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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_buf.h"
   6 #include "btree_cache.h"
   7 #include "btree_iter.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "debug.h"
  12 #include "error.h"
  13 #include "extents.h"
  14 #include "journal.h"
  15 #include "recovery.h"
  16 #include "replicas.h"
  17 #include "subvolume.h"
  18
  19 #include <linux/prandom.h>
  20 #include <linux/prefetch.h>
  21 #include <trace/events/bcachefs.h>
  22
  23 static void btree_trans_verify_sorted(struct btree_trans *);
  24 inline void bch2_btree_path_check_sort(struct btree_trans *, struct btree_path *, int);
  25 static __always_inline void bch2_btree_path_check_sort_fast(struct btree_trans *,
  26                                                    struct btree_path *, int);
  27
  28 static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *);
  29 static inline void btree_path_list_add(struct btree_trans *, struct btree_path *,
  30                                        struct btree_path *);
  31
  32 static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter)
  33 {
  34 #ifdef CONFIG_BCACHEFS_DEBUG
  35         return iter->ip_allocated;
  36 #else
  37         return 0;
  38 #endif
  39 }
  40
  41 static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *);
  42
  43 /*
  44  * Unlocks before scheduling
  45  * Note: does not revalidate iterator
  46  */
  47 static inline int bch2_trans_cond_resched(struct btree_trans *trans)
  48 {
  49         if (need_resched() || race_fault()) {
  50                 bch2_trans_unlock(trans);
  51                 schedule();
  52                 return bch2_trans_relock(trans);
  53         } else {
  54                 return 0;
  55         }
  56 }
  57
  58 static inline int __btree_path_cmp(const struct btree_path *l,
  59                                    enum btree_id        r_btree_id,
  60                                    bool                 r_cached,
  61                                    struct bpos          r_pos,
  62                                    unsigned             r_level)
  63 {
  64         /*
  65          * Must match lock ordering as defined by __bch2_btree_node_lock:
  66          */
  67         return   cmp_int(l->btree_id,   r_btree_id) ?:
  68                  cmp_int((int) l->cached,       (int) r_cached) ?:
  69                  bpos_cmp(l->pos,       r_pos) ?:
  70                 -cmp_int(l->level,      r_level);
  71 }
  72
  73 static inline int btree_path_cmp(const struct btree_path *l,
  74                                  const struct btree_path *r)
  75 {
  76         return __btree_path_cmp(l, r->btree_id, r->cached, r->pos, r->level);
  77 }
  78
  79 static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p)
  80 {
  81         /* Are we iterating over keys in all snapshots? */
  82         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  83                 p = bpos_successor(p);
  84         } else {
  85                 p = bpos_nosnap_successor(p);
  86                 p.snapshot = iter->snapshot;
  87         }
  88
  89         return p;
  90 }
  91
  92 static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p)
  93 {
  94         /* Are we iterating over keys in all snapshots? */
  95         if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) {
  96                 p = bpos_predecessor(p);
  97         } else {
  98                 p = bpos_nosnap_predecessor(p);
  99                 p.snapshot = iter->snapshot;
 100         }
 101
 102         return p;
 103 }
 104
 105 static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
 106 {
 107         struct bpos pos = iter->pos;
 108
 109         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 110             bkey_cmp(pos, POS_MAX))
 111                 pos = bkey_successor(iter, pos);
 112         return pos;
 113 }
 114
 115 static inline bool btree_path_pos_before_node(struct btree_path *path,
 116                                               struct btree *b)
 117 {
 118         return bpos_cmp(path->pos, b->data->min_key) < 0;
 119 }
 120
 121 static inline bool btree_path_pos_after_node(struct btree_path *path,
 122                                              struct btree *b)
 123 {
 124         return bpos_cmp(b->key.k.p, path->pos) < 0;
 125 }
 126
 127 static inline bool btree_path_pos_in_node(struct btree_path *path,
 128                                           struct btree *b)
 129 {
 130         return path->btree_id == b->c.btree_id &&
 131                 !btree_path_pos_before_node(path, b) &&
 132                 !btree_path_pos_after_node(path, b);
 133 }
 134
 135 /* Btree iterator: */
 136
 137 #ifdef CONFIG_BCACHEFS_DEBUG
 138
 139 static void bch2_btree_path_verify_cached(struct btree_trans *trans,
 140                                           struct btree_path *path)
 141 {
 142         struct bkey_cached *ck;
 143         bool locked = btree_node_locked(path, 0);
 144
 145         if (!bch2_btree_node_relock(trans, path, 0))
 146                 return;
 147
 148         ck = (void *) path->l[0].b;
 149         BUG_ON(ck->key.btree_id != path->btree_id ||
 150                bkey_cmp(ck->key.pos, path->pos));
 151
 152         if (!locked)
 153                 btree_node_unlock(trans, path, 0);
 154 }
 155
 156 static void bch2_btree_path_verify_level(struct btree_trans *trans,
 157                                 struct btree_path *path, unsigned level)
 158 {
 159         struct btree_path_level *l;
 160         struct btree_node_iter tmp;
 161         bool locked;
 162         struct bkey_packed *p, *k;
 163         struct printbuf buf1 = PRINTBUF;
 164         struct printbuf buf2 = PRINTBUF;
 165         struct printbuf buf3 = PRINTBUF;
 166         const char *msg;
 167
 168         if (!bch2_debug_check_iterators)
 169                 return;
 170
 171         l       = &path->l[level];
 172         tmp     = l->iter;
 173         locked  = btree_node_locked(path, level);
 174
 175         if (path->cached) {
 176                 if (!level)
 177                         bch2_btree_path_verify_cached(trans, path);
 178                 return;
 179         }
 180
 181         if (!btree_path_node(path, level))
 182                 return;
 183
 184         if (!bch2_btree_node_relock_notrace(trans, path, level))
 185                 return;
 186
 187         BUG_ON(!btree_path_pos_in_node(path, l->b));
 188
 189         bch2_btree_node_iter_verify(&l->iter, l->b);
 190
 191         /*
 192          * For interior nodes, the iterator will have skipped past deleted keys:
 193          */
 194         p = level
 195                 ? bch2_btree_node_iter_prev(&tmp, l->b)
 196                 : bch2_btree_node_iter_prev_all(&tmp, l->b);
 197         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 198
 199         if (p && bkey_iter_pos_cmp(l->b, p, &path->pos) >= 0) {
 200                 msg = "before";
 201                 goto err;
 202         }
 203
 204         if (k && bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 205                 msg = "after";
 206                 goto err;
 207         }
 208
 209         if (!locked)
 210                 btree_node_unlock(trans, path, level);
 211         return;
 212 err:
 213         bch2_bpos_to_text(&buf1, path->pos);
 214
 215         if (p) {
 216                 struct bkey uk = bkey_unpack_key(l->b, p);
 217
 218                 bch2_bkey_to_text(&buf2, &uk);
 219         } else {
 220                 prt_printf(&buf2, "(none)");
 221         }
 222
 223         if (k) {
 224                 struct bkey uk = bkey_unpack_key(l->b, k);
 225
 226                 bch2_bkey_to_text(&buf3, &uk);
 227         } else {
 228                 prt_printf(&buf3, "(none)");
 229         }
 230
 231         panic("path should be %s key at level %u:\n"
 232               "path pos %s\n"
 233               "prev key %s\n"
 234               "cur  key %s\n",
 235               msg, level, buf1.buf, buf2.buf, buf3.buf);
 236 }
 237
 238 static void bch2_btree_path_verify(struct btree_trans *trans,
 239                                    struct btree_path *path)
 240 {
 241         struct bch_fs *c = trans->c;
 242         unsigned i;
 243
 244         EBUG_ON(path->btree_id >= BTREE_ID_NR);
 245
 246         for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) {
 247                 if (!path->l[i].b) {
 248                         BUG_ON(!path->cached &&
 249                                c->btree_roots[path->btree_id].b->c.level > i);
 250                         break;
 251                 }
 252
 253                 bch2_btree_path_verify_level(trans, path, i);
 254         }
 255
 256         bch2_btree_path_verify_locks(path);
 257 }
 258
 259 void bch2_trans_verify_paths(struct btree_trans *trans)
 260 {
 261         struct btree_path *path;
 262
 263         trans_for_each_path(trans, path)
 264                 bch2_btree_path_verify(trans, path);
 265 }
 266
 267 static void bch2_btree_iter_verify(struct btree_iter *iter)
 268 {
 269         struct btree_trans *trans = iter->trans;
 270
 271         BUG_ON(iter->btree_id >= BTREE_ID_NR);
 272
 273         BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached);
 274
 275         BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 276                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
 277
 278         BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 279                (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 280                !btree_type_has_snapshots(iter->btree_id));
 281
 282         if (iter->update_path)
 283                 bch2_btree_path_verify(trans, iter->update_path);
 284         bch2_btree_path_verify(trans, iter->path);
 285 }
 286
 287 static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 288 {
 289         BUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 290                !iter->pos.snapshot);
 291
 292         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 293                iter->pos.snapshot != iter->snapshot);
 294
 295         BUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
 296                bkey_cmp(iter->pos, iter->k.p) > 0);
 297 }
 298
 299 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
 300 {
 301         struct btree_trans *trans = iter->trans;
 302         struct btree_iter copy;
 303         struct bkey_s_c prev;
 304         int ret = 0;
 305
 306         if (!bch2_debug_check_iterators)
 307                 return 0;
 308
 309         if (!(iter->flags & BTREE_ITER_FILTER_SNAPSHOTS))
 310                 return 0;
 311
 312         if (bkey_err(k) || !k.k)
 313                 return 0;
 314
 315         BUG_ON(!bch2_snapshot_is_ancestor(trans->c,
 316                                           iter->snapshot,
 317                                           k.k->p.snapshot));
 318
 319         bch2_trans_iter_init(trans, &copy, iter->btree_id, iter->pos,
 320                              BTREE_ITER_NOPRESERVE|
 321                              BTREE_ITER_ALL_SNAPSHOTS);
 322         prev = bch2_btree_iter_prev(&copy);
 323         if (!prev.k)
 324                 goto out;
 325
 326         ret = bkey_err(prev);
 327         if (ret)
 328                 goto out;
 329
 330         if (!bkey_cmp(prev.k->p, k.k->p) &&
 331             bch2_snapshot_is_ancestor(trans->c, iter->snapshot,
 332                                       prev.k->p.snapshot) > 0) {
 333                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 334
 335                 bch2_bkey_to_text(&buf1, k.k);
 336                 bch2_bkey_to_text(&buf2, prev.k);
 337
 338                 panic("iter snap %u\n"
 339                       "k    %s\n"
 340                       "prev %s\n",
 341                       iter->snapshot,
 342                       buf1.buf, buf2.buf);
 343         }
 344 out:
 345         bch2_trans_iter_exit(trans, &copy);
 346         return ret;
 347 }
 348
 349 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 350                             struct bpos pos, bool key_cache)
 351 {
 352         struct btree_path *path;
 353         unsigned idx;
 354         struct printbuf buf = PRINTBUF;
 355
 356         trans_for_each_path_inorder(trans, path, idx) {
 357                 int cmp = cmp_int(path->btree_id, id) ?:
 358                         cmp_int(path->cached, key_cache);
 359
 360                 if (cmp > 0)
 361                         break;
 362                 if (cmp < 0)
 363                         continue;
 364
 365                 if (!btree_node_locked(path, 0) ||
 366                     !path->should_be_locked)
 367                         continue;
 368
 369                 if (!key_cache) {
 370                         if (bkey_cmp(pos, path->l[0].b->data->min_key) >= 0 &&
 371                             bkey_cmp(pos, path->l[0].b->key.k.p) <= 0)
 372                                 return;
 373                 } else {
 374                         if (!bkey_cmp(pos, path->pos))
 375                                 return;
 376                 }
 377         }
 378
 379         bch2_dump_trans_paths_updates(trans);
 380         bch2_bpos_to_text(&buf, pos);
 381
 382         panic("not locked: %s %s%s\n",
 383               bch2_btree_ids[id], buf.buf,
 384               key_cache ? " cached" : "");
 385 }
 386
 387 #else
 388
 389 static inline void bch2_btree_path_verify_level(struct btree_trans *trans,
 390                                                 struct btree_path *path, unsigned l) {}
 391 static inline void bch2_btree_path_verify(struct btree_trans *trans,
 392                                           struct btree_path *path) {}
 393 static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 394 static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {}
 395 static inline int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) { return 0; }
 396
 397 #endif
 398
 399 /* Btree path: fixups after btree updates */
 400
 401 static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
 402                                         struct btree *b,
 403                                         struct bset_tree *t,
 404                                         struct bkey_packed *k)
 405 {
 406         struct btree_node_iter_set *set;
 407
 408         btree_node_iter_for_each(iter, set)
 409                 if (set->end == t->end_offset) {
 410                         set->k = __btree_node_key_to_offset(b, k);
 411                         bch2_btree_node_iter_sort(iter, b);
 412                         return;
 413                 }
 414
 415         bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
 416 }
 417
 418 static void __bch2_btree_path_fix_key_modified(struct btree_path *path,
 419                                                struct btree *b,
 420                                                struct bkey_packed *where)
 421 {
 422         struct btree_path_level *l = &path->l[b->c.level];
 423
 424         if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
 425                 return;
 426
 427         if (bkey_iter_pos_cmp(l->b, where, &path->pos) < 0)
 428                 bch2_btree_node_iter_advance(&l->iter, l->b);
 429 }
 430
 431 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 432                                       struct btree *b,
 433                                       struct bkey_packed *where)
 434 {
 435         struct btree_path *path;
 436
 437         trans_for_each_path_with_node(trans, b, path) {
 438                 __bch2_btree_path_fix_key_modified(path, b, where);
 439                 bch2_btree_path_verify_level(trans, path, b->c.level);
 440         }
 441 }
 442
 443 static void __bch2_btree_node_iter_fix(struct btree_path *path,
 444                                        struct btree *b,
 445                                        struct btree_node_iter *node_iter,
 446                                        struct bset_tree *t,
 447                                        struct bkey_packed *where,
 448                                        unsigned clobber_u64s,
 449                                        unsigned new_u64s)
 450 {
 451         const struct bkey_packed *end = btree_bkey_last(b, t);
 452         struct btree_node_iter_set *set;
 453         unsigned offset = __btree_node_key_to_offset(b, where);
 454         int shift = new_u64s - clobber_u64s;
 455         unsigned old_end = t->end_offset - shift;
 456         unsigned orig_iter_pos = node_iter->data[0].k;
 457         bool iter_current_key_modified =
 458                 orig_iter_pos >= offset &&
 459                 orig_iter_pos <= offset + clobber_u64s;
 460
 461         btree_node_iter_for_each(node_iter, set)
 462                 if (set->end == old_end)
 463                         goto found;
 464
 465         /* didn't find the bset in the iterator - might have to readd it: */
 466         if (new_u64s &&
 467             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 468                 bch2_btree_node_iter_push(node_iter, b, where, end);
 469                 goto fixup_done;
 470         } else {
 471                 /* Iterator is after key that changed */
 472                 return;
 473         }
 474 found:
 475         set->end = t->end_offset;
 476
 477         /* Iterator hasn't gotten to the key that changed yet: */
 478         if (set->k < offset)
 479                 return;
 480
 481         if (new_u64s &&
 482             bkey_iter_pos_cmp(b, where, &path->pos) >= 0) {
 483                 set->k = offset;
 484         } else if (set->k < offset + clobber_u64s) {
 485                 set->k = offset + new_u64s;
 486                 if (set->k == set->end)
 487                         bch2_btree_node_iter_set_drop(node_iter, set);
 488         } else {
 489                 /* Iterator is after key that changed */
 490                 set->k = (int) set->k + shift;
 491                 return;
 492         }
 493
 494         bch2_btree_node_iter_sort(node_iter, b);
 495 fixup_done:
 496         if (node_iter->data[0].k != orig_iter_pos)
 497                 iter_current_key_modified = true;
 498
 499         /*
 500          * When a new key is added, and the node iterator now points to that
 501          * key, the iterator might have skipped past deleted keys that should
 502          * come after the key the iterator now points to. We have to rewind to
 503          * before those deleted keys - otherwise
 504          * bch2_btree_node_iter_prev_all() breaks:
 505          */
 506         if (!bch2_btree_node_iter_end(node_iter) &&
 507             iter_current_key_modified &&
 508             b->c.level) {
 509                 struct bset_tree *t;
 510                 struct bkey_packed *k, *k2, *p;
 511
 512                 k = bch2_btree_node_iter_peek_all(node_iter, b);
 513
 514                 for_each_bset(b, t) {
 515                         bool set_pos = false;
 516
 517                         if (node_iter->data[0].end == t->end_offset)
 518                                 continue;
 519
 520                         k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
 521
 522                         while ((p = bch2_bkey_prev_all(b, t, k2)) &&
 523                                bkey_iter_cmp(b, k, p) < 0) {
 524                                 k2 = p;
 525                                 set_pos = true;
 526                         }
 527
 528                         if (set_pos)
 529                                 btree_node_iter_set_set_pos(node_iter,
 530                                                             b, t, k2);
 531                 }
 532         }
 533 }
 534
 535 void bch2_btree_node_iter_fix(struct btree_trans *trans,
 536                               struct btree_path *path,
 537                               struct btree *b,
 538                               struct btree_node_iter *node_iter,
 539                               struct bkey_packed *where,
 540                               unsigned clobber_u64s,
 541                               unsigned new_u64s)
 542 {
 543         struct bset_tree *t = bch2_bkey_to_bset(b, where);
 544         struct btree_path *linked;
 545
 546         if (node_iter != &path->l[b->c.level].iter) {
 547                 __bch2_btree_node_iter_fix(path, b, node_iter, t,
 548                                            where, clobber_u64s, new_u64s);
 549
 550                 if (bch2_debug_check_iterators)
 551                         bch2_btree_node_iter_verify(node_iter, b);
 552         }
 553
 554         trans_for_each_path_with_node(trans, b, linked) {
 555                 __bch2_btree_node_iter_fix(linked, b,
 556                                            &linked->l[b->c.level].iter, t,
 557                                            where, clobber_u64s, new_u64s);
 558                 bch2_btree_path_verify_level(trans, linked, b->c.level);
 559         }
 560 }
 561
 562 /* Btree path level: pointer to a particular btree node and node iter */
 563
 564 static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c,
 565                                                   struct btree_path_level *l,
 566                                                   struct bkey *u,
 567                                                   struct bkey_packed *k)
 568 {
 569         if (unlikely(!k)) {
 570                 /*
 571                  * signal to bch2_btree_iter_peek_slot() that we're currently at
 572                  * a hole
 573                  */
 574                 u->type = KEY_TYPE_deleted;
 575                 return bkey_s_c_null;
 576         }
 577
 578         return bkey_disassemble(l->b, k, u);
 579 }
 580
 581 static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c,
 582                                                         struct btree_path_level *l,
 583                                                         struct bkey *u)
 584 {
 585         return __btree_iter_unpack(c, l, u,
 586                         bch2_btree_node_iter_peek_all(&l->iter, l->b));
 587 }
 588
 589 static inline struct bkey_s_c btree_path_level_peek(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_peek(&l->iter, l->b));
 596
 597         path->pos = k.k ? k.k->p : l->b->key.k.p;
 598         bch2_btree_path_verify_level(trans, path, l - path->l);
 599         return k;
 600 }
 601
 602 static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans,
 603                                                     struct btree_path *path,
 604                                                     struct btree_path_level *l,
 605                                                     struct bkey *u)
 606 {
 607         struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u,
 608                         bch2_btree_node_iter_prev(&l->iter, l->b));
 609
 610         path->pos = k.k ? k.k->p : l->b->data->min_key;
 611         bch2_btree_path_verify_level(trans, path, l - path->l);
 612         return k;
 613 }
 614
 615 static inline bool btree_path_advance_to_pos(struct btree_path *path,
 616                                              struct btree_path_level *l,
 617                                              int max_advance)
 618 {
 619         struct bkey_packed *k;
 620         int nr_advanced = 0;
 621
 622         while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
 623                bkey_iter_pos_cmp(l->b, k, &path->pos) < 0) {
 624                 if (max_advance > 0 && nr_advanced >= max_advance)
 625                         return false;
 626
 627                 bch2_btree_node_iter_advance(&l->iter, l->b);
 628                 nr_advanced++;
 629         }
 630
 631         return true;
 632 }
 633
 634 static inline void __btree_path_level_init(struct btree_path *path,
 635                                            unsigned level)
 636 {
 637         struct btree_path_level *l = &path->l[level];
 638
 639         bch2_btree_node_iter_init(&l->iter, l->b, &path->pos);
 640
 641         /*
 642          * Iterators to interior nodes should always be pointed at the first non
 643          * whiteout:
 644          */
 645         if (level)
 646                 bch2_btree_node_iter_peek(&l->iter, l->b);
 647 }
 648
 649 inline void bch2_btree_path_level_init(struct btree_trans *trans,
 650                                        struct btree_path *path,
 651                                        struct btree *b)
 652 {
 653         BUG_ON(path->cached);
 654
 655         EBUG_ON(!btree_path_pos_in_node(path, b));
 656         EBUG_ON(b->c.lock.state.seq & 1);
 657
 658         path->l[b->c.level].lock_seq = b->c.lock.state.seq;
 659         path->l[b->c.level].b = b;
 660         __btree_path_level_init(path, b->c.level);
 661 }
 662
 663 /* Btree path: fixups after btree node updates: */
 664
 665 /*
 666  * A btree node is being replaced - update the iterator to point to the new
 667  * node:
 668  */
 669 void bch2_trans_node_add(struct btree_trans *trans, struct btree *b)
 670 {
 671         struct btree_path *path;
 672
 673         trans_for_each_path(trans, path)
 674                 if (path->uptodate == BTREE_ITER_UPTODATE &&
 675                     !path->cached &&
 676                     btree_path_pos_in_node(path, b)) {
 677                         enum btree_node_locked_type t =
 678                                 btree_lock_want(path, b->c.level);
 679
 680                         if (t != BTREE_NODE_UNLOCKED) {
 681                                 btree_node_unlock(trans, path, b->c.level);
 682                                 six_lock_increment(&b->c.lock, t);
 683                                 mark_btree_node_locked(trans, path, b->c.level, t);
 684                         }
 685
 686                         bch2_btree_path_level_init(trans, path, b);
 687                 }
 688 }
 689
 690 /*
 691  * A btree node has been modified in such a way as to invalidate iterators - fix
 692  * them:
 693  */
 694 void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b)
 695 {
 696         struct btree_path *path;
 697
 698         trans_for_each_path_with_node(trans, b, path)
 699                 __btree_path_level_init(path, b->c.level);
 700 }
 701
 702 /* Btree path: traverse, set_pos: */
 703
 704 static inline int btree_path_lock_root(struct btree_trans *trans,
 705                                        struct btree_path *path,
 706                                        unsigned depth_want,
 707                                        unsigned long trace_ip)
 708 {
 709         struct bch_fs *c = trans->c;
 710         struct btree *b, **rootp = &c->btree_roots[path->btree_id].b;
 711         enum six_lock_type lock_type;
 712         unsigned i;
 713         int ret;
 714
 715         EBUG_ON(path->nodes_locked);
 716
 717         while (1) {
 718                 b = READ_ONCE(*rootp);
 719                 path->level = READ_ONCE(b->c.level);
 720
 721                 if (unlikely(path->level < depth_want)) {
 722                         /*
 723                          * the root is at a lower depth than the depth we want:
 724                          * got to the end of the btree, or we're walking nodes
 725                          * greater than some depth and there are no nodes >=
 726                          * that depth
 727                          */
 728                         path->level = depth_want;
 729                         for (i = path->level; i < BTREE_MAX_DEPTH; i++)
 730                                 path->l[i].b = NULL;
 731                         return 1;
 732                 }
 733
 734                 lock_type = __btree_lock_want(path, path->level);
 735                 ret = btree_node_lock(trans, path, &b->c,
 736                                       path->level, lock_type, trace_ip);
 737                 if (unlikely(ret)) {
 738                         if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
 739                                 continue;
 740                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 741                                 return ret;
 742                         BUG();
 743                 }
 744
 745                 if (likely(b == READ_ONCE(*rootp) &&
 746                            b->c.level == path->level &&
 747                            !race_fault())) {
 748                         for (i = 0; i < path->level; i++)
 749                                 path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root);
 750                         path->l[path->level].b = b;
 751                         for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++)
 752                                 path->l[i].b = NULL;
 753
 754                         mark_btree_node_locked(trans, path, path->level, lock_type);
 755                         bch2_btree_path_level_init(trans, path, b);
 756                         return 0;
 757                 }
 758
 759                 six_unlock_type(&b->c.lock, lock_type);
 760         }
 761 }
 762
 763 noinline
 764 static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *path)
 765 {
 766         struct bch_fs *c = trans->c;
 767         struct btree_path_level *l = path_l(path);
 768         struct btree_node_iter node_iter = l->iter;
 769         struct bkey_packed *k;
 770         struct bkey_buf tmp;
 771         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 772                 ? (path->level > 1 ? 0 :  2)
 773                 : (path->level > 1 ? 1 : 16);
 774         bool was_locked = btree_node_locked(path, path->level);
 775         int ret = 0;
 776
 777         bch2_bkey_buf_init(&tmp);
 778
 779         while (nr-- && !ret) {
 780                 if (!bch2_btree_node_relock(trans, path, path->level))
 781                         break;
 782
 783                 bch2_btree_node_iter_advance(&node_iter, l->b);
 784                 k = bch2_btree_node_iter_peek(&node_iter, l->b);
 785                 if (!k)
 786                         break;
 787
 788                 bch2_bkey_buf_unpack(&tmp, c, l->b, k);
 789                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
 790                                                path->level - 1);
 791         }
 792
 793         if (!was_locked)
 794                 btree_node_unlock(trans, path, path->level);
 795
 796         bch2_bkey_buf_exit(&tmp, c);
 797         return ret;
 798 }
 799
 800 static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path,
 801                                  struct btree_and_journal_iter *jiter)
 802 {
 803         struct bch_fs *c = trans->c;
 804         struct bkey_s_c k;
 805         struct bkey_buf tmp;
 806         unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
 807                 ? (path->level > 1 ? 0 :  2)
 808                 : (path->level > 1 ? 1 : 16);
 809         bool was_locked = btree_node_locked(path, path->level);
 810         int ret = 0;
 811
 812         bch2_bkey_buf_init(&tmp);
 813
 814         while (nr-- && !ret) {
 815                 if (!bch2_btree_node_relock(trans, path, path->level))
 816                         break;
 817
 818                 bch2_btree_and_journal_iter_advance(jiter);
 819                 k = bch2_btree_and_journal_iter_peek(jiter);
 820                 if (!k.k)
 821                         break;
 822
 823                 bch2_bkey_buf_reassemble(&tmp, c, k);
 824                 ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id,
 825                                                path->level - 1);
 826         }
 827
 828         if (!was_locked)
 829                 btree_node_unlock(trans, path, path->level);
 830
 831         bch2_bkey_buf_exit(&tmp, c);
 832         return ret;
 833 }
 834
 835 static noinline void btree_node_mem_ptr_set(struct btree_trans *trans,
 836                                             struct btree_path *path,
 837                                             unsigned plevel, struct btree *b)
 838 {
 839         struct btree_path_level *l = &path->l[plevel];
 840         bool locked = btree_node_locked(path, plevel);
 841         struct bkey_packed *k;
 842         struct bch_btree_ptr_v2 *bp;
 843
 844         if (!bch2_btree_node_relock(trans, path, plevel))
 845                 return;
 846
 847         k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 848         BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
 849
 850         bp = (void *) bkeyp_val(&l->b->format, k);
 851         bp->mem_ptr = (unsigned long)b;
 852
 853         if (!locked)
 854                 btree_node_unlock(trans, path, plevel);
 855 }
 856
 857 static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans,
 858                                                      struct btree_path *path,
 859                                                      unsigned flags,
 860                                                      struct bkey_buf *out)
 861 {
 862         struct bch_fs *c = trans->c;
 863         struct btree_path_level *l = path_l(path);
 864         struct btree_and_journal_iter jiter;
 865         struct bkey_s_c k;
 866         int ret = 0;
 867
 868         __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos);
 869
 870         k = bch2_btree_and_journal_iter_peek(&jiter);
 871
 872         bch2_bkey_buf_reassemble(out, c, k);
 873
 874         if (flags & BTREE_ITER_PREFETCH)
 875                 ret = btree_path_prefetch_j(trans, path, &jiter);
 876
 877         bch2_btree_and_journal_iter_exit(&jiter);
 878         return ret;
 879 }
 880
 881 static __always_inline int btree_path_down(struct btree_trans *trans,
 882                                            struct btree_path *path,
 883                                            unsigned flags,
 884                                            unsigned long trace_ip)
 885 {
 886         struct bch_fs *c = trans->c;
 887         struct btree_path_level *l = path_l(path);
 888         struct btree *b;
 889         unsigned level = path->level - 1;
 890         enum six_lock_type lock_type = __btree_lock_want(path, level);
 891         bool replay_done = test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
 892         struct bkey_buf tmp;
 893         int ret;
 894
 895         EBUG_ON(!btree_node_locked(path, path->level));
 896
 897         bch2_bkey_buf_init(&tmp);
 898
 899         if (unlikely(!replay_done)) {
 900                 ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp);
 901                 if (ret)
 902                         goto err;
 903         } else {
 904                 bch2_bkey_buf_unpack(&tmp, c, l->b,
 905                                  bch2_btree_node_iter_peek(&l->iter, l->b));
 906
 907                 if (flags & BTREE_ITER_PREFETCH) {
 908                         ret = btree_path_prefetch(trans, path);
 909                         if (ret)
 910                                 goto err;
 911                 }
 912         }
 913
 914         b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip);
 915         ret = PTR_ERR_OR_ZERO(b);
 916         if (unlikely(ret))
 917                 goto err;
 918
 919         if (likely(replay_done && tmp.k->k.type == KEY_TYPE_btree_ptr_v2) &&
 920             unlikely(b != btree_node_mem_ptr(tmp.k)))
 921                 btree_node_mem_ptr_set(trans, path, level + 1, b);
 922
 923         if (btree_node_read_locked(path, level + 1))
 924                 btree_node_unlock(trans, path, level + 1);
 925
 926         mark_btree_node_locked(trans, path, level, lock_type);
 927         path->level = level;
 928         bch2_btree_path_level_init(trans, path, b);
 929
 930         bch2_btree_path_verify_locks(path);
 931 err:
 932         bch2_bkey_buf_exit(&tmp, c);
 933         return ret;
 934 }
 935
 936 static int btree_path_traverse_one(struct btree_trans *, struct btree_path *,
 937                                    unsigned, unsigned long);
 938
 939 static int bch2_btree_path_traverse_all(struct btree_trans *trans)
 940 {
 941         struct bch_fs *c = trans->c;
 942         struct btree_path *path;
 943         unsigned long trace_ip = _RET_IP_;
 944         int ret = 0;
 945
 946         if (trans->in_traverse_all)
 947                 return -BCH_ERR_transaction_restart_in_traverse_all;
 948
 949         trans->in_traverse_all = true;
 950 retry_all:
 951         trans->restarted = 0;
 952         trans->traverse_all_idx = U8_MAX;
 953
 954         trans_for_each_path(trans, path)
 955                 path->should_be_locked = false;
 956
 957         btree_trans_verify_sorted(trans);
 958
 959         bch2_trans_unlock(trans);
 960         cond_resched();
 961
 962         if (unlikely(trans->memory_allocation_failure)) {
 963                 struct closure cl;
 964
 965                 closure_init_stack(&cl);
 966
 967                 do {
 968                         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 969                         closure_sync(&cl);
 970                 } while (ret);
 971         }
 972
 973         /* Now, redo traversals in correct order: */
 974         trans->traverse_all_idx = 0;
 975         while (trans->traverse_all_idx < trans->nr_sorted) {
 976                 path = trans->paths + trans->sorted[trans->traverse_all_idx];
 977
 978                 /*
 979                  * Traversing a path can cause another path to be added at about
 980                  * the same position:
 981                  */
 982                 if (path->uptodate) {
 983                         ret = btree_path_traverse_one(trans, path, 0, _THIS_IP_);
 984                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
 985                             ret == -ENOMEM)
 986                                 goto retry_all;
 987                         if (ret)
 988                                 goto err;
 989                         BUG_ON(path->uptodate);
 990                 } else {
 991                         trans->traverse_all_idx++;
 992                 }
 993         }
 994
 995         /*
 996          * BTREE_ITER_NEED_RELOCK is ok here - if we called bch2_trans_unlock()
 997          * and relock(), relock() won't relock since path->should_be_locked
 998          * isn't set yet, which is all fine
 999          */
1000         trans_for_each_path(trans, path)
1001                 BUG_ON(path->uptodate >= BTREE_ITER_NEED_TRAVERSE);
1002 err:
1003         bch2_btree_cache_cannibalize_unlock(c);
1004
1005         trans->in_traverse_all = false;
1006
1007         trace_and_count(c, trans_traverse_all, trans, trace_ip);
1008         return ret;
1009 }
1010
1011 static inline bool btree_path_check_pos_in_node(struct btree_path *path,
1012                                                 unsigned l, int check_pos)
1013 {
1014         if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b))
1015                 return false;
1016         if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b))
1017                 return false;
1018         return true;
1019 }
1020
1021 static inline bool btree_path_good_node(struct btree_trans *trans,
1022                                         struct btree_path *path,
1023                                         unsigned l, int check_pos)
1024 {
1025         return is_btree_node(path, l) &&
1026                 bch2_btree_node_relock(trans, path, l) &&
1027                 btree_path_check_pos_in_node(path, l, check_pos);
1028 }
1029
1030 static void btree_path_set_level_down(struct btree_trans *trans,
1031                                       struct btree_path *path,
1032                                       unsigned new_level)
1033 {
1034         unsigned l;
1035
1036         path->level = new_level;
1037
1038         for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++)
1039                 if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED)
1040                         btree_node_unlock(trans, path, l);
1041
1042         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1043         bch2_btree_path_verify(trans, path);
1044 }
1045
1046 static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans,
1047                                                          struct btree_path *path,
1048                                                          int check_pos)
1049 {
1050         unsigned i, l = path->level;
1051 again:
1052         while (btree_path_node(path, l) &&
1053                !btree_path_good_node(trans, path, l, check_pos))
1054                 __btree_path_set_level_up(trans, path, l++);
1055
1056         /* If we need intent locks, take them too: */
1057         for (i = l + 1;
1058              i < path->locks_want && btree_path_node(path, i);
1059              i++)
1060                 if (!bch2_btree_node_relock(trans, path, i)) {
1061                         while (l <= i)
1062                                 __btree_path_set_level_up(trans, path, l++);
1063                         goto again;
1064                 }
1065
1066         return l;
1067 }
1068
1069 static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans,
1070                                                      struct btree_path *path,
1071                                                      int check_pos)
1072 {
1073         return likely(btree_node_locked(path, path->level) &&
1074                       btree_path_check_pos_in_node(path, path->level, check_pos))
1075                 ? path->level
1076                 : __btree_path_up_until_good_node(trans, path, check_pos);
1077 }
1078
1079 /*
1080  * This is the main state machine for walking down the btree - walks down to a
1081  * specified depth
1082  *
1083  * Returns 0 on success, -EIO on error (error reading in a btree node).
1084  *
1085  * On error, caller (peek_node()/peek_key()) must return NULL; the error is
1086  * stashed in the iterator and returned from bch2_trans_exit().
1087  */
1088 static int btree_path_traverse_one(struct btree_trans *trans,
1089                                    struct btree_path *path,
1090                                    unsigned flags,
1091                                    unsigned long trace_ip)
1092 {
1093         unsigned depth_want = path->level;
1094         int ret = trans->restarted;
1095
1096         if (unlikely(ret))
1097                 goto out;
1098
1099         /*
1100          * Ensure we obey path->should_be_locked: if it's set, we can't unlock
1101          * and re-traverse the path without a transaction restart:
1102          */
1103         if (path->should_be_locked) {
1104                 ret = bch2_btree_path_relock(trans, path, trace_ip);
1105                 goto out;
1106         }
1107
1108         if (path->cached) {
1109                 ret = bch2_btree_path_traverse_cached(trans, path, flags);
1110                 goto out;
1111         }
1112
1113         if (unlikely(path->level >= BTREE_MAX_DEPTH))
1114                 goto out;
1115
1116         path->level = btree_path_up_until_good_node(trans, path, 0);
1117
1118         EBUG_ON(btree_path_node(path, path->level) &&
1119                 !btree_node_locked(path, path->level));
1120
1121         /*
1122          * Note: path->nodes[path->level] may be temporarily NULL here - that
1123          * would indicate to other code that we got to the end of the btree,
1124          * here it indicates that relocking the root failed - it's critical that
1125          * btree_path_lock_root() comes next and that it can't fail
1126          */
1127         while (path->level > depth_want) {
1128                 ret = btree_path_node(path, path->level)
1129                         ? btree_path_down(trans, path, flags, trace_ip)
1130                         : btree_path_lock_root(trans, path, depth_want, trace_ip);
1131                 if (unlikely(ret)) {
1132                         if (ret == 1) {
1133                                 /*
1134                                  * No nodes at this level - got to the end of
1135                                  * the btree:
1136                                  */
1137                                 ret = 0;
1138                                 goto out;
1139                         }
1140
1141                         __bch2_btree_path_unlock(trans, path);
1142                         path->level = depth_want;
1143                         path->l[path->level].b = ERR_PTR(ret);
1144                         goto out;
1145                 }
1146         }
1147
1148         path->uptodate = BTREE_ITER_UPTODATE;
1149 out:
1150         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1151         bch2_btree_path_verify(trans, path);
1152         return ret;
1153 }
1154
1155 int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
1156                                           struct btree_path *path, unsigned flags)
1157 {
1158         if (0 && IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
1159                 unsigned restart_probability_bits = 4 << min(trans->restart_count, 32U);
1160                 u64 mask = ~(~0ULL << restart_probability_bits);
1161
1162                 if ((prandom_u32() & mask) == mask) {
1163                         trace_and_count(trans->c, trans_restart_injected, trans, _RET_IP_);
1164                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_fault_inject);
1165                 }
1166         }
1167
1168         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
1169                 return 0;
1170
1171         return  bch2_trans_cond_resched(trans) ?:
1172                 btree_path_traverse_one(trans, path, flags, _RET_IP_);
1173 }
1174
1175 static void btree_path_copy(struct btree_trans *trans, struct btree_path *dst,
1176                             struct btree_path *src)
1177 {
1178         unsigned i, offset = offsetof(struct btree_path, pos);
1179         int cmp = btree_path_cmp(dst, src);
1180
1181         memcpy((void *) dst + offset,
1182                (void *) src + offset,
1183                sizeof(struct btree_path) - offset);
1184
1185         for (i = 0; i < BTREE_MAX_DEPTH; i++) {
1186                 unsigned t = btree_node_locked_type(dst, i);
1187
1188                 if (t != BTREE_NODE_UNLOCKED)
1189                         six_lock_increment(&dst->l[i].b->c.lock, t);
1190         }
1191
1192         if (cmp)
1193                 bch2_btree_path_check_sort_fast(trans, dst, cmp);
1194 }
1195
1196 static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src,
1197                                            bool intent)
1198 {
1199         struct btree_path *new = btree_path_alloc(trans, src);
1200
1201         btree_path_copy(trans, new, src);
1202         __btree_path_get(new, intent);
1203         return new;
1204 }
1205
1206 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *trans,
1207                          struct btree_path *path, bool intent,
1208                          unsigned long ip)
1209 {
1210         if (path->ref > 1 || path->preserve) {
1211                 __btree_path_put(path, intent);
1212                 path = btree_path_clone(trans, path, intent);
1213                 path->preserve = false;
1214 #ifdef CONFIG_BCACHEFS_DEBUG
1215                 path->ip_allocated = ip;
1216 #endif
1217                 btree_trans_verify_sorted(trans);
1218         }
1219
1220         path->should_be_locked = false;
1221         return path;
1222 }
1223
1224 struct btree_path * __must_check
1225 bch2_btree_path_set_pos(struct btree_trans *trans,
1226                    struct btree_path *path, struct bpos new_pos,
1227                    bool intent, unsigned long ip)
1228 {
1229         int cmp = bpos_cmp(new_pos, path->pos);
1230         unsigned l = path->level;
1231
1232         EBUG_ON(trans->restarted);
1233         EBUG_ON(!path->ref);
1234
1235         if (!cmp)
1236                 return path;
1237
1238         path = bch2_btree_path_make_mut(trans, path, intent, ip);
1239
1240         path->pos = new_pos;
1241
1242         bch2_btree_path_check_sort_fast(trans, path, cmp);
1243
1244         if (unlikely(path->cached)) {
1245                 btree_node_unlock(trans, path, 0);
1246                 path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up);
1247                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1248                 goto out;
1249         }
1250
1251         l = btree_path_up_until_good_node(trans, path, cmp);
1252
1253         if (btree_path_node(path, l)) {
1254                 BUG_ON(!btree_node_locked(path, l));
1255                 /*
1256                  * We might have to skip over many keys, or just a few: try
1257                  * advancing the node iterator, and if we have to skip over too
1258                  * many keys just reinit it (or if we're rewinding, since that
1259                  * is expensive).
1260                  */
1261                 if (cmp < 0 ||
1262                     !btree_path_advance_to_pos(path, &path->l[l], 8))
1263                         __btree_path_level_init(path, l);
1264         }
1265
1266         if (unlikely(l != path->level)) {
1267                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1268                 __bch2_btree_path_unlock(trans, path);
1269         }
1270 out:
1271         bch2_btree_path_verify(trans, path);
1272         return path;
1273 }
1274
1275 /* Btree path: main interface: */
1276
1277 static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path)
1278 {
1279         struct btree_path *sib;
1280
1281         sib = prev_btree_path(trans, path);
1282         if (sib && !btree_path_cmp(sib, path))
1283                 return sib;
1284
1285         sib = next_btree_path(trans, path);
1286         if (sib && !btree_path_cmp(sib, path))
1287                 return sib;
1288
1289         return NULL;
1290 }
1291
1292 static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path)
1293 {
1294         struct btree_path *sib;
1295
1296         sib = prev_btree_path(trans, path);
1297         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1298                 return sib;
1299
1300         sib = next_btree_path(trans, path);
1301         if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b)
1302                 return sib;
1303
1304         return NULL;
1305 }
1306
1307 static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path)
1308 {
1309         __bch2_btree_path_unlock(trans, path);
1310         btree_path_list_remove(trans, path);
1311         trans->paths_allocated &= ~(1ULL << path->idx);
1312 }
1313
1314 void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
1315 {
1316         struct btree_path *dup;
1317
1318         EBUG_ON(trans->paths + path->idx != path);
1319         EBUG_ON(!path->ref);
1320
1321         if (!__btree_path_put(path, intent))
1322                 return;
1323
1324         dup = path->preserve
1325                 ? have_path_at_pos(trans, path)
1326                 : have_node_at_pos(trans, path);
1327
1328         if (!dup && !(!path->preserve && !is_btree_node(path, path->level)))
1329                 return;
1330
1331         if (path->should_be_locked &&
1332             !trans->restarted &&
1333             (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_)))
1334                 return;
1335
1336         if (dup) {
1337                 dup->preserve           |= path->preserve;
1338                 dup->should_be_locked   |= path->should_be_locked;
1339         }
1340
1341         __bch2_path_free(trans, path);
1342 }
1343
1344 static void bch2_path_put_nokeep(struct btree_trans *trans, struct btree_path *path,
1345                                  bool intent)
1346 {
1347         EBUG_ON(trans->paths + path->idx != path);
1348         EBUG_ON(!path->ref);
1349
1350         if (!__btree_path_put(path, intent))
1351                 return;
1352
1353         __bch2_path_free(trans, path);
1354 }
1355
1356 void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
1357 {
1358         struct btree_insert_entry *i;
1359
1360         prt_printf(buf, "transaction updates for %s journal seq %llu",
1361                trans->fn, trans->journal_res.seq);
1362         prt_newline(buf);
1363         printbuf_indent_add(buf, 2);
1364
1365         trans_for_each_update(trans, i) {
1366                 struct bkey_s_c old = { &i->old_k, i->old_v };
1367
1368                 prt_printf(buf, "update: btree=%s cached=%u %pS",
1369                        bch2_btree_ids[i->btree_id],
1370                        i->cached,
1371                        (void *) i->ip_allocated);
1372                 prt_newline(buf);
1373
1374                 prt_printf(buf, "  old ");
1375                 bch2_bkey_val_to_text(buf, trans->c, old);
1376                 prt_newline(buf);
1377
1378                 prt_printf(buf, "  new ");
1379                 bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k));
1380                 prt_newline(buf);
1381         }
1382
1383         printbuf_indent_sub(buf, 2);
1384 }
1385
1386 noinline __cold
1387 void bch2_dump_trans_updates(struct btree_trans *trans)
1388 {
1389         struct printbuf buf = PRINTBUF;
1390
1391         bch2_trans_updates_to_text(&buf, trans);
1392         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1393         printbuf_exit(&buf);
1394 }
1395
1396 void bch2_btree_path_to_text(struct printbuf *out, struct btree_path *path)
1397 {
1398         prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ",
1399                    path->idx, path->ref, path->intent_ref,
1400                    path->preserve ? 'P' : ' ',
1401                    path->should_be_locked ? 'S' : ' ',
1402                    bch2_btree_ids[path->btree_id],
1403                    path->level);
1404         bch2_bpos_to_text(out, path->pos);
1405
1406         prt_printf(out, " locks %u", path->nodes_locked);
1407 #ifdef CONFIG_BCACHEFS_DEBUG
1408         prt_printf(out, " %pS", (void *) path->ip_allocated);
1409 #endif
1410         prt_newline(out);
1411 }
1412
1413 void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans)
1414 {
1415         struct btree_path *path;
1416         unsigned idx;
1417
1418         trans_for_each_path_inorder(trans, path, idx)
1419                 bch2_btree_path_to_text(out, path);
1420 }
1421
1422 noinline __cold
1423 void bch2_dump_trans_paths_updates(struct btree_trans *trans)
1424 {
1425         struct printbuf buf = PRINTBUF;
1426
1427         bch2_trans_paths_to_text(&buf, trans);
1428         bch2_trans_updates_to_text(&buf, trans);
1429
1430         bch2_print_string_as_lines(KERN_ERR, buf.buf);
1431         printbuf_exit(&buf);
1432 }
1433
1434 noinline
1435 static void bch2_trans_update_max_paths(struct btree_trans *trans)
1436 {
1437         struct btree_transaction_stats *s = btree_trans_stats(trans);
1438         struct printbuf buf = PRINTBUF;
1439
1440         bch2_trans_paths_to_text(&buf, trans);
1441
1442         if (!buf.allocation_failure) {
1443                 mutex_lock(&s->lock);
1444                 if (s->nr_max_paths < hweight64(trans->paths_allocated)) {
1445                         s->nr_max_paths = trans->nr_max_paths =
1446                                 hweight64(trans->paths_allocated);
1447                         swap(s->max_paths_text, buf.buf);
1448                 }
1449                 mutex_unlock(&s->lock);
1450         }
1451
1452         printbuf_exit(&buf);
1453 }
1454
1455 static noinline void btree_path_overflow(struct btree_trans *trans)
1456 {
1457         bch2_dump_trans_paths_updates(trans);
1458         panic("trans path oveflow\n");
1459 }
1460
1461 static inline struct btree_path *btree_path_alloc(struct btree_trans *trans,
1462                                                   struct btree_path *pos)
1463 {
1464         struct btree_path *path;
1465         unsigned idx;
1466
1467         if (unlikely(trans->paths_allocated ==
1468                      ~((~0ULL << 1) << (BTREE_ITER_MAX - 1))))
1469                 btree_path_overflow(trans);
1470
1471         idx = __ffs64(~trans->paths_allocated);
1472         trans->paths_allocated |= 1ULL << idx;
1473
1474         if (unlikely(idx > trans->nr_max_paths))
1475                 bch2_trans_update_max_paths(trans);
1476
1477         path = &trans->paths[idx];
1478
1479         path->idx               = idx;
1480         path->ref               = 0;
1481         path->intent_ref        = 0;
1482         path->nodes_locked      = 0;
1483
1484         btree_path_list_add(trans, pos, path);
1485         return path;
1486 }
1487
1488 struct btree_path *bch2_path_get(struct btree_trans *trans,
1489                                  enum btree_id btree_id, struct bpos pos,
1490                                  unsigned locks_want, unsigned level,
1491                                  unsigned flags, unsigned long ip)
1492 {
1493         struct btree_path *path, *path_pos = NULL;
1494         bool cached = flags & BTREE_ITER_CACHED;
1495         bool intent = flags & BTREE_ITER_INTENT;
1496         int i;
1497
1498         BUG_ON(trans->restarted);
1499         btree_trans_verify_sorted(trans);
1500         bch2_trans_verify_locks(trans);
1501
1502         trans_for_each_path_inorder(trans, path, i) {
1503                 if (__btree_path_cmp(path,
1504                                      btree_id,
1505                                      cached,
1506                                      pos,
1507                                      level) > 0)
1508                         break;
1509
1510                 path_pos = path;
1511         }
1512
1513         if (path_pos &&
1514             path_pos->cached    == cached &&
1515             path_pos->btree_id  == btree_id &&
1516             path_pos->level     == level) {
1517                 __btree_path_get(path_pos, intent);
1518                 path = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip);
1519         } else {
1520                 path = btree_path_alloc(trans, path_pos);
1521                 path_pos = NULL;
1522
1523                 __btree_path_get(path, intent);
1524                 path->pos                       = pos;
1525                 path->btree_id                  = btree_id;
1526                 path->cached                    = cached;
1527                 path->uptodate                  = BTREE_ITER_NEED_TRAVERSE;
1528                 path->should_be_locked          = false;
1529                 path->level                     = level;
1530                 path->locks_want                = locks_want;
1531                 path->nodes_locked              = 0;
1532                 for (i = 0; i < ARRAY_SIZE(path->l); i++)
1533                         path->l[i].b            = ERR_PTR(-BCH_ERR_no_btree_node_init);
1534 #ifdef CONFIG_BCACHEFS_DEBUG
1535                 path->ip_allocated              = ip;
1536 #endif
1537                 btree_trans_verify_sorted(trans);
1538         }
1539
1540         if (!(flags & BTREE_ITER_NOPRESERVE))
1541                 path->preserve = true;
1542
1543         if (path->intent_ref)
1544                 locks_want = max(locks_want, level + 1);
1545
1546         /*
1547          * If the path has locks_want greater than requested, we don't downgrade
1548          * it here - on transaction restart because btree node split needs to
1549          * upgrade locks, we might be putting/getting the iterator again.
1550          * Downgrading iterators only happens via bch2_trans_downgrade(), after
1551          * a successful transaction commit.
1552          */
1553
1554         locks_want = min(locks_want, BTREE_MAX_DEPTH);
1555         if (locks_want > path->locks_want)
1556                 bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want);
1557
1558         return path;
1559 }
1560
1561 inline struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u)
1562 {
1563
1564         struct btree_path_level *l = path_l(path);
1565         struct bkey_packed *_k;
1566         struct bkey_s_c k;
1567
1568         if (unlikely(!l->b))
1569                 return bkey_s_c_null;
1570
1571         EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE);
1572         EBUG_ON(!btree_node_locked(path, path->level));
1573
1574         if (!path->cached) {
1575                 _k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
1576                 k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null;
1577
1578                 EBUG_ON(k.k && bkey_deleted(k.k) && bpos_cmp(k.k->p, path->pos) == 0);
1579
1580                 if (!k.k || bpos_cmp(path->pos, k.k->p))
1581                         goto hole;
1582         } else {
1583                 struct bkey_cached *ck = (void *) path->l[0].b;
1584
1585                 EBUG_ON(ck &&
1586                         (path->btree_id != ck->key.btree_id ||
1587                          bkey_cmp(path->pos, ck->key.pos)));
1588                 EBUG_ON(!ck || !ck->valid);
1589
1590                 *u = ck->k->k;
1591                 k = bkey_i_to_s_c(ck->k);
1592         }
1593
1594         return k;
1595 hole:
1596         bkey_init(u);
1597         u->p = path->pos;
1598         return (struct bkey_s_c) { u, NULL };
1599 }
1600
1601 /* Btree iterators: */
1602
1603 int __must_check
1604 __bch2_btree_iter_traverse(struct btree_iter *iter)
1605 {
1606         return bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1607 }
1608
1609 int __must_check
1610 bch2_btree_iter_traverse(struct btree_iter *iter)
1611 {
1612         int ret;
1613
1614         iter->path = bch2_btree_path_set_pos(iter->trans, iter->path,
1615                                         btree_iter_search_key(iter),
1616                                         iter->flags & BTREE_ITER_INTENT,
1617                                         btree_iter_ip_allocated(iter));
1618
1619         ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags);
1620         if (ret)
1621                 return ret;
1622
1623         btree_path_set_should_be_locked(iter->path);
1624         return 0;
1625 }
1626
1627 /* Iterate across nodes (leaf and interior nodes) */
1628
1629 struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
1630 {
1631         struct btree_trans *trans = iter->trans;
1632         struct btree *b = NULL;
1633         int ret;
1634
1635         EBUG_ON(iter->path->cached);
1636         bch2_btree_iter_verify(iter);
1637
1638         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1639         if (ret)
1640                 goto err;
1641
1642         b = btree_path_node(iter->path, iter->path->level);
1643         if (!b)
1644                 goto out;
1645
1646         BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0);
1647
1648         bkey_init(&iter->k);
1649         iter->k.p = iter->pos = b->key.k.p;
1650
1651         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1652                                         iter->flags & BTREE_ITER_INTENT,
1653                                         btree_iter_ip_allocated(iter));
1654         btree_path_set_should_be_locked(iter->path);
1655 out:
1656         bch2_btree_iter_verify_entry_exit(iter);
1657         bch2_btree_iter_verify(iter);
1658
1659         return b;
1660 err:
1661         b = ERR_PTR(ret);
1662         goto out;
1663 }
1664
1665 struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
1666 {
1667         struct btree_trans *trans = iter->trans;
1668         struct btree_path *path = iter->path;
1669         struct btree *b = NULL;
1670         int ret;
1671
1672         BUG_ON(trans->restarted);
1673         EBUG_ON(iter->path->cached);
1674         bch2_btree_iter_verify(iter);
1675
1676         /* already at end? */
1677         if (!btree_path_node(path, path->level))
1678                 return NULL;
1679
1680         /* got to end? */
1681         if (!btree_path_node(path, path->level + 1)) {
1682                 btree_path_set_level_up(trans, path);
1683                 return NULL;
1684         }
1685
1686         if (!bch2_btree_node_relock(trans, path, path->level + 1)) {
1687                 __bch2_btree_path_unlock(trans, path);
1688                 path->l[path->level].b          = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1689                 path->l[path->level + 1].b      = ERR_PTR(-BCH_ERR_no_btree_node_relock);
1690                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
1691                 trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path);
1692                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
1693                 goto err;
1694         }
1695
1696         b = btree_path_node(path, path->level + 1);
1697
1698         if (!bpos_cmp(iter->pos, b->key.k.p)) {
1699                 __btree_path_set_level_up(trans, path, path->level++);
1700         } else {
1701                 /*
1702                  * Haven't gotten to the end of the parent node: go back down to
1703                  * the next child node
1704                  */
1705                 path = iter->path =
1706                         bch2_btree_path_set_pos(trans, path, bpos_successor(iter->pos),
1707                                            iter->flags & BTREE_ITER_INTENT,
1708                                            btree_iter_ip_allocated(iter));
1709
1710                 btree_path_set_level_down(trans, path, iter->min_depth);
1711
1712                 ret = bch2_btree_path_traverse(trans, path, iter->flags);
1713                 if (ret)
1714                         goto err;
1715
1716                 b = path->l[path->level].b;
1717         }
1718
1719         bkey_init(&iter->k);
1720         iter->k.p = iter->pos = b->key.k.p;
1721
1722         iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p,
1723                                         iter->flags & BTREE_ITER_INTENT,
1724                                         btree_iter_ip_allocated(iter));
1725         btree_path_set_should_be_locked(iter->path);
1726         BUG_ON(iter->path->uptodate);
1727 out:
1728         bch2_btree_iter_verify_entry_exit(iter);
1729         bch2_btree_iter_verify(iter);
1730
1731         return b;
1732 err:
1733         b = ERR_PTR(ret);
1734         goto out;
1735 }
1736
1737 /* Iterate across keys (in leaf nodes only) */
1738
1739 inline bool bch2_btree_iter_advance(struct btree_iter *iter)
1740 {
1741         if (likely(!(iter->flags & BTREE_ITER_ALL_LEVELS))) {
1742                 struct bpos pos = iter->k.p;
1743                 bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1744                             ? bpos_cmp(pos, SPOS_MAX)
1745                             : bkey_cmp(pos, SPOS_MAX)) != 0;
1746
1747                 if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1748                         pos = bkey_successor(iter, pos);
1749                 bch2_btree_iter_set_pos(iter, pos);
1750                 return ret;
1751         } else {
1752                 if (!btree_path_node(iter->path, iter->path->level))
1753                         return true;
1754
1755                 iter->advanced = true;
1756                 return false;
1757         }
1758 }
1759
1760 inline bool bch2_btree_iter_rewind(struct btree_iter *iter)
1761 {
1762         struct bpos pos = bkey_start_pos(&iter->k);
1763         bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS
1764                     ? bpos_cmp(pos, POS_MIN)
1765                     : bkey_cmp(pos, POS_MIN)) != 0;
1766
1767         if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
1768                 pos = bkey_predecessor(iter, pos);
1769         bch2_btree_iter_set_pos(iter, pos);
1770         return ret;
1771 }
1772
1773 static inline struct bkey_i *btree_trans_peek_updates(struct btree_trans *trans,
1774                                                       enum btree_id btree_id,
1775                                                       struct bpos pos)
1776 {
1777         struct btree_insert_entry *i;
1778         struct bkey_i *ret = NULL;
1779
1780         trans_for_each_update(trans, i) {
1781                 if (i->btree_id < btree_id)
1782                         continue;
1783                 if (i->btree_id > btree_id)
1784                         break;
1785                 if (bpos_cmp(i->k->k.p, pos) < 0)
1786                         continue;
1787                 if (i->key_cache_already_flushed)
1788                         continue;
1789                 if (!ret || bpos_cmp(i->k->k.p, ret->k.p) < 0)
1790                         ret = i->k;
1791         }
1792
1793         return ret;
1794 }
1795
1796 struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
1797                                        struct btree_iter *iter,
1798                                        struct bpos start_pos,
1799                                        struct bpos end_pos)
1800 {
1801         struct bkey_i *k;
1802
1803         if (bpos_cmp(start_pos, iter->journal_pos) < 0)
1804                 iter->journal_idx = 0;
1805
1806         k = bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
1807                                         iter->path->level,
1808                                         start_pos, end_pos,
1809                                         &iter->journal_idx);
1810
1811         iter->journal_pos = k ? k->k.p : end_pos;
1812         return k;
1813 }
1814
1815 struct bkey_i *bch2_btree_journal_peek_slot(struct btree_trans *trans,
1816                                             struct btree_iter *iter,
1817                                             struct bpos pos)
1818 {
1819         return bch2_btree_journal_peek(trans, iter, pos, pos);
1820 }
1821
1822 static noinline
1823 struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
1824                                          struct btree_iter *iter,
1825                                          struct bkey_s_c k)
1826 {
1827         struct bkey_i *next_journal =
1828                 bch2_btree_journal_peek(trans, iter, iter->path->pos,
1829                                 k.k ? k.k->p : path_l(iter->path)->b->key.k.p);
1830
1831         if (next_journal) {
1832                 iter->k = next_journal->k;
1833                 k = bkey_i_to_s_c(next_journal);
1834         }
1835
1836         return k;
1837 }
1838
1839 /*
1840  * Checks btree key cache for key at iter->pos and returns it if present, or
1841  * bkey_s_c_null:
1842  */
1843 static noinline
1844 struct bkey_s_c __btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1845 {
1846         struct btree_trans *trans = iter->trans;
1847         struct bch_fs *c = trans->c;
1848         struct bkey u;
1849         int ret;
1850
1851         if (!bch2_btree_key_cache_find(c, iter->btree_id, pos))
1852                 return bkey_s_c_null;
1853
1854         if (!iter->key_cache_path)
1855                 iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos,
1856                                                      iter->flags & BTREE_ITER_INTENT, 0,
1857                                                      iter->flags|BTREE_ITER_CACHED,
1858                                                      _THIS_IP_);
1859
1860         iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos,
1861                                         iter->flags & BTREE_ITER_INTENT,
1862                                         btree_iter_ip_allocated(iter));
1863
1864         ret = bch2_btree_path_traverse(trans, iter->key_cache_path, iter->flags|BTREE_ITER_CACHED);
1865         if (unlikely(ret))
1866                 return bkey_s_c_err(ret);
1867
1868         btree_path_set_should_be_locked(iter->key_cache_path);
1869
1870         return bch2_btree_path_peek_slot(iter->key_cache_path, &u);
1871 }
1872
1873 static noinline
1874 struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos)
1875 {
1876         struct bkey_s_c ret = __btree_trans_peek_key_cache(iter, pos);
1877         int err = bkey_err(ret) ?: bch2_btree_path_relock(iter->trans, iter->path, _THIS_IP_);
1878
1879         return err ? bkey_s_c_err(err) : ret;
1880 }
1881
1882 static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key)
1883 {
1884         struct btree_trans *trans = iter->trans;
1885         struct bkey_i *next_update;
1886         struct bkey_s_c k, k2;
1887         int ret;
1888
1889         EBUG_ON(iter->path->cached);
1890         bch2_btree_iter_verify(iter);
1891
1892         while (1) {
1893                 struct btree_path_level *l;
1894
1895                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
1896                                         iter->flags & BTREE_ITER_INTENT,
1897                                         btree_iter_ip_allocated(iter));
1898
1899                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
1900                 if (unlikely(ret)) {
1901                         /* ensure that iter->k is consistent with iter->pos: */
1902                         bch2_btree_iter_set_pos(iter, iter->pos);
1903                         k = bkey_s_c_err(ret);
1904                         goto out;
1905                 }
1906
1907                 l = path_l(iter->path);
1908
1909                 if (unlikely(!l->b)) {
1910                         /* No btree nodes at requested level: */
1911                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
1912                         k = bkey_s_c_null;
1913                         goto out;
1914                 }
1915
1916                 btree_path_set_should_be_locked(iter->path);
1917
1918                 k = btree_path_level_peek_all(trans->c, l, &iter->k);
1919
1920                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
1921                     k.k &&
1922                     (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
1923                         k = k2;
1924                         ret = bkey_err(k);
1925                         if (ret) {
1926                                 bch2_btree_iter_set_pos(iter, iter->pos);
1927                                 goto out;
1928                         }
1929                 }
1930
1931                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL))
1932                         k = btree_trans_peek_journal(trans, iter, k);
1933
1934                 next_update = iter->flags & BTREE_ITER_WITH_UPDATES
1935                         ? btree_trans_peek_updates(trans, iter->btree_id, search_key)
1936                         : NULL;
1937                 if (next_update &&
1938                     bpos_cmp(next_update->k.p,
1939                              k.k ? k.k->p : l->b->key.k.p) <= 0) {
1940                         iter->k = next_update->k;
1941                         k = bkey_i_to_s_c(next_update);
1942                 }
1943
1944                 if (k.k && bkey_deleted(k.k)) {
1945                         /*
1946                          * If we've got a whiteout, and it's after the search
1947                          * key, advance the search key to the whiteout instead
1948                          * of just after the whiteout - it might be a btree
1949                          * whiteout, with a real key at the same position, since
1950                          * in the btree deleted keys sort before non deleted.
1951                          */
1952                         search_key = bpos_cmp(search_key, k.k->p)
1953                                 ? k.k->p
1954                                 : bpos_successor(k.k->p);
1955                         continue;
1956                 }
1957
1958                 if (likely(k.k)) {
1959                         break;
1960                 } else if (likely(bpos_cmp(l->b->key.k.p, SPOS_MAX))) {
1961                         /* Advance to next leaf node: */
1962                         search_key = bpos_successor(l->b->key.k.p);
1963                 } else {
1964                         /* End of btree: */
1965                         bch2_btree_iter_set_pos(iter, SPOS_MAX);
1966                         k = bkey_s_c_null;
1967                         goto out;
1968                 }
1969         }
1970 out:
1971         bch2_btree_iter_verify(iter);
1972
1973         return k;
1974 }
1975
1976 /**
1977  * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
1978  * current position
1979  */
1980 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
1981 {
1982         struct btree_trans *trans = iter->trans;
1983         struct bpos search_key = btree_iter_search_key(iter);
1984         struct bkey_s_c k;
1985         struct bpos iter_pos;
1986         int ret;
1987
1988         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
1989
1990         if (iter->update_path) {
1991                 bch2_path_put_nokeep(trans, iter->update_path,
1992                                      iter->flags & BTREE_ITER_INTENT);
1993                 iter->update_path = NULL;
1994         }
1995
1996         bch2_btree_iter_verify_entry_exit(iter);
1997
1998         while (1) {
1999                 k = __bch2_btree_iter_peek(iter, search_key);
2000                 if (!k.k || bkey_err(k))
2001                         goto out_no_locked;
2002
2003                 /*
2004                  * iter->pos should be mononotically increasing, and always be
2005                  * equal to the key we just returned - except extents can
2006                  * straddle iter->pos:
2007                  */
2008                 if (!(iter->flags & BTREE_ITER_IS_EXTENTS))
2009                         iter_pos = k.k->p;
2010                 else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
2011                         iter_pos = bkey_start_pos(k.k);
2012                 else
2013                         iter_pos = iter->pos;
2014
2015                 if (bkey_cmp(iter_pos, end) > 0) {
2016                         bch2_btree_iter_set_pos(iter, end);
2017                         k = bkey_s_c_null;
2018                         goto out_no_locked;
2019                 }
2020
2021                 if (iter->update_path &&
2022                     bkey_cmp(iter->update_path->pos, k.k->p)) {
2023                         bch2_path_put_nokeep(trans, iter->update_path,
2024                                              iter->flags & BTREE_ITER_INTENT);
2025                         iter->update_path = NULL;
2026                 }
2027
2028                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2029                     (iter->flags & BTREE_ITER_INTENT) &&
2030                     !(iter->flags & BTREE_ITER_IS_EXTENTS) &&
2031                     !iter->update_path) {
2032                         struct bpos pos = k.k->p;
2033
2034                         if (pos.snapshot < iter->snapshot) {
2035                                 search_key = bpos_successor(k.k->p);
2036                                 continue;
2037                         }
2038
2039                         pos.snapshot = iter->snapshot;
2040
2041                         /*
2042                          * advance, same as on exit for iter->path, but only up
2043                          * to snapshot
2044                          */
2045                         __btree_path_get(iter->path, iter->flags & BTREE_ITER_INTENT);
2046                         iter->update_path = iter->path;
2047
2048                         iter->update_path = bch2_btree_path_set_pos(trans,
2049                                                 iter->update_path, pos,
2050                                                 iter->flags & BTREE_ITER_INTENT,
2051                                                 _THIS_IP_);
2052                 }
2053
2054                 /*
2055                  * We can never have a key in a leaf node at POS_MAX, so
2056                  * we don't have to check these successor() calls:
2057                  */
2058                 if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
2059                     !bch2_snapshot_is_ancestor(trans->c,
2060                                                iter->snapshot,
2061                                                k.k->p.snapshot)) {
2062                         search_key = bpos_successor(k.k->p);
2063                         continue;
2064                 }
2065
2066                 if (bkey_whiteout(k.k) &&
2067                     !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2068                         search_key = bkey_successor(iter, k.k->p);
2069                         continue;
2070                 }
2071
2072                 break;
2073         }
2074
2075         iter->pos = iter_pos;
2076
2077         iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p,
2078                                 iter->flags & BTREE_ITER_INTENT,
2079                                 btree_iter_ip_allocated(iter));
2080
2081         btree_path_set_should_be_locked(iter->path);
2082 out_no_locked:
2083         if (iter->update_path) {
2084                 if (iter->update_path->uptodate &&
2085                     (ret = bch2_btree_path_relock(trans, iter->update_path, _THIS_IP_)))
2086                         k = bkey_s_c_err(ret);
2087                 else
2088                         btree_path_set_should_be_locked(iter->update_path);
2089         }
2090
2091         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
2092                 iter->pos.snapshot = iter->snapshot;
2093
2094         ret = bch2_btree_iter_verify_ret(iter, k);
2095         if (unlikely(ret)) {
2096                 bch2_btree_iter_set_pos(iter, iter->pos);
2097                 k = bkey_s_c_err(ret);
2098         }
2099
2100         bch2_btree_iter_verify_entry_exit(iter);
2101
2102         return k;
2103 }
2104
2105 /**
2106  * bch2_btree_iter_peek_all_levels: returns the first key greater than or equal
2107  * to iterator's current position, returning keys from every level of the btree.
2108  * For keys at different levels of the btree that compare equal, the key from
2109  * the lower level (leaf) is returned first.
2110  */
2111 struct bkey_s_c bch2_btree_iter_peek_all_levels(struct btree_iter *iter)
2112 {
2113         struct btree_trans *trans = iter->trans;
2114         struct bkey_s_c k;
2115         int ret;
2116
2117         EBUG_ON(iter->path->cached);
2118         bch2_btree_iter_verify(iter);
2119         BUG_ON(iter->path->level < iter->min_depth);
2120         BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS));
2121         EBUG_ON(!(iter->flags & BTREE_ITER_ALL_LEVELS));
2122
2123         while (1) {
2124                 iter->path = bch2_btree_path_set_pos(trans, iter->path, iter->pos,
2125                                         iter->flags & BTREE_ITER_INTENT,
2126                                         btree_iter_ip_allocated(iter));
2127
2128                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2129                 if (unlikely(ret)) {
2130                         /* ensure that iter->k is consistent with iter->pos: */
2131                         bch2_btree_iter_set_pos(iter, iter->pos);
2132                         k = bkey_s_c_err(ret);
2133                         goto out_no_locked;
2134                 }
2135
2136                 /* Already at end? */
2137                 if (!btree_path_node(iter->path, iter->path->level)) {
2138                         k = bkey_s_c_null;
2139                         goto out_no_locked;
2140                 }
2141
2142                 k = btree_path_level_peek_all(trans->c,
2143                                 &iter->path->l[iter->path->level], &iter->k);
2144
2145                 /* Check if we should go up to the parent node: */
2146                 if (!k.k ||
2147                     (iter->advanced &&
2148                      !bpos_cmp(path_l(iter->path)->b->key.k.p, iter->pos))) {
2149                         iter->pos = path_l(iter->path)->b->key.k.p;
2150                         btree_path_set_level_up(trans, iter->path);
2151                         iter->advanced = false;
2152                         continue;
2153                 }
2154
2155                 /*
2156                  * Check if we should go back down to a leaf:
2157                  * If we're not in a leaf node, we only return the current key
2158                  * if it exactly matches iter->pos - otherwise we first have to
2159                  * go back to the leaf:
2160                  */
2161                 if (iter->path->level != iter->min_depth &&
2162                     (iter->advanced ||
2163                      !k.k ||
2164                      bpos_cmp(iter->pos, k.k->p))) {
2165                         btree_path_set_level_down(trans, iter->path, iter->min_depth);
2166                         iter->pos = bpos_successor(iter->pos);
2167                         iter->advanced = false;
2168                         continue;
2169                 }
2170
2171                 /* Check if we should go to the next key: */
2172                 if (iter->path->level == iter->min_depth &&
2173                     iter->advanced &&
2174                     k.k &&
2175                     !bpos_cmp(iter->pos, k.k->p)) {
2176                         iter->pos = bpos_successor(iter->pos);
2177                         iter->advanced = false;
2178                         continue;
2179                 }
2180
2181                 if (iter->advanced &&
2182                     iter->path->level == iter->min_depth &&
2183                     bpos_cmp(k.k->p, iter->pos))
2184                         iter->advanced = false;
2185
2186                 BUG_ON(iter->advanced);
2187                 BUG_ON(!k.k);
2188                 break;
2189         }
2190
2191         iter->pos = k.k->p;
2192         btree_path_set_should_be_locked(iter->path);
2193 out_no_locked:
2194         bch2_btree_iter_verify(iter);
2195
2196         return k;
2197 }
2198
2199 /**
2200  * bch2_btree_iter_next: returns first key greater than iterator's current
2201  * position
2202  */
2203 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
2204 {
2205         if (!bch2_btree_iter_advance(iter))
2206                 return bkey_s_c_null;
2207
2208         return bch2_btree_iter_peek(iter);
2209 }
2210
2211 /**
2212  * bch2_btree_iter_peek_prev: returns first key less than or equal to
2213  * iterator's current position
2214  */
2215 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
2216 {
2217         struct btree_trans *trans = iter->trans;
2218         struct bpos search_key = iter->pos;
2219         struct btree_path *saved_path = NULL;
2220         struct bkey_s_c k;
2221         struct bkey saved_k;
2222         const struct bch_val *saved_v;
2223         int ret;
2224
2225         EBUG_ON(iter->path->cached || iter->path->level);
2226         EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES);
2227
2228         if (iter->flags & BTREE_ITER_WITH_JOURNAL)
2229                 return bkey_s_c_err(-EIO);
2230
2231         bch2_btree_iter_verify(iter);
2232         bch2_btree_iter_verify_entry_exit(iter);
2233
2234         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2235                 search_key.snapshot = U32_MAX;
2236
2237         while (1) {
2238                 iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2239                                                 iter->flags & BTREE_ITER_INTENT,
2240                                                 btree_iter_ip_allocated(iter));
2241
2242                 ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2243                 if (unlikely(ret)) {
2244                         /* ensure that iter->k is consistent with iter->pos: */
2245                         bch2_btree_iter_set_pos(iter, iter->pos);
2246                         k = bkey_s_c_err(ret);
2247                         goto out_no_locked;
2248                 }
2249
2250                 k = btree_path_level_peek(trans, iter->path,
2251                                           &iter->path->l[0], &iter->k);
2252                 if (!k.k ||
2253                     ((iter->flags & BTREE_ITER_IS_EXTENTS)
2254                      ? bpos_cmp(bkey_start_pos(k.k), search_key) >= 0
2255                      : bpos_cmp(k.k->p, search_key) > 0))
2256                         k = btree_path_level_prev(trans, iter->path,
2257                                                   &iter->path->l[0], &iter->k);
2258
2259                 bch2_btree_path_check_sort(trans, iter->path, 0);
2260
2261                 if (likely(k.k)) {
2262                         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) {
2263                                 if (k.k->p.snapshot == iter->snapshot)
2264                                         goto got_key;
2265
2266                                 /*
2267                                  * If we have a saved candidate, and we're no
2268                                  * longer at the same _key_ (not pos), return
2269                                  * that candidate
2270                                  */
2271                                 if (saved_path && bkey_cmp(k.k->p, saved_k.p)) {
2272                                         bch2_path_put_nokeep(trans, iter->path,
2273                                                       iter->flags & BTREE_ITER_INTENT);
2274                                         iter->path = saved_path;
2275                                         saved_path = NULL;
2276                                         iter->k = saved_k;
2277                                         k.v     = saved_v;
2278                                         goto got_key;
2279                                 }
2280
2281                                 if (bch2_snapshot_is_ancestor(iter->trans->c,
2282                                                               iter->snapshot,
2283                                                               k.k->p.snapshot)) {
2284                                         if (saved_path)
2285                                                 bch2_path_put_nokeep(trans, saved_path,
2286                                                       iter->flags & BTREE_ITER_INTENT);
2287                                         saved_path = btree_path_clone(trans, iter->path,
2288                                                                 iter->flags & BTREE_ITER_INTENT);
2289                                         saved_k = *k.k;
2290                                         saved_v = k.v;
2291                                 }
2292
2293                                 search_key = bpos_predecessor(k.k->p);
2294                                 continue;
2295                         }
2296 got_key:
2297                         if (bkey_whiteout(k.k) &&
2298                             !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) {
2299                                 search_key = bkey_predecessor(iter, k.k->p);
2300                                 if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2301                                         search_key.snapshot = U32_MAX;
2302                                 continue;
2303                         }
2304
2305                         break;
2306                 } else if (likely(bpos_cmp(iter->path->l[0].b->data->min_key, POS_MIN))) {
2307                         /* Advance to previous leaf node: */
2308                         search_key = bpos_predecessor(iter->path->l[0].b->data->min_key);
2309                 } else {
2310                         /* Start of btree: */
2311                         bch2_btree_iter_set_pos(iter, POS_MIN);
2312                         k = bkey_s_c_null;
2313                         goto out_no_locked;
2314                 }
2315         }
2316
2317         EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
2318
2319         /* Extents can straddle iter->pos: */
2320         if (bkey_cmp(k.k->p, iter->pos) < 0)
2321                 iter->pos = k.k->p;
2322
2323         if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)
2324                 iter->pos.snapshot = iter->snapshot;
2325
2326         btree_path_set_should_be_locked(iter->path);
2327 out_no_locked:
2328         if (saved_path)
2329                 bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT);
2330
2331         bch2_btree_iter_verify_entry_exit(iter);
2332         bch2_btree_iter_verify(iter);
2333
2334         return k;
2335 }
2336
2337 /**
2338  * bch2_btree_iter_prev: returns first key less than iterator's current
2339  * position
2340  */
2341 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
2342 {
2343         if (!bch2_btree_iter_rewind(iter))
2344                 return bkey_s_c_null;
2345
2346         return bch2_btree_iter_peek_prev(iter);
2347 }
2348
2349 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
2350 {
2351         struct btree_trans *trans = iter->trans;
2352         struct bpos search_key;
2353         struct bkey_s_c k;
2354         int ret;
2355
2356         bch2_btree_iter_verify(iter);
2357         bch2_btree_iter_verify_entry_exit(iter);
2358         EBUG_ON(iter->flags & BTREE_ITER_ALL_LEVELS);
2359         EBUG_ON(iter->path->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE));
2360
2361         /* extents can't span inode numbers: */
2362         if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
2363             unlikely(iter->pos.offset == KEY_OFFSET_MAX)) {
2364                 if (iter->pos.inode == KEY_INODE_MAX)
2365                         return bkey_s_c_null;
2366
2367                 bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
2368         }
2369
2370         search_key = btree_iter_search_key(iter);
2371         iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
2372                                         iter->flags & BTREE_ITER_INTENT,
2373                                         btree_iter_ip_allocated(iter));
2374
2375         ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
2376         if (unlikely(ret)) {
2377                 k = bkey_s_c_err(ret);
2378                 goto out_no_locked;
2379         }
2380
2381         if ((iter->flags & BTREE_ITER_CACHED) ||
2382             !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) {
2383                 struct bkey_i *next_update;
2384
2385                 if ((iter->flags & BTREE_ITER_WITH_UPDATES) &&
2386                     (next_update = btree_trans_peek_updates(trans,
2387                                                 iter->btree_id, search_key)) &&
2388                     !bpos_cmp(next_update->k.p, iter->pos)) {
2389                         iter->k = next_update->k;
2390                         k = bkey_i_to_s_c(next_update);
2391                         goto out;
2392                 }
2393
2394                 if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) &&
2395                     (next_update = bch2_btree_journal_peek_slot(trans,
2396                                         iter, iter->pos))) {
2397                         iter->k = next_update->k;
2398                         k = bkey_i_to_s_c(next_update);
2399                         goto out;
2400                 }
2401
2402                 if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) &&
2403                     (k = __btree_trans_peek_key_cache(iter, iter->pos)).k) {
2404                         if (!bkey_err(k))
2405                                 iter->k = *k.k;
2406                         /* We're not returning a key from iter->path: */
2407                         goto out_no_locked;
2408                 }
2409
2410                 k = bch2_btree_path_peek_slot(iter->path, &iter->k);
2411                 if (unlikely(!k.k))
2412                         goto out_no_locked;
2413         } else {
2414                 struct bpos next;
2415
2416                 EBUG_ON(iter->path->level);
2417
2418                 if (iter->flags & BTREE_ITER_INTENT) {
2419                         struct btree_iter iter2;
2420                         struct bpos end = iter->pos;
2421
2422                         if (iter->flags & BTREE_ITER_IS_EXTENTS)
2423                                 end.offset = U64_MAX;
2424
2425                         bch2_trans_copy_iter(&iter2, iter);
2426                         k = bch2_btree_iter_peek_upto(&iter2, end);
2427
2428                         if (k.k && !bkey_err(k)) {
2429                                 iter->k = iter2.k;
2430                                 k.k = &iter->k;
2431                         }
2432                         bch2_trans_iter_exit(trans, &iter2);
2433                 } else {
2434                         struct bpos pos = iter->pos;
2435
2436                         k = bch2_btree_iter_peek(iter);
2437                         if (unlikely(bkey_err(k)))
2438                                 bch2_btree_iter_set_pos(iter, pos);
2439                         else
2440                                 iter->pos = pos;
2441                 }
2442
2443                 if (unlikely(bkey_err(k)))
2444                         goto out_no_locked;
2445
2446                 next = k.k ? bkey_start_pos(k.k) : POS_MAX;
2447
2448                 if (bkey_cmp(iter->pos, next) < 0) {
2449                         bkey_init(&iter->k);
2450                         iter->k.p = iter->pos;
2451
2452                         if (iter->flags & BTREE_ITER_IS_EXTENTS) {
2453                                 bch2_key_resize(&iter->k,
2454                                                 min_t(u64, KEY_SIZE_MAX,
2455                                                       (next.inode == iter->pos.inode
2456                                                        ? next.offset
2457                                                        : KEY_OFFSET_MAX) -
2458                                                       iter->pos.offset));
2459                                 EBUG_ON(!iter->k.size);
2460                         }
2461
2462                         k = (struct bkey_s_c) { &iter->k, NULL };
2463                 }
2464         }
2465 out:
2466         btree_path_set_should_be_locked(iter->path);
2467 out_no_locked:
2468         bch2_btree_iter_verify_entry_exit(iter);
2469         bch2_btree_iter_verify(iter);
2470         ret = bch2_btree_iter_verify_ret(iter, k);
2471         if (unlikely(ret))
2472                 return bkey_s_c_err(ret);
2473
2474         return k;
2475 }
2476
2477 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
2478 {
2479         if (!bch2_btree_iter_advance(iter))
2480                 return bkey_s_c_null;
2481
2482         return bch2_btree_iter_peek_slot(iter);
2483 }
2484
2485 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter)
2486 {
2487         if (!bch2_btree_iter_rewind(iter))
2488                 return bkey_s_c_null;
2489
2490         return bch2_btree_iter_peek_slot(iter);
2491 }
2492
2493 /* new transactional stuff: */
2494
2495 static inline void btree_path_verify_sorted_ref(struct btree_trans *trans,
2496                                                 struct btree_path *path)
2497 {
2498         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2499         EBUG_ON(trans->sorted[path->sorted_idx] != path->idx);
2500         EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx)));
2501 }
2502
2503 static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans)
2504 {
2505 #ifdef CONFIG_BCACHEFS_DEBUG
2506         unsigned i;
2507
2508         for (i = 0; i < trans->nr_sorted; i++)
2509                 btree_path_verify_sorted_ref(trans, trans->paths + trans->sorted[i]);
2510 #endif
2511 }
2512
2513 static void btree_trans_verify_sorted(struct btree_trans *trans)
2514 {
2515 #ifdef CONFIG_BCACHEFS_DEBUG
2516         struct btree_path *path, *prev = NULL;
2517         unsigned i;
2518
2519         if (!bch2_debug_check_iterators)
2520                 return;
2521
2522         trans_for_each_path_inorder(trans, path, i) {
2523                 if (prev && btree_path_cmp(prev, path) > 0) {
2524                         bch2_dump_trans_paths_updates(trans);
2525                         panic("trans paths out of order!\n");
2526                 }
2527                 prev = path;
2528         }
2529 #endif
2530 }
2531
2532 static inline void btree_path_swap(struct btree_trans *trans,
2533                                    struct btree_path *l, struct btree_path *r)
2534 {
2535         swap(l->sorted_idx, r->sorted_idx);
2536         swap(trans->sorted[l->sorted_idx],
2537              trans->sorted[r->sorted_idx]);
2538
2539         btree_path_verify_sorted_ref(trans, l);
2540         btree_path_verify_sorted_ref(trans, r);
2541 }
2542
2543 static __always_inline void bch2_btree_path_check_sort_fast(struct btree_trans *trans,
2544                                                    struct btree_path *path,
2545                                                    int cmp)
2546 {
2547         struct btree_path *n;
2548         int cmp2;
2549
2550         EBUG_ON(!cmp);
2551
2552         while ((n = cmp < 0
2553                 ? prev_btree_path(trans, path)
2554                 : next_btree_path(trans, path)) &&
2555                (cmp2 = btree_path_cmp(n, path)) &&
2556                cmp2 != cmp)
2557                 btree_path_swap(trans, n, path);
2558
2559         btree_trans_verify_sorted(trans);
2560 }
2561
2562 inline void bch2_btree_path_check_sort(struct btree_trans *trans, struct btree_path *path,
2563                                        int cmp)
2564 {
2565         struct btree_path *n;
2566
2567         if (cmp <= 0) {
2568                 n = prev_btree_path(trans, path);
2569                 if (n && btree_path_cmp(n, path) > 0) {
2570                         do {
2571                                 btree_path_swap(trans, n, path);
2572                                 n = prev_btree_path(trans, path);
2573                         } while (n && btree_path_cmp(n, path) > 0);
2574
2575                         goto out;
2576                 }
2577         }
2578
2579         if (cmp >= 0) {
2580                 n = next_btree_path(trans, path);
2581                 if (n && btree_path_cmp(path, n) > 0) {
2582                         do {
2583                                 btree_path_swap(trans, path, n);
2584                                 n = next_btree_path(trans, path);
2585                         } while (n && btree_path_cmp(path, n) > 0);
2586                 }
2587         }
2588 out:
2589         btree_trans_verify_sorted(trans);
2590 }
2591
2592 static inline void btree_path_list_remove(struct btree_trans *trans,
2593                                           struct btree_path *path)
2594 {
2595         unsigned i;
2596
2597         EBUG_ON(path->sorted_idx >= trans->nr_sorted);
2598
2599         array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx);
2600
2601         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2602                 trans->paths[trans->sorted[i]].sorted_idx = i;
2603
2604         path->sorted_idx = U8_MAX;
2605
2606         btree_trans_verify_sorted_refs(trans);
2607 }
2608
2609 static inline void btree_path_list_add(struct btree_trans *trans,
2610                                        struct btree_path *pos,
2611                                        struct btree_path *path)
2612 {
2613         unsigned i;
2614
2615         btree_trans_verify_sorted_refs(trans);
2616
2617         path->sorted_idx = pos ? pos->sorted_idx + 1 : 0;
2618
2619         if (trans->in_traverse_all &&
2620             trans->traverse_all_idx != U8_MAX &&
2621             trans->traverse_all_idx >= path->sorted_idx)
2622                 trans->traverse_all_idx++;
2623
2624         array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx);
2625
2626         for (i = path->sorted_idx; i < trans->nr_sorted; i++)
2627                 trans->paths[trans->sorted[i]].sorted_idx = i;
2628
2629         btree_trans_verify_sorted_refs(trans);
2630 }
2631
2632 void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter)
2633 {
2634         if (iter->path)
2635                 bch2_path_put(trans, iter->path,
2636                               iter->flags & BTREE_ITER_INTENT);
2637         if (iter->update_path)
2638                 bch2_path_put_nokeep(trans, iter->update_path,
2639                               iter->flags & BTREE_ITER_INTENT);
2640         if (iter->key_cache_path)
2641                 bch2_path_put(trans, iter->key_cache_path,
2642                               iter->flags & BTREE_ITER_INTENT);
2643         iter->path = NULL;
2644         iter->update_path = NULL;
2645         iter->key_cache_path = NULL;
2646 }
2647
2648 static inline void __bch2_trans_iter_init(struct btree_trans *trans,
2649                                           struct btree_iter *iter,
2650                                           unsigned btree_id, struct bpos pos,
2651                                           unsigned locks_want,
2652                                           unsigned depth,
2653                                           unsigned flags,
2654                                           unsigned long ip)
2655 {
2656         if (unlikely(trans->restarted))
2657                 panic("bch2_trans_iter_init(): in transaction restart, %s by %pS\n",
2658                       bch2_err_str(trans->restarted),
2659                       (void *) trans->last_restarted_ip);
2660
2661         if (flags & BTREE_ITER_ALL_LEVELS)
2662                 flags |= BTREE_ITER_ALL_SNAPSHOTS|__BTREE_ITER_ALL_SNAPSHOTS;
2663
2664         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
2665             btree_node_type_is_extents(btree_id))
2666                 flags |= BTREE_ITER_IS_EXTENTS;
2667
2668         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
2669             !btree_type_has_snapshots(btree_id))
2670                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
2671
2672         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
2673             btree_type_has_snapshots(btree_id))
2674                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
2675
2676         if (trans->journal_replay_not_finished)
2677                 flags |= BTREE_ITER_WITH_JOURNAL;
2678
2679         iter->trans     = trans;
2680         iter->path      = NULL;
2681         iter->update_path = NULL;
2682         iter->key_cache_path = NULL;
2683         iter->btree_id  = btree_id;
2684         iter->min_depth = depth;
2685         iter->flags     = flags;
2686         iter->snapshot  = pos.snapshot;
2687         iter->pos       = pos;
2688         iter->k.type    = KEY_TYPE_deleted;
2689         iter->k.p       = pos;
2690         iter->k.size    = 0;
2691         iter->journal_idx = 0;
2692         iter->journal_pos = POS_MIN;
2693 #ifdef CONFIG_BCACHEFS_DEBUG
2694         iter->ip_allocated = ip;
2695 #endif
2696
2697         iter->path = bch2_path_get(trans, btree_id, iter->pos,
2698                                    locks_want, depth, flags, ip);
2699 }
2700
2701 void bch2_trans_iter_init(struct btree_trans *trans,
2702                           struct btree_iter *iter,
2703                           unsigned btree_id, struct bpos pos,
2704                           unsigned flags)
2705 {
2706         if (!btree_id_cached(trans->c, btree_id)) {
2707                 flags &= ~BTREE_ITER_CACHED;
2708                 flags &= ~BTREE_ITER_WITH_KEY_CACHE;
2709         } else if (!(flags & BTREE_ITER_CACHED))
2710                 flags |= BTREE_ITER_WITH_KEY_CACHE;
2711
2712         __bch2_trans_iter_init(trans, iter, btree_id, pos,
2713                                0, 0, flags, _RET_IP_);
2714 }
2715
2716 void bch2_trans_node_iter_init(struct btree_trans *trans,
2717                                struct btree_iter *iter,
2718                                enum btree_id btree_id,
2719                                struct bpos pos,
2720                                unsigned locks_want,
2721                                unsigned depth,
2722                                unsigned flags)
2723 {
2724         __bch2_trans_iter_init(trans, iter, btree_id, pos, locks_want, depth,
2725                                BTREE_ITER_NOT_EXTENTS|
2726                                __BTREE_ITER_ALL_SNAPSHOTS|
2727                                BTREE_ITER_ALL_SNAPSHOTS|
2728                                flags, _RET_IP_);
2729         BUG_ON(iter->path->locks_want    < min(locks_want, BTREE_MAX_DEPTH));
2730         BUG_ON(iter->path->level        != depth);
2731         BUG_ON(iter->min_depth          != depth);
2732 }
2733
2734 void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src)
2735 {
2736         *dst = *src;
2737         if (src->path)
2738                 __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT);
2739         if (src->update_path)
2740                 __btree_path_get(src->update_path, src->flags & BTREE_ITER_INTENT);
2741         dst->key_cache_path = NULL;
2742 }
2743
2744 void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
2745 {
2746         unsigned new_top = trans->mem_top + size;
2747         size_t old_bytes = trans->mem_bytes;
2748         size_t new_bytes = roundup_pow_of_two(new_top);
2749         void *new_mem;
2750         void *p;
2751
2752         trans->mem_max = max(trans->mem_max, new_top);
2753
2754         WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX);
2755
2756         new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
2757         if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) {
2758                 new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL);
2759                 new_bytes = BTREE_TRANS_MEM_MAX;
2760                 kfree(trans->mem);
2761         }
2762
2763         if (!new_mem)
2764                 return ERR_PTR(-ENOMEM);
2765
2766         trans->mem = new_mem;
2767         trans->mem_bytes = new_bytes;
2768
2769         if (old_bytes) {
2770                 trace_and_count(trans->c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes);
2771                 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced));
2772         }
2773
2774         p = trans->mem + trans->mem_top;
2775         trans->mem_top += size;
2776         memset(p, 0, size);
2777         return p;
2778 }
2779
2780 /**
2781  * bch2_trans_begin() - reset a transaction after a interrupted attempt
2782  * @trans: transaction to reset
2783  *
2784  * While iterating over nodes or updating nodes a attempt to lock a btree node
2785  * may return BCH_ERR_transaction_restart when the trylock fails. When this
2786  * occurs bch2_trans_begin() should be called and the transaction retried.
2787  */
2788 u32 bch2_trans_begin(struct btree_trans *trans)
2789 {
2790         struct btree_path *path;
2791
2792         bch2_trans_reset_updates(trans);
2793
2794         trans->restart_count++;
2795         trans->mem_top                  = 0;
2796
2797         if (trans->fs_usage_deltas) {
2798                 trans->fs_usage_deltas->used = 0;
2799                 memset((void *) trans->fs_usage_deltas +
2800                        offsetof(struct replicas_delta_list, memset_start), 0,
2801                        (void *) &trans->fs_usage_deltas->memset_end -
2802                        (void *) &trans->fs_usage_deltas->memset_start);
2803         }
2804
2805         trans_for_each_path(trans, path) {
2806                 path->should_be_locked = false;
2807
2808                 /*
2809                  * If the transaction wasn't restarted, we're presuming to be
2810                  * doing something new: dont keep iterators excpt the ones that
2811                  * are in use - except for the subvolumes btree:
2812                  */
2813                 if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes)
2814                         path->preserve = false;
2815
2816                 /*
2817                  * XXX: we probably shouldn't be doing this if the transaction
2818                  * was restarted, but currently we still overflow transaction
2819                  * iterators if we do that
2820                  */
2821                 if (!path->ref && !path->preserve)
2822                         __bch2_path_free(trans, path);
2823                 else
2824                         path->preserve = false;
2825         }
2826
2827         if (!trans->restarted &&
2828             (need_resched() ||
2829              local_clock() - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) {
2830                 bch2_trans_unlock(trans);
2831                 cond_resched();
2832                 bch2_trans_relock(trans);
2833         }
2834
2835         trans->last_restarted_ip = _RET_IP_;
2836         if (trans->restarted)
2837                 bch2_btree_path_traverse_all(trans);
2838
2839         trans->last_begin_time = local_clock();
2840         return trans->restart_count;
2841 }
2842
2843 void bch2_trans_verify_not_restarted(struct btree_trans *trans, u32 restart_count)
2844 {
2845         if (trans_was_restarted(trans, restart_count))
2846                 panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
2847                       trans->restart_count, restart_count,
2848                       (void *) trans->last_restarted_ip);
2849 }
2850
2851 static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c)
2852 {
2853         size_t paths_bytes      = sizeof(struct btree_path) * BTREE_ITER_MAX;
2854         size_t updates_bytes    = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX;
2855         void *p = NULL;
2856
2857         BUG_ON(trans->used_mempool);
2858
2859 #ifdef __KERNEL__
2860         p = this_cpu_xchg(c->btree_paths_bufs->path, NULL);
2861 #endif
2862         if (!p)
2863                 p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS);
2864
2865         trans->paths            = p; p += paths_bytes;
2866         trans->updates          = p; p += updates_bytes;
2867 }
2868
2869 const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
2870
2871 unsigned bch2_trans_get_fn_idx(const char *fn)
2872 {
2873         unsigned i;
2874
2875         for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++)
2876                 if (!bch2_btree_transaction_fns[i] ||
2877                     bch2_btree_transaction_fns[i] == fn) {
2878                         bch2_btree_transaction_fns[i] = fn;
2879                         return i;
2880                 }
2881
2882         pr_warn_once("BCH_TRANSACTIONS_NR not big enough!");
2883         return i;
2884 }
2885
2886 void __bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, unsigned fn_idx)
2887         __acquires(&c->btree_trans_barrier)
2888 {
2889         struct btree_transaction_stats *s;
2890         struct btree_trans *pos;
2891
2892         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
2893
2894         memset(trans, 0, sizeof(*trans));
2895         trans->c                = c;
2896         trans->fn               = fn_idx < ARRAY_SIZE(bch2_btree_transaction_fns)
2897                 ? bch2_btree_transaction_fns[fn_idx] : NULL;
2898         trans->last_begin_time  = local_clock();
2899         trans->fn_idx           = fn_idx;
2900         trans->locking_wait.task = current;
2901         trans->journal_replay_not_finished =
2902                 !test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags);
2903         closure_init_stack(&trans->ref);
2904
2905         bch2_trans_alloc_paths(trans, c);
2906
2907         s = btree_trans_stats(trans);
2908         if (s && s->max_mem) {
2909                 unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem);
2910
2911                 trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL);
2912
2913                 if (!unlikely(trans->mem)) {
2914                         trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL);
2915                         trans->mem_bytes = BTREE_TRANS_MEM_MAX;
2916                 } else {
2917                         trans->mem_bytes = expected_mem_bytes;
2918                 }
2919         }
2920         if (s)
2921                 trans->nr_max_paths = s->nr_max_paths;
2922
2923         trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
2924
2925         mutex_lock(&c->btree_trans_lock);
2926         list_for_each_entry(pos, &c->btree_trans_list, list) {
2927                 if (trans->locking_wait.task->pid < pos->locking_wait.task->pid) {
2928                         list_add_tail(&trans->list, &pos->list);
2929                         goto list_add_done;
2930                 }
2931         }
2932         list_add_tail(&trans->list, &c->btree_trans_list);
2933 list_add_done:
2934         mutex_unlock(&c->btree_trans_lock);
2935 }
2936
2937 static void check_btree_paths_leaked(struct btree_trans *trans)
2938 {
2939 #ifdef CONFIG_BCACHEFS_DEBUG
2940         struct bch_fs *c = trans->c;
2941         struct btree_path *path;
2942
2943         trans_for_each_path(trans, path)
2944                 if (path->ref)
2945                         goto leaked;
2946         return;
2947 leaked:
2948         bch_err(c, "btree paths leaked from %s!", trans->fn);
2949         trans_for_each_path(trans, path)
2950                 if (path->ref)
2951                         printk(KERN_ERR "  btree %s %pS\n",
2952                                bch2_btree_ids[path->btree_id],
2953                                (void *) path->ip_allocated);
2954         /* Be noisy about this: */
2955         bch2_fatal_error(c);
2956 #endif
2957 }
2958
2959 void bch2_trans_exit(struct btree_trans *trans)
2960         __releases(&c->btree_trans_barrier)
2961 {
2962         struct btree_insert_entry *i;
2963         struct bch_fs *c = trans->c;
2964         struct btree_transaction_stats *s = btree_trans_stats(trans);
2965
2966         bch2_trans_unlock(trans);
2967
2968         closure_sync(&trans->ref);
2969
2970         if (s)
2971                 s->max_mem = max(s->max_mem, trans->mem_max);
2972
2973         trans_for_each_update(trans, i)
2974                 __btree_path_put(i->path, true);
2975         trans->nr_updates               = 0;
2976
2977         check_btree_paths_leaked(trans);
2978
2979         mutex_lock(&c->btree_trans_lock);
2980         list_del(&trans->list);
2981         mutex_unlock(&c->btree_trans_lock);
2982
2983         srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx);
2984
2985         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
2986
2987         kfree(trans->extra_journal_entries.data);
2988
2989         if (trans->fs_usage_deltas) {
2990                 if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) ==
2991                     REPLICAS_DELTA_LIST_MAX)
2992                         mempool_free(trans->fs_usage_deltas,
2993                                      &c->replicas_delta_pool);
2994                 else
2995                         kfree(trans->fs_usage_deltas);
2996         }
2997
2998         if (trans->mem_bytes == BTREE_TRANS_MEM_MAX)
2999                 mempool_free(trans->mem, &c->btree_trans_mem_pool);
3000         else
3001                 kfree(trans->mem);
3002
3003 #ifdef __KERNEL__
3004         /*
3005          * Userspace doesn't have a real percpu implementation:
3006          */
3007         trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths);
3008 #endif
3009
3010         if (trans->paths)
3011                 mempool_free(trans->paths, &c->btree_paths_pool);
3012
3013         trans->mem      = (void *) 0x1;
3014         trans->paths    = (void *) 0x1;
3015 }
3016
3017 static void __maybe_unused
3018 bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
3019                                       struct btree_bkey_cached_common *b)
3020 {
3021         struct six_lock_count c = six_lock_counts(&b->lock);
3022         struct task_struct *owner;
3023         pid_t pid;
3024
3025         rcu_read_lock();
3026         owner = READ_ONCE(b->lock.owner);
3027         pid = owner ? owner->pid : 0;
3028         rcu_read_unlock();
3029
3030         prt_tab(out);
3031         prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
3032                    b->level, bch2_btree_ids[b->btree_id]);
3033         bch2_bpos_to_text(out, btree_node_pos(b));
3034
3035         prt_tab(out);
3036         prt_printf(out, " locks %u:%u:%u held by pid %u",
3037                    c.n[0], c.n[1], c.n[2], pid);
3038 }
3039
3040 void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
3041 {
3042         struct btree_path *path;
3043         struct btree_bkey_cached_common *b;
3044         static char lock_types[] = { 'r', 'i', 'w' };
3045         unsigned l;
3046
3047         if (!out->nr_tabstops) {
3048                 printbuf_tabstop_push(out, 16);
3049                 printbuf_tabstop_push(out, 32);
3050         }
3051
3052         prt_printf(out, "%i %s\n", trans->locking_wait.task->pid, trans->fn);
3053
3054         trans_for_each_path(trans, path) {
3055                 if (!path->nodes_locked)
3056                         continue;
3057
3058                 prt_printf(out, "  path %u %c l=%u %s:",
3059                        path->idx,
3060                        path->cached ? 'c' : 'b',
3061                        path->level,
3062                        bch2_btree_ids[path->btree_id]);
3063                 bch2_bpos_to_text(out, path->pos);
3064                 prt_newline(out);
3065
3066                 for (l = 0; l < BTREE_MAX_DEPTH; l++) {
3067                         if (btree_node_locked(path, l) &&
3068                             !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) {
3069                                 prt_printf(out, "    %c l=%u ",
3070                                            lock_types[btree_node_locked_type(path, l)], l);
3071                                 bch2_btree_bkey_cached_common_to_text(out, b);
3072                                 prt_newline(out);
3073                         }
3074                 }
3075         }
3076
3077         b = READ_ONCE(trans->locking);
3078         if (b) {
3079                 prt_str(out, "  want");
3080                 prt_newline(out);
3081                 prt_printf(out, "    %c", lock_types[trans->locking_wait.lock_want]);
3082                 bch2_btree_bkey_cached_common_to_text(out, b);
3083                 prt_newline(out);
3084         }
3085 }
3086
3087 void bch2_fs_btree_iter_exit(struct bch_fs *c)
3088 {
3089         struct btree_transaction_stats *s;
3090
3091         for (s = c->btree_transaction_stats;
3092              s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats);
3093              s++)
3094                 kfree(s->max_paths_text);
3095
3096         if (c->btree_trans_barrier_initialized)
3097                 cleanup_srcu_struct(&c->btree_trans_barrier);
3098         mempool_exit(&c->btree_trans_mem_pool);
3099         mempool_exit(&c->btree_paths_pool);
3100 }
3101
3102 int bch2_fs_btree_iter_init(struct bch_fs *c)
3103 {
3104         unsigned i, nr = BTREE_ITER_MAX;
3105         int ret;
3106
3107         for (i = 0; i < ARRAY_SIZE(c->btree_transaction_stats); i++)
3108                 mutex_init(&c->btree_transaction_stats[i].lock);
3109
3110         INIT_LIST_HEAD(&c->btree_trans_list);
3111         mutex_init(&c->btree_trans_lock);
3112
3113         ret   = mempool_init_kmalloc_pool(&c->btree_paths_pool, 1,
3114                         sizeof(struct btree_path) * nr +
3115                         sizeof(struct btree_insert_entry) * nr) ?:
3116                 mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1,
3117                                           BTREE_TRANS_MEM_MAX) ?:
3118                 init_srcu_struct(&c->btree_trans_barrier);
3119         if (!ret)
3120                 c->btree_trans_barrier_initialized = true;
3121         return ret;
3122 }