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