fs/bcachefs/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_snapshot_field) &&
 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(trans, &jiter, 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             c->opts.btree_node_prefetch)
 902                 ret = btree_path_prefetch_j(trans, path, &jiter);
 903
 904         bch2_btree_and_journal_iter_exit(&jiter);
 905         return ret;
 906 }
 907
 908 static __always_inline int btree_path_down(struct btree_trans *trans,
 909                                            struct btree_path *path,
 910                                            unsigned flags,
 911                                            unsigned long trace_ip)
 912 {
 913         struct bch_fs *c = trans->c;
 914         struct btree_path_level *l = path_l(path);
 915         struct btree *b;
 916         unsigned level = path->level - 1;
 917         enum six_lock_type lock_type = __btree_lock_want(path, level);
 918         struct bkey_buf tmp;
 919         int ret;
 920
 921         EBUG_ON(!btree_node_locked(path, path->level));
 922
 923         bch2_bkey_buf_init(&tmp);
 924
 925         if (unlikely(trans->journal_replay_not_finished)) {
 926                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
 927                 if (ret)
 928                         goto err;
 929         } else {
 930                 struct bkey_packed *k = bch2_btree_node_iter_peek(&l->iter, l->b);
 931                 if (!k) {
 932                         struct printbuf buf = PRINTBUF;
 933
 934                         prt_str(&buf, "node not found at pos ");
 935                         bch2_bpos_to_text(&buf, path->pos);
 936                         prt_str(&buf, " within parent node ");
 937                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&l->b->key));
 938
 939                         bch2_fs_fatal_error(c, "%s", buf.buf);
 940                         printbuf_exit(&buf);
 941                         ret = -BCH_ERR_btree_need_topology_repair;
 942                         goto err;
 943                 }
 944
 945                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
 946
 947                 if ((flags & BTREE_ITER_prefetch) &&
 948                     c->opts.btree_node_prefetch) {
 949                         ret = btree_path_prefetch(trans, path);
 950                         if (ret)
 951                                 goto err;
 952                 }
 953         }
 954
 955         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
 956         ret = PTR_ERR_OR_ZERO(b);
 957         if (unlikely(ret))
 958                 goto err;
 959
 960         if (likely(!trans->journal_replay_not_finished &&
 961                    tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
 962             unlikely(b != btree_node_mem_ptr(tmp.k)))
 963                 btree_node_mem_ptr_set(trans, path, level + 1, b);
 964
 965         if (btree_node_read_locked(path, level + 1))
 966                 btree_node_unlock(trans, path, level + 1);
 967
 968         mark_btree_node_locked(trans, path, level,
 969                                (enum btree_node_locked_type) lock_type);
 970         path->level = level;
 971         bch2_btree_path_level_init(trans, path, b);
 972
 973         bch2_btree_path_verify_locks(path);
 974 err:
 975         bch2_bkey_buf_exit(&tmp, c);
 976         return ret;
 977 }
 978
 979 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
 980 {
 981         struct bch_fs *c = trans->c;
 982         struct btree_path *path;
 983         unsigned long trace_ip = _RET_IP_;
 984         unsigned i;
 985         int ret = 0;
 986
 987         if (trans->in_traverse_all)
 988                 return -BCH_ERR_transaction_restart_in_traverse_all;
 989
 990         trans->in_traverse_all = true;
 991 retry_all:
 992         trans->restarted = 0;
 993         trans->last_restarted_ip = 0;
 994
 995         trans_for_each_path(trans, path, i)
 996                 path->should_be_locked = false;
 997
 998         btree_trans_sort_paths(trans);
 999
1000         bch2_trans_unlock(trans);
1001         cond_resched();
1002         trans->locked = true;
1003
1004         if (unlikely(trans->memory_allocation_failure)) {
1005                 struct closure cl;
1006
1007                 closure_init_stack(&cl);
1008
1009                 do {
1010                         ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
1011                         closure_sync(&cl);
1012                 } while (ret);
1013         }
1014
1015         /* Now, redo traversals in correct order: */
1016         i = 0;
1017         while (i < trans->nr_sorted) {
1018                 btree_path_idx_t idx = trans->sorted[i];
1019
1020                 /*
1021                  * Traversing a path can cause another path to be added at about
1022                  * the same position:
1023                  */
1024                 if (trans->paths[idx].uptodate) {
1025                         __btree_path_get(&trans->paths[idx], false);
1026                         ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_);
1027                         __btree_path_put(&trans->paths[idx], false);
1028
1029                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1030                             bch2_err_matches(ret, ENOMEM))
1031                                 goto retry_all;
1032                         if (ret)
1033                                 goto err;
1034                 } else {
1035                         i++;
1036                 }
1037         }
1038
1039         /*
1040          * We used to assert that all paths had been traversed here
1041          * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since
1042          * path->should_be_locked is not set yet, we might have unlocked and
1043          * then failed to relock a path - that's fine.
1044          */
1045 err:
1046         bch2_btree_cache_cannibalize_unlock(trans);
1047
1048         trans->in_traverse_all = false;
1049
1050         trace_and_count(c, trans_traverse_all, trans, trace_ip);
1051         return ret;
1052 }
1053
1054 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1055                                                 unsigned l, int check_pos)
1056 {
1057         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1058                 return false;
1059         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1060                 return false;
1061         return true;
1062 }
1063
1064 static inline bool btree_path_good_node(struct btree_trans *trans,
1065                                         struct btree_path *path,
1066                                         unsigned l, int check_pos)
1067 {
1068         return is_btree_node(path, l) &&
1069                 bch2_btree_node_relock(trans, path, l) &&
1070                 btree_path_check_pos_in_node(path, l, check_pos);
1071 }
1072
1073 static void btree_path_set_level_down(struct btree_trans *trans,
1074                                       struct btree_path *path,
1075                                       unsigned new_level)
1076 {
1077         unsigned l;
1078
1079         path->level = new_level;
1080
1081         for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1082                 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1083                         btree_node_unlock(trans, path, l);
1084
1085         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1086         bch2_btree_path_verify(trans, path);
1087 }
1088
1089 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1090                                                          struct btree_path *path,
1091                                                          int check_pos)
1092 {
1093         unsigned i, l = path->level;
1094 again:
1095         while (btree_path_node(path, l) &&
1096                !btree_path_good_node(trans, path, l, check_pos))
1097                 __btree_path_set_level_up(trans, path, l++);
1098
1099         /* If we need intent locks, take them too: */
1100         for (i = l + 1;
1101              i < path->locks_want && btree_path_node(path, i);
1102              i++)
1103                 if (!bch2_btree_node_relock(trans, path, i)) {
1104                         while (l <= i)
1105                                 __btree_path_set_level_up(trans, path, l++);
1106                         goto again;
1107                 }
1108
1109         return l;
1110 }
1111
1112 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1113                                                      struct btree_path *path,
1114                                                      int check_pos)
1115 {
1116         return likely(btree_node_locked(path, path->level) &&
1117                       btree_path_check_pos_in_node(path, path->level, check_pos))
1118                 ? path->level
1119                 : __btree_path_up_until_good_node(trans, path, check_pos);
1120 }
1121
1122 /*
1123  * This is the main state machine for walking down the btree - walks down to a
1124  * specified depth
1125  *
1126  * Returns 0 on success, -EIO on error (error reading in a btree node).
1127  *
1128  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1129  * stashed in the iterator and returned from bch2_trans_exit().
1130  */
1131 int bch2_btree_path_traverse_one(struct btree_trans *trans,
1132                                  btree_path_idx_t path_idx,
1133                                  unsigned flags,
1134                                  unsigned long trace_ip)
1135 {
1136         struct btree_path *path = &trans->paths[path_idx];
1137         unsigned depth_want = path->level;
1138         int ret = -((int) trans->restarted);
1139
1140         if (unlikely(ret))
1141                 goto out;
1142
1143         if (unlikely(!trans->srcu_held))
1144                 bch2_trans_srcu_lock(trans);
1145
1146         /*
1147          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1148          * and re-traverse the path without a transaction restart:
1149          */
1150         if (path->should_be_locked) {
1151                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1152                 goto out;
1153         }
1154
1155         if (path->cached) {
1156                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1157                 goto out;
1158         }
1159
1160         path = &trans->paths[path_idx];
1161
1162         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1163                 goto out_uptodate;
1164
1165         path->level = btree_path_up_until_good_node(trans, path, 0);
1166         unsigned max_level = path->level;
1167
1168         EBUG_ON(btree_path_node(path, path->level) &&
1169                 !btree_node_locked(path, path->level));
1170
1171         /*
1172          * Note: path->nodes[path->level] may be temporarily NULL here - that
1173          * would indicate to other code that we got to the end of the btree,
1174          * here it indicates that relocking the root failed - it's critical that
1175          * btree_path_lock_root() comes next and that it can't fail
1176          */
1177         while (path->level > depth_want) {
1178                 ret = btree_path_node(path, path->level)
1179                         ? btree_path_down(trans, path, flags, trace_ip)
1180                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1181                 if (unlikely(ret)) {
1182                         if (ret == 1) {
1183                                 /*
1184                                  * No nodes at this level - got to the end of
1185                                  * the btree:
1186                                  */
1187                                 ret = 0;
1188                                 goto out;
1189                         }
1190
1191                         __bch2_btree_path_unlock(trans, path);
1192                         path->level = depth_want;
1193                         path->l[path->level].b = ERR_PTR(ret);
1194                         goto out;
1195                 }
1196         }
1197
1198         if (unlikely(max_level > path->level)) {
1199                 struct btree_path *linked;
1200                 unsigned iter;
1201
1202                 trans_for_each_path_with_node(trans, path_l(path)->b, linked, iter)
1203                         for (unsigned j = path->level + 1; j < max_level; j++)
1204                                 linked->l[j] = path->l[j];
1205         }
1206
1207 out_uptodate:
1208         path->uptodate = BTREE_ITER_UPTODATE;
1209 out:
1210         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
1211                 panic("ret %s (%i) trans->restarted %s (%i)\n",
1212                       bch2_err_str(ret), ret,
1213                       bch2_err_str(trans->restarted), trans->restarted);
1214         bch2_btree_path_verify(trans, path);
1215         return ret;
1216 }
1217
1218 static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1219                             struct btree_path *src)
1220 {
1221         unsigned i, offset = offsetof(struct btree_path, pos);
1222
1223         memcpy((void *) dst + offset,
1224                (void *) src + offset,
1225                sizeof(struct btree_path) - offset);
1226
1227         for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1228                 unsigned t = btree_node_locked_type(dst, i);
1229
1230                 if (t != BTREE_NODE_UNLOCKED)
1231                         six_lock_increment(&dst->l[i].b->c.lock, t);
1232         }
1233 }
1234
1235 static btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src,
1236                                          bool intent, unsigned long ip)
1237 {
1238         btree_path_idx_t new = btree_path_alloc(trans, src);
1239         btree_path_copy(trans, trans->paths + new, trans->paths + src);
1240         __btree_path_get(trans->paths + new, intent);
1241 #ifdef TRACK_PATH_ALLOCATED
1242         trans->paths[new].ip_allocated = ip;
1243 #endif
1244         return new;
1245 }
1246
1247 __flatten
1248 btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans,
1249                         btree_path_idx_t path, bool intent, unsigned long ip)
1250 {
1251         __btree_path_put(trans->paths + path, intent);
1252         path = btree_path_clone(trans, path, intent, ip);
1253         trans->paths[path].preserve = false;
1254         return path;
1255 }
1256
1257 btree_path_idx_t __must_check
1258 __bch2_btree_path_set_pos(struct btree_trans *trans,
1259                           btree_path_idx_t path_idx, struct bpos new_pos,
1260                           bool intent, unsigned long ip)
1261 {
1262         int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos);
1263
1264         bch2_trans_verify_not_in_restart(trans);
1265         EBUG_ON(!trans->paths[path_idx].ref);
1266
1267         path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip);
1268
1269         struct btree_path *path = trans->paths + path_idx;
1270         path->pos               = new_pos;
1271         trans->paths_sorted     = false;
1272
1273         if (unlikely(path->cached)) {
1274                 btree_node_unlock(trans, path, 0);
1275                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1276                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1277                 goto out;
1278         }
1279
1280         unsigned level = btree_path_up_until_good_node(trans, path, cmp);
1281
1282         if (btree_path_node(path, level)) {
1283                 struct btree_path_level *l = &path->l[level];
1284
1285                 BUG_ON(!btree_node_locked(path, level));
1286                 /*
1287                  * We might have to skip over many keys, or just a few: try
1288                  * advancing the node iterator, and if we have to skip over too
1289                  * many keys just reinit it (or if we're rewinding, since that
1290                  * is expensive).
1291                  */
1292                 if (cmp < 0 ||
1293                     !btree_path_advance_to_pos(path, l, 8))
1294                         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
1295
1296                 /*
1297                  * Iterators to interior nodes should always be pointed at the first non
1298                  * whiteout:
1299                  */
1300                 if (unlikely(level))
1301                         bch2_btree_node_iter_peek(&l->iter, l->b);
1302         }
1303
1304         if (unlikely(level != path->level)) {
1305                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1306                 __bch2_btree_path_unlock(trans, path);
1307         }
1308 out:
1309         bch2_btree_path_verify(trans, path);
1310         return path_idx;
1311 }
1312
1313 /* Btree path: main interface: */
1314
1315 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1316 {
1317         struct btree_path *sib;
1318
1319         sib = prev_btree_path(trans, path);
1320         if (sib && !btree_path_cmp(sib, path))
1321                 return sib;
1322
1323         sib = next_btree_path(trans, path);
1324         if (sib && !btree_path_cmp(sib, path))
1325                 return sib;
1326
1327         return NULL;
1328 }
1329
1330 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1331 {
1332         struct btree_path *sib;
1333
1334         sib = prev_btree_path(trans, path);
1335         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1336                 return sib;
1337
1338         sib = next_btree_path(trans, path);
1339         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1340                 return sib;
1341
1342         return NULL;
1343 }
1344
1345 static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path)
1346 {
1347         __bch2_btree_path_unlock(trans, trans->paths + path);
1348         btree_path_list_remove(trans, trans->paths + path);
1349         __clear_bit(path, trans->paths_allocated);
1350 }
1351
1352 static bool bch2_btree_path_can_relock(struct btree_trans *trans, struct btree_path *path)
1353 {
1354         unsigned l = path->level;
1355
1356         do {
1357                 if (!btree_path_node(path, l))
1358                         break;
1359
1360                 if (!is_btree_node(path, l))
1361                         return false;
1362
1363                 if (path->l[l].lock_seq != path->l[l].b->c.lock.seq)
1364                         return false;
1365
1366                 l++;
1367         } while (l < path->locks_want);
1368
1369         return true;
1370 }
1371
1372 void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent)
1373 {
1374         struct btree_path *path = trans->paths + path_idx, *dup;
1375
1376         if (!__btree_path_put(path, intent))
1377                 return;
1378
1379         dup = path->preserve
1380                 ? have_path_at_pos(trans, path)
1381                 : have_node_at_pos(trans, path);
1382
1383         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1384                 return;
1385
1386         if (path->should_be_locked && !trans->restarted) {
1387                 if (!dup)
1388                         return;
1389
1390                 if (!(trans->locked
1391                       ? bch2_btree_path_relock_norestart(trans, dup)
1392                       : bch2_btree_path_can_relock(trans, dup)))
1393                         return;
1394         }
1395
1396         if (dup) {
1397                 dup->preserve           |= path->preserve;
1398                 dup->should_be_locked   |= path->should_be_locked;
1399         }
1400
1401         __bch2_path_free(trans, path_idx);
1402 }
1403
1404 static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path,
1405                                  bool intent)
1406 {
1407         if (!__btree_path_put(trans->paths + path, intent))
1408                 return;
1409
1410         __bch2_path_free(trans, path);
1411 }
1412
1413 void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
1414 {
1415         panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
1416               trans->restart_count, restart_count,
1417               (void *) trans->last_begin_ip);
1418 }
1419
1420 void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
1421 {
1422         panic("in transaction restart: %s, last restarted by %pS\n",
1423               bch2_err_str(trans->restarted),
1424               (void *) trans->last_restarted_ip);
1425 }
1426
1427 void __noreturn bch2_trans_unlocked_error(struct btree_trans *trans)
1428 {
1429         panic("trans should be locked, unlocked by %pS\n",
1430               (void *) trans->last_unlock_ip);
1431 }
1432
1433 noinline __cold
1434 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1435 {
1436         prt_printf(buf, "transaction updates for %s journal seq %llu\n",
1437                trans->fn, trans->journal_res.seq);
1438         printbuf_indent_add(buf, 2);
1439
1440         trans_for_each_update(trans, i) {
1441                 struct bkey_s_c old = { &i->old_k, i->old_v };
1442
1443                 prt_printf(buf, "update: btree=%s cached=%u %pS\n",
1444                        bch2_btree_id_str(i->btree_id),
1445                        i->cached,
1446                        (void *) i->ip_allocated);
1447
1448                 prt_printf(buf, "  old ");
1449                 bch2_bkey_val_to_text(buf, trans->c, old);
1450                 prt_newline(buf);
1451
1452                 prt_printf(buf, "  new ");
1453                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1454                 prt_newline(buf);
1455         }
1456
1457         for (struct jset_entry *e = trans->journal_entries;
1458              e != btree_trans_journal_entries_top(trans);
1459              e = vstruct_next(e))
1460                 bch2_journal_entry_to_text(buf, trans->c, e);
1461
1462         printbuf_indent_sub(buf, 2);
1463 }
1464
1465 noinline __cold
1466 void bch2_dump_trans_updates(struct btree_trans *trans)
1467 {
1468         struct printbuf buf = PRINTBUF;
1469
1470         bch2_trans_updates_to_text(&buf, trans);
1471         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1472         printbuf_exit(&buf);
1473 }
1474
1475 static void bch2_btree_path_to_text_short(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
1476 {
1477         struct btree_path *path = trans->paths + path_idx;
1478
1479         prt_printf(out, "path: idx %2u ref %u:%u %c %c %c btree=%s l=%u pos ",
1480                    path_idx, path->ref, path->intent_ref,
1481                    path->preserve ? 'P' : ' ',
1482                    path->should_be_locked ? 'S' : ' ',
1483                    path->cached ? 'C' : 'B',
1484                    bch2_btree_id_str(path->btree_id),
1485                    path->level);
1486         bch2_bpos_to_text(out, path->pos);
1487
1488 #ifdef TRACK_PATH_ALLOCATED
1489         prt_printf(out, " %pS", (void *) path->ip_allocated);
1490 #endif
1491 }
1492
1493 static const char *btree_node_locked_str(enum btree_node_locked_type t)
1494 {
1495         switch (t) {
1496         case BTREE_NODE_UNLOCKED:
1497                 return "unlocked";
1498         case BTREE_NODE_READ_LOCKED:
1499                 return "read";
1500         case BTREE_NODE_INTENT_LOCKED:
1501                 return "intent";
1502         case BTREE_NODE_WRITE_LOCKED:
1503                 return "write";
1504         default:
1505                 return NULL;
1506         }
1507 }
1508
1509 void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx)
1510 {
1511         bch2_btree_path_to_text_short(out, trans, path_idx);
1512
1513         struct btree_path *path = trans->paths + path_idx;
1514
1515         prt_printf(out, " uptodate %u locks_want %u", path->uptodate, path->locks_want);
1516         prt_newline(out);
1517
1518         printbuf_indent_add(out, 2);
1519         for (unsigned l = 0; l < BTREE_MAX_DEPTH; l++) {
1520                 prt_printf(out, "l=%u locks %s seq %u node ", l,
1521                            btree_node_locked_str(btree_node_locked_type(path, l)),
1522                            path->l[l].lock_seq);
1523
1524                 int ret = PTR_ERR_OR_ZERO(path->l[l].b);
1525                 if (ret)
1526                         prt_str(out, bch2_err_str(ret));
1527                 else
1528                         prt_printf(out, "%px", path->l[l].b);
1529                 prt_newline(out);
1530         }
1531         printbuf_indent_sub(out, 2);
1532 }
1533
1534 static noinline __cold
1535 void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans,
1536                                 bool nosort)
1537 {
1538         struct trans_for_each_path_inorder_iter iter;
1539
1540         if (!nosort)
1541                 btree_trans_sort_paths(trans);
1542
1543         trans_for_each_path_idx_inorder(trans, iter) {
1544                 bch2_btree_path_to_text_short(out, trans, iter.path_idx);
1545                 prt_newline(out);
1546         }
1547 }
1548
1549 noinline __cold
1550 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1551 {
1552         __bch2_trans_paths_to_text(out, trans, false);
1553 }
1554
1555 static noinline __cold
1556 void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort)
1557 {
1558         struct printbuf buf = PRINTBUF;
1559
1560         __bch2_trans_paths_to_text(&buf, trans, nosort);
1561         bch2_trans_updates_to_text(&buf, trans);
1562
1563         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1564         printbuf_exit(&buf);
1565 }
1566
1567 noinline __cold
1568 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1569 {
1570         __bch2_dump_trans_paths_updates(trans, false);
1571 }
1572
1573 noinline __cold
1574 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1575 {
1576         struct btree_transaction_stats *s = btree_trans_stats(trans);
1577         struct printbuf buf = PRINTBUF;
1578         size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths);
1579
1580         bch2_trans_paths_to_text(&buf, trans);
1581
1582         if (!buf.allocation_failure) {
1583                 mutex_lock(&s->lock);
1584                 if (nr > s->nr_max_paths) {
1585                         s->nr_max_paths = nr;
1586                         swap(s->max_paths_text, buf.buf);
1587                 }
1588                 mutex_unlock(&s->lock);
1589         }
1590
1591         printbuf_exit(&buf);
1592
1593         trans->nr_paths_max = nr;
1594 }
1595
1596 noinline __cold
1597 int __bch2_btree_trans_too_many_iters(struct btree_trans *trans)
1598 {
1599         if (trace_trans_restart_too_many_iters_enabled()) {
1600                 struct printbuf buf = PRINTBUF;
1601
1602                 bch2_trans_paths_to_text(&buf, trans);
1603                 trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf);
1604                 printbuf_exit(&buf);
1605         }
1606
1607         count_event(trans->c, trans_restart_too_many_iters);
1608
1609         return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters);
1610 }
1611
1612 static noinline void btree_path_overflow(struct btree_trans *trans)
1613 {
1614         bch2_dump_trans_paths_updates(trans);
1615         bch_err(trans->c, "trans path overflow");
1616 }
1617
1618 static noinline void btree_paths_realloc(struct btree_trans *trans)
1619 {
1620         unsigned nr = trans->nr_paths * 2;
1621
1622         void *p = kvzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) +
1623                           sizeof(struct btree_trans_paths) +
1624                           nr * sizeof(struct btree_path) +
1625                           nr * sizeof(btree_path_idx_t) + 8 +
1626                           nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL);
1627
1628         unsigned long *paths_allocated = p;
1629         memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long));
1630         p += BITS_TO_LONGS(nr) * sizeof(unsigned long);
1631
1632         p += sizeof(struct btree_trans_paths);
1633         struct btree_path *paths = p;
1634         *trans_paths_nr(paths) = nr;
1635         memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path));
1636         p += nr * sizeof(struct btree_path);
1637
1638         btree_path_idx_t *sorted = p;
1639         memcpy(sorted, trans->sorted, trans->nr_sorted * sizeof(btree_path_idx_t));
1640         p += nr * sizeof(btree_path_idx_t) + 8;
1641
1642         struct btree_insert_entry *updates = p;
1643         memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_insert_entry));
1644
1645         unsigned long *old = trans->paths_allocated;
1646
1647         rcu_assign_pointer(trans->paths_allocated,      paths_allocated);
1648         rcu_assign_pointer(trans->paths,                paths);
1649         rcu_assign_pointer(trans->sorted,               sorted);
1650         rcu_assign_pointer(trans->updates,              updates);
1651
1652         trans->nr_paths         = nr;
1653
1654         if (old != trans->_paths_allocated)
1655                 kfree_rcu_mightsleep(old);
1656 }
1657
1658 static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans,
1659                                                 btree_path_idx_t pos)
1660 {
1661         btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths);
1662
1663         if (unlikely(idx == trans->nr_paths)) {
1664                 if (trans->nr_paths == BTREE_ITER_MAX) {
1665                         btree_path_overflow(trans);
1666                         return 0;
1667                 }
1668
1669                 btree_paths_realloc(trans);
1670         }
1671
1672         /*
1673          * Do this before marking the new path as allocated, since it won't be
1674          * initialized yet:
1675          */
1676         if (unlikely(idx > trans->nr_paths_max))
1677                 bch2_trans_update_max_paths(trans);
1678
1679         __set_bit(idx, trans->paths_allocated);
1680
1681         struct btree_path *path = &trans->paths[idx];
1682         path->ref               = 0;
1683         path->intent_ref        = 0;
1684         path->nodes_locked      = 0;
1685
1686         btree_path_list_add(trans, pos, idx);
1687         trans->paths_sorted = false;
1688         return idx;
1689 }
1690
1691 btree_path_idx_t bch2_path_get(struct btree_trans *trans,
1692                              enum btree_id btree_id, struct bpos pos,
1693                              unsigned locks_want, unsigned level,
1694                              unsigned flags, unsigned long ip)
1695 {
1696         struct btree_path *path;
1697         bool cached = flags & BTREE_ITER_cached;
1698         bool intent = flags & BTREE_ITER_intent;
1699         struct trans_for_each_path_inorder_iter iter;
1700         btree_path_idx_t path_pos = 0, path_idx;
1701
1702         bch2_trans_verify_not_unlocked(trans);
1703         bch2_trans_verify_not_in_restart(trans);
1704         bch2_trans_verify_locks(trans);
1705
1706         btree_trans_sort_paths(trans);
1707
1708         trans_for_each_path_inorder(trans, path, iter) {
1709                 if (__btree_path_cmp(path,
1710                                      btree_id,
1711                                      cached,
1712                                      pos,
1713                                      level) > 0)
1714                         break;
1715
1716                 path_pos = iter.path_idx;
1717         }
1718
1719         if (path_pos &&
1720             trans->paths[path_pos].cached       == cached &&
1721             trans->paths[path_pos].btree_id     == btree_id &&
1722             trans->paths[path_pos].level        == level) {
1723                 __btree_path_get(trans->paths + path_pos, intent);
1724                 path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1725                 path = trans->paths + path_idx;
1726         } else {
1727                 path_idx = btree_path_alloc(trans, path_pos);
1728                 path = trans->paths + path_idx;
1729
1730                 __btree_path_get(path, intent);
1731                 path->pos                       = pos;
1732                 path->btree_id                  = btree_id;
1733                 path->cached                    = cached;
1734                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1735                 path->should_be_locked          = false;
1736                 path->level                     = level;
1737                 path->locks_want                = locks_want;
1738                 path->nodes_locked              = 0;
1739                 for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++)
1740                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1741 #ifdef TRACK_PATH_ALLOCATED
1742                 path->ip_allocated              = ip;
1743 #endif
1744                 trans->paths_sorted             = false;
1745         }
1746
1747         if (!(flags & BTREE_ITER_nopreserve))
1748                 path->preserve = true;
1749
1750         if (path->intent_ref)
1751                 locks_want = max(locks_want, level + 1);
1752
1753         /*
1754          * If the path has locks_want greater than requested, we don't downgrade
1755          * it here - on transaction restart because btree node split needs to
1756          * upgrade locks, we might be putting/getting the iterator again.
1757          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1758          * a successful transaction commit.
1759          */
1760
1761         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1762         if (locks_want > path->locks_want)
1763                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL);
1764
1765         return path_idx;
1766 }
1767
1768 btree_path_idx_t bch2_path_get_unlocked_mut(struct btree_trans *trans,
1769                                             enum btree_id btree_id,
1770                                             unsigned level,
1771                                             struct bpos pos)
1772 {
1773         btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level,
1774                              BTREE_ITER_nopreserve|
1775                              BTREE_ITER_intent, _RET_IP_);
1776         path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_);
1777
1778         struct btree_path *path = trans->paths + path_idx;
1779         bch2_btree_path_downgrade(trans, path);
1780         __bch2_btree_path_unlock(trans, path);
1781         return path_idx;
1782 }
1783
1784 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1785 {
1786
1787         struct btree_path_level *l = path_l(path);
1788         struct bkey_packed *_k;
1789         struct bkey_s_c k;
1790
1791         if (unlikely(!l->b))
1792                 return bkey_s_c_null;
1793
1794         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1795         EBUG_ON(!btree_node_locked(path, path->level));
1796
1797         if (!path->cached) {
1798                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1799                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1800
1801                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos));
1802
1803                 if (!k.k || !bpos_eq(path->pos, k.k->p))
1804                         goto hole;
1805         } else {
1806                 struct bkey_cached *ck = (void *) path->l[0].b;
1807
1808                 EBUG_ON(ck &&
1809                         (path->btree_id != ck->key.btree_id ||
1810                          !bkey_eq(path->pos, ck->key.pos)));
1811                 if (!ck || !ck->valid)
1812                         return bkey_s_c_null;
1813
1814                 *u = ck->k->k;
1815                 k = bkey_i_to_s_c(ck->k);
1816         }
1817
1818         return k;
1819 hole:
1820         bkey_init(u);
1821         u->p = path->pos;
1822         return (struct bkey_s_c) { u, NULL };
1823 }
1824
1825
1826 void bch2_set_btree_iter_dontneed(struct btree_iter *iter)
1827 {
1828         struct btree_trans *trans = iter->trans;
1829
1830         if (!iter->path || trans->restarted)
1831                 return;
1832
1833         struct btree_path *path = btree_iter_path(trans, iter);
1834         path->preserve          = false;
1835         if (path->ref == 1)
1836                 path->should_be_locked  = false;
1837 }
1838 /* Btree iterators: */
1839
1840 int __must_check
1841 __bch2_btree_iter_traverse(struct btree_iter *iter)
1842 {
1843         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1844 }
1845
1846 int __must_check
1847 bch2_btree_iter_traverse(struct btree_iter *iter)
1848 {
1849         struct btree_trans *trans = iter->trans;
1850         int ret;
1851
1852         bch2_trans_verify_not_unlocked(trans);
1853
1854         iter->path = bch2_btree_path_set_pos(trans, iter->path,
1855                                         btree_iter_search_key(iter),
1856                                         iter->flags & BTREE_ITER_intent,
1857                                         btree_iter_ip_allocated(iter));
1858
1859         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1860         if (ret)
1861                 return ret;
1862
1863         struct btree_path *path = btree_iter_path(trans, iter);
1864         if (btree_path_node(path, path->level))
1865                 btree_path_set_should_be_locked(path);
1866         return 0;
1867 }
1868
1869 /* Iterate across nodes (leaf and interior nodes) */
1870
1871 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1872 {
1873         struct btree_trans *trans = iter->trans;
1874         struct btree *b = NULL;
1875         int ret;
1876
1877         EBUG_ON(trans->paths[iter->path].cached);
1878         bch2_btree_iter_verify(iter);
1879
1880         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1881         if (ret)
1882                 goto err;
1883
1884         struct btree_path *path = btree_iter_path(trans, iter);
1885         b = btree_path_node(path, path->level);
1886         if (!b)
1887                 goto out;
1888
1889         BUG_ON(bpos_lt(b->key.k.p, iter->pos));
1890
1891         bkey_init(&iter->k);
1892         iter->k.p = iter->pos = b->key.k.p;
1893
1894         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1895                                         iter->flags & BTREE_ITER_intent,
1896                                         btree_iter_ip_allocated(iter));
1897         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1898 out:
1899         bch2_btree_iter_verify_entry_exit(iter);
1900         bch2_btree_iter_verify(iter);
1901
1902         return b;
1903 err:
1904         b = ERR_PTR(ret);
1905         goto out;
1906 }
1907
1908 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter)
1909 {
1910         struct btree *b;
1911
1912         while (b = bch2_btree_iter_peek_node(iter),
1913                bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart))
1914                 bch2_trans_begin(iter->trans);
1915
1916         return b;
1917 }
1918
1919 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1920 {
1921         struct btree_trans *trans = iter->trans;
1922         struct btree *b = NULL;
1923         int ret;
1924
1925         EBUG_ON(trans->paths[iter->path].cached);
1926         bch2_trans_verify_not_in_restart(trans);
1927         bch2_btree_iter_verify(iter);
1928
1929         struct btree_path *path = btree_iter_path(trans, iter);
1930
1931         /* already at end? */
1932         if (!btree_path_node(path, path->level))
1933                 return NULL;
1934
1935         /* got to end? */
1936         if (!btree_path_node(path, path->level + 1)) {
1937                 btree_path_set_level_up(trans, path);
1938                 return NULL;
1939         }
1940
1941         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1942                 __bch2_btree_path_unlock(trans, path);
1943                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1944                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1945                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1946                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1947                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1948                 goto err;
1949         }
1950
1951         b = btree_path_node(path, path->level + 1);
1952
1953         if (bpos_eq(iter->pos, b->key.k.p)) {
1954                 __btree_path_set_level_up(trans, path, path->level++);
1955         } else {
1956                 if (btree_lock_want(path, path->level + 1) == BTREE_NODE_UNLOCKED)
1957                         btree_node_unlock(trans, path, path->level + 1);
1958
1959                 /*
1960                  * Haven't gotten to the end of the parent node: go back down to
1961                  * the next child node
1962                  */
1963                 iter->path = bch2_btree_path_set_pos(trans, iter->path,
1964                                         bpos_successor(iter->pos),
1965                                         iter->flags & BTREE_ITER_intent,
1966                                         btree_iter_ip_allocated(iter));
1967
1968                 path = btree_iter_path(trans, iter);
1969                 btree_path_set_level_down(trans, path, iter->min_depth);
1970
1971                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1972                 if (ret)
1973                         goto err;
1974
1975                 path = btree_iter_path(trans, iter);
1976                 b = path->l[path->level].b;
1977         }
1978
1979         bkey_init(&iter->k);
1980         iter->k.p = iter->pos = b->key.k.p;
1981
1982         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1983                                         iter->flags & BTREE_ITER_intent,
1984                                         btree_iter_ip_allocated(iter));
1985         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
1986         EBUG_ON(btree_iter_path(trans, iter)->uptodate);
1987 out:
1988         bch2_btree_iter_verify_entry_exit(iter);
1989         bch2_btree_iter_verify(iter);
1990
1991         return b;
1992 err:
1993         b = ERR_PTR(ret);
1994         goto out;
1995 }
1996
1997 /* Iterate across keys (in leaf nodes only) */
1998
1999 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
2000 {
2001         struct bpos pos = iter->k.p;
2002         bool ret = !(iter->flags & BTREE_ITER_all_snapshots
2003                      ? bpos_eq(pos, SPOS_MAX)
2004                      : bkey_eq(pos, SPOS_MAX));
2005
2006         if (ret && !(iter->flags & BTREE_ITER_is_extents))
2007                 pos = bkey_successor(iter, pos);
2008         bch2_btree_iter_set_pos(iter, pos);
2009         return ret;
2010 }
2011
2012 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
2013 {
2014         struct bpos pos = bkey_start_pos(&iter->k);
2015         bool ret = !(iter->flags & BTREE_ITER_all_snapshots
2016                      ? bpos_eq(pos, POS_MIN)
2017                      : bkey_eq(pos, POS_MIN));
2018
2019         if (ret && !(iter->flags & BTREE_ITER_is_extents))
2020                 pos = bkey_predecessor(iter, pos);
2021         bch2_btree_iter_set_pos(iter, pos);
2022         return ret;
2023 }
2024
2025 static noinline
2026 void bch2_btree_trans_peek_prev_updates(struct btree_trans *trans, struct btree_iter *iter,
2027                                         struct bkey_s_c *k)
2028 {
2029         struct bpos end = path_l(btree_iter_path(trans, iter))->b->data->min_key;
2030
2031         trans_for_each_update(trans, i)
2032                 if (!i->key_cache_already_flushed &&
2033                     i->btree_id == iter->btree_id &&
2034                     bpos_le(i->k->k.p, iter->pos) &&
2035                     bpos_ge(i->k->k.p, k->k ? k->k->p : end)) {
2036                         iter->k = i->k->k;
2037                         *k = bkey_i_to_s_c(i->k);
2038                 }
2039 }
2040
2041 static noinline
2042 void bch2_btree_trans_peek_updates(struct btree_trans *trans, struct btree_iter *iter,
2043                                    struct bkey_s_c *k)
2044 {
2045         struct btree_path *path = btree_iter_path(trans, iter);
2046         struct bpos end = path_l(path)->b->key.k.p;
2047
2048         trans_for_each_update(trans, i)
2049                 if (!i->key_cache_already_flushed &&
2050                     i->btree_id == iter->btree_id &&
2051                     bpos_ge(i->k->k.p, path->pos) &&
2052                     bpos_le(i->k->k.p, k->k ? k->k->p : end)) {
2053                         iter->k = i->k->k;
2054                         *k = bkey_i_to_s_c(i->k);
2055                 }
2056 }
2057
2058 static noinline
2059 void bch2_btree_trans_peek_slot_updates(struct btree_trans *trans, struct btree_iter *iter,
2060                                         struct bkey_s_c *k)
2061 {
2062         trans_for_each_update(trans, i)
2063                 if (!i->key_cache_already_flushed &&
2064                     i->btree_id == iter->btree_id &&
2065                     bpos_eq(i->k->k.p, iter->pos)) {
2066                         iter->k = i->k->k;
2067                         *k = bkey_i_to_s_c(i->k);
2068                 }
2069 }
2070
2071 static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
2072                                               struct btree_iter *iter,
2073                                               struct bpos end_pos)
2074 {
2075         struct btree_path *path = btree_iter_path(trans, iter);
2076
2077         return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
2078                                            path->level,
2079                                            path->pos,
2080                                            end_pos,
2081                                            &iter->journal_idx);
2082 }
2083
2084 static noinline
2085 struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
2086                                               struct btree_iter *iter)
2087 {
2088         struct btree_path *path = btree_iter_path(trans, iter);
2089         struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos);
2090
2091         if (k) {
2092                 iter->k = k->k;
2093                 return bkey_i_to_s_c(k);
2094         } else {
2095                 return bkey_s_c_null;
2096         }
2097 }
2098
2099 static noinline
2100 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
2101                                          struct btree_iter *iter,
2102                                          struct bkey_s_c k)
2103 {
2104         struct btree_path *path = btree_iter_path(trans, iter);
2105         struct bkey_i *next_journal =
2106                 bch2_btree_journal_peek(trans, iter,
2107                                 k.k ? k.k->p : path_l(path)->b->key.k.p);
2108
2109         if (next_journal) {
2110                 iter->k = next_journal->k;
2111                 k = bkey_i_to_s_c(next_journal);
2112         }
2113
2114         return k;
2115 }
2116
2117 /*
2118  * Checks btree key cache for key at iter->pos and returns it if present, or
2119  * bkey_s_c_null:
2120  */
2121 static noinline
2122 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
2123 {
2124         struct btree_trans *trans = iter->trans;
2125         struct bch_fs *c = trans->c;
2126         struct bkey u;
2127         struct bkey_s_c k;
2128         int ret;
2129
2130         bch2_trans_verify_not_in_restart(trans);
2131         bch2_trans_verify_not_unlocked(trans);
2132
2133         if ((iter->flags & BTREE_ITER_key_cache_fill) &&
2134             bpos_eq(iter->pos, pos))
2135                 return bkey_s_c_null;
2136
2137         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
2138                 return bkey_s_c_null;
2139
2140         if (!iter->key_cache_path)
2141                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
2142                                                      iter->flags & BTREE_ITER_intent, 0,
2143                                                      iter->flags|BTREE_ITER_cached|
2144                                                      BTREE_ITER_cached_nofill,
2145                                                      _THIS_IP_);
2146
2147         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
2148                                         iter->flags & BTREE_ITER_intent,
2149                                         btree_iter_ip_allocated(iter));
2150
2151         ret =   bch2_btree_path_traverse(trans, iter->key_cache_path,
2152                                          iter->flags|BTREE_ITER_cached) ?:
2153                 bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_);
2154         if (unlikely(ret))
2155                 return bkey_s_c_err(ret);
2156
2157         btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
2158
2159         k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
2160         if (k.k && !bkey_err(k)) {
2161                 iter->k = u;
2162                 k.k = &iter->k;
2163         }
2164         return k;
2165 }
2166
2167 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
2168 {
2169         struct btree_trans *trans = iter->trans;
2170         struct bkey_s_c k, k2;
2171         int ret;
2172
2173         EBUG_ON(btree_iter_path(trans, iter)->cached);
2174         bch2_btree_iter_verify(iter);
2175
2176         while (1) {
2177                 struct btree_path_level *l;
2178
2179                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2180                                         iter->flags & BTREE_ITER_intent,
2181                                         btree_iter_ip_allocated(iter));
2182
2183                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2184                 if (unlikely(ret)) {
2185                         /* ensure that iter->k is consistent with iter->pos: */
2186                         bch2_btree_iter_set_pos(iter, iter->pos);
2187                         k = bkey_s_c_err(ret);
2188                         goto out;
2189                 }
2190
2191                 struct btree_path *path = btree_iter_path(trans, iter);
2192                 l = path_l(path);
2193
2194                 if (unlikely(!l->b)) {
2195                         /* No btree nodes at requested level: */
2196                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2197                         k = bkey_s_c_null;
2198                         goto out;
2199                 }
2200
2201                 btree_path_set_should_be_locked(path);
2202
2203                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
2204
2205                 if (unlikely(iter->flags & BTREE_ITER_with_key_cache) &&
2206                     k.k &&
2207                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
2208                         k = k2;
2209                         ret = bkey_err(k);
2210                         if (ret) {
2211                                 bch2_btree_iter_set_pos(iter, iter->pos);
2212                                 goto out;
2213                         }
2214                 }
2215
2216                 if (unlikely(iter->flags & BTREE_ITER_with_journal))
2217                         k = btree_trans_peek_journal(trans, iter, k);
2218
2219                 if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
2220                              trans->nr_updates))
2221                         bch2_btree_trans_peek_updates(trans, iter, &k);
2222
2223                 if (k.k && bkey_deleted(k.k)) {
2224                         /*
2225                          * If we've got a whiteout, and it's after the search
2226                          * key, advance the search key to the whiteout instead
2227                          * of just after the whiteout - it might be a btree
2228                          * whiteout, with a real key at the same position, since
2229                          * in the btree deleted keys sort before non deleted.
2230                          */
2231                         search_key = !bpos_eq(search_key, k.k->p)
2232                                 ? k.k->p
2233                                 : bpos_successor(k.k->p);
2234                         continue;
2235                 }
2236
2237                 if (likely(k.k)) {
2238                         break;
2239                 } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) {
2240                         /* Advance to next leaf node: */
2241                         search_key = bpos_successor(l->b->key.k.p);
2242                 } else {
2243                         /* End of btree: */
2244                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
2245                         k = bkey_s_c_null;
2246                         goto out;
2247                 }
2248         }
2249 out:
2250         bch2_btree_iter_verify(iter);
2251
2252         return k;
2253 }
2254
2255 /**
2256  * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
2257  * iterator's current position
2258  * @iter:       iterator to peek from
2259  * @end:        search limit: returns keys less than or equal to @end
2260  *
2261  * Returns:     key if found, or an error extractable with bkey_err().
2262  */
2263 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
2264 {
2265         struct btree_trans *trans = iter->trans;
2266         struct bpos search_key = btree_iter_search_key(iter);
2267         struct bkey_s_c k;
2268         struct bpos iter_pos;
2269         int ret;
2270
2271         bch2_trans_verify_not_unlocked(trans);
2272         EBUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && bkey_eq(end, POS_MAX));
2273
2274         if (iter->update_path) {
2275                 bch2_path_put_nokeep(trans, iter->update_path,
2276                                      iter->flags & BTREE_ITER_intent);
2277                 iter->update_path = 0;
2278         }
2279
2280         bch2_btree_iter_verify_entry_exit(iter);
2281
2282         while (1) {
2283                 k = __bch2_btree_iter_peek(iter, search_key);
2284                 if (unlikely(!k.k))
2285                         goto end;
2286                 if (unlikely(bkey_err(k)))
2287                         goto out_no_locked;
2288
2289                 /*
2290                  * We need to check against @end before FILTER_SNAPSHOTS because
2291                  * if we get to a different inode that requested we might be
2292                  * seeing keys for a different snapshot tree that will all be
2293                  * filtered out.
2294                  *
2295                  * But we can't do the full check here, because bkey_start_pos()
2296                  * isn't monotonically increasing before FILTER_SNAPSHOTS, and
2297                  * that's what we check against in extents mode:
2298                  */
2299                 if (unlikely(!(iter->flags & BTREE_ITER_is_extents)
2300                              ? bkey_gt(k.k->p, end)
2301                              : k.k->p.inode > end.inode))
2302                         goto end;
2303
2304                 if (iter->update_path &&
2305                     !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
2306                         bch2_path_put_nokeep(trans, iter->update_path,
2307                                              iter->flags & BTREE_ITER_intent);
2308                         iter->update_path = 0;
2309                 }
2310
2311                 if ((iter->flags & BTREE_ITER_filter_snapshots) &&
2312                     (iter->flags & BTREE_ITER_intent) &&
2313                     !(iter->flags & BTREE_ITER_is_extents) &&
2314                     !iter->update_path) {
2315                         struct bpos pos = k.k->p;
2316
2317                         if (pos.snapshot < iter->snapshot) {
2318                                 search_key = bpos_successor(k.k->p);
2319                                 continue;
2320                         }
2321
2322                         pos.snapshot = iter->snapshot;
2323
2324                         /*
2325                          * advance, same as on exit for iter->path, but only up
2326                          * to snapshot
2327                          */
2328                         __btree_path_get(trans->paths + iter->path, iter->flags & BTREE_ITER_intent);
2329                         iter->update_path = iter->path;
2330
2331                         iter->update_path = bch2_btree_path_set_pos(trans,
2332                                                 iter->update_path, pos,
2333                                                 iter->flags & BTREE_ITER_intent,
2334                                                 _THIS_IP_);
2335                         ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
2336                         if (unlikely(ret)) {
2337                                 k = bkey_s_c_err(ret);
2338                                 goto out_no_locked;
2339                         }
2340                 }
2341
2342                 /*
2343                  * We can never have a key in a leaf node at POS_MAX, so
2344                  * we don't have to check these successor() calls:
2345                  */
2346                 if ((iter->flags & BTREE_ITER_filter_snapshots) &&
2347                     !bch2_snapshot_is_ancestor(trans->c,
2348                                                iter->snapshot,
2349                                                k.k->p.snapshot)) {
2350                         search_key = bpos_successor(k.k->p);
2351                         continue;
2352                 }
2353
2354                 if (bkey_whiteout(k.k) &&
2355                     !(iter->flags & BTREE_ITER_all_snapshots)) {
2356                         search_key = bkey_successor(iter, k.k->p);
2357                         continue;
2358                 }
2359
2360                 /*
2361                  * iter->pos should be mononotically increasing, and always be
2362                  * equal to the key we just returned - except extents can
2363                  * straddle iter->pos:
2364                  */
2365                 if (!(iter->flags & BTREE_ITER_is_extents))
2366                         iter_pos = k.k->p;
2367                 else
2368                         iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k));
2369
2370                 if (unlikely(!(iter->flags & BTREE_ITER_is_extents)
2371                              ? bkey_gt(iter_pos, end)
2372                              : bkey_ge(iter_pos, end)))
2373                         goto end;
2374
2375                 break;
2376         }
2377
2378         iter->pos = iter_pos;
2379
2380         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2381                                 iter->flags & BTREE_ITER_intent,
2382                                 btree_iter_ip_allocated(iter));
2383
2384         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2385 out_no_locked:
2386         if (iter->update_path) {
2387                 ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_);
2388                 if (unlikely(ret))
2389                         k = bkey_s_c_err(ret);
2390                 else
2391                         btree_path_set_should_be_locked(trans->paths + iter->update_path);
2392         }
2393
2394         if (!(iter->flags & BTREE_ITER_all_snapshots))
2395                 iter->pos.snapshot = iter->snapshot;
2396
2397         ret = bch2_btree_iter_verify_ret(iter, k);
2398         if (unlikely(ret)) {
2399                 bch2_btree_iter_set_pos(iter, iter->pos);
2400                 k = bkey_s_c_err(ret);
2401         }
2402
2403         bch2_btree_iter_verify_entry_exit(iter);
2404
2405         return k;
2406 end:
2407         bch2_btree_iter_set_pos(iter, end);
2408         k = bkey_s_c_null;
2409         goto out_no_locked;
2410 }
2411
2412 /**
2413  * bch2_btree_iter_next() - returns first key greater than iterator's current
2414  * position
2415  * @iter:       iterator to peek from
2416  *
2417  * Returns:     key if found, or an error extractable with bkey_err().
2418  */
2419 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2420 {
2421         if (!bch2_btree_iter_advance(iter))
2422                 return bkey_s_c_null;
2423
2424         return bch2_btree_iter_peek(iter);
2425 }
2426
2427 /**
2428  * bch2_btree_iter_peek_prev() - returns first key less than or equal to
2429  * iterator's current position
2430  * @iter:       iterator to peek from
2431  *
2432  * Returns:     key if found, or an error extractable with bkey_err().
2433  */
2434 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2435 {
2436         struct btree_trans *trans = iter->trans;
2437         struct bpos search_key = iter->pos;
2438         struct bkey_s_c k;
2439         struct bkey saved_k;
2440         const struct bch_val *saved_v;
2441         btree_path_idx_t saved_path = 0;
2442         int ret;
2443
2444         bch2_trans_verify_not_unlocked(trans);
2445         EBUG_ON(btree_iter_path(trans, iter)->cached ||
2446                 btree_iter_path(trans, iter)->level);
2447
2448         if (iter->flags & BTREE_ITER_with_journal)
2449                 return bkey_s_c_err(-BCH_ERR_btree_iter_with_journal_not_supported);
2450
2451         bch2_btree_iter_verify(iter);
2452         bch2_btree_iter_verify_entry_exit(iter);
2453
2454         if (iter->flags & BTREE_ITER_filter_snapshots)
2455                 search_key.snapshot = U32_MAX;
2456
2457         while (1) {
2458                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2459                                                 iter->flags & BTREE_ITER_intent,
2460                                                 btree_iter_ip_allocated(iter));
2461
2462                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2463                 if (unlikely(ret)) {
2464                         /* ensure that iter->k is consistent with iter->pos: */
2465                         bch2_btree_iter_set_pos(iter, iter->pos);
2466                         k = bkey_s_c_err(ret);
2467                         goto out_no_locked;
2468                 }
2469
2470                 struct btree_path *path = btree_iter_path(trans, iter);
2471
2472                 k = btree_path_level_peek(trans, path, &path->l[0], &iter->k);
2473                 if (!k.k ||
2474                     ((iter->flags & BTREE_ITER_is_extents)
2475                      ? bpos_ge(bkey_start_pos(k.k), search_key)
2476                      : bpos_gt(k.k->p, search_key)))
2477                         k = btree_path_level_prev(trans, path, &path->l[0], &iter->k);
2478
2479                 if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
2480                              trans->nr_updates))
2481                         bch2_btree_trans_peek_prev_updates(trans, iter, &k);
2482
2483                 if (likely(k.k)) {
2484                         if (iter->flags & BTREE_ITER_filter_snapshots) {
2485                                 if (k.k->p.snapshot == iter->snapshot)
2486                                         goto got_key;
2487
2488                                 /*
2489                                  * If we have a saved candidate, and we're no
2490                                  * longer at the same _key_ (not pos), return
2491                                  * that candidate
2492                                  */
2493                                 if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
2494                                         bch2_path_put_nokeep(trans, iter->path,
2495                                                       iter->flags & BTREE_ITER_intent);
2496                                         iter->path = saved_path;
2497                                         saved_path = 0;
2498                                         iter->k = saved_k;
2499                                         k.v     = saved_v;
2500                                         goto got_key;
2501                                 }
2502
2503                                 if (bch2_snapshot_is_ancestor(trans->c,
2504                                                               iter->snapshot,
2505                                                               k.k->p.snapshot)) {
2506                                         if (saved_path)
2507                                                 bch2_path_put_nokeep(trans, saved_path,
2508                                                       iter->flags & BTREE_ITER_intent);
2509                                         saved_path = btree_path_clone(trans, iter->path,
2510                                                                 iter->flags & BTREE_ITER_intent,
2511                                                                 _THIS_IP_);
2512                                         path = btree_iter_path(trans, iter);
2513                                         saved_k = *k.k;
2514                                         saved_v = k.v;
2515                                 }
2516
2517                                 search_key = bpos_predecessor(k.k->p);
2518                                 continue;
2519                         }
2520 got_key:
2521                         if (bkey_whiteout(k.k) &&
2522                             !(iter->flags & BTREE_ITER_all_snapshots)) {
2523                                 search_key = bkey_predecessor(iter, k.k->p);
2524                                 if (iter->flags & BTREE_ITER_filter_snapshots)
2525                                         search_key.snapshot = U32_MAX;
2526                                 continue;
2527                         }
2528
2529                         btree_path_set_should_be_locked(path);
2530                         break;
2531                 } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
2532                         /* Advance to previous leaf node: */
2533                         search_key = bpos_predecessor(path->l[0].b->data->min_key);
2534                 } else {
2535                         /* Start of btree: */
2536                         bch2_btree_iter_set_pos(iter, POS_MIN);
2537                         k = bkey_s_c_null;
2538                         goto out_no_locked;
2539                 }
2540         }
2541
2542         EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
2543
2544         /* Extents can straddle iter->pos: */
2545         if (bkey_lt(k.k->p, iter->pos))
2546                 iter->pos = k.k->p;
2547
2548         if (iter->flags & BTREE_ITER_filter_snapshots)
2549                 iter->pos.snapshot = iter->snapshot;
2550 out_no_locked:
2551         if (saved_path)
2552                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_intent);
2553
2554         bch2_btree_iter_verify_entry_exit(iter);
2555         bch2_btree_iter_verify(iter);
2556
2557         return k;
2558 }
2559
2560 /**
2561  * bch2_btree_iter_prev() - returns first key less than iterator's current
2562  * position
2563  * @iter:       iterator to peek from
2564  *
2565  * Returns:     key if found, or an error extractable with bkey_err().
2566  */
2567 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2568 {
2569         if (!bch2_btree_iter_rewind(iter))
2570                 return bkey_s_c_null;
2571
2572         return bch2_btree_iter_peek_prev(iter);
2573 }
2574
2575 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2576 {
2577         struct btree_trans *trans = iter->trans;
2578         struct bpos search_key;
2579         struct bkey_s_c k;
2580         int ret;
2581
2582         bch2_trans_verify_not_unlocked(trans);
2583         bch2_btree_iter_verify(iter);
2584         bch2_btree_iter_verify_entry_exit(iter);
2585         EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_with_key_cache));
2586
2587         /* extents can't span inode numbers: */
2588         if ((iter->flags & BTREE_ITER_is_extents) &&
2589             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2590                 if (iter->pos.inode == KEY_INODE_MAX)
2591                         return bkey_s_c_null;
2592
2593                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2594         }
2595
2596         search_key = btree_iter_search_key(iter);
2597         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2598                                         iter->flags & BTREE_ITER_intent,
2599                                         btree_iter_ip_allocated(iter));
2600
2601         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2602         if (unlikely(ret)) {
2603                 k = bkey_s_c_err(ret);
2604                 goto out_no_locked;
2605         }
2606
2607         if ((iter->flags & BTREE_ITER_cached) ||
2608             !(iter->flags & (BTREE_ITER_is_extents|BTREE_ITER_filter_snapshots))) {
2609                 k = bkey_s_c_null;
2610
2611                 if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
2612                              trans->nr_updates)) {
2613                         bch2_btree_trans_peek_slot_updates(trans, iter, &k);
2614                         if (k.k)
2615                                 goto out;
2616                 }
2617
2618                 if (unlikely(iter->flags & BTREE_ITER_with_journal) &&
2619                     (k = btree_trans_peek_slot_journal(trans, iter)).k)
2620                         goto out;
2621
2622                 if (unlikely(iter->flags & BTREE_ITER_with_key_cache) &&
2623                     (k = btree_trans_peek_key_cache(iter, iter->pos)).k) {
2624                         if (!bkey_err(k))
2625                                 iter->k = *k.k;
2626                         /* We're not returning a key from iter->path: */
2627                         goto out_no_locked;
2628                 }
2629
2630                 k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k);
2631                 if (unlikely(!k.k))
2632                         goto out_no_locked;
2633         } else {
2634                 struct bpos next;
2635                 struct bpos end = iter->pos;
2636
2637                 if (iter->flags & BTREE_ITER_is_extents)
2638                         end.offset = U64_MAX;
2639
2640                 EBUG_ON(btree_iter_path(trans, iter)->level);
2641
2642                 if (iter->flags & BTREE_ITER_intent) {
2643                         struct btree_iter iter2;
2644
2645                         bch2_trans_copy_iter(&iter2, iter);
2646                         k = bch2_btree_iter_peek_upto(&iter2, end);
2647
2648                         if (k.k && !bkey_err(k)) {
2649                                 swap(iter->key_cache_path, iter2.key_cache_path);
2650                                 iter->k = iter2.k;
2651                                 k.k = &iter->k;
2652                         }
2653                         bch2_trans_iter_exit(trans, &iter2);
2654                 } else {
2655                         struct bpos pos = iter->pos;
2656
2657                         k = bch2_btree_iter_peek_upto(iter, end);
2658                         if (unlikely(bkey_err(k)))
2659                                 bch2_btree_iter_set_pos(iter, pos);
2660                         else
2661                                 iter->pos = pos;
2662                 }
2663
2664                 if (unlikely(bkey_err(k)))
2665                         goto out_no_locked;
2666
2667                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2668
2669                 if (bkey_lt(iter->pos, next)) {
2670                         bkey_init(&iter->k);
2671                         iter->k.p = iter->pos;
2672
2673                         if (iter->flags & BTREE_ITER_is_extents) {
2674                                 bch2_key_resize(&iter->k,
2675                                                 min_t(u64, KEY_SIZE_MAX,
2676                                                       (next.inode == iter->pos.inode
2677                                                        ? next.offset
2678                                                        : KEY_OFFSET_MAX) -
2679                                                       iter->pos.offset));
2680                                 EBUG_ON(!iter->k.size);
2681                         }
2682
2683                         k = (struct bkey_s_c) { &iter->k, NULL };
2684                 }
2685         }
2686 out:
2687         btree_path_set_should_be_locked(btree_iter_path(trans, iter));
2688 out_no_locked:
2689         bch2_btree_iter_verify_entry_exit(iter);
2690         bch2_btree_iter_verify(iter);
2691         ret = bch2_btree_iter_verify_ret(iter, k);
2692         if (unlikely(ret))
2693                 return bkey_s_c_err(ret);
2694
2695         return k;
2696 }
2697
2698 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2699 {
2700         if (!bch2_btree_iter_advance(iter))
2701                 return bkey_s_c_null;
2702
2703         return bch2_btree_iter_peek_slot(iter);
2704 }
2705
2706 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2707 {
2708         if (!bch2_btree_iter_rewind(iter))
2709                 return bkey_s_c_null;
2710
2711         return bch2_btree_iter_peek_slot(iter);
2712 }
2713
2714 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter)
2715 {
2716         struct bkey_s_c k;
2717
2718         while (btree_trans_too_many_iters(iter->trans) ||
2719                (k = bch2_btree_iter_peek_type(iter, iter->flags),
2720                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
2721                 bch2_trans_begin(iter->trans);
2722
2723         return k;
2724 }
2725
2726 /* new transactional stuff: */
2727
2728 #ifdef CONFIG_BCACHEFS_DEBUG
2729 static void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2730 {
2731         struct btree_path *path;
2732         unsigned i;
2733
2734         BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1);
2735
2736         trans_for_each_path(trans, path, i) {
2737                 BUG_ON(path->sorted_idx >= trans->nr_sorted);
2738                 BUG_ON(trans->sorted[path->sorted_idx] != i);
2739         }
2740
2741         for (i = 0; i < trans->nr_sorted; i++) {
2742                 unsigned idx = trans->sorted[i];
2743
2744                 BUG_ON(!test_bit(idx, trans->paths_allocated));
2745                 BUG_ON(trans->paths[idx].sorted_idx != i);
2746         }
2747 }
2748
2749 static void btree_trans_verify_sorted(struct btree_trans *trans)
2750 {
2751         struct btree_path *path, *prev = NULL;
2752         struct trans_for_each_path_inorder_iter iter;
2753
2754         if (!bch2_debug_check_iterators)
2755                 return;
2756
2757         trans_for_each_path_inorder(trans, path, iter) {
2758                 if (prev && btree_path_cmp(prev, path) > 0) {
2759                         __bch2_dump_trans_paths_updates(trans, true);
2760                         panic("trans paths out of order!\n");
2761                 }
2762                 prev = path;
2763         }
2764 }
2765 #else
2766 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {}
2767 static inline void btree_trans_verify_sorted(struct btree_trans *trans) {}
2768 #endif
2769
2770 void __bch2_btree_trans_sort_paths(struct btree_trans *trans)
2771 {
2772         int i, l = 0, r = trans->nr_sorted, inc = 1;
2773         bool swapped;
2774
2775         btree_trans_verify_sorted_refs(trans);
2776
2777         if (trans->paths_sorted)
2778                 goto out;
2779
2780         /*
2781          * Cocktail shaker sort: this is efficient because iterators will be
2782          * mostly sorted.
2783          */
2784         do {
2785                 swapped = false;
2786
2787                 for (i = inc > 0 ? l : r - 2;
2788                      i + 1 < r && i >= l;
2789                      i += inc) {
2790                         if (btree_path_cmp(trans->paths + trans->sorted[i],
2791                                            trans->paths + trans->sorted[i + 1]) > 0) {
2792                                 swap(trans->sorted[i], trans->sorted[i + 1]);
2793                                 trans->paths[trans->sorted[i]].sorted_idx = i;
2794                                 trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1;
2795                                 swapped = true;
2796                         }
2797                 }
2798
2799                 if (inc > 0)
2800                         --r;
2801                 else
2802                         l++;
2803                 inc = -inc;
2804         } while (swapped);
2805
2806         trans->paths_sorted = true;
2807 out:
2808         btree_trans_verify_sorted(trans);
2809 }
2810
2811 static inline void btree_path_list_remove(struct btree_trans *trans,
2812                                           struct btree_path *path)
2813 {
2814         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2815 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2816         trans->nr_sorted--;
2817         memmove_u64s_down_small(trans->sorted + path->sorted_idx,
2818                                 trans->sorted + path->sorted_idx + 1,
2819                                 DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2820                                              sizeof(u64) / sizeof(btree_path_idx_t)));
2821 #else
2822         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2823 #endif
2824         for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
2825                 trans->paths[trans->sorted[i]].sorted_idx = i;
2826 }
2827
2828 static inline void btree_path_list_add(struct btree_trans *trans,
2829                                        btree_path_idx_t pos,
2830                                        btree_path_idx_t path_idx)
2831 {
2832         struct btree_path *path = trans->paths + path_idx;
2833
2834         path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted;
2835
2836 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
2837         memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1,
2838                               trans->sorted + path->sorted_idx,
2839                               DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx,
2840                                            sizeof(u64) / sizeof(btree_path_idx_t)));
2841         trans->nr_sorted++;
2842         trans->sorted[path->sorted_idx] = path_idx;
2843 #else
2844         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx);
2845 #endif
2846
2847         for (unsigned i = path->sorted_idx; i < trans->nr_sorted; i++)
2848                 trans->paths[trans->sorted[i]].sorted_idx = i;
2849
2850         btree_trans_verify_sorted_refs(trans);
2851 }
2852
2853 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2854 {
2855         if (iter->update_path)
2856                 bch2_path_put_nokeep(trans, iter->update_path,
2857                               iter->flags & BTREE_ITER_intent);
2858         if (iter->path)
2859                 bch2_path_put(trans, iter->path,
2860                               iter->flags & BTREE_ITER_intent);
2861         if (iter->key_cache_path)
2862                 bch2_path_put(trans, iter->key_cache_path,
2863                               iter->flags & BTREE_ITER_intent);
2864         iter->path              = 0;
2865         iter->update_path       = 0;
2866         iter->key_cache_path    = 0;
2867         iter->trans             = NULL;
2868 }
2869
2870 void bch2_trans_iter_init_outlined(struct btree_trans *trans,
2871                           struct btree_iter *iter,
2872                           enum btree_id btree_id, struct bpos pos,
2873                           unsigned flags)
2874 {
2875         bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
2876                                bch2_btree_iter_flags(trans, btree_id, flags),
2877                                _RET_IP_);
2878 }
2879
2880 void bch2_trans_node_iter_init(struct btree_trans *trans,
2881                                struct btree_iter *iter,
2882                                enum btree_id btree_id,
2883                                struct bpos pos,
2884                                unsigned locks_want,
2885                                unsigned depth,
2886                                unsigned flags)
2887 {
2888         flags |= BTREE_ITER_not_extents;
2889         flags |= BTREE_ITER_snapshot_field;
2890         flags |= BTREE_ITER_all_snapshots;
2891
2892         bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
2893                                __bch2_btree_iter_flags(trans, btree_id, flags),
2894                                _RET_IP_);
2895
2896         iter->min_depth = depth;
2897
2898         struct btree_path *path = btree_iter_path(trans, iter);
2899         BUG_ON(path->locks_want  < min(locks_want, BTREE_MAX_DEPTH));
2900         BUG_ON(path->level      != depth);
2901         BUG_ON(iter->min_depth  != depth);
2902 }
2903
2904 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2905 {
2906         struct btree_trans *trans = src->trans;
2907
2908         *dst = *src;
2909 #ifdef TRACK_PATH_ALLOCATED
2910         dst->ip_allocated = _RET_IP_;
2911 #endif
2912         if (src->path)
2913                 __btree_path_get(trans->paths + src->path, src->flags & BTREE_ITER_intent);
2914         if (src->update_path)
2915                 __btree_path_get(trans->paths + src->update_path, src->flags & BTREE_ITER_intent);
2916         dst->key_cache_path = 0;
2917 }
2918
2919 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2920 {
2921         struct bch_fs *c = trans->c;
2922         unsigned new_top = trans->mem_top + size;
2923         unsigned old_bytes = trans->mem_bytes;
2924         unsigned new_bytes = roundup_pow_of_two(new_top);
2925         int ret;
2926         void *new_mem;
2927         void *p;
2928
2929         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2930
2931         struct btree_transaction_stats *s = btree_trans_stats(trans);
2932         s->max_mem = max(s->max_mem, new_bytes);
2933
2934         if (trans->used_mempool) {
2935                 if (trans->mem_bytes >= new_bytes)
2936                         goto out_change_top;
2937
2938                 /* No more space from mempool item, need malloc new one */
2939                 new_mem = kmalloc(new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2940                 if (unlikely(!new_mem)) {
2941                         bch2_trans_unlock(trans);
2942
2943                         new_mem = kmalloc(new_bytes, GFP_KERNEL);
2944                         if (!new_mem)
2945                                 return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2946
2947                         ret = bch2_trans_relock(trans);
2948                         if (ret) {
2949                                 kfree(new_mem);
2950                                 return ERR_PTR(ret);
2951                         }
2952                 }
2953                 memcpy(new_mem, trans->mem, trans->mem_top);
2954                 trans->used_mempool = false;
2955                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
2956                 goto out_new_mem;
2957         }
2958
2959         new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN);
2960         if (unlikely(!new_mem)) {
2961                 bch2_trans_unlock(trans);
2962
2963                 new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL);
2964                 if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2965                         new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2966                         new_bytes = BTREE_TRANS_MEM_MAX;
2967                         memcpy(new_mem, trans->mem, trans->mem_top);
2968                         trans->used_mempool = true;
2969                         kfree(trans->mem);
2970                 }
2971
2972                 if (!new_mem)
2973                         return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc);
2974
2975                 trans->mem = new_mem;
2976                 trans->mem_bytes = new_bytes;
2977
2978                 ret = bch2_trans_relock(trans);
2979                 if (ret)
2980                         return ERR_PTR(ret);
2981         }
2982 out_new_mem:
2983         trans->mem = new_mem;
2984         trans->mem_bytes = new_bytes;
2985
2986         if (old_bytes) {
2987                 trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2988                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2989         }
2990 out_change_top:
2991         p = trans->mem + trans->mem_top;
2992         trans->mem_top += size;
2993         memset(p, 0, size);
2994         return p;
2995 }
2996
2997 static inline void check_srcu_held_too_long(struct btree_trans *trans)
2998 {
2999         WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10),
3000              "btree trans held srcu lock (delaying memory reclaim) for %lu seconds",
3001              (jiffies - trans->srcu_lock_time) / HZ);
3002 }
3003
3004 void bch2_trans_srcu_unlock(struct btree_trans *trans)
3005 {
3006         if (trans->srcu_held) {
3007                 struct bch_fs *c = trans->c;
3008                 struct btree_path *path;
3009                 unsigned i;
3010
3011                 trans_for_each_path(trans, path, i)
3012                         if (path->cached && !btree_node_locked(path, 0))
3013                                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset);
3014
3015                 check_srcu_held_too_long(trans);
3016                 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3017                 trans->srcu_held = false;
3018         }
3019 }
3020
3021 static void bch2_trans_srcu_lock(struct btree_trans *trans)
3022 {
3023         if (!trans->srcu_held) {
3024                 trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier);
3025                 trans->srcu_lock_time   = jiffies;
3026                 trans->srcu_held = true;
3027         }
3028 }
3029
3030 /**
3031  * bch2_trans_begin() - reset a transaction after a interrupted attempt
3032  * @trans: transaction to reset
3033  *
3034  * Returns:     current restart counter, to be used with trans_was_restarted()
3035  *
3036  * While iterating over nodes or updating nodes a attempt to lock a btree node
3037  * may return BCH_ERR_transaction_restart when the trylock fails. When this
3038  * occurs bch2_trans_begin() should be called and the transaction retried.
3039  */
3040 u32 bch2_trans_begin(struct btree_trans *trans)
3041 {
3042         struct btree_path *path;
3043         unsigned i;
3044         u64 now;
3045
3046         bch2_trans_reset_updates(trans);
3047
3048         trans->restart_count++;
3049         trans->mem_top                  = 0;
3050         trans->journal_entries          = NULL;
3051
3052         trans_for_each_path(trans, path, i) {
3053                 path->should_be_locked = false;
3054
3055                 /*
3056                  * If the transaction wasn't restarted, we're presuming to be
3057                  * doing something new: dont keep iterators excpt the ones that
3058                  * are in use - except for the subvolumes btree:
3059                  */
3060                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
3061                         path->preserve = false;
3062
3063                 /*
3064                  * XXX: we probably shouldn't be doing this if the transaction
3065                  * was restarted, but currently we still overflow transaction
3066                  * iterators if we do that
3067                  */
3068                 if (!path->ref && !path->preserve)
3069                         __bch2_path_free(trans, i);
3070                 else
3071                         path->preserve = false;
3072         }
3073
3074         now = local_clock();
3075
3076         if (!IS_ENABLED(CONFIG_BCACHEFS_NO_LATENCY_ACCT) &&
3077             time_after64(now, trans->last_begin_time + 10))
3078                 __bch2_time_stats_update(&btree_trans_stats(trans)->duration,
3079                                          trans->last_begin_time, now);
3080
3081         if (!trans->restarted &&
3082             (need_resched() ||
3083              time_after64(now, trans->last_begin_time + BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS))) {
3084                 bch2_trans_unlock(trans);
3085                 cond_resched();
3086                 now = local_clock();
3087         }
3088         trans->last_begin_time = now;
3089
3090         if (unlikely(trans->srcu_held &&
3091                      time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
3092                 bch2_trans_srcu_unlock(trans);
3093
3094         trans->last_begin_ip = _RET_IP_;
3095         trans->locked  = true;
3096
3097         if (trans->restarted) {
3098                 bch2_btree_path_traverse_all(trans);
3099                 trans->notrace_relock_fail = false;
3100         }
3101
3102         bch2_trans_verify_not_unlocked(trans);
3103         return trans->restart_count;
3104 }
3105
3106 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR] = { "(unknown)" };
3107
3108 unsigned bch2_trans_get_fn_idx(const char *fn)
3109 {
3110         for (unsigned i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
3111                 if (!bch2_btree_transaction_fns[i] ||
3112                     bch2_btree_transaction_fns[i] == fn) {
3113                         bch2_btree_transaction_fns[i] = fn;
3114                         return i;
3115                 }
3116
3117         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
3118         return 0;
3119 }
3120
3121 struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx)
3122         __acquires(&c->btree_trans_barrier)
3123 {
3124         struct btree_trans *trans;
3125
3126         if (IS_ENABLED(__KERNEL__)) {
3127                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL);
3128                 if (trans) {
3129                         memset(trans, 0, offsetof(struct btree_trans, list));
3130                         goto got_trans;
3131                 }
3132         }
3133
3134         trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS);
3135         memset(trans, 0, sizeof(*trans));
3136         closure_init_stack(&trans->ref);
3137
3138         seqmutex_lock(&c->btree_trans_lock);
3139         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
3140                 struct btree_trans *pos;
3141                 pid_t pid = current->pid;
3142
3143                 trans->locking_wait.task = current;
3144
3145                 list_for_each_entry(pos, &c->btree_trans_list, list) {
3146                         struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task);
3147                         /*
3148                          * We'd much prefer to be stricter here and completely
3149                          * disallow multiple btree_trans in the same thread -
3150                          * but the data move path calls bch2_write when we
3151                          * already have a btree_trans initialized.
3152                          */
3153                         BUG_ON(pos_task &&
3154                                pid == pos_task->pid &&
3155                                pos->locked);
3156
3157                         if (pos_task && pid < pos_task->pid) {
3158                                 list_add_tail(&trans->list, &pos->list);
3159                                 goto list_add_done;
3160                         }
3161                 }
3162         }
3163         list_add_tail(&trans->list, &c->btree_trans_list);
3164 list_add_done:
3165         seqmutex_unlock(&c->btree_trans_lock);
3166 got_trans:
3167         trans->c                = c;
3168         trans->last_begin_time  = local_clock();
3169         trans->fn_idx           = fn_idx;
3170         trans->locking_wait.task = current;
3171         trans->locked           = true;
3172         trans->journal_replay_not_finished =
3173                 unlikely(!test_bit(JOURNAL_replay_done, &c->journal.flags)) &&
3174                 atomic_inc_not_zero(&c->journal_keys.ref);
3175         trans->nr_paths         = ARRAY_SIZE(trans->_paths);
3176         trans->paths_allocated  = trans->_paths_allocated;
3177         trans->sorted           = trans->_sorted;
3178         trans->paths            = trans->_paths;
3179         trans->updates          = trans->_updates;
3180
3181         *trans_paths_nr(trans->paths) = BTREE_ITER_INITIAL;
3182
3183         trans->paths_allocated[0] = 1;
3184
3185         if (fn_idx < BCH_TRANSACTIONS_NR) {
3186                 trans->fn = bch2_btree_transaction_fns[fn_idx];
3187
3188                 struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx];
3189
3190                 if (s->max_mem) {
3191                         unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
3192
3193                         trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
3194                         if (likely(trans->mem))
3195                                 trans->mem_bytes = expected_mem_bytes;
3196                 }
3197
3198                 trans->nr_paths_max = s->nr_max_paths;
3199                 trans->journal_entries_size = s->journal_entries_size;
3200         }
3201
3202         trans->srcu_idx         = srcu_read_lock(&c->btree_trans_barrier);
3203         trans->srcu_lock_time   = jiffies;
3204         trans->srcu_held        = true;
3205         return trans;
3206 }
3207
3208 static void check_btree_paths_leaked(struct btree_trans *trans)
3209 {
3210 #ifdef CONFIG_BCACHEFS_DEBUG
3211         struct bch_fs *c = trans->c;
3212         struct btree_path *path;
3213         unsigned i;
3214
3215         trans_for_each_path(trans, path, i)
3216                 if (path->ref)
3217                         goto leaked;
3218         return;
3219 leaked:
3220         bch_err(c, "btree paths leaked from %s!", trans->fn);
3221         trans_for_each_path(trans, path, i)
3222                 if (path->ref)
3223                         printk(KERN_ERR "  btree %s %pS\n",
3224                                bch2_btree_id_str(path->btree_id),
3225                                (void *) path->ip_allocated);
3226         /* Be noisy about this: */
3227         bch2_fatal_error(c);
3228 #endif
3229 }
3230
3231 void bch2_trans_put(struct btree_trans *trans)
3232         __releases(&c->btree_trans_barrier)
3233 {
3234         struct bch_fs *c = trans->c;
3235
3236         bch2_trans_unlock(trans);
3237
3238         trans_for_each_update(trans, i)
3239                 __btree_path_put(trans->paths + i->path, true);
3240         trans->nr_updates       = 0;
3241         trans->locking_wait.task = NULL;
3242
3243         check_btree_paths_leaked(trans);
3244
3245         if (trans->srcu_held) {
3246                 check_srcu_held_too_long(trans);
3247                 srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
3248         }
3249
3250         if (trans->fs_usage_deltas) {
3251                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
3252                     REPLICAS_DELTA_LIST_MAX)
3253                         mempool_free(trans->fs_usage_deltas,
3254                                      &c->replicas_delta_pool);
3255                 else
3256                         kfree(trans->fs_usage_deltas);
3257         }
3258
3259         if (unlikely(trans->journal_replay_not_finished))
3260                 bch2_journal_keys_put(c);
3261
3262         unsigned long *paths_allocated = trans->paths_allocated;
3263         trans->paths_allocated  = NULL;
3264         trans->paths            = NULL;
3265
3266         if (paths_allocated != trans->_paths_allocated)
3267                 kvfree_rcu_mightsleep(paths_allocated);
3268
3269         if (trans->used_mempool)
3270                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3271         else
3272                 kfree(trans->mem);
3273
3274         /* Userspace doesn't have a real percpu implementation: */
3275         if (IS_ENABLED(__KERNEL__))
3276                 trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans);
3277
3278         if (trans) {
3279                 closure_sync(&trans->ref);
3280
3281                 seqmutex_lock(&c->btree_trans_lock);
3282                 list_del(&trans->list);
3283                 seqmutex_unlock(&c->btree_trans_lock);
3284
3285                 mempool_free(trans, &c->btree_trans_pool);
3286         }
3287 }
3288
3289 static void __maybe_unused
3290 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3291                                       struct btree_bkey_cached_common *b)
3292 {
3293         struct six_lock_count c = six_lock_counts(&b->lock);
3294         struct task_struct *owner;
3295         pid_t pid;
3296
3297         rcu_read_lock();
3298         owner = READ_ONCE(b->lock.owner);
3299         pid = owner ? owner->pid : 0;
3300         rcu_read_unlock();
3301
3302         prt_printf(out, "\t%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3303                    b->level, bch2_btree_id_str(b->btree_id));
3304         bch2_bpos_to_text(out, btree_node_pos(b));
3305
3306         prt_printf(out, "\t locks %u:%u:%u held by pid %u",
3307                    c.n[0], c.n[1], c.n[2], pid);
3308 }
3309
3310 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3311 {
3312         struct btree_bkey_cached_common *b;
3313         static char lock_types[] = { 'r', 'i', 'w' };
3314         struct task_struct *task = READ_ONCE(trans->locking_wait.task);
3315         unsigned l, idx;
3316
3317         /* before rcu_read_lock(): */
3318         bch2_printbuf_make_room(out, 4096);
3319
3320         if (!out->nr_tabstops) {
3321                 printbuf_tabstop_push(out, 16);
3322                 printbuf_tabstop_push(out, 32);
3323         }
3324
3325         prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn);
3326
3327         /* trans->paths is rcu protected vs. freeing */
3328         rcu_read_lock();
3329         out->atomic++;
3330
3331         struct btree_path *paths = rcu_dereference(trans->paths);
3332         if (!paths)
3333                 goto out;
3334
3335         unsigned long *paths_allocated = trans_paths_allocated(paths);
3336
3337         trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) {
3338                 struct btree_path *path = paths + idx;
3339                 if (!path->nodes_locked)
3340                         continue;
3341
3342                 prt_printf(out, "  path %u %c l=%u %s:",
3343                        idx,
3344                        path->cached ? 'c' : 'b',
3345                        path->level,
3346                        bch2_btree_id_str(path->btree_id));
3347                 bch2_bpos_to_text(out, path->pos);
3348                 prt_newline(out);
3349
3350                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3351                         if (btree_node_locked(path, l) &&
3352                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3353                                 prt_printf(out, "    %c l=%u ",
3354                                            lock_types[btree_node_locked_type(path, l)], l);
3355                                 bch2_btree_bkey_cached_common_to_text(out, b);
3356                                 prt_newline(out);
3357                         }
3358                 }
3359         }
3360
3361         b = READ_ONCE(trans->locking);
3362         if (b) {
3363                 prt_printf(out, "  blocked for %lluus on\n",
3364                            div_u64(local_clock() - trans->locking_wait.start_time, 1000));
3365                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3366                 bch2_btree_bkey_cached_common_to_text(out, b);
3367                 prt_newline(out);
3368         }
3369 out:
3370         --out->atomic;
3371         rcu_read_unlock();
3372 }
3373
3374 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3375 {
3376         struct btree_transaction_stats *s;
3377         struct btree_trans *trans;
3378         int cpu;
3379
3380         if (c->btree_trans_bufs)
3381                 for_each_possible_cpu(cpu) {
3382                         struct btree_trans *trans =
3383                                 per_cpu_ptr(c->btree_trans_bufs, cpu)->trans;
3384
3385                         if (trans) {
3386                                 closure_sync(&trans->ref);
3387
3388                                 seqmutex_lock(&c->btree_trans_lock);
3389                                 list_del(&trans->list);
3390                                 seqmutex_unlock(&c->btree_trans_lock);
3391                         }
3392                         kfree(trans);
3393                 }
3394         free_percpu(c->btree_trans_bufs);
3395
3396         trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list);
3397         if (trans)
3398                 panic("%s leaked btree_trans\n", trans->fn);
3399
3400         for (s = c->btree_transaction_stats;
3401              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3402              s++) {
3403                 kfree(s->max_paths_text);
3404                 bch2_time_stats_exit(&s->lock_hold_times);
3405         }
3406
3407         if (c->btree_trans_barrier_initialized)
3408                 cleanup_srcu_struct(&c->btree_trans_barrier);
3409         mempool_exit(&c->btree_trans_mem_pool);
3410         mempool_exit(&c->btree_trans_pool);
3411 }
3412
3413 void bch2_fs_btree_iter_init_early(struct bch_fs *c)
3414 {
3415         struct btree_transaction_stats *s;
3416
3417         for (s = c->btree_transaction_stats;
3418              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3419              s++) {
3420                 bch2_time_stats_init(&s->duration);
3421                 bch2_time_stats_init(&s->lock_hold_times);
3422                 mutex_init(&s->lock);
3423         }
3424
3425         INIT_LIST_HEAD(&c->btree_trans_list);
3426         seqmutex_init(&c->btree_trans_lock);
3427 }
3428
3429 int bch2_fs_btree_iter_init(struct bch_fs *c)
3430 {
3431         int ret;
3432
3433         c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf);
3434         if (!c->btree_trans_bufs)
3435                 return -ENOMEM;
3436
3437         ret   = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1,
3438                                           sizeof(struct btree_trans)) ?:
3439                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3440                                           BTREE_TRANS_MEM_MAX) ?:
3441                 init_srcu_struct(&c->btree_trans_barrier);
3442         if (!ret)
3443                 c->btree_trans_barrier_initialized = true;
3444         return ret;
3445 }