_fs *c = trans->c;
3029
3030         bch2_trans_unlock(trans);
3031
3032         trans_for_each_update(trans, i)
3033                 __btree_path_put(trans->paths + i->path, true);
3034         trans->nr_updates       = 0;
3035         trans->locking_wait.task = NULL;
3036
3037         check_btree_paths_leaked(trans);
3038
3039         if (trans->srcu_held) {
3040                 check_srcu_held_too_long(trans);
3041                 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3042         }
3043
3044         if (trans->fs_usage_deltas) {
3045                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3046                     REPLICAS_DELTA_LIST_MAX)
3047                         mempool_free(trans->fs_usage_deltas,
3048                                      &c->replicas_delta_pool);
3049                 else
3050                         kfree(trans->fs_usage_deltas);
3051         }
3052
3053         if (unlikely(trans->journal_replay_not_finished))
3054                 bch2_journal_keys_put(c);
3055
3056         unsigned long *paths_allocated = trans->paths_allocated;
3057         trans->paths_allocated  = NULL;
3058         trans->paths            = NULL;
3059
3060         if (paths_allocated != trans->_paths_allocated)
3061                 kfree_rcu_mightsleep(paths_allocated);
3062
3063         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3064                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3065         else
3066                 kfree(trans->mem);
3067
3068         /* Userspace doesn't have a real percpu implementation: */
3069         if (IS_ENABLED(__KERNEL__))
3070                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
3071
3072         if (trans) {
3073                 closure_sync(&trans->ref);
3074
3075                 seqmutex_lock(&c->btree_trans_lock);
3076                 list_del(&trans->list);
3077                 seqmutex_unlock(&c->btree_trans_lock);
3078
3079                 mempool_free(trans, &c->btree_trans_pool);
3080         }
3081 }
3082
3083 static void __maybe_unused
3084 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3085                                       struct btree_bkey_cached_common *b)
3086 {
3087         struct six_lock_count c = six_lock_counts(&b->lock);
3088         struct task_struct *owner;
3089         pid_t pid;
3090
3091         rcu_read_lock();
3092         owner = READ_ONCE(b->lock.owner);
3093         pid = owner ? owner->pid : 0;
3094         rcu_read_unlock();
3095
3096         prt_tab(out);
3097         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3098                    b->level, bch2_btree_id_str(b->btree_id));
3099         bch2_bpos_to_text(out, btree_node_pos(b));
3100
3101         prt_tab(out);
3102         prt_printf(out, " locks %u:%u:%u held by pid %u",
3103                    c.n[0], c.n[1], c.n[2], pid);
3104 }
3105
3106 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3107 {
3108         struct btree_bkey_cached_common *b;
3109         static char lock_types[] = { 'r', 'i', 'w' };
3110         struct task_struct *task = READ_ONCE(trans->locking_wait.task);
3111         unsigned l, idx;
3112
3113         /* before rcu_read_lock(): */
3114         bch2_printbuf_make_room(out, 4096);
3115
3116         if (!out->nr_tabstops) {
3117                 printbuf_tabstop_push(out, 16);
3118                 printbuf_tabstop_push(out, 32);
3119         }
3120
3121         prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);
3122
3123         /* trans->paths is rcu protected vs. freeing */
3124         rcu_read_lock();
3125         out->atomic++;
3126
3127         struct btree_path *paths = rcu_dereference(trans->paths);
3128         if (!paths)
3129                 goto out;
3130
3131         unsigned long *paths_allocated = trans_paths_allocated(paths);
3132
3133         trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) {
3134                 struct btree_path *path = paths + idx;
3135                 if (!path->nodes_locked)
3136                         continue;
3137
3138                 prt_printf(out, "  path %u %c l=%u %s:",
3139                        idx,
3140                        path->cached ? 'c' : 'b',
3141                        path->level,
3142                        bch2_btree_id_str(path->btree_id));
3143                 bch2_bpos_to_text(out, path->pos);
3144                 prt_newline(out);
3145
3146                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3147                         if (btree_node_locked(path, l) &&
3148                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3149                                 prt_printf(out, "    %c l=%u ",
3150                                            lock_types[btree_node_locked_type(path, l)], l);
3151                                 bch2_btree_bkey_cached_common_to_text(out, b);
3152                                 prt_newline(out);
3153                         }
3154                 }
3155         }
3156
3157         b = READ_ONCE(trans->locking);
3158         if (b) {
3159                 prt_printf(out, "  blocked for %lluus on",
3160                            div_u64(local_clock() - trans->locking_wait.start_time,
3161                                    1000));
3162                 prt_newline(out);
3163                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3164                 bch2_btree_bkey_cached_common_to_text(out, b);
3165                 prt_newline(out);
3166         }
3167 out:
3168         --out->atomic;
3169         rcu_read_unlock();
3170 }
3171
3172 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3173 {
3174         struct btree_transaction_stats *s;
3175         struct btree_trans *trans;
3176         int cpu;
3177
3178         if (c->btree_trans_bufs)
3179                 for_each_possible_cpu(cpu) {
3180                         struct btree_trans *trans =
3181                                 per_cpu_ptr(c->btree_trans_bufs, cpu)->trans;
3182
3183                         if (trans) {
3184                                 closure_sync(&trans->ref);
3185
3186                                 seqmutex_lock(&c->btree_trans_lock);
3187                                 list_del(&trans->list);
3188                                 seqmutex_unlock(&c->btree_trans_lock);
3189                         }
3190                         kfree(trans);
3191                 }
3192         free_percpu(c->btree_trans_bufs);
3193
3194         trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3195         if (trans)
3196                 panic("%s leaked btree_trans\n", trans->fn);
3197
3198         for (s = c->btree_transaction_stats;
3199              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3200              s++) {
3201                 kfree(s->max_paths_text);
3202                 bch2_time_stats_exit(&s->lock_hold_times);
3203         }
3204
3205         if (c->btree_trans_barrier_initialized)
3206                 cleanup_srcu_struct(&c->btree_trans_barrier);
3207         mempool_exit(&c->btree_trans_mem_pool);
3208         mempool_exit(&c->btree_trans_pool);
3209 }
3210
3211 void bch2_fs_btree_iter_init_early(struct bch_fs *c)
3212 {
3213         struct btree_transaction_stats *s;
3214
3215         for (s = c->btree_transaction_stats;
3216              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3217              s++) {
3218                 bch2_time_stats_init(&s->lock_hold_times);
3219                 mutex_init(&s->lock);
3220         }
3221
3222         INIT_LIST_HEAD(&c->btree_trans_list);
3223         seqmutex_init(&c->btree_trans_lock);
3224 }
3225
3226 int bch2_fs_btree_iter_init(struct bch_fs *c)
3227 {
3228         int ret;
3229
3230         c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3231         if (!c->btree_trans_bufs)
3232                 return -ENOMEM;
3233
3234         ret   = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3235                                           sizeof(struct btree_trans)) ?:
3236                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3237                                           BTREE_TRANS_MEM_MAX) ?:
3238                 init_srcu_struct(&c->btree_trans_barrier);
3239         if (!ret)
3240                 c->btree_trans_barrier_initialized = true;
3241         return ret;
3242 }