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         btree_path_copy(trans, trans->paths + new, trans->paths + src);
1213         __btree_path_get(trans->paths + new, intent);
1214         return new;
1215 }
1216
1217 __flatten
1218 btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans,
1219                         btree_path_idx_t path, bool intent, unsigned long ip)
1220 {
1221         __btree_path_put(trans->paths + path, intent);
1222         path = btree_path_clone(trans, path, intent);
1223         trans->paths[path].preserve = false;
1224         return path;
1225 }
1226
1227 btree_path_idx_t __must_check
1228 __bch2_btree_path_set_pos(struct btree_trans *trans,
1229                           btree_path_idx_t path_idx, struct bpos new_pos,
1230                           bool intent, unsigned long ip)
1231 {
1232         int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos);
1233
1234         bch2_trans_verify_not_in_restart(trans);
1235         EBUG_ON(!trans->paths[path_idx].ref);
1236
1237         path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip);
1238
1239         struct btree_path *path = trans->paths + path_idx;
1240         path->pos               = new_pos;
1241         trans->paths_sorted     = false;
1242
1243         if (unlikely(path->cached)) {
1244                 btree_node_unlock(trans, path, 0);
1245                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1246                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1247                 goto out;
1248         }
1249
1250         unsigned level = btree_path_up_until_good_node(trans, path, cmp);
1251
1252         if (btree_path_node(path, level)) {
1253                 struct btree_path_level *l = &path->l[level];
1254
1255                 BUG_ON(!btree_node_locked(path, level));
1256                 /*
1257                  * We might have to skip over many keys, or just a few: try
1258                  * advancing the node iterator, and if we have to skip over too
1259                  * many keys just reinit it (or if we're rewinding, since that
1260                  * is expensive).
1261                  */
1262                 if (cmp < 0 ||
1263                     !btree_path_advance_to_pos(path, l, 8))
1264                         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1265
1266                 /*
1267                  * Iterators to interior nodes should always be pointed at the first non
1268                  * whiteout:
1269                  */
1270                 if (unlikely(level))
1271                         bch2_btree_node_iter_peek(&l->iter, l->b);
1272         }
1273
1274         if (unlikely(level != path->level)) {
1275                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1276                 __bch2_btree_path_unlock(trans, path);
1277         }
1278 out:
1279         bch2_btree_path_verify(trans, path);
1280         return path_idx;
1281 }
1282
1283 /* Btree path: main interface: */
1284
1285 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1286 {
1287         struct btree_path *sib;
1288
1289         sib = prev_btree_path(trans, path);
1290         if (sib && !btree_path_cmp(sib, path))
1291                 return sib;
1292
1293         sib = next_btree_path(trans, path);
1294         if (sib && !btree_path_cmp(sib, path))
1295                 return sib;
1296
1297         return NULL;
1298 }
1299
1300 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1301 {
1302         struct btree_path *sib;
1303
1304         sib = prev_btree_path(trans, path);
1305         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1306                 return sib;
1307
1308         sib = next_btree_path(trans, path);
1309         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1310                 return sib;
1311
1312         return NULL;
1313 }
1314
1315 static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path)
1316 {
1317         __bch2_btree_path_unlock(trans, trans->paths + path);
1318         btree_path_list_remove(trans, trans->paths + path);
1319         __clear_bit(path, trans->paths_allocated);
1320 }
1321
1322 void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent)
1323 {
1324         struct btree_path *path = trans->paths + path_idx, *dup;
1325
1326         if (!__btree_path_put(path, intent))
1327                 return;
1328
1329         dup = path->preserve
1330                 ? have_path_at_pos(trans, path)
1331                 : have_node_at_pos(trans, path);
1332
1333         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1334                 return;
1335
1336         if (path->should_be_locked &&
1337             !trans->restarted &&
1338             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1339                 return;
1340
1341         if (dup) {
1342                 dup->preserve           |= path->preserve;
1343                 dup->should_be_locked   |= path->should_be_locked;
1344         }
1345
1346         __bch2_path_free(trans, path_idx);
1347 }
1348
1349 static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path,
1350                                  bool intent)
1351 {
1352         if (!__btree_path_put(trans->paths + path, intent))
1353                 return;
1354
1355         __bch2_path_free(trans, path);
1356 }
1357
1358 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1359 {
1360         panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1361               trans->restart_count, restart_count,
1362               (void *) trans->last_begin_ip);
1363 }
1364
1365 void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1366 {
1367         panic("in transaction restart: %s, last restarted by %pS\n",
1368               bch2_err_str(trans->restarted),
1369               (void *) trans->last_restarted_ip);
1370 }
1371
1372 noinline __cold
1373 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1374 {
1375         prt_printf(buf, "transaction updates for %s journal seq %llu",
1376                trans->fn, trans->journal_res.seq);
1377         prt_newline(buf);
1378         printbuf_indent_add(buf, 2);
1379
1380         trans_for_each_update(trans, i) {
1381                 struct bkey_s_c old = { &i->old_k, i->old_v };
1382
1383                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1384                        bch2_btree_id_str(i->btree_id),
1385                        i->cached,
1386                        (void *) i->ip_allocated);
1387                 prt_newline(buf);
1388
1389                 prt_printf(buf, "  old ");
1390                 bch2_bkey_val_to_text(buf, trans->c, old);
1391                 prt_newline(buf);
1392
1393                 prt_printf(buf, "  new ");
1394                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1395                 prt_newline(buf);
1396         }
1397
1398         for (struct jset_entry *e = trans->journal_entries;
1399              e != btree_trans_journal_entries_top(trans);
1400              e = vstruct_next(e))
1401                 bch2_journal_entry_to_text(buf, trans->c, e);
1402
1403         printbuf_indent_sub(buf, 2);
1404 }
1405
1406 noinline __cold
1407 void bch2_dump_trans_updates(struct btree_trans *trans)
1408 {
1409         struct printbuf buf = PRINTBUF;
1410
1411         bch2_trans_updates_to_text(&buf, trans);
1412         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1413         printbuf_exit(&buf);
1414 }
1415
1416 static void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
1417 {
1418         struct btree_path *path = trans->paths + path_idx;
1419
1420         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1421                    path_idx, path->ref, path->intent_ref,
1422                    path->preserve ? 'P' : ' ',
1423                    path->should_be_locked ? 'S' : ' ',
1424                    bch2_btree_id_str(path->btree_id),
1425                    path->level);
1426         bch2_bpos_to_text(out, path->pos);
1427
1428         prt_printf(out, " locks %u", path->nodes_locked);
1429 #ifdef TRACK_PATH_ALLOCATED
1430         prt_printf(out, " %pS", (void *) path->ip_allocated);
1431 #endif
1432         prt_newline(out);
1433 }
1434
1435 static noinline __cold
1436 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1437                                 bool nosort)
1438 {
1439         struct trans_for_each_path_inorder_iter iter;
1440
1441         if (!nosort)
1442                 btree_trans_sort_paths(trans);
1443
1444         trans_for_each_path_idx_inorder(trans, iter)
1445                 bch2_btree_path_to_text(out, trans, iter.path_idx);
1446 }
1447
1448 noinline __cold
1449 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1450 {
1451         __bch2_trans_paths_to_text(out, trans, false);
1452 }
1453
1454 static noinline __cold
1455 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1456 {
1457         struct printbuf buf = PRINTBUF;
1458
1459         __bch2_trans_paths_to_text(&buf, trans, nosort);
1460         bch2_trans_updates_to_text(&buf, trans);
1461
1462         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1463         printbuf_exit(&buf);
1464 }
1465
1466 noinline __cold
1467 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1468 {
1469         __bch2_dump_trans_paths_updates(trans, false);
1470 }
1471
1472 noinline __cold
1473 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1474 {
1475         struct btree_transaction_stats *s = btree_trans_stats(trans);
1476         struct printbuf buf = PRINTBUF;
1477         size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths);
1478
1479         bch2_trans_paths_to_text(&buf, trans);
1480
1481         if (!buf.allocation_failure) {
1482                 mutex_lock(&s->lock);
1483                 if (nr > s->nr_max_paths) {
1484                         s->nr_max_paths = nr;
1485                         swap(s->max_paths_text, buf.buf);
1486                 }
1487                 mutex_unlock(&s->lock);
1488         }
1489
1490         printbuf_exit(&buf);
1491
1492         trans->nr_paths_max = nr;
1493 }
1494
1495 noinline __cold
1496 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
1497 {
1498         if (trace_trans_restart_too_many_iters_enabled()) {
1499                 struct printbuf buf = PRINTBUF;
1500
1501                 bch2_trans_paths_to_text(&buf, trans);
1502                 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
1503                 printbuf_exit(&buf);
1504         }
1505
1506         count_event(trans->c, trans_restart_too_many_iters);
1507
1508         return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
1509 }
1510
1511 static noinline void btree_path_overflow(struct btree_trans *trans)
1512 {
1513         bch2_dump_trans_paths_updates(trans);
1514         bch_err(trans->c, "trans path overflow");
1515 }
1516
1517 static noinline void btree_paths_realloc(struct btree_trans *trans)
1518 {
1519         unsigned nr = trans->nr_paths * 2;
1520
1521         void *p = kzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) +
1522                           sizeof(struct btree_trans_paths) +
1523                           nr * sizeof(struct btree_path) +
1524                           nr * sizeof(btree_path_idx_t) + 8 +
1525                           nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL);
1526
1527         unsigned long *paths_allocated = p;
1528         memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long));
1529         p += BITS_TO_LONGS(nr) * sizeof(unsigned long);
1530
1531         p += sizeof(struct btree_trans_paths);
1532         struct btree_path *paths = p;
1533         *trans_paths_nr(paths) = nr;
1534         memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path));
1535         p += nr * sizeof(struct btree_path);
1536
1537         btree_path_idx_t *sorted = p;
1538         memcpy(sorted, trans->sorted, trans->nr_sorted * sizeof(btree_path_idx_t));
1539         p += nr * sizeof(btree_path_idx_t) + 8;
1540
1541         struct btree_insert_entry *updates = p;
1542         memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_insert_entry));
1543
1544         unsigned long *old = trans->paths_allocated;
1545
1546         rcu_assign_pointer(trans->paths_allocated,      paths_allocated);
1547         rcu_assign_pointer(trans->paths,                paths);
1548         rcu_assign_pointer(trans->sorted,               sorted);
1549         rcu_assign_pointer(trans->updates,              updates);
1550
1551         trans->nr_paths         = nr;
1552
1553         if (old != trans->_paths_allocated)
1554                 kfree_rcu_mightsleep(old);
1555 }
1556
1557 static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans,
1558                                                 btree_path_idx_t pos)
1559 {
1560         btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths);
1561
1562         if (unlikely(idx == trans->nr_paths)) {
1563                 if (trans->nr_paths == BTREE_ITER_MAX) {
1564                         btree_path_overflow(trans);
1565                         return 0;
1566                 }
1567
1568                 btree_paths_realloc(trans);
1569         }
1570
1571         /*
1572          * Do this before marking the new path as allocated, since it won't be
1573          * initialized yet:
1574          */
1575         if (unlikely(idx > trans->nr_paths_max))
1576                 bch2_trans_update_max_paths(trans);
1577
1578         __set_bit(idx, trans->paths_allocated);
1579
1580         struct btree_path *path = &trans->paths[idx];
1581         path->ref               = 0;
1582         path->intent_ref        = 0;
1583         path->nodes_locked      = 0;
1584
1585         btree_path_list_add(trans, pos, idx);
1586         trans->paths_sorted = false;
1587         return idx;
1588 }
1589
1590 btree_path_idx_t bch2_path_get(struct btree_trans *trans,
1591                              enum btree_id btree_id, struct bpos pos,
1592                              unsigned locks_want, unsigned level,
1593                              unsigned flags, unsigned long ip)
1594 {
1595         struct btree_path *path;
1596         bool cached = flags & BTREE_ITER_CACHED;
1597         bool intent = flags & BTREE_ITER_INTENT;
1598         struct trans_for_each_path_inorder_iter iter;
1599         btree_path_idx_t path_pos = 0, path_idx;
1600
1601         bch2_trans_verify_not_in_restart(trans);
1602         bch2_trans_verify_locks(trans);
1603
1604         btree_trans_sort_paths(trans);
1605
1606         trans_for_each_path_inorder(trans, path, iter) {
1607                 if (__btree_path_cmp(path,
1608                                      btree_id,
1609                                      cached,
1610                                      pos,
1611                                      level) > 0)
1612                         break;
1613
1614                 path_pos = iter.path_idx;
1615         }
1616
1617         if (path_pos &&
1618             trans->paths[path_pos].cached       == cached &&
1619             trans->paths[path_pos].btree_id     == btree_id &&
1620             trans->paths[path_pos].level        == level) {
1621                 __btree_path_get(trans->paths + path_pos, intent);
1622                 path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1623                 path = trans->paths + path_idx;
1624         } else {
1625                 path_idx = btree_path_alloc(trans, path_pos);
1626                 path = trans->paths + path_idx;
1627
1628                 __btree_path_get(path, intent);
1629                 path->pos                       = pos;
1630                 path->btree_id                  = btree_id;
1631                 path->cached                    = cached;
1632                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1633                 path->should_be_locked          = false;
1634                 path->level                     = level;
1635                 path->locks_want                = locks_want;
1636                 path->nodes_locked              = 0;
1637                 for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++)
1638                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1639 #ifdef TRACK_PATH_ALLOCATED
1640                 path->ip_allocated              = ip;
1641 #endif
1642                 trans->paths_sorted             = false;
1643         }
1644
1645         if (!(flags & BTREE_ITER_NOPRESERVE))
1646                 path->preserve = true;
1647
1648         if (path->intent_ref)
1649                 locks_want = max(locks_want, level + 1);
1650
1651         /*
1652          * If the path has locks_want greater than requested, we don't downgrade
1653          * it here - on transaction restart because btree node split needs to
1654          * upgrade locks, we might be putting/getting the iterator again.
1655          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1656          * a successful transaction commit.
1657          */
1658
1659         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1660         if (locks_want > path->locks_want)
1661                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
1662
1663         return path_idx;
1664 }
1665
1666 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1667 {
1668
1669         struct btree_path_level *l = path_l(path);
1670         struct bkey_packed *_k;
1671         struct bkey_s_c k;
1672
1673         if (unlikely(!l->b))
1674                 return bkey_s_c_null;
1675
1676         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1677         EBUG_ON(!btree_node_locked(path, path->level));
1678
1679         if (!path->cached) {
1680                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1681                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1682
1683                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1684
1685                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1686                         goto hole;
1687         } else {
1688                 struct bkey_cached *ck = (void *) path->l[0].b;
1689
1690                 EBUG_ON(ck &&
1691                         (path->btree_id != ck->key.btree_id ||
1692                          !bkey_eq(path->pos, ck->key.pos)));
1693                 if (!ck || !ck->valid)
1694                         return bkey_s_c_null;
1695
1696                 *u = ck->k->k;
1697                 k = bkey_i_to_s_c(ck->k);
1698         }
1699
1700         return k;
1701 hole:
1702         bkey_init(u);
1703         u->p = path->pos;
1704         return (struct bkey_s_c) { u, NULL };
1705 }
1706
1707 /* Btree iterators: */
1708
1709 int __must_check
1710 __bch2_btree_iter_traverse(struct btree_iter *iter)
1711 {
1712         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1713 }
1714
1715 int __must_check
1716 bch2_btree_iter_traverse(struct btree_iter *iter)
1717 {
1718         struct btree_trans *trans = iter->trans;
1719         int ret;
1720
1721         iter->path = bch2_btree_path_set_pos(trans, iter->path,
1722                                         btree_iter_search_key(iter),
1723                                         iter->flags & BTREE_ITER_INTENT,
1724                                         btree_iter_ip_allocated(iter));
1725
1726         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1727         if (ret)
1728                 return ret;
1729
1730         btree_path_set_should_be_locked(trans->paths + iter->path);
1731         return 0;
1732 }
1733
1734 /* Iterate across nodes (leaf and interior nodes) */
1735
1736 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1737 {
1738         struct btree_trans *trans = iter->trans;
1739         struct btree *b = NULL;
1740         int ret;
1741
1742         EBUG_ON(trans->paths[iter->path].cached);
1743         bch2_btree_iter_verify(iter);
1744
1745         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1746         if (ret)
1747                 goto err;
1748
1749         struct btree_path *path = btree_iter_path(trans, iter);
1750         b = btree_path_node(path, path->level);
1751         if (!b)
1752                 goto out;
1753
1754         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1755
1756         bkey_init(&iter->k);
1757         iter->k.p = iter->pos = b->key.k.p;
1758
1759         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1760                                         iter->flags & BTREE_ITER_INTENT,
1761                                         btree_iter_ip_allocated(iter));
1762         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1763 out:
1764         bch2_btree_iter_verify_entry_exit(iter);
1765         bch2_btree_iter_verify(iter);
1766
1767         return b;
1768 err:
1769         b = ERR_PTR(ret);
1770         goto out;
1771 }
1772
1773 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1774 {
1775         struct btree *b;
1776
1777         while (b = bch2_btree_iter_peek_node(iter),
1778                bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1779                 bch2_trans_begin(iter->trans);
1780
1781         return b;
1782 }
1783
1784 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1785 {
1786         struct btree_trans *trans = iter->trans;
1787         struct btree *b = NULL;
1788         int ret;
1789
1790         EBUG_ON(trans->paths[iter->path].cached);
1791         bch2_trans_verify_not_in_restart(trans);
1792         bch2_btree_iter_verify(iter);
1793
1794         struct btree_path *path = btree_iter_path(trans, iter);
1795
1796         /* already at end? */
1797         if (!btree_path_node(path, path->level))
1798                 return NULL;
1799
1800         /* got to end? */
1801         if (!btree_path_node(path, path->level + 1)) {
1802                 btree_path_set_level_up(trans, path);
1803                 return NULL;
1804         }
1805
1806         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1807                 __bch2_btree_path_unlock(trans, path);
1808                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1809                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1810                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1811                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1812                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1813                 goto err;
1814         }
1815
1816         b = btree_path_node(path, path->level + 1);
1817
1818         if (bpos_eq(iter->pos, b->key.k.p)) {
1819                 __btree_path_set_level_up(trans, path, path->level++);
1820         } else {
1821                 /*
1822                  * Haven't gotten to the end of the parent node: go back down to
1823                  * the next child node
1824                  */
1825                 iter->path = bch2_btree_path_set_pos(trans, iter->path,
1826                                         bpos_successor(iter->pos),
1827                                         iter->flags & BTREE_ITER_INTENT,
1828                                         btree_iter_ip_allocated(iter));
1829
1830                 path = btree_iter_path(trans, iter);
1831                 btree_path_set_level_down(trans, path, iter->min_depth);
1832
1833                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1834                 if (ret)
1835                         goto err;
1836
1837                 path = btree_iter_path(trans, iter);
1838                 b = path->l[path->level].b;
1839         }
1840
1841         bkey_init(&iter->k);
1842         iter->k.p = iter->pos = b->key.k.p;
1843
1844         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1845                                         iter->flags & BTREE_ITER_INTENT,
1846                                         btree_iter_ip_allocated(iter));
1847         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1848         EBUG_ON(btree_iter_path(trans, iter)->uptodate);
1849 out:
1850         bch2_btree_iter_verify_entry_exit(iter);
1851         bch2_btree_iter_verify(iter);
1852
1853         return b;
1854 err:
1855         b = ERR_PTR(ret);
1856         goto out;
1857 }
1858
1859 /* Iterate across keys (in leaf nodes only) */
1860
1861 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1862 {
1863         struct bpos pos = iter->k.p;
1864         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1865                      ? bpos_eq(pos, SPOS_MAX)
1866                      : bkey_eq(pos, SPOS_MAX));
1867
1868         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1869                 pos = bkey_successor(iter, pos);
1870         bch2_btree_iter_set_pos(iter, pos);
1871         return ret;
1872 }
1873
1874 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1875 {
1876         struct bpos pos = bkey_start_pos(&iter->k);
1877         bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1878                      ? bpos_eq(pos, POS_MIN)
1879                      : bkey_eq(pos, POS_MIN));
1880
1881         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1882                 pos = bkey_predecessor(iter, pos);
1883         bch2_btree_iter_set_pos(iter, pos);
1884         return ret;
1885 }
1886
1887 static noinline
1888 void bch2_btree_trans_peek_prev_updates(struct btree_trans *trans, struct btree_iter *iter,
1889                                         struct bkey_s_c *k)
1890 {
1891         struct bpos end = path_l(btree_iter_path(trans, iter))->b->data->min_key;
1892
1893         trans_for_each_update(trans, i)
1894                 if (!i->key_cache_already_flushed &&
1895                     i->btree_id == iter->btree_id &&
1896                     bpos_le(i->k->k.p, iter->pos) &&
1897                     bpos_ge(i->k->k.p, k->k ? k->k->p : end)) {
1898                         iter->k = i->k->k;
1899                         *k = bkey_i_to_s_c(i->k);
1900                 }
1901 }
1902
1903 static noinline
1904 void bch2_btree_trans_peek_updates(struct btree_trans *trans, struct btree_iter *iter,
1905                                    struct bkey_s_c *k)
1906 {
1907         struct btree_path *path = btree_iter_path(trans, iter);
1908         struct bpos end = path_l(path)->b->key.k.p;
1909
1910         trans_for_each_update(trans, i)
1911                 if (!i->key_cache_already_flushed &&
1912                     i->btree_id == iter->btree_id &&
1913                     bpos_ge(i->k->k.p, path->pos) &&
1914                     bpos_le(i->k->k.p, k->k ? k->k->p : end)) {
1915                         iter->k = i->k->k;
1916                         *k = bkey_i_to_s_c(i->k);
1917                 }
1918 }
1919
1920 static noinline
1921 void bch2_btree_trans_peek_slot_updates(struct btree_trans *trans, struct btree_iter *iter,
1922                                         struct bkey_s_c *k)
1923 {
1924         trans_for_each_update(trans, i)
1925                 if (!i->key_cache_already_flushed &&
1926                     i->btree_id == iter->btree_id &&
1927                     bpos_eq(i->k->k.p, iter->pos)) {
1928                         iter->k = i->k->k;
1929                         *k = bkey_i_to_s_c(i->k);
1930                 }
1931 }
1932
1933 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1934                                               struct btree_iter *iter,
1935                                               struct bpos end_pos)
1936 {
1937         struct btree_path *path = btree_iter_path(trans, iter);
1938
1939         return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1940                                            path->level,
1941                                            path->pos,
1942                                            end_pos,
1943                                            &iter->journal_idx);
1944 }
1945
1946 static noinline
1947 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
1948                                               struct btree_iter *iter)
1949 {
1950         struct btree_path *path = btree_iter_path(trans, iter);
1951         struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos);
1952
1953         if (k) {
1954                 iter->k = k->k;
1955                 return bkey_i_to_s_c(k);
1956         } else {
1957                 return bkey_s_c_null;
1958         }
1959 }
1960
1961 static noinline
1962 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1963                                          struct btree_iter *iter,
1964                                          struct bkey_s_c k)
1965 {
1966         struct btree_path *path = btree_iter_path(trans, iter);
1967         struct bkey_i *next_journal =
1968                 bch2_btree_journal_peek(trans, iter,
1969                                 k.k ? k.k->p : path_l(path)->b->key.k.p);
1970
1971         if (next_journal) {
1972                 iter->k = next_journal->k;
1973                 k = bkey_i_to_s_c(next_journal);
1974         }
1975
1976         return k;
1977 }
1978
1979 /*
1980  * Checks btree key cache for key at iter->pos and returns it if present, or
1981  * bkey_s_c_null:
1982  */
1983 static noinline
1984 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1985 {
1986         struct btree_trans *trans = iter->trans;
1987         struct bch_fs *c = trans->c;
1988         struct bkey u;
1989         struct bkey_s_c k;
1990         int ret;
1991
1992         if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) &&
1993             bpos_eq(iter->pos, pos))
1994                 return bkey_s_c_null;
1995
1996         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1997                 return bkey_s_c_null;
1998
1999         if (!iter->key_cache_path)
2000                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
2001                                                      iter->flags & BTREE_ITER_INTENT, 0,
2002                                                      iter->flags|BTREE_ITER_CACHED|
2003                                                      BTREE_ITER_CACHED_NOFILL,
2004                                                      _THIS_IP_);
2005
2006         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
2007                                         iter->flags & BTREE_ITER_INTENT,
2008                                         btree_iter_ip_allocated(iter));
2009
2010         ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
2011                                          iter->flags|BTREE_ITER_CACHED) ?:
2012                 bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_);
2013         if (unlikely(ret))
2014                 return bkey_s_c_err(ret);
2015
2016         btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
2017
2018         k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
2019         if (k.k && !bkey_err(k)) {
2020                 iter->k = u;
2021                 k.k = &iter->k;
2022         }
2023         return k;
2024 }
2025
2026 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
2027 {
2028         struct btree_trans *trans = iter->trans;
2029         struct bkey_s_c k, k2;
2030         int ret;
2031
2032         EBUG_ON(btree_iter_path(trans, iter)->cached);
2033         bch2_btree_iter_verify(iter);
2034
2035         while (1) {
2036                 struct btree_path_level *l;
2037
2038                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2039                                         iter->flags & BTREE_ITER_INTENT,
2040                                         btree_iter_ip_allocated(iter));
2041
2042                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2043                 if (unlikely(ret)) {
2044                         /* ensure that iter->k is consistent with iter->pos: */
2045                         bch2_btree_iter_set_pos(iter, iter->pos);
2046                         k = bkey_s_c_err(ret);
2047                         goto out;
2048                 }
2049
2050                 struct btree_path *path = btree_iter_path(trans, iter);
2051                 l = path_l(path);
2052
2053                 if (unlikely(!l->b)) {
2054                         /* No btree nodes at requested level: */
2055                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2056                         k = bkey_s_c_null;
2057                         goto out;
2058                 }
2059
2060                 btree_path_set_should_be_locked(path);
2061
2062                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
2063
2064                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2065                     k.k &&
2066                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2067                         k = k2;
2068                         ret = bkey_err(k);
2069                         if (ret) {
2070                                 bch2_btree_iter_set_pos(iter, iter->pos);
2071                                 goto out;
2072                         }
2073                 }
2074
2075                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
2076                         k = btree_trans_peek_journal(trans, iter, k);
2077
2078                 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2079                              trans->nr_updates))
2080                         bch2_btree_trans_peek_updates(trans, iter, &k);
2081
2082                 if (k.k && bkey_deleted(k.k)) {
2083                         /*
2084                          * If we've got a whiteout, and it's after the search
2085                          * key, advance the search key to the whiteout instead
2086                          * of just after the whiteout - it might be a btree
2087                          * whiteout, with a real key at the same position, since
2088                          * in the btree deleted keys sort before non deleted.
2089                          */
2090                         search_key = !bpos_eq(search_key, k.k->p)
2091                                 ? k.k->p
2092                                 : bpos_successor(k.k->p);
2093                         continue;
2094                 }
2095
2096                 if (likely(k.k)) {
2097                         break;
2098                 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2099                         /* Advance to next leaf node: */
2100                         search_key = bpos_successor(l->b->key.k.p);
2101                 } else {
2102                         /* End of btree: */
2103                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2104                         k = bkey_s_c_null;
2105                         goto out;
2106                 }
2107         }
2108 out:
2109         bch2_btree_iter_verify(iter);
2110
2111         return k;
2112 }
2113
2114 /**
2115  * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
2116  * iterator's current position
2117  * @iter:       iterator to peek from
2118  * @end:        search limit: returns keys less than or equal to @end
2119  *
2120  * Returns:     key if found, or an error extractable with bkey_err().
2121  */
2122 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2123 {
2124         struct btree_trans *trans = iter->trans;
2125         struct bpos search_key = btree_iter_search_key(iter);
2126         struct bkey_s_c k;
2127         struct bpos iter_pos;
2128         int ret;
2129
2130         EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX));
2131
2132         if (iter->update_path) {
2133                 bch2_path_put_nokeep(trans, iter->update_path,
2134                                      iter->flags & BTREE_ITER_INTENT);
2135                 iter->update_path = 0;
2136         }
2137
2138         bch2_btree_iter_verify_entry_exit(iter);
2139
2140         while (1) {
2141                 k = __bch2_btree_iter_peek(iter, search_key);
2142                 if (unlikely(!k.k))
2143                         goto end;
2144                 if (unlikely(bkey_err(k)))
2145                         goto out_no_locked;
2146
2147                 /*
2148                  * We need to check against @end before FILTER_SNAPSHOTS because
2149                  * if we get to a different inode that requested we might be
2150                  * seeing keys for a different snapshot tree that will all be
2151                  * filtered out.
2152                  *
2153                  * But we can't do the full check here, because bkey_start_pos()
2154                  * isn't monotonically increasing before FILTER_SNAPSHOTS, and
2155                  * that's what we check against in extents mode:
2156                  */
2157                 if (k.k->p.inode > end.inode)
2158                         goto end;
2159
2160                 if (iter->update_path &&
2161                     !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
2162                         bch2_path_put_nokeep(trans, iter->update_path,
2163                                              iter->flags & BTREE_ITER_INTENT);
2164                         iter->update_path = 0;
2165                 }
2166
2167                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2168                     (iter->flags & BTREE_ITER_INTENT) &&
2169                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2170                     !iter->update_path) {
2171                         struct bpos pos = k.k->p;
2172
2173                         if (pos.snapshot < iter->snapshot) {
2174                                 search_key = bpos_successor(k.k->p);
2175                                 continue;
2176                         }
2177
2178                         pos.snapshot = iter->snapshot;
2179
2180                         /*
2181                          * advance, same as on exit for iter->path, but only up
2182                          * to snapshot
2183                          */
2184                         __btree_path_get(trans->paths + iter->path, iter->flags & BTREE_ITER_INTENT);
2185                         iter->update_path = iter->path;
2186
2187                         iter->update_path = bch2_btree_path_set_pos(trans,
2188                                                 iter->update_path, pos,
2189                                                 iter->flags & BTREE_ITER_INTENT,
2190                                                 _THIS_IP_);
2191                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2192                         if (unlikely(ret)) {
2193                                 k = bkey_s_c_err(ret);
2194                                 goto out_no_locked;
2195                         }
2196                 }
2197
2198                 /*
2199                  * We can never have a key in a leaf node at POS_MAX, so
2200                  * we don't have to check these successor() calls:
2201                  */
2202                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2203                     !bch2_snapshot_is_ancestor(trans->c,
2204                                                iter->snapshot,
2205                                                k.k->p.snapshot)) {
2206                         search_key = bpos_successor(k.k->p);
2207                         continue;
2208                 }
2209
2210                 if (bkey_whiteout(k.k) &&
2211                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2212                         search_key = bkey_successor(iter, k.k->p);
2213                         continue;
2214                 }
2215
2216                 /*
2217                  * iter->pos should be mononotically increasing, and always be
2218                  * equal to the key we just returned - except extents can
2219                  * straddle iter->pos:
2220                  */
2221                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2222                         iter_pos = k.k->p;
2223                 else
2224                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2225
2226                 if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS)
2227                              ? bkey_gt(iter_pos, end)
2228                              : bkey_ge(iter_pos, end)))
2229                         goto end;
2230
2231                 break;
2232         }
2233
2234         iter->pos = iter_pos;
2235
2236         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2237                                 iter->flags & BTREE_ITER_INTENT,
2238                                 btree_iter_ip_allocated(iter));
2239
2240         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2241 out_no_locked:
2242         if (iter->update_path) {
2243                 ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_);
2244                 if (unlikely(ret))
2245                         k = bkey_s_c_err(ret);
2246                 else
2247                         btree_path_set_should_be_locked(trans->paths + iter->update_path);
2248         }
2249
2250         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2251                 iter->pos.snapshot = iter->snapshot;
2252
2253         ret = bch2_btree_iter_verify_ret(iter, k);
2254         if (unlikely(ret)) {
2255                 bch2_btree_iter_set_pos(iter, iter->pos);
2256                 k = bkey_s_c_err(ret);
2257         }
2258
2259         bch2_btree_iter_verify_entry_exit(iter);
2260
2261         return k;
2262 end:
2263         bch2_btree_iter_set_pos(iter, end);
2264         k = bkey_s_c_null;
2265         goto out_no_locked;
2266 }
2267
2268 /**
2269  * bch2_btree_iter_next() - returns first key greater than iterator's current
2270  * position
2271  * @iter:       iterator to peek from
2272  *
2273  * Returns:     key if found, or an error extractable with bkey_err().
2274  */
2275 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2276 {
2277         if (!bch2_btree_iter_advance(iter))
2278                 return bkey_s_c_null;
2279
2280         return bch2_btree_iter_peek(iter);
2281 }
2282
2283 /**
2284  * bch2_btree_iter_peek_prev() - returns first key less than or equal to
2285  * iterator's current position
2286  * @iter:       iterator to peek from
2287  *
2288  * Returns:     key if found, or an error extractable with bkey_err().
2289  */
2290 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2291 {
2292         struct btree_trans *trans = iter->trans;
2293         struct bpos search_key = iter->pos;
2294         struct bkey_s_c k;
2295         struct bkey saved_k;
2296         const struct bch_val *saved_v;
2297         btree_path_idx_t saved_path = 0;
2298         int ret;
2299
2300         EBUG_ON(btree_iter_path(trans, iter)->cached ||
2301                 btree_iter_path(trans, iter)->level);
2302
2303         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2304                 return bkey_s_c_err(-EIO);
2305
2306         bch2_btree_iter_verify(iter);
2307         bch2_btree_iter_verify_entry_exit(iter);
2308
2309         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2310                 search_key.snapshot = U32_MAX;
2311
2312         while (1) {
2313                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2314                                                 iter->flags & BTREE_ITER_INTENT,
2315                                                 btree_iter_ip_allocated(iter));
2316
2317                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2318                 if (unlikely(ret)) {
2319                         /* ensure that iter->k is consistent with iter->pos: */
2320                         bch2_btree_iter_set_pos(iter, iter->pos);
2321                         k = bkey_s_c_err(ret);
2322                         goto out_no_locked;
2323                 }
2324
2325                 struct btree_path *path = btree_iter_path(trans, iter);
2326
2327                 k = btree_path_level_peek(trans, path, &path->l[0], &iter->k);
2328                 if (!k.k ||
2329                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2330                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2331                      : bpos_gt(k.k->p, search_key)))
2332                         k = btree_path_level_prev(trans, path, &path->l[0], &iter->k);
2333
2334                 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2335                              trans->nr_updates))
2336                         bch2_btree_trans_peek_prev_updates(trans, iter, &k);
2337
2338                 if (likely(k.k)) {
2339                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2340                                 if (k.k->p.snapshot == iter->snapshot)
2341                                         goto got_key;
2342
2343                                 /*
2344                                  * If we have a saved candidate, and we're no
2345                                  * longer at the same _key_ (not pos), return
2346                                  * that candidate
2347                                  */
2348                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2349                                         bch2_path_put_nokeep(trans, iter->path,
2350                                                       iter->flags & BTREE_ITER_INTENT);
2351                                         iter->path = saved_path;
2352                                         saved_path = 0;
2353                                         iter->k = saved_k;
2354                                         k.v     = saved_v;
2355                                         goto got_key;
2356                                 }
2357
2358                                 if (bch2_snapshot_is_ancestor(trans->c,
2359                                                               iter->snapshot,
2360                                                               k.k->p.snapshot)) {
2361                                         if (saved_path)
2362                                                 bch2_path_put_nokeep(trans, saved_path,
2363                                                       iter->flags & BTREE_ITER_INTENT);
2364                                         saved_path = btree_path_clone(trans, iter->path,
2365                                                                 iter->flags & BTREE_ITER_INTENT);
2366                                         path = btree_iter_path(trans, iter);
2367                                         saved_k = *k.k;
2368                                         saved_v = k.v;
2369                                 }
2370
2371                                 search_key = bpos_predecessor(k.k->p);
2372                                 continue;
2373                         }
2374 got_key:
2375                         if (bkey_whiteout(k.k) &&
2376                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2377                                 search_key = bkey_predecessor(iter, k.k->p);
2378                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2379                                         search_key.snapshot = U32_MAX;
2380                                 continue;
2381                         }
2382
2383                         btree_path_set_should_be_locked(path);
2384                         break;
2385                 } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
2386                         /* Advance to previous leaf node: */
2387                         search_key = bpos_predecessor(path->l[0].b->data->min_key);
2388                 } else {
2389                         /* Start of btree: */
2390                         bch2_btree_iter_set_pos(iter, POS_MIN);
2391                         k = bkey_s_c_null;
2392                         goto out_no_locked;
2393                 }
2394         }
2395
2396         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2397
2398         /* Extents can straddle iter->pos: */
2399         if (bkey_lt(k.k->p, iter->pos))
2400                 iter->pos = k.k->p;
2401
2402         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2403                 iter->pos.snapshot = iter->snapshot;
2404 out_no_locked:
2405         if (saved_path)
2406                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2407
2408         bch2_btree_iter_verify_entry_exit(iter);
2409         bch2_btree_iter_verify(iter);
2410
2411         return k;
2412 }
2413
2414 /**
2415  * bch2_btree_iter_prev() - returns first key less than iterator's current
2416  * position
2417  * @iter:       iterator to peek from
2418  *
2419  * Returns:     key if found, or an error extractable with bkey_err().
2420  */
2421 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2422 {
2423         if (!bch2_btree_iter_rewind(iter))
2424                 return bkey_s_c_null;
2425
2426         return bch2_btree_iter_peek_prev(iter);
2427 }
2428
2429 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2430 {
2431         struct btree_trans *trans = iter->trans;
2432         struct bpos search_key;
2433         struct bkey_s_c k;
2434         int ret;
2435
2436         bch2_btree_iter_verify(iter);
2437         bch2_btree_iter_verify_entry_exit(iter);
2438         EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2439
2440         /* extents can't span inode numbers: */
2441         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2442             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2443                 if (iter->pos.inode == KEY_INODE_MAX)
2444                         return bkey_s_c_null;
2445
2446                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2447         }
2448
2449         search_key = btree_iter_search_key(iter);
2450         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2451                                         iter->flags & BTREE_ITER_INTENT,
2452                                         btree_iter_ip_allocated(iter));
2453
2454         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2455         if (unlikely(ret)) {
2456                 k = bkey_s_c_err(ret);
2457                 goto out_no_locked;
2458         }
2459
2460         if ((iter->flags & BTREE_ITER_CACHED) ||
2461             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2462                 k = bkey_s_c_null;
2463
2464                 if (unlikely((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2465                              trans->nr_updates)) {
2466                         bch2_btree_trans_peek_slot_updates(trans, iter, &k);
2467                         if (k.k)
2468                                 goto out;
2469                 }
2470
2471                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2472                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2473                         goto out;
2474
2475                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2476                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2477                         if (!bkey_err(k))
2478                                 iter->k = *k.k;
2479                         /* We're not returning a key from iter->path: */
2480                         goto out_no_locked;
2481                 }
2482
2483                 k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k);
2484                 if (unlikely(!k.k))
2485                         goto out_no_locked;
2486         } else {
2487                 struct bpos next;
2488                 struct bpos end = iter->pos;
2489
2490                 if (iter->flags & BTREE_ITER_IS_EXTENTS)
2491                         end.offset = U64_MAX;
2492
2493                 EBUG_ON(btree_iter_path(trans, iter)->level);
2494
2495                 if (iter->flags & BTREE_ITER_INTENT) {
2496                         struct btree_iter iter2;
2497
2498                         bch2_trans_copy_iter(&iter2, iter);
2499                         k = bch2_btree_iter_peek_upto(&iter2, end);
2500
2501                         if (k.k && !bkey_err(k)) {
2502                                 iter->k = iter2.k;
2503                                 k.k = &iter->k;
2504                         }
2505                         bch2_trans_iter_exit(trans, &iter2);
2506                 } else {
2507                         struct bpos pos = iter->pos;
2508
2509                         k = bch2_btree_iter_peek_upto(iter, end);
2510                         if (unlikely(bkey_err(k)))
2511                                 bch2_btree_iter_set_pos(iter, pos);
2512                         else
2513                                 iter->pos = pos;
2514                 }
2515
2516                 if (unlikely(bkey_err(k)))
2517                         goto out_no_locked;
2518
2519                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2520
2521                 if (bkey_lt(iter->pos, next)) {
2522                         bkey_init(&iter->k);
2523                         iter->k.p = iter->pos;
2524
2525                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2526                                 bch2_key_resize(&iter->k,
2527                                                 min_t(u64, KEY_SIZE_MAX,
2528                                                       (next.inode == iter->pos.inode
2529                                                        ? next.offset
2530                                                        : KEY_OFFSET_MAX) -
2531                                                       iter->pos.offset));
2532                                 EBUG_ON(!iter->k.size);
2533                         }
2534
2535                         k = (struct bkey_s_c) { &iter->k, NULL };
2536                 }
2537         }
2538 out:
2539         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2540 out_no_locked:
2541         bch2_btree_iter_verify_entry_exit(iter);
2542         bch2_btree_iter_verify(iter);
2543         ret = bch2_btree_iter_verify_ret(iter, k);
2544         if (unlikely(ret))
2545                 return bkey_s_c_err(ret);
2546
2547         return k;
2548 }
2549
2550 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2551 {
2552         if (!bch2_btree_iter_advance(iter))
2553                 return bkey_s_c_null;
2554
2555         return bch2_btree_iter_peek_slot(iter);
2556 }
2557
2558 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2559 {
2560         if (!bch2_btree_iter_rewind(iter))
2561                 return bkey_s_c_null;
2562
2563         return bch2_btree_iter_peek_slot(iter);
2564 }
2565
2566 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2567 {
2568         struct bkey_s_c k;
2569
2570         while (btree_trans_too_many_iters(iter->trans) ||
2571                (k = bch2_btree_iter_peek_type(iter, iter->flags),
2572                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2573                 bch2_trans_begin(iter->trans);
2574
2575         return k;
2576 }
2577
2578 /* new transactional stuff: */
2579
2580 #ifdef CONFIG_BCACHEFS_DEBUG
2581 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2582 {
2583         struct btree_path *path;
2584         unsigned i;
2585
2586         BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1);
2587
2588         trans_for_each_path(trans, path, i) {
2589                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2590                 BUG_ON(trans->sorted[path->sorted_idx] != i);
2591         }
2592
2593         for (i = 0; i < trans->nr_sorted; i++) {
2594                 unsigned idx = trans->sorted[i];
2595
2596                 BUG_ON(!test_bit(idx, trans->paths_allocated));
2597                 BUG_ON(trans->paths[idx].sorted_idx != i);
2598         }
2599 }
2600
2601 static void btree_trans_verify_sorted(struct btree_trans *trans)
2602 {
2603         struct btree_path *path, *prev = NULL;
2604         struct trans_for_each_path_inorder_iter iter;
2605
2606         if (!bch2_debug_check_iterators)
2607                 return;
2608
2609         trans_for_each_path_inorder(trans, path, iter) {
2610                 if (prev && btree_path_cmp(prev, path) > 0) {
2611                         __bch2_dump_trans_paths_updates(trans, true);
2612                         panic("trans paths out of order!\n");
2613                 }
2614                 prev = path;
2615         }
2616 }
2617 #else
2618 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2619 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2620 #endif
2621
2622 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2623 {
2624         int i, l = 0, r = trans->nr_sorted, inc = 1;
2625         bool swapped;
2626
2627         btree_trans_verify_sorted_refs(trans);
2628
2629         if (trans->paths_sorted)
2630                 goto out;
2631
2632         /*
2633          * Cocktail shaker sort: this is efficient because iterators will be
2634          * mostly sorted.
2635          */
2636         do {
2637                 swapped = false;
2638
2639                 for (i = inc > 0 ? l : r - 2;
2640                      i + 1 < r && i >= l;
2641                      i += inc) {
2642                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2643                                            trans->paths + trans->sorted[i + 1]) > 0) {
2644                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2645                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2646                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2647                                 swapped = true;
2648                         }
2649                 }
2650
2651                 if (inc > 0)
2652                         --r;
2653                 else
2654                         l++;
2655                 inc = -inc;
2656         } while (swapped);
2657
2658         trans->paths_sorted = true;
2659 out:
2660         btree_trans_verify_sorted(trans);
2661 }
2662
2663 static inline void btree_path_list_remove(struct btree_trans *trans,
2664                                           struct btree_path *path)
2665 {
2666         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2667 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2668         trans->nr_sorted--;
2669         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2670                                 trans->sorted + path->sorted_idx + 1,
2671                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2672                                              sizeof(u64) / sizeof(btree_path_idx_t)));
2673 #else
2674         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2675 #endif
2676         for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
2677                 trans->paths[trans->sorted[i]].sorted_idx = i;
2678 }
2679
2680 static inline void btree_path_list_add(struct btree_trans *trans,
2681                                        btree_path_idx_t pos,
2682                                        btree_path_idx_t path_idx)
2683 {
2684         struct btree_path *path = trans->paths + path_idx;
2685
2686         path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted;
2687
2688 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2689         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2690                               trans->sorted + path->sorted_idx,
2691                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2692                                            sizeof(u64) / sizeof(btree_path_idx_t)));
2693         trans->nr_sorted++;
2694         trans->sorted[path->sorted_idx] = path_idx;
2695 #else
2696         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx);
2697 #endif
2698
2699         for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
2700                 trans->paths[trans->sorted[i]].sorted_idx = i;
2701
2702         btree_trans_verify_sorted_refs(trans);
2703 }
2704
2705 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2706 {
2707         if (iter->update_path)
2708                 bch2_path_put_nokeep(trans, iter->update_path,
2709                               iter->flags & BTREE_ITER_INTENT);
2710         if (iter->path)
2711                 bch2_path_put(trans, iter->path,
2712                               iter->flags & BTREE_ITER_INTENT);
2713         if (iter->key_cache_path)
2714                 bch2_path_put(trans, iter->key_cache_path,
2715                               iter->flags & BTREE_ITER_INTENT);
2716         iter->path              = 0;
2717         iter->update_path       = 0;
2718         iter->key_cache_path    = 0;
2719         iter->trans             = NULL;
2720 }
2721
2722 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2723                           struct btree_iter *iter,
2724                           enum btree_id btree_id, struct bpos pos,
2725                           unsigned flags)
2726 {
2727         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2728                                bch2_btree_iter_flags(trans, btree_id, flags),
2729                                _RET_IP_);
2730 }
2731
2732 void bch2_trans_node_iter_init(struct btree_trans *trans,
2733                                struct btree_iter *iter,
2734                                enum btree_id btree_id,
2735                                struct bpos pos,
2736                                unsigned locks_want,
2737                                unsigned depth,
2738                                unsigned flags)
2739 {
2740         flags |= BTREE_ITER_NOT_EXTENTS;
2741         flags |= __BTREE_ITER_ALL_SNAPSHOTS;
2742         flags |= BTREE_ITER_ALL_SNAPSHOTS;
2743
2744         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2745                                __bch2_btree_iter_flags(trans, btree_id, flags),
2746                                _RET_IP_);
2747
2748         iter->min_depth = depth;
2749
2750         struct btree_path *path = btree_iter_path(trans, iter);
2751         BUG_ON(path->locks_want  < min(locks_want, BTREE_MAX_DEPTH));
2752         BUG_ON(path->level      != depth);
2753         BUG_ON(iter->min_depth  != depth);
2754 }
2755
2756 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2757 {
2758         struct btree_trans *trans = src->trans;
2759
2760         *dst = *src;
2761         if (src->path)
2762                 __btree_path_get(trans->paths + src->path, src->flags & BTREE_ITER_INTENT);
2763         if (src->update_path)
2764                 __btree_path_get(trans->paths + src->update_path, src->flags & BTREE_ITER_INTENT);
2765         dst->key_cache_path = 0;
2766 }
2767
2768 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2769 {
2770         struct bch_fs *c = trans->c;
2771         unsigned new_top = trans->mem_top + size;
2772         unsigned old_bytes = trans->mem_bytes;
2773         unsigned new_bytes = roundup_pow_of_two(new_top);
2774         int ret;
2775         void *new_mem;
2776         void *p;
2777
2778         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2779
2780         struct btree_transaction_stats *s = btree_trans_stats(trans);
2781         s->max_mem = max(s->max_mem, new_bytes);
2782
2783         new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2784         if (unlikely(!new_mem)) {
2785                 bch2_trans_unlock(trans);
2786
2787                 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2788                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2789                         new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2790                         new_bytes = BTREE_TRANS_MEM_MAX;
2791                         kfree(trans->mem);
2792                 }
2793
2794                 if (!new_mem)
2795                         return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2796
2797                 trans->mem = new_mem;
2798                 trans->mem_bytes = new_bytes;
2799
2800                 ret = bch2_trans_relock(trans);
2801                 if (ret)
2802                         return ERR_PTR(ret);
2803         }
2804
2805         trans->mem = new_mem;
2806         trans->mem_bytes = new_bytes;
2807
2808         if (old_bytes) {
2809                 trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2810                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2811         }
2812
2813         p = trans->mem + trans->mem_top;
2814         trans->mem_top += size;
2815         memset(p, 0, size);
2816         return p;
2817 }
2818
2819 static inline void check_srcu_held_too_long(struct btree_trans *trans)
2820 {
2821         WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
2822              "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
2823              (jiffies - trans->srcu_lock_time) / HZ);
2824 }
2825
2826 void bch2_trans_srcu_unlock(struct btree_trans *trans)
2827 {
2828         if (trans->srcu_held) {
2829                 struct bch_fs *c = trans->c;
2830                 struct btree_path *path;
2831                 unsigned i;
2832
2833                 trans_for_each_path(trans, path, i)
2834                         if (path->cached && !btree_node_locked(path, 0))
2835                                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
2836
2837                 check_srcu_held_too_long(trans);
2838                 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2839                 trans->srcu_held = false;
2840         }
2841 }
2842
2843 static void bch2_trans_srcu_lock(struct btree_trans *trans)
2844 {
2845         if (!trans->srcu_held) {
2846                 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
2847                 trans->srcu_lock_time   = jiffies;
2848                 trans->srcu_held = true;
2849         }
2850 }
2851
2852 /**
2853  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2854  * @trans: transaction to reset
2855  *
2856  * Returns:     current restart counter, to be used with trans_was_restarted()
2857  *
2858  * While iterating over nodes or updating nodes a attempt to lock a btree node
2859  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2860  * occurs bch2_trans_begin() should be called and the transaction retried.
2861  */
2862 u32 bch2_trans_begin(struct btree_trans *trans)
2863 {
2864         struct btree_path *path;
2865         unsigned i;
2866         u64 now;
2867
2868         bch2_trans_reset_updates(trans);
2869
2870         trans->restart_count++;
2871         trans->mem_top                  = 0;
2872         trans->journal_entries          = NULL;
2873
2874         trans_for_each_path(trans, path, i) {
2875                 path->should_be_locked = false;
2876
2877                 /*
2878                  * If the transaction wasn't restarted, we're presuming to be
2879                  * doing something new: dont keep iterators excpt the ones that
2880                  * are in use - except for the subvolumes btree:
2881                  */
2882                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2883                         path->preserve = false;
2884
2885                 /*
2886                  * XXX: we probably shouldn't be doing this if the transaction
2887                  * was restarted, but currently we still overflow transaction
2888                  * iterators if we do that
2889                  */
2890                 if (!path->ref && !path->preserve)
2891                         __bch2_path_free(trans, i);
2892                 else
2893                         path->preserve = false;
2894         }
2895
2896         now = local_clock();
2897
2898         if (!IS_ENABLED(CONFIG_BCACHEFS_NO_LATENCY_ACCT) &&
2899             time_after64(now, trans->last_begin_time + 10))
2900                 __bch2_time_stats_update(&btree_trans_stats(trans)->duration,
2901                                          trans->last_begin_time, now);
2902
2903         if (!trans->restarted &&
2904             (need_resched() ||
2905              time_after64(now, trans->last_begin_time + BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS))) {
2906                 drop_locks_do(trans, (cond_resched(), 0));
2907                 now = local_clock();
2908         }
2909         trans->last_begin_time = now;
2910
2911         if (unlikely(trans->srcu_held &&
2912                      time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
2913                 bch2_trans_srcu_unlock(trans);
2914
2915         trans->last_begin_ip = _RET_IP_;
2916         if (trans->restarted) {
2917                 bch2_btree_path_traverse_all(trans);
2918                 trans->notrace_relock_fail = false;
2919         }
2920
2921         return trans->restart_count;
2922 }
2923
2924 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR] = { "(unknown)" };
2925
2926 unsigned bch2_trans_get_fn_idx(const char *fn)
2927 {
2928         for (unsigned i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2929                 if (!bch2_btree_transaction_fns[i] ||
2930                     bch2_btree_transaction_fns[i] == fn) {
2931                         bch2_btree_transaction_fns[i] = fn;
2932                         return i;
2933                 }
2934
2935         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2936         return 0;
2937 }
2938
2939 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
2940         __acquires(&c->btree_trans_barrier)
2941 {
2942         struct btree_trans *trans;
2943
2944         if (IS_ENABLED(__KERNEL__)) {
2945                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
2946                 if (trans) {
2947                         memset(trans, 0, offsetof(struct btree_trans, list));
2948                         goto got_trans;
2949                 }
2950         }
2951
2952         trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
2953         memset(trans, 0, sizeof(*trans));
2954         closure_init_stack(&trans->ref);
2955
2956         seqmutex_lock(&c->btree_trans_lock);
2957         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
2958                 struct btree_trans *pos;
2959                 pid_t pid = current->pid;
2960
2961                 trans->locking_wait.task = current;
2962
2963                 list_for_each_entry(pos, &c->btree_trans_list, list) {
2964                         struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task);
2965                         /*
2966                          * We'd much prefer to be stricter here and completely
2967                          * disallow multiple btree_trans in the same thread -
2968                          * but the data move path calls bch2_write when we
2969                          * already have a btree_trans initialized.
2970                          */
2971                         BUG_ON(pos_task &&
2972                                pid == pos_task->pid &&
2973                                bch2_trans_locked(pos));
2974
2975                         if (pos_task && pid < pos_task->pid) {
2976                                 list_add_tail(&trans->list, &pos->list);
2977                                 goto list_add_done;
2978                         }
2979                 }
2980         }
2981         list_add_tail(&trans->list, &c->btree_trans_list);
2982 list_add_done:
2983         seqmutex_unlock(&c->btree_trans_lock);
2984 got_trans:
2985         trans->c                = c;
2986         trans->last_begin_time  = local_clock();
2987         trans->fn_idx           = fn_idx;
2988         trans->locking_wait.task = current;
2989         trans->journal_replay_not_finished =
2990                 unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) &&
2991                 atomic_inc_not_zero(&c->journal_keys.ref);
2992         trans->nr_paths         = ARRAY_SIZE(trans->_paths);
2993         trans->paths_allocated  = trans->_paths_allocated;
2994         trans->sorted           = trans->_sorted;
2995         trans->paths            = trans->_paths;
2996         trans->updates          = trans->_updates;
2997
2998         *trans_paths_nr(trans->paths) = BTREE_ITER_INITIAL;
2999
3000         trans->paths_allocated[0] = 1;
3001
3002         if (fn_idx < BCH_TRANSACTIONS_NR) {
3003                 trans->fn = bch2_btree_transaction_fns[fn_idx];
3004
3005                 struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx];
3006
3007                 if (s->max_mem) {
3008                         unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
3009
3010                         trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
3011                         if (likely(trans->mem))
3012                                 trans->mem_bytes = expected_mem_bytes;
3013                 }
3014
3015                 trans->nr_paths_max = s->nr_max_paths;
3016                 trans->journal_entries_size = s->journal_entries_size;
3017         }
3018
3019         trans->srcu_idx         = srcu_read_lock(&c->btree_trans_barrier);
3020         trans->srcu_lock_time   = jiffies;
3021         trans->srcu_held        = true;
3022         return trans;
3023 }
3024
3025 static void check_btree_paths_leaked(struct btree_trans *trans)
3026 {
3027 #ifdef CONFIG_BCACHEFS_DEBUG
3028         struct bch_fs *c = trans->c;
3029         struct btree_path *path;
3030         unsigned i;
3031
3032         trans_for_each_path(trans, path, i)
3033                 if (path->ref)
3034                         goto leaked;
3035         return;
3036 leaked:
3037         bch_err(c, "btree paths leaked from %s!", trans->fn);
3038         trans_for_each_path(trans, path, i)
3039                 if (path->ref)
3040                         printk(KERN_ERR "  btree %s %pS\n",
3041                                bch2_btree_id_str(path->btree_id),
3042                                (void *) path->ip_allocated);
3043         /* Be noisy about this: */
3044         bch2_fatal_error(c);
3045 #endif
3046 }
3047
3048 void bch2_trans_put(struct btree_trans *trans)
3049         __releases(&c->btree_trans_barrier)
3050 {
3051         struct bch_fs *c = trans->c;
3052
3053         bch2_trans_unlock(trans);
3054
3055         trans_for_each_update(trans, i)
3056                 __btree_path_put(trans->paths + i->path, true);
3057         trans->nr_updates       = 0;
3058         trans->locking_wait.task = NULL;
3059
3060         check_btree_paths_leaked(trans);
3061
3062         if (trans->srcu_held) {
3063                 check_srcu_held_too_long(trans);
3064                 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3065         }
3066
3067         if (trans->fs_usage_deltas) {
3068                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3069                     REPLICAS_DELTA_LIST_MAX)
3070                         mempool_free(trans->fs_usage_deltas,
3071                                      &c->replicas_delta_pool);
3072                 else
3073                         kfree(trans->fs_usage_deltas);
3074         }
3075
3076         if (unlikely(trans->journal_replay_not_finished))
3077                 bch2_journal_keys_put(c);
3078
3079         unsigned long *paths_allocated = trans->paths_allocated;
3080         trans->paths_allocated  = NULL;
3081         trans->paths            = NULL;
3082
3083         if (paths_allocated != trans->_paths_allocated)
3084                 kfree_rcu_mightsleep(paths_allocated);
3085
3086         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
3087                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3088         else
3089                 kfree(trans->mem);
3090
3091         /* Userspace doesn't have a real percpu implementation: */
3092         if (IS_ENABLED(__KERNEL__))
3093                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
3094
3095         if (trans) {
3096                 closure_sync(&trans->ref);
3097
3098                 seqmutex_lock(&c->btree_trans_lock);
3099                 list_del(&trans->list);
3100                 seqmutex_unlock(&c->btree_trans_lock);
3101
3102                 mempool_free(trans, &c->btree_trans_pool);
3103         }
3104 }
3105
3106 static void __maybe_unused
3107 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3108                                       struct btree_bkey_cached_common *b)
3109 {
3110         struct six_lock_count c = six_lock_counts(&b->lock);
3111         struct task_struct *owner;
3112         pid_t pid;
3113
3114         rcu_read_lock();
3115         owner = READ_ONCE(b->lock.owner);
3116         pid = owner ? owner->pid : 0;
3117         rcu_read_unlock();
3118
3119         prt_tab(out);
3120         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3121                    b->level, bch2_btree_id_str(b->btree_id));
3122         bch2_bpos_to_text(out, btree_node_pos(b));
3123
3124         prt_tab(out);
3125         prt_printf(out, " locks %u:%u:%u held by pid %u",
3126                    c.n[0], c.n[1], c.n[2], pid);
3127 }
3128
3129 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3130 {
3131         struct btree_bkey_cached_common *b;
3132         static char lock_types[] = { 'r', 'i', 'w' };
3133         struct task_struct *task = READ_ONCE(trans->locking_wait.task);
3134         unsigned l, idx;
3135
3136         /* before rcu_read_lock(): */
3137         bch2_printbuf_make_room(out, 4096);
3138
3139         if (!out->nr_tabstops) {
3140                 printbuf_tabstop_push(out, 16);
3141                 printbuf_tabstop_push(out, 32);
3142         }
3143
3144         prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);
3145
3146         /* trans->paths is rcu protected vs. freeing */
3147         rcu_read_lock();
3148         out->atomic++;
3149
3150         struct btree_path *paths = rcu_dereference(trans->paths);
3151         if (!paths)
3152                 goto out;
3153
3154         unsigned long *paths_allocated = trans_paths_allocated(paths);
3155
3156         trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) {
3157                 struct btree_path *path = paths + idx;
3158                 if (!path->nodes_locked)
3159                         continue;
3160
3161                 prt_printf(out, "  path %u %c l=%u %s:",
3162                        idx,
3163                        path->cached ? 'c' : 'b',
3164                        path->level,
3165                        bch2_btree_id_str(path->btree_id));
3166                 bch2_bpos_to_text(out, path->pos);
3167                 prt_newline(out);
3168
3169                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3170                         if (btree_node_locked(path, l) &&
3171                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3172                                 prt_printf(out, "    %c l=%u ",
3173                                            lock_types[btree_node_locked_type(path, l)], l);
3174                                 bch2_btree_bkey_cached_common_to_text(out, b);
3175                                 prt_newline(out);
3176                         }
3177                 }
3178         }
3179
3180         b = READ_ONCE(trans->locking);
3181         if (b) {
3182                 prt_printf(out, "  blocked for %lluus on",
3183                            div_u64(local_clock() - trans->locking_wait.start_time,
3184                                    1000));
3185                 prt_newline(out);
3186                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3187                 bch2_btree_bkey_cached_common_to_text(out, b);
3188                 prt_newline(out);
3189         }
3190 out:
3191         --out->atomic;
3192         rcu_read_unlock();
3193 }
3194
3195 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3196 {
3197         struct btree_transaction_stats *s;
3198         struct btree_trans *trans;
3199         int cpu;
3200
3201         if (c->btree_trans_bufs)
3202                 for_each_possible_cpu(cpu) {
3203                         struct btree_trans *trans =
3204                                 per_cpu_ptr(c->btree_trans_bufs, cpu)->trans;
3205
3206                         if (trans) {
3207                                 closure_sync(&trans->ref);
3208
3209                                 seqmutex_lock(&c->btree_trans_lock);
3210                                 list_del(&trans->list);
3211                                 seqmutex_unlock(&c->btree_trans_lock);
3212                         }
3213                         kfree(trans);
3214                 }
3215         free_percpu(c->btree_trans_bufs);
3216
3217         trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3218         if (trans)
3219                 panic("%s leaked btree_trans\n", trans->fn);
3220
3221         for (s = c->btree_transaction_stats;
3222              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3223              s++) {
3224                 kfree(s->max_paths_text);
3225                 bch2_time_stats_exit(&s->lock_hold_times);
3226         }
3227
3228         if (c->btree_trans_barrier_initialized)
3229                 cleanup_srcu_struct(&c->btree_trans_barrier);
3230         mempool_exit(&c->btree_trans_mem_pool);
3231         mempool_exit(&c->btree_trans_pool);
3232 }
3233
3234 void bch2_fs_btree_iter_init_early(struct bch_fs *c)
3235 {
3236         struct btree_transaction_stats *s;
3237
3238         for (s = c->btree_transaction_stats;
3239              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3240              s++) {
3241                 bch2_time_stats_init(&s->duration);
3242                 bch2_time_stats_init(&s->lock_hold_times);
3243                 mutex_init(&s->lock);
3244         }
3245
3246         INIT_LIST_HEAD(&c->btree_trans_list);
3247         seqmutex_init(&c->btree_trans_lock);
3248 }
3249
3250 int bch2_fs_btree_iter_init(struct bch_fs *c)
3251 {
3252         int ret;
3253
3254         c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3255         if (!c->btree_trans_bufs)
3256                 return -ENOMEM;
3257
3258         ret   = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3259                                           sizeof(struct btree_trans)) ?:
3260                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3261                                           BTREE_TRANS_MEM_MAX) ?:
3262                 init_srcu_struct(&c->btree_trans_barrier);
3263         if (!ret)
3264                 c->btree_trans_barrier_initialized = true;
3265         return ret;
3266 }