fs/bcachefs/fsck.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_buf.h"
   5 #include "btree_cache.h"
   6 #include "btree_update.h"
   7 #include "buckets.h"
   8 #include "darray.h"
   9 #include "dirent.h"
  10 #include "error.h"
  11 #include "fs-common.h"
  12 #include "fsck.h"
  13 #include "inode.h"
  14 #include "keylist.h"
  15 #include "recovery_passes.h"
  16 #include "snapshot.h"
  17 #include "super.h"
  18 #include "xattr.h"
  19
  20 #include <linux/bsearch.h>
  21 #include <linux/dcache.h> /* struct qstr */
  22
  23 /*
  24  * XXX: this is handling transaction restarts without returning
  25  * -BCH_ERR_transaction_restart_nested, this is not how we do things anymore:
  26  */
  27 static s64 bch2_count_inode_sectors(struct btree_trans *trans, u64 inum,
  28                                     u32 snapshot)
  29 {
  30         u64 sectors = 0;
  31
  32         int ret = for_each_btree_key_upto(trans, iter, BTREE_ID_extents,
  33                                 SPOS(inum, 0, snapshot),
  34                                 POS(inum, U64_MAX),
  35                                 0, k, ({
  36                 if (bkey_extent_is_allocation(k.k))
  37                         sectors += k.k->size;
  38                 0;
  39         }));
  40
  41         return ret ?: sectors;
  42 }
  43
  44 static s64 bch2_count_subdirs(struct btree_trans *trans, u64 inum,
  45                                     u32 snapshot)
  46 {
  47         u64 subdirs = 0;
  48
  49         int ret = for_each_btree_key_upto(trans, iter, BTREE_ID_dirents,
  50                                     SPOS(inum, 0, snapshot),
  51                                     POS(inum, U64_MAX),
  52                                     0, k, ({
  53                 if (k.k->type == KEY_TYPE_dirent &&
  54                     bkey_s_c_to_dirent(k).v->d_type == DT_DIR)
  55                         subdirs++;
  56                 0;
  57         }));
  58
  59         return ret ?: subdirs;
  60 }
  61
  62 static int subvol_lookup(struct btree_trans *trans, u32 subvol,
  63                          u32 *snapshot, u64 *inum)
  64 {
  65         struct bch_subvolume s;
  66         int ret = bch2_subvolume_get(trans, subvol, false, 0, &s);
  67
  68         *snapshot = le32_to_cpu(s.snapshot);
  69         *inum = le64_to_cpu(s.inode);
  70         return ret;
  71 }
  72
  73 static int lookup_first_inode(struct btree_trans *trans, u64 inode_nr,
  74                               struct bch_inode_unpacked *inode)
  75 {
  76         struct btree_iter iter;
  77         struct bkey_s_c k;
  78         int ret;
  79
  80         bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes,
  81                              POS(0, inode_nr),
  82                              BTREE_ITER_ALL_SNAPSHOTS);
  83         k = bch2_btree_iter_peek(&iter);
  84         ret = bkey_err(k);
  85         if (ret)
  86                 goto err;
  87
  88         if (!k.k || !bkey_eq(k.k->p, POS(0, inode_nr))) {
  89                 ret = -BCH_ERR_ENOENT_inode;
  90                 goto err;
  91         }
  92
  93         ret = bch2_inode_unpack(k, inode);
  94 err:
  95         bch_err_msg(trans->c, ret, "fetching inode %llu", inode_nr);
  96         bch2_trans_iter_exit(trans, &iter);
  97         return ret;
  98 }
  99
 100 static int lookup_inode(struct btree_trans *trans, u64 inode_nr,
 101                         struct bch_inode_unpacked *inode,
 102                         u32 *snapshot)
 103 {
 104         struct btree_iter iter;
 105         struct bkey_s_c k;
 106         int ret;
 107
 108         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
 109                                SPOS(0, inode_nr, *snapshot), 0);
 110         ret = bkey_err(k);
 111         if (ret)
 112                 goto err;
 113
 114         ret = bkey_is_inode(k.k)
 115                 ? bch2_inode_unpack(k, inode)
 116                 : -BCH_ERR_ENOENT_inode;
 117         if (!ret)
 118                 *snapshot = iter.pos.snapshot;
 119 err:
 120         bch2_trans_iter_exit(trans, &iter);
 121         return ret;
 122 }
 123
 124 static int lookup_dirent_in_snapshot(struct btree_trans *trans,
 125                            struct bch_hash_info hash_info,
 126                            subvol_inum dir, struct qstr *name,
 127                            u64 *target, unsigned *type, u32 snapshot)
 128 {
 129         struct btree_iter iter;
 130         struct bkey_s_c_dirent d;
 131         int ret = bch2_hash_lookup_in_snapshot(trans, &iter, bch2_dirent_hash_desc,
 132                                &hash_info, dir, name, 0, snapshot);
 133         if (ret)
 134                 return ret;
 135
 136         d = bkey_s_c_to_dirent(bch2_btree_iter_peek_slot(&iter));
 137         *target = le64_to_cpu(d.v->d_inum);
 138         *type = d.v->d_type;
 139         bch2_trans_iter_exit(trans, &iter);
 140         return 0;
 141 }
 142
 143 static int __remove_dirent(struct btree_trans *trans, struct bpos pos)
 144 {
 145         struct bch_fs *c = trans->c;
 146         struct btree_iter iter;
 147         struct bch_inode_unpacked dir_inode;
 148         struct bch_hash_info dir_hash_info;
 149         int ret;
 150
 151         ret = lookup_first_inode(trans, pos.inode, &dir_inode);
 152         if (ret)
 153                 goto err;
 154
 155         dir_hash_info = bch2_hash_info_init(c, &dir_inode);
 156
 157         bch2_trans_iter_init(trans, &iter, BTREE_ID_dirents, pos, BTREE_ITER_INTENT);
 158
 159         ret =   bch2_btree_iter_traverse(&iter) ?:
 160                 bch2_hash_delete_at(trans, bch2_dirent_hash_desc,
 161                                     &dir_hash_info, &iter,
 162                                     BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 163         bch2_trans_iter_exit(trans, &iter);
 164 err:
 165         bch_err_fn(c, ret);
 166         return ret;
 167 }
 168
 169 /* Get lost+found, create if it doesn't exist: */
 170 static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 171                             struct bch_inode_unpacked *lostfound,
 172                             u64 reattaching_inum)
 173 {
 174         struct bch_fs *c = trans->c;
 175         struct qstr lostfound_str = QSTR("lost+found");
 176         u64 inum = 0;
 177         unsigned d_type = 0;
 178         int ret;
 179
 180         struct bch_snapshot_tree st;
 181         ret = bch2_snapshot_tree_lookup(trans,
 182                         bch2_snapshot_tree(c, snapshot), &st);
 183         if (ret)
 184                 return ret;
 185
 186         subvol_inum root_inum = { .subvol = le32_to_cpu(st.master_subvol) };
 187
 188         struct bch_subvolume subvol;
 189         ret = bch2_subvolume_get(trans, le32_to_cpu(st.master_subvol),
 190                                  false, 0, &subvol);
 191         bch_err_msg(c, ret, "looking up root subvol %u for snapshot %u",
 192                     le32_to_cpu(st.master_subvol), snapshot);
 193         if (ret)
 194                 return ret;
 195
 196         if (!subvol.inode) {
 197                 struct btree_iter iter;
 198                 struct bkey_i_subvolume *subvol = bch2_bkey_get_mut_typed(trans, &iter,
 199                                 BTREE_ID_subvolumes, POS(0, le32_to_cpu(st.master_subvol)),
 200                                 0, subvolume);
 201                 ret = PTR_ERR_OR_ZERO(subvol);
 202                 if (ret)
 203                         return ret;
 204
 205                 subvol->v.inode = cpu_to_le64(reattaching_inum);
 206                 bch2_trans_iter_exit(trans, &iter);
 207         }
 208
 209         root_inum.inum = le64_to_cpu(subvol.inode);
 210
 211         struct bch_inode_unpacked root_inode;
 212         struct bch_hash_info root_hash_info;
 213         u32 root_inode_snapshot = snapshot;
 214         ret = lookup_inode(trans, root_inum.inum, &root_inode, &root_inode_snapshot);
 215         bch_err_msg(c, ret, "looking up root inode %llu for subvol %u",
 216                     root_inum.inum, le32_to_cpu(st.master_subvol));
 217         if (ret)
 218                 return ret;
 219
 220         root_hash_info = bch2_hash_info_init(c, &root_inode);
 221
 222         ret = lookup_dirent_in_snapshot(trans, root_hash_info, root_inum,
 223                               &lostfound_str, &inum, &d_type, snapshot);
 224         if (bch2_err_matches(ret, ENOENT))
 225                 goto create_lostfound;
 226
 227         bch_err_fn(c, ret);
 228         if (ret)
 229                 return ret;
 230
 231         if (d_type != DT_DIR) {
 232                 bch_err(c, "error looking up lost+found: not a directory");
 233                 return -BCH_ERR_ENOENT_not_directory;
 234         }
 235
 236         /*
 237          * The bch2_check_dirents pass has already run, dangling dirents
 238          * shouldn't exist here:
 239          */
 240         ret = lookup_inode(trans, inum, lostfound, &snapshot);
 241         bch_err_msg(c, ret, "looking up lost+found %llu:%u in (root inode %llu, snapshot root %u)",
 242                     inum, snapshot, root_inum.inum, bch2_snapshot_root(c, snapshot));
 243         return ret;
 244
 245 create_lostfound:
 246         /*
 247          * XXX: we could have a nicer log message here  if we had a nice way to
 248          * walk backpointers to print a path
 249          */
 250         bch_notice(c, "creating lost+found in snapshot %u", le32_to_cpu(st.root_snapshot));
 251
 252         u64 now = bch2_current_time(c);
 253         struct btree_iter lostfound_iter = { NULL };
 254         u64 cpu = raw_smp_processor_id();
 255
 256         bch2_inode_init_early(c, lostfound);
 257         bch2_inode_init_late(lostfound, now, 0, 0, S_IFDIR|0700, 0, &root_inode);
 258         lostfound->bi_dir = root_inode.bi_inum;
 259
 260         root_inode.bi_nlink++;
 261
 262         ret = bch2_inode_create(trans, &lostfound_iter, lostfound, snapshot, cpu);
 263         if (ret)
 264                 goto err;
 265
 266         bch2_btree_iter_set_snapshot(&lostfound_iter, snapshot);
 267         ret = bch2_btree_iter_traverse(&lostfound_iter);
 268         if (ret)
 269                 goto err;
 270
 271         ret =   bch2_dirent_create_snapshot(trans,
 272                                 0, root_inode.bi_inum, snapshot, &root_hash_info,
 273                                 mode_to_type(lostfound->bi_mode),
 274                                 &lostfound_str,
 275                                 lostfound->bi_inum,
 276                                 &lostfound->bi_dir_offset,
 277                                 BCH_HASH_SET_MUST_CREATE) ?:
 278                 bch2_inode_write_flags(trans, &lostfound_iter, lostfound,
 279                                        BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 280 err:
 281         bch_err_msg(c, ret, "creating lost+found");
 282         bch2_trans_iter_exit(trans, &lostfound_iter);
 283         return ret;
 284 }
 285
 286 static int reattach_inode(struct btree_trans *trans,
 287                           struct bch_inode_unpacked *inode,
 288                           u32 inode_snapshot)
 289 {
 290         struct bch_hash_info dir_hash;
 291         struct bch_inode_unpacked lostfound;
 292         char name_buf[20];
 293         struct qstr name;
 294         u64 dir_offset = 0;
 295         u32 dirent_snapshot = inode_snapshot;
 296         int ret;
 297
 298         if (inode->bi_subvol) {
 299                 inode->bi_parent_subvol = BCACHEFS_ROOT_SUBVOL;
 300
 301                 u64 root_inum;
 302                 ret = subvol_lookup(trans, inode->bi_parent_subvol,
 303                                     &dirent_snapshot, &root_inum);
 304                 if (ret)
 305                         return ret;
 306
 307                 snprintf(name_buf, sizeof(name_buf), "subvol-%u", inode->bi_subvol);
 308         } else {
 309                 snprintf(name_buf, sizeof(name_buf), "%llu", inode->bi_inum);
 310         }
 311
 312         ret = lookup_lostfound(trans, dirent_snapshot, &lostfound, inode->bi_inum);
 313         if (ret)
 314                 return ret;
 315
 316         if (S_ISDIR(inode->bi_mode)) {
 317                 lostfound.bi_nlink++;
 318
 319                 ret = __bch2_fsck_write_inode(trans, &lostfound, U32_MAX);
 320                 if (ret)
 321                         return ret;
 322         }
 323
 324         dir_hash = bch2_hash_info_init(trans->c, &lostfound);
 325
 326         name = (struct qstr) QSTR(name_buf);
 327
 328         ret = bch2_dirent_create_snapshot(trans,
 329                                 inode->bi_parent_subvol, lostfound.bi_inum,
 330                                 dirent_snapshot,
 331                                 &dir_hash,
 332                                 inode_d_type(inode),
 333                                 &name,
 334                                 inode->bi_subvol ?: inode->bi_inum,
 335                                 &dir_offset,
 336                                 BCH_HASH_SET_MUST_CREATE);
 337         if (ret)
 338                 return ret;
 339
 340         inode->bi_dir           = lostfound.bi_inum;
 341         inode->bi_dir_offset    = dir_offset;
 342
 343         return __bch2_fsck_write_inode(trans, inode, inode_snapshot);
 344 }
 345
 346 static int remove_backpointer(struct btree_trans *trans,
 347                               struct bch_inode_unpacked *inode)
 348 {
 349         struct btree_iter iter;
 350         struct bkey_s_c_dirent d;
 351         int ret;
 352
 353         d = bch2_bkey_get_iter_typed(trans, &iter, BTREE_ID_dirents,
 354                                      POS(inode->bi_dir, inode->bi_dir_offset), 0,
 355                                      dirent);
 356         ret =   bkey_err(d) ?:
 357                 __remove_dirent(trans, d.k->p);
 358         bch2_trans_iter_exit(trans, &iter);
 359         return ret;
 360 }
 361
 362 static int reattach_subvol(struct btree_trans *trans, struct bkey_s_c_subvolume s)
 363 {
 364         struct bch_fs *c = trans->c;
 365
 366         struct bch_inode_unpacked inode;
 367         int ret = bch2_inode_find_by_inum_trans(trans,
 368                                 (subvol_inum) { s.k->p.offset, le64_to_cpu(s.v->inode) },
 369                                 &inode);
 370         if (ret)
 371                 return ret;
 372
 373         ret = remove_backpointer(trans, &inode);
 374         bch_err_msg(c, ret, "removing dirent");
 375         if (ret)
 376                 return ret;
 377
 378         ret = reattach_inode(trans, &inode, le32_to_cpu(s.v->snapshot));
 379         bch_err_msg(c, ret, "reattaching inode %llu", inode.bi_inum);
 380         return ret;
 381 }
 382
 383 static int reconstruct_subvol(struct btree_trans *trans, u32 snapshotid, u32 subvolid, u64 inum)
 384 {
 385         struct bch_fs *c = trans->c;
 386
 387         if (!bch2_snapshot_is_leaf(c, snapshotid)) {
 388                 bch_err(c, "need to reconstruct subvol, but have interior node snapshot");
 389                 return -BCH_ERR_fsck_repair_unimplemented;
 390         }
 391
 392         /*
 393          * If inum isn't set, that means we're being called from check_dirents,
 394          * not check_inodes - the root of this subvolume doesn't exist or we
 395          * would have found it there:
 396          */
 397         if (!inum) {
 398                 struct btree_iter inode_iter = {};
 399                 struct bch_inode_unpacked new_inode;
 400                 u64 cpu = raw_smp_processor_id();
 401
 402                 bch2_inode_init_early(c, &new_inode);
 403                 bch2_inode_init_late(&new_inode, bch2_current_time(c), 0, 0, S_IFDIR|0755, 0, NULL);
 404
 405                 new_inode.bi_subvol = subvolid;
 406
 407                 int ret = bch2_inode_create(trans, &inode_iter, &new_inode, snapshotid, cpu) ?:
 408                           bch2_btree_iter_traverse(&inode_iter) ?:
 409                           bch2_inode_write(trans, &inode_iter, &new_inode);
 410                 bch2_trans_iter_exit(trans, &inode_iter);
 411                 if (ret)
 412                         return ret;
 413
 414                 inum = new_inode.bi_inum;
 415         }
 416
 417         bch_info(c, "reconstructing subvol %u with root inode %llu", subvolid, inum);
 418
 419         struct bkey_i_subvolume *new_subvol = bch2_trans_kmalloc(trans, sizeof(*new_subvol));
 420         int ret = PTR_ERR_OR_ZERO(new_subvol);
 421         if (ret)
 422                 return ret;
 423
 424         bkey_subvolume_init(&new_subvol->k_i);
 425         new_subvol->k.p.offset  = subvolid;
 426         new_subvol->v.snapshot  = cpu_to_le32(snapshotid);
 427         new_subvol->v.inode     = cpu_to_le64(inum);
 428         ret = bch2_btree_insert_trans(trans, BTREE_ID_subvolumes, &new_subvol->k_i, 0);
 429         if (ret)
 430                 return ret;
 431
 432         struct btree_iter iter;
 433         struct bkey_i_snapshot *s = bch2_bkey_get_mut_typed(trans, &iter,
 434                         BTREE_ID_snapshots, POS(0, snapshotid),
 435                         0, snapshot);
 436         ret = PTR_ERR_OR_ZERO(s);
 437         bch_err_msg(c, ret, "getting snapshot %u", snapshotid);
 438         if (ret)
 439                 return ret;
 440
 441         u32 snapshot_tree = le32_to_cpu(s->v.tree);
 442
 443         s->v.subvol = cpu_to_le32(subvolid);
 444         SET_BCH_SNAPSHOT_SUBVOL(&s->v, true);
 445         bch2_trans_iter_exit(trans, &iter);
 446
 447         struct bkey_i_snapshot_tree *st = bch2_bkey_get_mut_typed(trans, &iter,
 448                         BTREE_ID_snapshot_trees, POS(0, snapshot_tree),
 449                         0, snapshot_tree);
 450         ret = PTR_ERR_OR_ZERO(st);
 451         bch_err_msg(c, ret, "getting snapshot tree %u", snapshot_tree);
 452         if (ret)
 453                 return ret;
 454
 455         if (!st->v.master_subvol)
 456                 st->v.master_subvol = cpu_to_le32(subvolid);
 457
 458         bch2_trans_iter_exit(trans, &iter);
 459         return 0;
 460 }
 461
 462 static int reconstruct_inode(struct btree_trans *trans, u32 snapshot, u64 inum, u64 size, unsigned mode)
 463 {
 464         struct bch_fs *c = trans->c;
 465         struct bch_inode_unpacked new_inode;
 466
 467         bch2_inode_init_early(c, &new_inode);
 468         bch2_inode_init_late(&new_inode, bch2_current_time(c), 0, 0, mode|0755, 0, NULL);
 469         new_inode.bi_size = size;
 470         new_inode.bi_inum = inum;
 471
 472         return __bch2_fsck_write_inode(trans, &new_inode, snapshot);
 473 }
 474
 475 static int reconstruct_reg_inode(struct btree_trans *trans, u32 snapshot, u64 inum)
 476 {
 477         struct btree_iter iter = {};
 478
 479         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents, SPOS(inum, U64_MAX, snapshot), 0);
 480         struct bkey_s_c k = bch2_btree_iter_peek_prev(&iter);
 481         bch2_trans_iter_exit(trans, &iter);
 482         int ret = bkey_err(k);
 483         if (ret)
 484                 return ret;
 485
 486         return reconstruct_inode(trans, snapshot, inum, k.k->p.offset << 9, S_IFREG);
 487 }
 488
 489 struct snapshots_seen_entry {
 490         u32                             id;
 491         u32                             equiv;
 492 };
 493
 494 struct snapshots_seen {
 495         struct bpos                     pos;
 496         DARRAY(struct snapshots_seen_entry) ids;
 497 };
 498
 499 static inline void snapshots_seen_exit(struct snapshots_seen *s)
 500 {
 501         darray_exit(&s->ids);
 502 }
 503
 504 static inline void snapshots_seen_init(struct snapshots_seen *s)
 505 {
 506         memset(s, 0, sizeof(*s));
 507 }
 508
 509 static int snapshots_seen_add_inorder(struct bch_fs *c, struct snapshots_seen *s, u32 id)
 510 {
 511         struct snapshots_seen_entry *i, n = {
 512                 .id     = id,
 513                 .equiv  = bch2_snapshot_equiv(c, id),
 514         };
 515         int ret = 0;
 516
 517         __darray_for_each(s->ids, i) {
 518                 if (i->id == id)
 519                         return 0;
 520                 if (i->id > id)
 521                         break;
 522         }
 523
 524         ret = darray_insert_item(&s->ids, i - s->ids.data, n);
 525         if (ret)
 526                 bch_err(c, "error reallocating snapshots_seen table (size %zu)",
 527                         s->ids.size);
 528         return ret;
 529 }
 530
 531 static int snapshots_seen_update(struct bch_fs *c, struct snapshots_seen *s,
 532                                  enum btree_id btree_id, struct bpos pos)
 533 {
 534         struct snapshots_seen_entry n = {
 535                 .id     = pos.snapshot,
 536                 .equiv  = bch2_snapshot_equiv(c, pos.snapshot),
 537         };
 538         int ret = 0;
 539
 540         if (!bkey_eq(s->pos, pos))
 541                 s->ids.nr = 0;
 542
 543         s->pos = pos;
 544         s->pos.snapshot = n.equiv;
 545
 546         darray_for_each(s->ids, i) {
 547                 if (i->id == n.id)
 548                         return 0;
 549
 550                 /*
 551                  * We currently don't rigorously track for snapshot cleanup
 552                  * needing to be run, so it shouldn't be a fsck error yet:
 553                  */
 554                 if (i->equiv == n.equiv) {
 555                         bch_err(c, "snapshot deletion did not finish:\n"
 556                                 "  duplicate keys in btree %s at %llu:%llu snapshots %u, %u (equiv %u)\n",
 557                                 bch2_btree_id_str(btree_id),
 558                                 pos.inode, pos.offset,
 559                                 i->id, n.id, n.equiv);
 560                         set_bit(BCH_FS_need_delete_dead_snapshots, &c->flags);
 561                         return bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_delete_dead_snapshots);
 562                 }
 563         }
 564
 565         ret = darray_push(&s->ids, n);
 566         if (ret)
 567                 bch_err(c, "error reallocating snapshots_seen table (size %zu)",
 568                         s->ids.size);
 569         return ret;
 570 }
 571
 572 /**
 573  * key_visible_in_snapshot - returns true if @id is a descendent of @ancestor,
 574  * and @ancestor hasn't been overwritten in @seen
 575  *
 576  * @c:          filesystem handle
 577  * @seen:       list of snapshot ids already seen at current position
 578  * @id:         descendent snapshot id
 579  * @ancestor:   ancestor snapshot id
 580  *
 581  * Returns:     whether key in @ancestor snapshot is visible in @id snapshot
 582  */
 583 static bool key_visible_in_snapshot(struct bch_fs *c, struct snapshots_seen *seen,
 584                                     u32 id, u32 ancestor)
 585 {
 586         ssize_t i;
 587
 588         EBUG_ON(id > ancestor);
 589         EBUG_ON(!bch2_snapshot_is_equiv(c, id));
 590         EBUG_ON(!bch2_snapshot_is_equiv(c, ancestor));
 591
 592         /* @ancestor should be the snapshot most recently added to @seen */
 593         EBUG_ON(ancestor != seen->pos.snapshot);
 594         EBUG_ON(ancestor != seen->ids.data[seen->ids.nr - 1].equiv);
 595
 596         if (id == ancestor)
 597                 return true;
 598
 599         if (!bch2_snapshot_is_ancestor(c, id, ancestor))
 600                 return false;
 601
 602         /*
 603          * We know that @id is a descendant of @ancestor, we're checking if
 604          * we've seen a key that overwrote @ancestor - i.e. also a descendent of
 605          * @ascestor and with @id as a descendent.
 606          *
 607          * But we already know that we're scanning IDs between @id and @ancestor
 608          * numerically, since snapshot ID lists are kept sorted, so if we find
 609          * an id that's an ancestor of @id we're done:
 610          */
 611
 612         for (i = seen->ids.nr - 2;
 613              i >= 0 && seen->ids.data[i].equiv >= id;
 614              --i)
 615                 if (bch2_snapshot_is_ancestor(c, id, seen->ids.data[i].equiv))
 616                         return false;
 617
 618         return true;
 619 }
 620
 621 /**
 622  * ref_visible - given a key with snapshot id @src that points to a key with
 623  * snapshot id @dst, test whether there is some snapshot in which @dst is
 624  * visible.
 625  *
 626  * @c:          filesystem handle
 627  * @s:          list of snapshot IDs already seen at @src
 628  * @src:        snapshot ID of src key
 629  * @dst:        snapshot ID of dst key
 630  * Returns:     true if there is some snapshot in which @dst is visible
 631  *
 632  * Assumes we're visiting @src keys in natural key order
 633  */
 634 static bool ref_visible(struct bch_fs *c, struct snapshots_seen *s,
 635                         u32 src, u32 dst)
 636 {
 637         return dst <= src
 638                 ? key_visible_in_snapshot(c, s, dst, src)
 639                 : bch2_snapshot_is_ancestor(c, src, dst);
 640 }
 641
 642 static int ref_visible2(struct bch_fs *c,
 643                         u32 src, struct snapshots_seen *src_seen,
 644                         u32 dst, struct snapshots_seen *dst_seen)
 645 {
 646         src = bch2_snapshot_equiv(c, src);
 647         dst = bch2_snapshot_equiv(c, dst);
 648
 649         if (dst > src) {
 650                 swap(dst, src);
 651                 swap(dst_seen, src_seen);
 652         }
 653         return key_visible_in_snapshot(c, src_seen, dst, src);
 654 }
 655
 656 #define for_each_visible_inode(_c, _s, _w, _snapshot, _i)                               \
 657         for (_i = (_w)->inodes.data; _i < (_w)->inodes.data + (_w)->inodes.nr &&        \
 658              (_i)->snapshot <= (_snapshot); _i++)                                       \
 659                 if (key_visible_in_snapshot(_c, _s, _i->snapshot, _snapshot))
 660
 661 struct inode_walker_entry {
 662         struct bch_inode_unpacked inode;
 663         u32                     snapshot;
 664         bool                    seen_this_pos;
 665         u64                     count;
 666 };
 667
 668 struct inode_walker {
 669         bool                            first_this_inode;
 670         bool                            recalculate_sums;
 671         struct bpos                     last_pos;
 672
 673         DARRAY(struct inode_walker_entry) inodes;
 674 };
 675
 676 static void inode_walker_exit(struct inode_walker *w)
 677 {
 678         darray_exit(&w->inodes);
 679 }
 680
 681 static struct inode_walker inode_walker_init(void)
 682 {
 683         return (struct inode_walker) { 0, };
 684 }
 685
 686 static int add_inode(struct bch_fs *c, struct inode_walker *w,
 687                      struct bkey_s_c inode)
 688 {
 689         struct bch_inode_unpacked u;
 690
 691         BUG_ON(bch2_inode_unpack(inode, &u));
 692
 693         return darray_push(&w->inodes, ((struct inode_walker_entry) {
 694                 .inode          = u,
 695                 .snapshot       = bch2_snapshot_equiv(c, inode.k->p.snapshot),
 696         }));
 697 }
 698
 699 static int get_inodes_all_snapshots(struct btree_trans *trans,
 700                                     struct inode_walker *w, u64 inum)
 701 {
 702         struct bch_fs *c = trans->c;
 703         struct btree_iter iter;
 704         struct bkey_s_c k;
 705         int ret;
 706
 707         w->recalculate_sums = false;
 708         w->inodes.nr = 0;
 709
 710         for_each_btree_key_norestart(trans, iter, BTREE_ID_inodes, POS(0, inum),
 711                                      BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
 712                 if (k.k->p.offset != inum)
 713                         break;
 714
 715                 if (bkey_is_inode(k.k))
 716                         add_inode(c, w, k);
 717         }
 718         bch2_trans_iter_exit(trans, &iter);
 719
 720         if (ret)
 721                 return ret;
 722
 723         w->first_this_inode = true;
 724         return 0;
 725 }
 726
 727 static struct inode_walker_entry *
 728 lookup_inode_for_snapshot(struct bch_fs *c, struct inode_walker *w, struct bkey_s_c k)
 729 {
 730         bool is_whiteout = k.k->type == KEY_TYPE_whiteout;
 731         u32 snapshot = bch2_snapshot_equiv(c, k.k->p.snapshot);
 732
 733         struct inode_walker_entry *i;
 734         __darray_for_each(w->inodes, i)
 735                 if (bch2_snapshot_is_ancestor(c, snapshot, i->snapshot))
 736                         goto found;
 737
 738         return NULL;
 739 found:
 740         BUG_ON(snapshot > i->snapshot);
 741
 742         if (snapshot != i->snapshot && !is_whiteout) {
 743                 struct inode_walker_entry new = *i;
 744
 745                 new.snapshot = snapshot;
 746                 new.count = 0;
 747
 748                 struct printbuf buf = PRINTBUF;
 749                 bch2_bkey_val_to_text(&buf, c, k);
 750
 751                 bch_info(c, "have key for inode %llu:%u but have inode in ancestor snapshot %u\n"
 752                          "unexpected because we should always update the inode when we update a key in that inode\n"
 753                          "%s",
 754                          w->last_pos.inode, snapshot, i->snapshot, buf.buf);
 755                 printbuf_exit(&buf);
 756
 757                 while (i > w->inodes.data && i[-1].snapshot > snapshot)
 758                         --i;
 759
 760                 size_t pos = i - w->inodes.data;
 761                 int ret = darray_insert_item(&w->inodes, pos, new);
 762                 if (ret)
 763                         return ERR_PTR(ret);
 764
 765                 i = w->inodes.data + pos;
 766         }
 767
 768         return i;
 769 }
 770
 771 static struct inode_walker_entry *walk_inode(struct btree_trans *trans,
 772                                              struct inode_walker *w,
 773                                              struct bkey_s_c k)
 774 {
 775         if (w->last_pos.inode != k.k->p.inode) {
 776                 int ret = get_inodes_all_snapshots(trans, w, k.k->p.inode);
 777                 if (ret)
 778                         return ERR_PTR(ret);
 779         } else if (bkey_cmp(w->last_pos, k.k->p)) {
 780                 darray_for_each(w->inodes, i)
 781                         i->seen_this_pos = false;
 782         }
 783
 784         w->last_pos = k.k->p;
 785
 786         return lookup_inode_for_snapshot(trans->c, w, k);
 787 }
 788
 789 static int __get_visible_inodes(struct btree_trans *trans,
 790                                 struct inode_walker *w,
 791                                 struct snapshots_seen *s,
 792                                 u64 inum)
 793 {
 794         struct bch_fs *c = trans->c;
 795         struct btree_iter iter;
 796         struct bkey_s_c k;
 797         int ret;
 798
 799         w->inodes.nr = 0;
 800
 801         for_each_btree_key_norestart(trans, iter, BTREE_ID_inodes, POS(0, inum),
 802                            BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
 803                 u32 equiv = bch2_snapshot_equiv(c, k.k->p.snapshot);
 804
 805                 if (k.k->p.offset != inum)
 806                         break;
 807
 808                 if (!ref_visible(c, s, s->pos.snapshot, equiv))
 809                         continue;
 810
 811                 if (bkey_is_inode(k.k))
 812                         add_inode(c, w, k);
 813
 814                 if (equiv >= s->pos.snapshot)
 815                         break;
 816         }
 817         bch2_trans_iter_exit(trans, &iter);
 818
 819         return ret;
 820 }
 821
 822 static int check_key_has_snapshot(struct btree_trans *trans,
 823                                   struct btree_iter *iter,
 824                                   struct bkey_s_c k)
 825 {
 826         struct bch_fs *c = trans->c;
 827         struct printbuf buf = PRINTBUF;
 828         int ret = 0;
 829
 830         if (mustfix_fsck_err_on(!bch2_snapshot_equiv(c, k.k->p.snapshot), c,
 831                                 bkey_in_missing_snapshot,
 832                                 "key in missing snapshot: %s",
 833                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
 834                 ret = bch2_btree_delete_at(trans, iter,
 835                                             BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?: 1;
 836 fsck_err:
 837         printbuf_exit(&buf);
 838         return ret;
 839 }
 840
 841 static int hash_redo_key(struct btree_trans *trans,
 842                          const struct bch_hash_desc desc,
 843                          struct bch_hash_info *hash_info,
 844                          struct btree_iter *k_iter, struct bkey_s_c k)
 845 {
 846         struct bkey_i *delete;
 847         struct bkey_i *tmp;
 848
 849         delete = bch2_trans_kmalloc(trans, sizeof(*delete));
 850         if (IS_ERR(delete))
 851                 return PTR_ERR(delete);
 852
 853         tmp = bch2_bkey_make_mut_noupdate(trans, k);
 854         if (IS_ERR(tmp))
 855                 return PTR_ERR(tmp);
 856
 857         bkey_init(&delete->k);
 858         delete->k.p = k_iter->pos;
 859         return  bch2_btree_iter_traverse(k_iter) ?:
 860                 bch2_trans_update(trans, k_iter, delete, 0) ?:
 861                 bch2_hash_set_in_snapshot(trans, desc, hash_info,
 862                                        (subvol_inum) { 0, k.k->p.inode },
 863                                        k.k->p.snapshot, tmp,
 864                                        BCH_HASH_SET_MUST_CREATE,
 865                                        BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 866                 bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
 867 }
 868
 869 static int hash_check_key(struct btree_trans *trans,
 870                           const struct bch_hash_desc desc,
 871                           struct bch_hash_info *hash_info,
 872                           struct btree_iter *k_iter, struct bkey_s_c hash_k)
 873 {
 874         struct bch_fs *c = trans->c;
 875         struct btree_iter iter = { NULL };
 876         struct printbuf buf = PRINTBUF;
 877         struct bkey_s_c k;
 878         u64 hash;
 879         int ret = 0;
 880
 881         if (hash_k.k->type != desc.key_type)
 882                 return 0;
 883
 884         hash = desc.hash_bkey(hash_info, hash_k);
 885
 886         if (likely(hash == hash_k.k->p.offset))
 887                 return 0;
 888
 889         if (hash_k.k->p.offset < hash)
 890                 goto bad_hash;
 891
 892         for_each_btree_key_norestart(trans, iter, desc.btree_id,
 893                                      SPOS(hash_k.k->p.inode, hash, hash_k.k->p.snapshot),
 894                                      BTREE_ITER_SLOTS, k, ret) {
 895                 if (bkey_eq(k.k->p, hash_k.k->p))
 896                         break;
 897
 898                 if (fsck_err_on(k.k->type == desc.key_type &&
 899                                 !desc.cmp_bkey(k, hash_k), c,
 900                                 hash_table_key_duplicate,
 901                                 "duplicate hash table keys:\n%s",
 902                                 (printbuf_reset(&buf),
 903                                  bch2_bkey_val_to_text(&buf, c, hash_k),
 904                                  buf.buf))) {
 905                         ret = bch2_hash_delete_at(trans, desc, hash_info, k_iter, 0) ?: 1;
 906                         break;
 907                 }
 908
 909                 if (bkey_deleted(k.k)) {
 910                         bch2_trans_iter_exit(trans, &iter);
 911                         goto bad_hash;
 912                 }
 913         }
 914 out:
 915         bch2_trans_iter_exit(trans, &iter);
 916         printbuf_exit(&buf);
 917         return ret;
 918 bad_hash:
 919         if (fsck_err(c, hash_table_key_wrong_offset,
 920                      "hash table key at wrong offset: btree %s inode %llu offset %llu, hashed to %llu\n%s",
 921                      bch2_btree_id_str(desc.btree_id), hash_k.k->p.inode, hash_k.k->p.offset, hash,
 922                      (printbuf_reset(&buf),
 923                       bch2_bkey_val_to_text(&buf, c, hash_k), buf.buf))) {
 924                 ret = hash_redo_key(trans, desc, hash_info, k_iter, hash_k);
 925                 bch_err_fn(c, ret);
 926                 if (ret)
 927                         return ret;
 928                 ret = -BCH_ERR_transaction_restart_nested;
 929         }
 930 fsck_err:
 931         goto out;
 932 }
 933
 934 static struct bkey_s_c_dirent dirent_get_by_pos(struct btree_trans *trans,
 935                                                 struct btree_iter *iter,
 936                                                 struct bpos pos)
 937 {
 938         return bch2_bkey_get_iter_typed(trans, iter, BTREE_ID_dirents, pos, 0, dirent);
 939 }
 940
 941 static struct bkey_s_c_dirent inode_get_dirent(struct btree_trans *trans,
 942                                                struct btree_iter *iter,
 943                                                struct bch_inode_unpacked *inode,
 944                                                u32 *snapshot)
 945 {
 946         if (inode->bi_subvol) {
 947                 u64 inum;
 948                 int ret = subvol_lookup(trans, inode->bi_parent_subvol, snapshot, &inum);
 949                 if (ret)
 950                         return ((struct bkey_s_c_dirent) { .k = ERR_PTR(ret) });
 951         }
 952
 953         return dirent_get_by_pos(trans, iter, SPOS(inode->bi_dir, inode->bi_dir_offset, *snapshot));
 954 }
 955
 956 static bool inode_points_to_dirent(struct bch_inode_unpacked *inode,
 957                                    struct bkey_s_c_dirent d)
 958 {
 959         return  inode->bi_dir           == d.k->p.inode &&
 960                 inode->bi_dir_offset    == d.k->p.offset;
 961 }
 962
 963 static bool dirent_points_to_inode(struct bkey_s_c_dirent d,
 964                                    struct bch_inode_unpacked *inode)
 965 {
 966         return d.v->d_type == DT_SUBVOL
 967                 ? le32_to_cpu(d.v->d_child_subvol)      == inode->bi_subvol
 968                 : le64_to_cpu(d.v->d_inum)              == inode->bi_inum;
 969 }
 970
 971 static int check_inode_deleted_list(struct btree_trans *trans, struct bpos p)
 972 {
 973         struct btree_iter iter;
 974         struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_deleted_inodes, p, 0);
 975         int ret = bkey_err(k) ?: k.k->type == KEY_TYPE_set;
 976         bch2_trans_iter_exit(trans, &iter);
 977         return ret;
 978 }
 979
 980 static int check_inode_dirent_inode(struct btree_trans *trans, struct bkey_s_c inode_k,
 981                                     struct bch_inode_unpacked *inode,
 982                                     u32 inode_snapshot, bool *write_inode)
 983 {
 984         struct bch_fs *c = trans->c;
 985         struct printbuf buf = PRINTBUF;
 986
 987         struct btree_iter dirent_iter = {};
 988         struct bkey_s_c_dirent d = inode_get_dirent(trans, &dirent_iter, inode, &inode_snapshot);
 989         int ret = bkey_err(d);
 990         if (ret && !bch2_err_matches(ret, ENOENT))
 991                 return ret;
 992
 993         if (fsck_err_on(ret,
 994                         c, inode_points_to_missing_dirent,
 995                         "inode points to missing dirent\n%s",
 996                         (bch2_bkey_val_to_text(&buf, c, inode_k), buf.buf)) ||
 997             fsck_err_on(!ret && !dirent_points_to_inode(d, inode),
 998                         c, inode_points_to_wrong_dirent,
 999                         "inode points to dirent that does not point back:\n%s",
1000                         (bch2_bkey_val_to_text(&buf, c, inode_k),
1001                          prt_newline(&buf),
1002                          bch2_bkey_val_to_text(&buf, c, d.s_c), buf.buf))) {
1003                 /*
1004                  * We just clear the backpointer fields for now. If we find a
1005                  * dirent that points to this inode in check_dirents(), we'll
1006                  * update it then; then when we get to check_path() if the
1007                  * backpointer is still 0 we'll reattach it.
1008                  */
1009                 inode->bi_dir = 0;
1010                 inode->bi_dir_offset = 0;
1011                 inode->bi_flags &= ~BCH_INODE_backptr_untrusted;
1012                 *write_inode = true;
1013         }
1014
1015         ret = 0;
1016 fsck_err:
1017         bch2_trans_iter_exit(trans, &dirent_iter);
1018         printbuf_exit(&buf);
1019         bch_err_fn(c, ret);
1020         return ret;
1021 }
1022
1023 static int check_inode(struct btree_trans *trans,
1024                        struct btree_iter *iter,
1025                        struct bkey_s_c k,
1026                        struct bch_inode_unpacked *prev,
1027                        struct snapshots_seen *s,
1028                        bool full)
1029 {
1030         struct bch_fs *c = trans->c;
1031         struct bch_inode_unpacked u;
1032         bool do_update = false;
1033         int ret;
1034
1035         ret = check_key_has_snapshot(trans, iter, k);
1036         if (ret < 0)
1037                 goto err;
1038         if (ret)
1039                 return 0;
1040
1041         ret = snapshots_seen_update(c, s, iter->btree_id, k.k->p);
1042         if (ret)
1043                 goto err;
1044
1045         if (!bkey_is_inode(k.k))
1046                 return 0;
1047
1048         BUG_ON(bch2_inode_unpack(k, &u));
1049
1050         if (!full &&
1051             !(u.bi_flags & (BCH_INODE_i_size_dirty|
1052                             BCH_INODE_i_sectors_dirty|
1053                             BCH_INODE_unlinked)))
1054                 return 0;
1055
1056         if (prev->bi_inum != u.bi_inum)
1057                 *prev = u;
1058
1059         if (fsck_err_on(prev->bi_hash_seed      != u.bi_hash_seed ||
1060                         inode_d_type(prev)      != inode_d_type(&u),
1061                         c, inode_snapshot_mismatch,
1062                         "inodes in different snapshots don't match")) {
1063                 bch_err(c, "repair not implemented yet");
1064                 return -BCH_ERR_fsck_repair_unimplemented;
1065         }
1066
1067         if ((u.bi_flags & (BCH_INODE_i_size_dirty|BCH_INODE_unlinked)) &&
1068             bch2_key_has_snapshot_overwrites(trans, BTREE_ID_inodes, k.k->p)) {
1069                 struct bpos new_min_pos;
1070
1071                 ret = bch2_propagate_key_to_snapshot_leaves(trans, iter->btree_id, k, &new_min_pos);
1072                 if (ret)
1073                         goto err;
1074
1075                 u.bi_flags &= ~BCH_INODE_i_size_dirty|BCH_INODE_unlinked;
1076
1077                 ret = __bch2_fsck_write_inode(trans, &u, iter->pos.snapshot);
1078
1079                 bch_err_msg(c, ret, "in fsck updating inode");
1080                 if (ret)
1081                         return ret;
1082
1083                 if (!bpos_eq(new_min_pos, POS_MIN))
1084                         bch2_btree_iter_set_pos(iter, bpos_predecessor(new_min_pos));
1085                 return 0;
1086         }
1087
1088         if (u.bi_flags & BCH_INODE_unlinked) {
1089                 ret = check_inode_deleted_list(trans, k.k->p);
1090                 if (ret < 0)
1091                         return ret;
1092
1093                 fsck_err_on(!ret, c, unlinked_inode_not_on_deleted_list,
1094                             "inode %llu:%u unlinked, but not on deleted list",
1095                             u.bi_inum, k.k->p.snapshot);
1096                 ret = 0;
1097         }
1098
1099         if (u.bi_flags & BCH_INODE_unlinked &&
1100             (!c->sb.clean ||
1101              fsck_err(c, inode_unlinked_but_clean,
1102                       "filesystem marked clean, but inode %llu unlinked",
1103                       u.bi_inum))) {
1104                 ret = bch2_inode_rm_snapshot(trans, u.bi_inum, iter->pos.snapshot);
1105                 bch_err_msg(c, ret, "in fsck deleting inode");
1106                 return ret;
1107         }
1108
1109         if (u.bi_flags & BCH_INODE_i_size_dirty &&
1110             (!c->sb.clean ||
1111              fsck_err(c, inode_i_size_dirty_but_clean,
1112                       "filesystem marked clean, but inode %llu has i_size dirty",
1113                       u.bi_inum))) {
1114                 bch_verbose(c, "truncating inode %llu", u.bi_inum);
1115
1116                 /*
1117                  * XXX: need to truncate partial blocks too here - or ideally
1118                  * just switch units to bytes and that issue goes away
1119                  */
1120                 ret = bch2_btree_delete_range_trans(trans, BTREE_ID_extents,
1121                                 SPOS(u.bi_inum, round_up(u.bi_size, block_bytes(c)) >> 9,
1122                                      iter->pos.snapshot),
1123                                 POS(u.bi_inum, U64_MAX),
1124                                 0, NULL);
1125                 bch_err_msg(c, ret, "in fsck truncating inode");
1126                 if (ret)
1127                         return ret;
1128
1129                 /*
1130                  * We truncated without our normal sector accounting hook, just
1131                  * make sure we recalculate it:
1132                  */
1133                 u.bi_flags |= BCH_INODE_i_sectors_dirty;
1134
1135                 u.bi_flags &= ~BCH_INODE_i_size_dirty;
1136                 do_update = true;
1137         }
1138
1139         if (u.bi_flags & BCH_INODE_i_sectors_dirty &&
1140             (!c->sb.clean ||
1141              fsck_err(c, inode_i_sectors_dirty_but_clean,
1142                       "filesystem marked clean, but inode %llu has i_sectors dirty",
1143                       u.bi_inum))) {
1144                 s64 sectors;
1145
1146                 bch_verbose(c, "recounting sectors for inode %llu",
1147                             u.bi_inum);
1148
1149                 sectors = bch2_count_inode_sectors(trans, u.bi_inum, iter->pos.snapshot);
1150                 if (sectors < 0) {
1151                         bch_err_msg(c, sectors, "in fsck recounting inode sectors");
1152                         return sectors;
1153                 }
1154
1155                 u.bi_sectors = sectors;
1156                 u.bi_flags &= ~BCH_INODE_i_sectors_dirty;
1157                 do_update = true;
1158         }
1159
1160         if (u.bi_flags & BCH_INODE_backptr_untrusted) {
1161                 u.bi_dir = 0;
1162                 u.bi_dir_offset = 0;
1163                 u.bi_flags &= ~BCH_INODE_backptr_untrusted;
1164                 do_update = true;
1165         }
1166
1167         if (u.bi_dir || u.bi_dir_offset) {
1168                 ret = check_inode_dirent_inode(trans, k, &u, k.k->p.snapshot, &do_update);
1169                 if (ret)
1170                         goto err;
1171         }
1172
1173         if (fsck_err_on(u.bi_parent_subvol &&
1174                         (u.bi_subvol == 0 ||
1175                          u.bi_subvol == BCACHEFS_ROOT_SUBVOL),
1176                         c, inode_bi_parent_nonzero,
1177                         "inode %llu:%u has subvol %u but nonzero parent subvol %u",
1178                         u.bi_inum, k.k->p.snapshot, u.bi_subvol, u.bi_parent_subvol)) {
1179                 u.bi_parent_subvol = 0;
1180                 do_update = true;
1181         }
1182
1183         if (u.bi_subvol) {
1184                 struct bch_subvolume s;
1185
1186                 ret = bch2_subvolume_get(trans, u.bi_subvol, false, 0, &s);
1187                 if (ret && !bch2_err_matches(ret, ENOENT))
1188                         goto err;
1189
1190                 if (ret && (c->sb.btrees_lost_data & BIT_ULL(BTREE_ID_subvolumes))) {
1191                         ret = reconstruct_subvol(trans, k.k->p.snapshot, u.bi_subvol, u.bi_inum);
1192                         goto do_update;
1193                 }
1194
1195                 if (fsck_err_on(ret,
1196                                 c, inode_bi_subvol_missing,
1197                                 "inode %llu:%u bi_subvol points to missing subvolume %u",
1198                                 u.bi_inum, k.k->p.snapshot, u.bi_subvol) ||
1199                     fsck_err_on(le64_to_cpu(s.inode) != u.bi_inum ||
1200                                 !bch2_snapshot_is_ancestor(c, le32_to_cpu(s.snapshot),
1201                                                            k.k->p.snapshot),
1202                                 c, inode_bi_subvol_wrong,
1203                                 "inode %llu:%u points to subvol %u, but subvol points to %llu:%u",
1204                                 u.bi_inum, k.k->p.snapshot, u.bi_subvol,
1205                                 le64_to_cpu(s.inode),
1206                                 le32_to_cpu(s.snapshot))) {
1207                         u.bi_subvol = 0;
1208                         u.bi_parent_subvol = 0;
1209                         do_update = true;
1210                 }
1211         }
1212 do_update:
1213         if (do_update) {
1214                 ret = __bch2_fsck_write_inode(trans, &u, iter->pos.snapshot);
1215                 bch_err_msg(c, ret, "in fsck updating inode");
1216                 if (ret)
1217                         return ret;
1218         }
1219 err:
1220 fsck_err:
1221         bch_err_fn(c, ret);
1222         return ret;
1223 }
1224
1225 int bch2_check_inodes(struct bch_fs *c)
1226 {
1227         bool full = c->opts.fsck;
1228         struct bch_inode_unpacked prev = { 0 };
1229         struct snapshots_seen s;
1230
1231         snapshots_seen_init(&s);
1232
1233         int ret = bch2_trans_run(c,
1234                 for_each_btree_key_commit(trans, iter, BTREE_ID_inodes,
1235                                 POS_MIN,
1236                                 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS, k,
1237                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1238                         check_inode(trans, &iter, k, &prev, &s, full)));
1239
1240         snapshots_seen_exit(&s);
1241         bch_err_fn(c, ret);
1242         return ret;
1243 }
1244
1245 static int check_i_sectors_notnested(struct btree_trans *trans, struct inode_walker *w)
1246 {
1247         struct bch_fs *c = trans->c;
1248         int ret = 0;
1249         s64 count2;
1250
1251         darray_for_each(w->inodes, i) {
1252                 if (i->inode.bi_sectors == i->count)
1253                         continue;
1254
1255                 count2 = bch2_count_inode_sectors(trans, w->last_pos.inode, i->snapshot);
1256
1257                 if (w->recalculate_sums)
1258                         i->count = count2;
1259
1260                 if (i->count != count2) {
1261                         bch_err_ratelimited(c, "fsck counted i_sectors wrong for inode %llu:%u: got %llu should be %llu",
1262                                             w->last_pos.inode, i->snapshot, i->count, count2);
1263                         return -BCH_ERR_internal_fsck_err;
1264                 }
1265
1266                 if (fsck_err_on(!(i->inode.bi_flags & BCH_INODE_i_sectors_dirty),
1267                                 c, inode_i_sectors_wrong,
1268                                 "inode %llu:%u has incorrect i_sectors: got %llu, should be %llu",
1269                                 w->last_pos.inode, i->snapshot,
1270                                 i->inode.bi_sectors, i->count)) {
1271                         i->inode.bi_sectors = i->count;
1272                         ret = bch2_fsck_write_inode(trans, &i->inode, i->snapshot);
1273                         if (ret)
1274                                 break;
1275                 }
1276         }
1277 fsck_err:
1278         bch_err_fn(c, ret);
1279         return ret;
1280 }
1281
1282 static int check_i_sectors(struct btree_trans *trans, struct inode_walker *w)
1283 {
1284         u32 restart_count = trans->restart_count;
1285         return check_i_sectors_notnested(trans, w) ?:
1286                 trans_was_restarted(trans, restart_count);
1287 }
1288
1289 struct extent_end {
1290         u32                     snapshot;
1291         u64                     offset;
1292         struct snapshots_seen   seen;
1293 };
1294
1295 struct extent_ends {
1296         struct bpos                     last_pos;
1297         DARRAY(struct extent_end)       e;
1298 };
1299
1300 static void extent_ends_reset(struct extent_ends *extent_ends)
1301 {
1302         darray_for_each(extent_ends->e, i)
1303                 snapshots_seen_exit(&i->seen);
1304         extent_ends->e.nr = 0;
1305 }
1306
1307 static void extent_ends_exit(struct extent_ends *extent_ends)
1308 {
1309         extent_ends_reset(extent_ends);
1310         darray_exit(&extent_ends->e);
1311 }
1312
1313 static void extent_ends_init(struct extent_ends *extent_ends)
1314 {
1315         memset(extent_ends, 0, sizeof(*extent_ends));
1316 }
1317
1318 static int extent_ends_at(struct bch_fs *c,
1319                           struct extent_ends *extent_ends,
1320                           struct snapshots_seen *seen,
1321                           struct bkey_s_c k)
1322 {
1323         struct extent_end *i, n = (struct extent_end) {
1324                 .offset         = k.k->p.offset,
1325                 .snapshot       = k.k->p.snapshot,
1326                 .seen           = *seen,
1327         };
1328
1329         n.seen.ids.data = kmemdup(seen->ids.data,
1330                               sizeof(seen->ids.data[0]) * seen->ids.size,
1331                               GFP_KERNEL);
1332         if (!n.seen.ids.data)
1333                 return -BCH_ERR_ENOMEM_fsck_extent_ends_at;
1334
1335         __darray_for_each(extent_ends->e, i) {
1336                 if (i->snapshot == k.k->p.snapshot) {
1337                         snapshots_seen_exit(&i->seen);
1338                         *i = n;
1339                         return 0;
1340                 }
1341
1342                 if (i->snapshot >= k.k->p.snapshot)
1343                         break;
1344         }
1345
1346         return darray_insert_item(&extent_ends->e, i - extent_ends->e.data, n);
1347 }
1348
1349 static int overlapping_extents_found(struct btree_trans *trans,
1350                                      enum btree_id btree,
1351                                      struct bpos pos1, struct snapshots_seen *pos1_seen,
1352                                      struct bkey pos2,
1353                                      bool *fixed,
1354                                      struct extent_end *extent_end)
1355 {
1356         struct bch_fs *c = trans->c;
1357         struct printbuf buf = PRINTBUF;
1358         struct btree_iter iter1, iter2 = { NULL };
1359         struct bkey_s_c k1, k2;
1360         int ret;
1361
1362         BUG_ON(bkey_le(pos1, bkey_start_pos(&pos2)));
1363
1364         bch2_trans_iter_init(trans, &iter1, btree, pos1,
1365                              BTREE_ITER_ALL_SNAPSHOTS|
1366                              BTREE_ITER_NOT_EXTENTS);
1367         k1 = bch2_btree_iter_peek_upto(&iter1, POS(pos1.inode, U64_MAX));
1368         ret = bkey_err(k1);
1369         if (ret)
1370                 goto err;
1371
1372         prt_str(&buf, "\n  ");
1373         bch2_bkey_val_to_text(&buf, c, k1);
1374
1375         if (!bpos_eq(pos1, k1.k->p)) {
1376                 prt_str(&buf, "\n  wanted\n  ");
1377                 bch2_bpos_to_text(&buf, pos1);
1378                 prt_str(&buf, "\n  ");
1379                 bch2_bkey_to_text(&buf, &pos2);
1380
1381                 bch_err(c, "%s: error finding first overlapping extent when repairing, got%s",
1382                         __func__, buf.buf);
1383                 ret = -BCH_ERR_internal_fsck_err;
1384                 goto err;
1385         }
1386
1387         bch2_trans_copy_iter(&iter2, &iter1);
1388
1389         while (1) {
1390                 bch2_btree_iter_advance(&iter2);
1391
1392                 k2 = bch2_btree_iter_peek_upto(&iter2, POS(pos1.inode, U64_MAX));
1393                 ret = bkey_err(k2);
1394                 if (ret)
1395                         goto err;
1396
1397                 if (bpos_ge(k2.k->p, pos2.p))
1398                         break;
1399         }
1400
1401         prt_str(&buf, "\n  ");
1402         bch2_bkey_val_to_text(&buf, c, k2);
1403
1404         if (bpos_gt(k2.k->p, pos2.p) ||
1405             pos2.size != k2.k->size) {
1406                 bch_err(c, "%s: error finding seconding overlapping extent when repairing%s",
1407                         __func__, buf.buf);
1408                 ret = -BCH_ERR_internal_fsck_err;
1409                 goto err;
1410         }
1411
1412         prt_printf(&buf, "\n  overwriting %s extent",
1413                    pos1.snapshot >= pos2.p.snapshot ? "first" : "second");
1414
1415         if (fsck_err(c, extent_overlapping,
1416                      "overlapping extents%s", buf.buf)) {
1417                 struct btree_iter *old_iter = &iter1;
1418                 struct disk_reservation res = { 0 };
1419
1420                 if (pos1.snapshot < pos2.p.snapshot) {
1421                         old_iter = &iter2;
1422                         swap(k1, k2);
1423                 }
1424
1425                 trans->extra_disk_res += bch2_bkey_sectors_compressed(k2);
1426
1427                 ret =   bch2_trans_update_extent_overwrite(trans, old_iter,
1428                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE,
1429                                 k1, k2) ?:
1430                         bch2_trans_commit(trans, &res, NULL, BCH_TRANS_COMMIT_no_enospc);
1431                 bch2_disk_reservation_put(c, &res);
1432
1433                 if (ret)
1434                         goto err;
1435
1436                 *fixed = true;
1437
1438                 if (pos1.snapshot == pos2.p.snapshot) {
1439                         /*
1440                          * We overwrote the first extent, and did the overwrite
1441                          * in the same snapshot:
1442                          */
1443                         extent_end->offset = bkey_start_offset(&pos2);
1444                 } else if (pos1.snapshot > pos2.p.snapshot) {
1445                         /*
1446                          * We overwrote the first extent in pos2's snapshot:
1447                          */
1448                         ret = snapshots_seen_add_inorder(c, pos1_seen, pos2.p.snapshot);
1449                 } else {
1450                         /*
1451                          * We overwrote the second extent - restart
1452                          * check_extent() from the top:
1453                          */
1454                         ret = -BCH_ERR_transaction_restart_nested;
1455                 }
1456         }
1457 fsck_err:
1458 err:
1459         bch2_trans_iter_exit(trans, &iter2);
1460         bch2_trans_iter_exit(trans, &iter1);
1461         printbuf_exit(&buf);
1462         return ret;
1463 }
1464
1465 static int check_overlapping_extents(struct btree_trans *trans,
1466                               struct snapshots_seen *seen,
1467                               struct extent_ends *extent_ends,
1468                               struct bkey_s_c k,
1469                               u32 equiv,
1470                               struct btree_iter *iter,
1471                               bool *fixed)
1472 {
1473         struct bch_fs *c = trans->c;
1474         int ret = 0;
1475
1476         /* transaction restart, running again */
1477         if (bpos_eq(extent_ends->last_pos, k.k->p))
1478                 return 0;
1479
1480         if (extent_ends->last_pos.inode != k.k->p.inode)
1481                 extent_ends_reset(extent_ends);
1482
1483         darray_for_each(extent_ends->e, i) {
1484                 if (i->offset <= bkey_start_offset(k.k))
1485                         continue;
1486
1487                 if (!ref_visible2(c,
1488                                   k.k->p.snapshot, seen,
1489                                   i->snapshot, &i->seen))
1490                         continue;
1491
1492                 ret = overlapping_extents_found(trans, iter->btree_id,
1493                                                 SPOS(iter->pos.inode,
1494                                                      i->offset,
1495                                                      i->snapshot),
1496                                                 &i->seen,
1497                                                 *k.k, fixed, i);
1498                 if (ret)
1499                         goto err;
1500         }
1501
1502         extent_ends->last_pos = k.k->p;
1503 err:
1504         return ret;
1505 }
1506
1507 static int check_extent_overbig(struct btree_trans *trans, struct btree_iter *iter,
1508                                 struct bkey_s_c k)
1509 {
1510         struct bch_fs *c = trans->c;
1511         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1512         struct bch_extent_crc_unpacked crc;
1513         const union bch_extent_entry *i;
1514         unsigned encoded_extent_max_sectors = c->opts.encoded_extent_max >> 9;
1515
1516         bkey_for_each_crc(k.k, ptrs, crc, i)
1517                 if (crc_is_encoded(crc) &&
1518                     crc.uncompressed_size > encoded_extent_max_sectors) {
1519                         struct printbuf buf = PRINTBUF;
1520
1521                         bch2_bkey_val_to_text(&buf, c, k);
1522                         bch_err(c, "overbig encoded extent, please report this:\n  %s", buf.buf);
1523                         printbuf_exit(&buf);
1524                 }
1525
1526         return 0;
1527 }
1528
1529 static int check_extent(struct btree_trans *trans, struct btree_iter *iter,
1530                         struct bkey_s_c k,
1531                         struct inode_walker *inode,
1532                         struct snapshots_seen *s,
1533                         struct extent_ends *extent_ends)
1534 {
1535         struct bch_fs *c = trans->c;
1536         struct inode_walker_entry *i;
1537         struct printbuf buf = PRINTBUF;
1538         struct bpos equiv = k.k->p;
1539         int ret = 0;
1540
1541         equiv.snapshot = bch2_snapshot_equiv(c, k.k->p.snapshot);
1542
1543         ret = check_key_has_snapshot(trans, iter, k);
1544         if (ret) {
1545                 ret = ret < 0 ? ret : 0;
1546                 goto out;
1547         }
1548
1549         if (inode->last_pos.inode != k.k->p.inode) {
1550                 ret = check_i_sectors(trans, inode);
1551                 if (ret)
1552                         goto err;
1553         }
1554
1555         i = walk_inode(trans, inode, k);
1556         ret = PTR_ERR_OR_ZERO(i);
1557         if (ret)
1558                 goto err;
1559
1560         ret = snapshots_seen_update(c, s, iter->btree_id, k.k->p);
1561         if (ret)
1562                 goto err;
1563
1564         if (k.k->type != KEY_TYPE_whiteout) {
1565                 if (!i && (c->sb.btrees_lost_data & BIT_ULL(BTREE_ID_inodes))) {
1566                         ret =   reconstruct_reg_inode(trans, k.k->p.snapshot, k.k->p.inode) ?:
1567                                 bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
1568                         if (ret)
1569                                 goto err;
1570
1571                         inode->last_pos.inode--;
1572                         ret = -BCH_ERR_transaction_restart_nested;
1573                         goto err;
1574                 }
1575
1576                 if (fsck_err_on(!i, c, extent_in_missing_inode,
1577                                 "extent in missing inode:\n  %s",
1578                                 (printbuf_reset(&buf),
1579                                  bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
1580                         goto delete;
1581
1582                 if (fsck_err_on(i &&
1583                                 !S_ISREG(i->inode.bi_mode) &&
1584                                 !S_ISLNK(i->inode.bi_mode),
1585                                 c, extent_in_non_reg_inode,
1586                                 "extent in non regular inode mode %o:\n  %s",
1587                                 i->inode.bi_mode,
1588                                 (printbuf_reset(&buf),
1589                                  bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
1590                         goto delete;
1591
1592                 ret = check_overlapping_extents(trans, s, extent_ends, k,
1593                                                 equiv.snapshot, iter,
1594                                                 &inode->recalculate_sums);
1595                 if (ret)
1596                         goto err;
1597         }
1598
1599         /*
1600          * Check inodes in reverse order, from oldest snapshots to newest,
1601          * starting from the inode that matches this extent's snapshot. If we
1602          * didn't have one, iterate over all inodes:
1603          */
1604         if (!i)
1605                 i = inode->inodes.data + inode->inodes.nr - 1;
1606
1607         for (;
1608              inode->inodes.data && i >= inode->inodes.data;
1609              --i) {
1610                 if (i->snapshot > equiv.snapshot ||
1611                     !key_visible_in_snapshot(c, s, i->snapshot, equiv.snapshot))
1612                         continue;
1613
1614                 if (k.k->type != KEY_TYPE_whiteout) {
1615                         if (fsck_err_on(!(i->inode.bi_flags & BCH_INODE_i_size_dirty) &&
1616                                         k.k->p.offset > round_up(i->inode.bi_size, block_bytes(c)) >> 9 &&
1617                                         !bkey_extent_is_reservation(k),
1618                                         c, extent_past_end_of_inode,
1619                                         "extent type past end of inode %llu:%u, i_size %llu\n  %s",
1620                                         i->inode.bi_inum, i->snapshot, i->inode.bi_size,
1621                                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1622                                 struct btree_iter iter2;
1623
1624                                 bch2_trans_copy_iter(&iter2, iter);
1625                                 bch2_btree_iter_set_snapshot(&iter2, i->snapshot);
1626                                 ret =   bch2_btree_iter_traverse(&iter2) ?:
1627                                         bch2_btree_delete_at(trans, &iter2,
1628                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1629                                 bch2_trans_iter_exit(trans, &iter2);
1630                                 if (ret)
1631                                         goto err;
1632
1633                                 iter->k.type = KEY_TYPE_whiteout;
1634                         }
1635
1636                         if (bkey_extent_is_allocation(k.k))
1637                                 i->count += k.k->size;
1638                 }
1639
1640                 i->seen_this_pos = true;
1641         }
1642
1643         if (k.k->type != KEY_TYPE_whiteout) {
1644                 ret = extent_ends_at(c, extent_ends, s, k);
1645                 if (ret)
1646                         goto err;
1647         }
1648 out:
1649 err:
1650 fsck_err:
1651         printbuf_exit(&buf);
1652         bch_err_fn(c, ret);
1653         return ret;
1654 delete:
1655         ret = bch2_btree_delete_at(trans, iter, BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1656         goto out;
1657 }
1658
1659 /*
1660  * Walk extents: verify that extents have a corresponding S_ISREG inode, and
1661  * that i_size an i_sectors are consistent
1662  */
1663 int bch2_check_extents(struct bch_fs *c)
1664 {
1665         struct inode_walker w = inode_walker_init();
1666         struct snapshots_seen s;
1667         struct extent_ends extent_ends;
1668         struct disk_reservation res = { 0 };
1669
1670         snapshots_seen_init(&s);
1671         extent_ends_init(&extent_ends);
1672
1673         int ret = bch2_trans_run(c,
1674                 for_each_btree_key_commit(trans, iter, BTREE_ID_extents,
1675                                 POS(BCACHEFS_ROOT_INO, 0),
1676                                 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS, k,
1677                                 &res, NULL,
1678                                 BCH_TRANS_COMMIT_no_enospc, ({
1679                         bch2_disk_reservation_put(c, &res);
1680                         check_extent(trans, &iter, k, &w, &s, &extent_ends) ?:
1681                         check_extent_overbig(trans, &iter, k);
1682                 })) ?:
1683                 check_i_sectors_notnested(trans, &w));
1684
1685         bch2_disk_reservation_put(c, &res);
1686         extent_ends_exit(&extent_ends);
1687         inode_walker_exit(&w);
1688         snapshots_seen_exit(&s);
1689
1690         bch_err_fn(c, ret);
1691         return ret;
1692 }
1693
1694 int bch2_check_indirect_extents(struct bch_fs *c)
1695 {
1696         struct disk_reservation res = { 0 };
1697
1698         int ret = bch2_trans_run(c,
1699                 for_each_btree_key_commit(trans, iter, BTREE_ID_reflink,
1700                                 POS_MIN,
1701                                 BTREE_ITER_PREFETCH, k,
1702                                 &res, NULL,
1703                                 BCH_TRANS_COMMIT_no_enospc, ({
1704                         bch2_disk_reservation_put(c, &res);
1705                         check_extent_overbig(trans, &iter, k);
1706                 })));
1707
1708         bch2_disk_reservation_put(c, &res);
1709         bch_err_fn(c, ret);
1710         return ret;
1711 }
1712
1713 static int check_subdir_count_notnested(struct btree_trans *trans, struct inode_walker *w)
1714 {
1715         struct bch_fs *c = trans->c;
1716         int ret = 0;
1717         s64 count2;
1718
1719         darray_for_each(w->inodes, i) {
1720                 if (i->inode.bi_nlink == i->count)
1721                         continue;
1722
1723                 count2 = bch2_count_subdirs(trans, w->last_pos.inode, i->snapshot);
1724                 if (count2 < 0)
1725                         return count2;
1726
1727                 if (i->count != count2) {
1728                         bch_err_ratelimited(c, "fsck counted subdirectories wrong for inum %llu:%u: got %llu should be %llu",
1729                                             w->last_pos.inode, i->snapshot, i->count, count2);
1730                         i->count = count2;
1731                         if (i->inode.bi_nlink == i->count)
1732                                 continue;
1733                 }
1734
1735                 if (fsck_err_on(i->inode.bi_nlink != i->count,
1736                                 c, inode_dir_wrong_nlink,
1737                                 "directory %llu:%u with wrong i_nlink: got %u, should be %llu",
1738                                 w->last_pos.inode, i->snapshot, i->inode.bi_nlink, i->count)) {
1739                         i->inode.bi_nlink = i->count;
1740                         ret = bch2_fsck_write_inode(trans, &i->inode, i->snapshot);
1741                         if (ret)
1742                                 break;
1743                 }
1744         }
1745 fsck_err:
1746         bch_err_fn(c, ret);
1747         return ret;
1748 }
1749
1750 static int check_subdir_count(struct btree_trans *trans, struct inode_walker *w)
1751 {
1752         u32 restart_count = trans->restart_count;
1753         return check_subdir_count_notnested(trans, w) ?:
1754                 trans_was_restarted(trans, restart_count);
1755 }
1756
1757 static int check_dirent_inode_dirent(struct btree_trans *trans,
1758                                    struct btree_iter *iter,
1759                                    struct bkey_s_c_dirent d,
1760                                    struct bch_inode_unpacked *target,
1761                                    u32 target_snapshot)
1762 {
1763         struct bch_fs *c = trans->c;
1764         struct printbuf buf = PRINTBUF;
1765         int ret = 0;
1766
1767         if (inode_points_to_dirent(target, d))
1768                 return 0;
1769
1770         if (!target->bi_dir &&
1771             !target->bi_dir_offset) {
1772                 target->bi_dir          = d.k->p.inode;
1773                 target->bi_dir_offset   = d.k->p.offset;
1774                 return __bch2_fsck_write_inode(trans, target, target_snapshot);
1775         }
1776
1777         struct btree_iter bp_iter = { NULL };
1778         struct bkey_s_c_dirent bp_dirent = dirent_get_by_pos(trans, &bp_iter,
1779                               SPOS(target->bi_dir, target->bi_dir_offset, target_snapshot));
1780         ret = bkey_err(bp_dirent);
1781         if (ret && !bch2_err_matches(ret, ENOENT))
1782                 goto err;
1783
1784         bool backpointer_exists = !ret;
1785         ret = 0;
1786
1787         if (fsck_err_on(!backpointer_exists,
1788                         c, inode_wrong_backpointer,
1789                         "inode %llu:%u has wrong backpointer:\n"
1790                         "got       %llu:%llu\n"
1791                         "should be %llu:%llu",
1792                         target->bi_inum, target_snapshot,
1793                         target->bi_dir,
1794                         target->bi_dir_offset,
1795                         d.k->p.inode,
1796                         d.k->p.offset)) {
1797                 target->bi_dir          = d.k->p.inode;
1798                 target->bi_dir_offset   = d.k->p.offset;
1799                 ret = __bch2_fsck_write_inode(trans, target, target_snapshot);
1800                 goto out;
1801         }
1802
1803         bch2_bkey_val_to_text(&buf, c, d.s_c);
1804         prt_newline(&buf);
1805         if (backpointer_exists)
1806                 bch2_bkey_val_to_text(&buf, c, bp_dirent.s_c);
1807
1808         if (fsck_err_on(backpointer_exists &&
1809                         (S_ISDIR(target->bi_mode) ||
1810                          target->bi_subvol),
1811                         c, inode_dir_multiple_links,
1812                         "%s %llu:%u with multiple links\n%s",
1813                         S_ISDIR(target->bi_mode) ? "directory" : "subvolume",
1814                         target->bi_inum, target_snapshot, buf.buf)) {
1815                 ret = __remove_dirent(trans, d.k->p);
1816                 goto out;
1817         }
1818
1819         /*
1820          * hardlinked file with nlink 0:
1821          * We're just adjusting nlink here so check_nlinks() will pick
1822          * it up, it ignores inodes with nlink 0
1823          */
1824         if (fsck_err_on(backpointer_exists && !target->bi_nlink,
1825                         c, inode_multiple_links_but_nlink_0,
1826                         "inode %llu:%u type %s has multiple links but i_nlink 0\n%s",
1827                         target->bi_inum, target_snapshot, bch2_d_types[d.v->d_type], buf.buf)) {
1828                 target->bi_nlink++;
1829                 target->bi_flags &= ~BCH_INODE_unlinked;
1830                 ret = __bch2_fsck_write_inode(trans, target, target_snapshot);
1831                 if (ret)
1832                         goto err;
1833         }
1834 out:
1835 err:
1836 fsck_err:
1837         bch2_trans_iter_exit(trans, &bp_iter);
1838         printbuf_exit(&buf);
1839         bch_err_fn(c, ret);
1840         return ret;
1841 }
1842
1843 static int check_dirent_target(struct btree_trans *trans,
1844                                struct btree_iter *iter,
1845                                struct bkey_s_c_dirent d,
1846                                struct bch_inode_unpacked *target,
1847                                u32 target_snapshot)
1848 {
1849         struct bch_fs *c = trans->c;
1850         struct bkey_i_dirent *n;
1851         struct printbuf buf = PRINTBUF;
1852         int ret = 0;
1853
1854         ret = check_dirent_inode_dirent(trans, iter, d, target, target_snapshot);
1855         if (ret)
1856                 goto err;
1857
1858         if (fsck_err_on(d.v->d_type != inode_d_type(target),
1859                         c, dirent_d_type_wrong,
1860                         "incorrect d_type: got %s, should be %s:\n%s",
1861                         bch2_d_type_str(d.v->d_type),
1862                         bch2_d_type_str(inode_d_type(target)),
1863                         (printbuf_reset(&buf),
1864                          bch2_bkey_val_to_text(&buf, c, d.s_c), buf.buf))) {
1865                 n = bch2_trans_kmalloc(trans, bkey_bytes(d.k));
1866                 ret = PTR_ERR_OR_ZERO(n);
1867                 if (ret)
1868                         goto err;
1869
1870                 bkey_reassemble(&n->k_i, d.s_c);
1871                 n->v.d_type = inode_d_type(target);
1872                 if (n->v.d_type == DT_SUBVOL) {
1873                         n->v.d_parent_subvol = cpu_to_le32(target->bi_parent_subvol);
1874                         n->v.d_child_subvol = cpu_to_le32(target->bi_subvol);
1875                 } else {
1876                         n->v.d_inum = cpu_to_le64(target->bi_inum);
1877                 }
1878
1879                 ret = bch2_trans_update(trans, iter, &n->k_i, 0);
1880                 if (ret)
1881                         goto err;
1882
1883                 d = dirent_i_to_s_c(n);
1884         }
1885 err:
1886 fsck_err:
1887         printbuf_exit(&buf);
1888         bch_err_fn(c, ret);
1889         return ret;
1890 }
1891
1892 /* find a subvolume that's a descendent of @snapshot: */
1893 static int find_snapshot_subvol(struct btree_trans *trans, u32 snapshot, u32 *subvolid)
1894 {
1895         struct btree_iter iter;
1896         struct bkey_s_c k;
1897         int ret;
1898
1899         for_each_btree_key_norestart(trans, iter, BTREE_ID_subvolumes, POS_MIN, 0, k, ret) {
1900                 if (k.k->type != KEY_TYPE_subvolume)
1901                         continue;
1902
1903                 struct bkey_s_c_subvolume s = bkey_s_c_to_subvolume(k);
1904                 if (bch2_snapshot_is_ancestor(trans->c, le32_to_cpu(s.v->snapshot), snapshot)) {
1905                         bch2_trans_iter_exit(trans, &iter);
1906                         *subvolid = k.k->p.offset;
1907                         goto found;
1908                 }
1909         }
1910         if (!ret)
1911                 ret = -ENOENT;
1912 found:
1913         bch2_trans_iter_exit(trans, &iter);
1914         return ret;
1915 }
1916
1917 static int check_dirent_to_subvol(struct btree_trans *trans, struct btree_iter *iter,
1918                                   struct bkey_s_c_dirent d)
1919 {
1920         struct bch_fs *c = trans->c;
1921         struct btree_iter subvol_iter = {};
1922         struct bch_inode_unpacked subvol_root;
1923         u32 parent_subvol = le32_to_cpu(d.v->d_parent_subvol);
1924         u32 target_subvol = le32_to_cpu(d.v->d_child_subvol);
1925         u32 parent_snapshot;
1926         u32 new_parent_subvol = 0;
1927         u64 parent_inum;
1928         struct printbuf buf = PRINTBUF;
1929         int ret = 0;
1930
1931         ret = subvol_lookup(trans, parent_subvol, &parent_snapshot, &parent_inum);
1932         if (ret && !bch2_err_matches(ret, ENOENT))
1933                 return ret;
1934
1935         if (ret ||
1936             (!ret && !bch2_snapshot_is_ancestor(c, parent_snapshot, d.k->p.snapshot))) {
1937                 int ret2 = find_snapshot_subvol(trans, d.k->p.snapshot, &new_parent_subvol);
1938                 if (ret2 && !bch2_err_matches(ret, ENOENT))
1939                         return ret2;
1940         }
1941
1942         if (ret &&
1943             !new_parent_subvol &&
1944             (c->sb.btrees_lost_data & BIT_ULL(BTREE_ID_subvolumes))) {
1945                 /*
1946                  * Couldn't find a subvol for dirent's snapshot - but we lost
1947                  * subvols, so we need to reconstruct:
1948                  */
1949                 ret = reconstruct_subvol(trans, d.k->p.snapshot, parent_subvol, 0);
1950                 if (ret)
1951                         return ret;
1952
1953                 parent_snapshot = d.k->p.snapshot;
1954         }
1955
1956         if (fsck_err_on(ret, c, dirent_to_missing_parent_subvol,
1957                         "dirent parent_subvol points to missing subvolume\n%s",
1958                         (bch2_bkey_val_to_text(&buf, c, d.s_c), buf.buf)) ||
1959             fsck_err_on(!ret && !bch2_snapshot_is_ancestor(c, parent_snapshot, d.k->p.snapshot),
1960                         c, dirent_not_visible_in_parent_subvol,
1961                         "dirent not visible in parent_subvol (not an ancestor of subvol snap %u)\n%s",
1962                         parent_snapshot,
1963                         (bch2_bkey_val_to_text(&buf, c, d.s_c), buf.buf))) {
1964                 if (!new_parent_subvol) {
1965                         bch_err(c, "could not find a subvol for snapshot %u", d.k->p.snapshot);
1966                         return -BCH_ERR_fsck_repair_unimplemented;
1967                 }
1968
1969                 struct bkey_i_dirent *new_dirent = bch2_bkey_make_mut_typed(trans, iter, &d.s_c, 0, dirent);
1970                 ret = PTR_ERR_OR_ZERO(new_dirent);
1971                 if (ret)
1972                         goto err;
1973
1974                 new_dirent->v.d_parent_subvol = cpu_to_le32(new_parent_subvol);
1975         }
1976
1977         struct bkey_s_c_subvolume s =
1978                 bch2_bkey_get_iter_typed(trans, &subvol_iter,
1979                                          BTREE_ID_subvolumes, POS(0, target_subvol),
1980                                          0, subvolume);
1981         ret = bkey_err(s.s_c);
1982         if (ret && !bch2_err_matches(ret, ENOENT))
1983                 return ret;
1984
1985         if (ret) {
1986                 if (fsck_err(c, dirent_to_missing_subvol,
1987                              "dirent points to missing subvolume\n%s",
1988                              (bch2_bkey_val_to_text(&buf, c, d.s_c), buf.buf)))
1989                         return __remove_dirent(trans, d.k->p);
1990                 ret = 0;
1991                 goto out;
1992         }
1993
1994         if (fsck_err_on(le32_to_cpu(s.v->fs_path_parent) != parent_subvol,
1995                         c, subvol_fs_path_parent_wrong,
1996                         "subvol with wrong fs_path_parent, should be be %u\n%s",
1997                         parent_subvol,
1998                         (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
1999                 struct bkey_i_subvolume *n =
2000                         bch2_bkey_make_mut_typed(trans, &subvol_iter, &s.s_c, 0, subvolume);
2001                 ret = PTR_ERR_OR_ZERO(n);
2002                 if (ret)
2003                         goto err;
2004
2005                 n->v.fs_path_parent = cpu_to_le32(parent_subvol);
2006         }
2007
2008         u64 target_inum = le64_to_cpu(s.v->inode);
2009         u32 target_snapshot = le32_to_cpu(s.v->snapshot);
2010
2011         ret = lookup_inode(trans, target_inum, &subvol_root, &target_snapshot);
2012         if (ret && !bch2_err_matches(ret, ENOENT))
2013                 goto err;
2014
2015         if (ret) {
2016                 bch_err(c, "subvol %u points to missing inode root %llu", target_subvol, target_inum);
2017                 ret = -BCH_ERR_fsck_repair_unimplemented;
2018                 ret = 0;
2019                 goto err;
2020         }
2021
2022         if (fsck_err_on(!ret && parent_subvol != subvol_root.bi_parent_subvol,
2023                         c, inode_bi_parent_wrong,
2024                         "subvol root %llu has wrong bi_parent_subvol: got %u, should be %u",
2025                         target_inum,
2026                         subvol_root.bi_parent_subvol, parent_subvol)) {
2027                 subvol_root.bi_parent_subvol = parent_subvol;
2028                 ret = __bch2_fsck_write_inode(trans, &subvol_root, target_snapshot);
2029                 if (ret)
2030                         goto err;
2031         }
2032
2033         ret = check_dirent_target(trans, iter, d, &subvol_root,
2034                                   target_snapshot);
2035         if (ret)
2036                 goto err;
2037 out:
2038 err:
2039 fsck_err:
2040         bch2_trans_iter_exit(trans, &subvol_iter);
2041         printbuf_exit(&buf);
2042         return ret;
2043 }
2044
2045 static int check_dirent(struct btree_trans *trans, struct btree_iter *iter,
2046                         struct bkey_s_c k,
2047                         struct bch_hash_info *hash_info,
2048                         struct inode_walker *dir,
2049                         struct inode_walker *target,
2050                         struct snapshots_seen *s)
2051 {
2052         struct bch_fs *c = trans->c;
2053         struct inode_walker_entry *i;
2054         struct printbuf buf = PRINTBUF;
2055         struct bpos equiv;
2056         int ret = 0;
2057
2058         ret = check_key_has_snapshot(trans, iter, k);
2059         if (ret) {
2060                 ret = ret < 0 ? ret : 0;
2061                 goto out;
2062         }
2063
2064         equiv = k.k->p;
2065         equiv.snapshot = bch2_snapshot_equiv(c, k.k->p.snapshot);
2066
2067         ret = snapshots_seen_update(c, s, iter->btree_id, k.k->p);
2068         if (ret)
2069                 goto err;
2070
2071         if (k.k->type == KEY_TYPE_whiteout)
2072                 goto out;
2073
2074         if (dir->last_pos.inode != k.k->p.inode) {
2075                 ret = check_subdir_count(trans, dir);
2076                 if (ret)
2077                         goto err;
2078         }
2079
2080         BUG_ON(!btree_iter_path(trans, iter)->should_be_locked);
2081
2082         i = walk_inode(trans, dir, k);
2083         ret = PTR_ERR_OR_ZERO(i);
2084         if (ret < 0)
2085                 goto err;
2086
2087         if (dir->first_this_inode && dir->inodes.nr)
2088                 *hash_info = bch2_hash_info_init(c, &dir->inodes.data[0].inode);
2089         dir->first_this_inode = false;
2090
2091         if (!i && (c->sb.btrees_lost_data & BIT_ULL(BTREE_ID_inodes))) {
2092                 ret =   reconstruct_inode(trans, k.k->p.snapshot, k.k->p.inode, 0, S_IFDIR) ?:
2093                         bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
2094                 if (ret)
2095                         goto err;
2096
2097                 dir->last_pos.inode--;
2098                 ret = -BCH_ERR_transaction_restart_nested;
2099                 goto err;
2100         }
2101
2102         if (fsck_err_on(!i, c, dirent_in_missing_dir_inode,
2103                         "dirent in nonexisting directory:\n%s",
2104                         (printbuf_reset(&buf),
2105                          bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
2106                 ret = bch2_btree_delete_at(trans, iter,
2107                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
2108                 goto out;
2109         }
2110
2111         if (!i)
2112                 goto out;
2113
2114         if (fsck_err_on(!S_ISDIR(i->inode.bi_mode),
2115                         c, dirent_in_non_dir_inode,
2116                         "dirent in non directory inode type %s:\n%s",
2117                         bch2_d_type_str(inode_d_type(&i->inode)),
2118                         (printbuf_reset(&buf),
2119                          bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
2120                 ret = bch2_btree_delete_at(trans, iter, 0);
2121                 goto out;
2122         }
2123
2124         ret = hash_check_key(trans, bch2_dirent_hash_desc, hash_info, iter, k);
2125         if (ret < 0)
2126                 goto err;
2127         if (ret) {
2128                 /* dirent has been deleted */
2129                 ret = 0;
2130                 goto out;
2131         }
2132
2133         if (k.k->type != KEY_TYPE_dirent)
2134                 goto out;
2135
2136         struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
2137
2138         if (d.v->d_type == DT_SUBVOL) {
2139                 ret = check_dirent_to_subvol(trans, iter, d);
2140                 if (ret)
2141                         goto err;
2142         } else {
2143                 ret = __get_visible_inodes(trans, target, s, le64_to_cpu(d.v->d_inum));
2144                 if (ret)
2145                         goto err;
2146
2147                 if (fsck_err_on(!target->inodes.nr,
2148                                 c, dirent_to_missing_inode,
2149                                 "dirent points to missing inode: (equiv %u)\n%s",
2150                                 equiv.snapshot,
2151                                 (printbuf_reset(&buf),
2152                                  bch2_bkey_val_to_text(&buf, c, k),
2153                                  buf.buf))) {
2154                         ret = __remove_dirent(trans, d.k->p);
2155                         if (ret)
2156                                 goto err;
2157                 }
2158
2159                 darray_for_each(target->inodes, i) {
2160                         ret = check_dirent_target(trans, iter, d,
2161                                                   &i->inode, i->snapshot);
2162                         if (ret)
2163                                 goto err;
2164                 }
2165
2166                 if (d.v->d_type == DT_DIR)
2167                         for_each_visible_inode(c, s, dir, equiv.snapshot, i)
2168                                 i->count++;
2169         }
2170 out:
2171 err:
2172 fsck_err:
2173         printbuf_exit(&buf);
2174         bch_err_fn(c, ret);
2175         return ret;
2176 }
2177
2178 /*
2179  * Walk dirents: verify that they all have a corresponding S_ISDIR inode,
2180  * validate d_type
2181  */
2182 int bch2_check_dirents(struct bch_fs *c)
2183 {
2184         struct inode_walker dir = inode_walker_init();
2185         struct inode_walker target = inode_walker_init();
2186         struct snapshots_seen s;
2187         struct bch_hash_info hash_info;
2188
2189         snapshots_seen_init(&s);
2190
2191         int ret = bch2_trans_run(c,
2192                 for_each_btree_key_commit(trans, iter, BTREE_ID_dirents,
2193                                 POS(BCACHEFS_ROOT_INO, 0),
2194                                 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
2195                                 k,
2196                                 NULL, NULL,
2197                                 BCH_TRANS_COMMIT_no_enospc,
2198                         check_dirent(trans, &iter, k, &hash_info, &dir, &target, &s)) ?:
2199                 check_subdir_count_notnested(trans, &dir));
2200
2201         snapshots_seen_exit(&s);
2202         inode_walker_exit(&dir);
2203         inode_walker_exit(&target);
2204         bch_err_fn(c, ret);
2205         return ret;
2206 }
2207
2208 static int check_xattr(struct btree_trans *trans, struct btree_iter *iter,
2209                        struct bkey_s_c k,
2210                        struct bch_hash_info *hash_info,
2211                        struct inode_walker *inode)
2212 {
2213         struct bch_fs *c = trans->c;
2214         struct inode_walker_entry *i;
2215         int ret;
2216
2217         ret = check_key_has_snapshot(trans, iter, k);
2218         if (ret < 0)
2219                 return ret;
2220         if (ret)
2221                 return 0;
2222
2223         i = walk_inode(trans, inode, k);
2224         ret = PTR_ERR_OR_ZERO(i);
2225         if (ret)
2226                 return ret;
2227
2228         if (inode->first_this_inode && inode->inodes.nr)
2229                 *hash_info = bch2_hash_info_init(c, &inode->inodes.data[0].inode);
2230         inode->first_this_inode = false;
2231
2232         if (fsck_err_on(!i, c, xattr_in_missing_inode,
2233                         "xattr for missing inode %llu",
2234                         k.k->p.inode))
2235                 return bch2_btree_delete_at(trans, iter, 0);
2236
2237         if (!i)
2238                 return 0;
2239
2240         ret = hash_check_key(trans, bch2_xattr_hash_desc, hash_info, iter, k);
2241 fsck_err:
2242         bch_err_fn(c, ret);
2243         return ret;
2244 }
2245
2246 /*
2247  * Walk xattrs: verify that they all have a corresponding inode
2248  */
2249 int bch2_check_xattrs(struct bch_fs *c)
2250 {
2251         struct inode_walker inode = inode_walker_init();
2252         struct bch_hash_info hash_info;
2253         int ret = 0;
2254
2255         ret = bch2_trans_run(c,
2256                 for_each_btree_key_commit(trans, iter, BTREE_ID_xattrs,
2257                         POS(BCACHEFS_ROOT_INO, 0),
2258                         BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
2259                         k,
2260                         NULL, NULL,
2261                         BCH_TRANS_COMMIT_no_enospc,
2262                 check_xattr(trans, &iter, k, &hash_info, &inode)));
2263         bch_err_fn(c, ret);
2264         return ret;
2265 }
2266
2267 static int check_root_trans(struct btree_trans *trans)
2268 {
2269         struct bch_fs *c = trans->c;
2270         struct bch_inode_unpacked root_inode;
2271         u32 snapshot;
2272         u64 inum;
2273         int ret;
2274
2275         ret = subvol_lookup(trans, BCACHEFS_ROOT_SUBVOL, &snapshot, &inum);
2276         if (ret && !bch2_err_matches(ret, ENOENT))
2277                 return ret;
2278
2279         if (mustfix_fsck_err_on(ret, c, root_subvol_missing,
2280                                 "root subvol missing")) {
2281                 struct bkey_i_subvolume *root_subvol =
2282                         bch2_trans_kmalloc(trans, sizeof(*root_subvol));
2283                 ret = PTR_ERR_OR_ZERO(root_subvol);
2284                 if (ret)
2285                         goto err;
2286
2287                 snapshot        = U32_MAX;
2288                 inum            = BCACHEFS_ROOT_INO;
2289
2290                 bkey_subvolume_init(&root_subvol->k_i);
2291                 root_subvol->k.p.offset = BCACHEFS_ROOT_SUBVOL;
2292                 root_subvol->v.flags    = 0;
2293                 root_subvol->v.snapshot = cpu_to_le32(snapshot);
2294                 root_subvol->v.inode    = cpu_to_le64(inum);
2295                 ret = bch2_btree_insert_trans(trans, BTREE_ID_subvolumes, &root_subvol->k_i, 0);
2296                 bch_err_msg(c, ret, "writing root subvol");
2297                 if (ret)
2298                         goto err;
2299         }
2300
2301         ret = lookup_inode(trans, BCACHEFS_ROOT_INO, &root_inode, &snapshot);
2302         if (ret && !bch2_err_matches(ret, ENOENT))
2303                 return ret;
2304
2305         if (mustfix_fsck_err_on(ret, c, root_dir_missing,
2306                                 "root directory missing") ||
2307             mustfix_fsck_err_on(!S_ISDIR(root_inode.bi_mode),
2308                                 c, root_inode_not_dir,
2309                                 "root inode not a directory")) {
2310                 bch2_inode_init(c, &root_inode, 0, 0, S_IFDIR|0755,
2311                                 0, NULL);
2312                 root_inode.bi_inum = inum;
2313
2314                 ret = __bch2_fsck_write_inode(trans, &root_inode, snapshot);
2315                 bch_err_msg(c, ret, "writing root inode");
2316         }
2317 err:
2318 fsck_err:
2319         return ret;
2320 }
2321
2322 /* Get root directory, create if it doesn't exist: */
2323 int bch2_check_root(struct bch_fs *c)
2324 {
2325         int ret = bch2_trans_do(c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
2326                 check_root_trans(trans));
2327         bch_err_fn(c, ret);
2328         return ret;
2329 }
2330
2331 typedef DARRAY(u32) darray_u32;
2332
2333 static bool darray_u32_has(darray_u32 *d, u32 v)
2334 {
2335         darray_for_each(*d, i)
2336                 if (*i == v)
2337                         return true;
2338         return false;
2339 }
2340
2341 /*
2342  * We've checked that inode backpointers point to valid dirents; here, it's
2343  * sufficient to check that the subvolume root has a dirent:
2344  */
2345 static int subvol_has_dirent(struct btree_trans *trans, struct bkey_s_c_subvolume s)
2346 {
2347         struct bch_inode_unpacked inode;
2348         int ret = bch2_inode_find_by_inum_trans(trans,
2349                                 (subvol_inum) { s.k->p.offset, le64_to_cpu(s.v->inode) },
2350                                 &inode);
2351         if (ret)
2352                 return ret;
2353
2354         return inode.bi_dir != 0;
2355 }
2356
2357 static int check_subvol_path(struct btree_trans *trans, struct btree_iter *iter, struct bkey_s_c k)
2358 {
2359         struct bch_fs *c = trans->c;
2360         struct btree_iter parent_iter = {};
2361         darray_u32 subvol_path = {};
2362         struct printbuf buf = PRINTBUF;
2363         int ret = 0;
2364
2365         if (k.k->type != KEY_TYPE_subvolume)
2366                 return 0;
2367
2368         while (k.k->p.offset != BCACHEFS_ROOT_SUBVOL) {
2369                 ret = darray_push(&subvol_path, k.k->p.offset);
2370                 if (ret)
2371                         goto err;
2372
2373                 struct bkey_s_c_subvolume s = bkey_s_c_to_subvolume(k);
2374
2375                 ret = subvol_has_dirent(trans, s);
2376                 if (ret < 0)
2377                         break;
2378
2379                 if (fsck_err_on(!ret,
2380                                 c, subvol_unreachable,
2381                                 "unreachable subvolume %s",
2382                                 (bch2_bkey_val_to_text(&buf, c, s.s_c),
2383                                  buf.buf))) {
2384                         ret = reattach_subvol(trans, s);
2385                         break;
2386                 }
2387
2388                 u32 parent = le32_to_cpu(s.v->fs_path_parent);
2389
2390                 if (darray_u32_has(&subvol_path, parent)) {
2391                         if (fsck_err(c, subvol_loop, "subvolume loop"))
2392                                 ret = reattach_subvol(trans, s);
2393                         break;
2394                 }
2395
2396                 bch2_trans_iter_exit(trans, &parent_iter);
2397                 bch2_trans_iter_init(trans, &parent_iter,
2398                                      BTREE_ID_subvolumes, POS(0, parent), 0);
2399                 k = bch2_btree_iter_peek_slot(&parent_iter);
2400                 ret = bkey_err(k);
2401                 if (ret)
2402                         goto err;
2403
2404                 if (fsck_err_on(k.k->type != KEY_TYPE_subvolume,
2405                                 c, subvol_unreachable,
2406                                 "unreachable subvolume %s",
2407                                 (bch2_bkey_val_to_text(&buf, c, s.s_c),
2408                                  buf.buf))) {
2409                         ret = reattach_subvol(trans, s);
2410                         break;
2411                 }
2412         }
2413 fsck_err:
2414 err:
2415         printbuf_exit(&buf);
2416         darray_exit(&subvol_path);
2417         bch2_trans_iter_exit(trans, &parent_iter);
2418         return ret;
2419 }
2420
2421 int bch2_check_subvolume_structure(struct bch_fs *c)
2422 {
2423         int ret = bch2_trans_run(c,
2424                 for_each_btree_key_commit(trans, iter,
2425                                 BTREE_ID_subvolumes, POS_MIN, BTREE_ITER_PREFETCH, k,
2426                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
2427                         check_subvol_path(trans, &iter, k)));
2428         bch_err_fn(c, ret);
2429         return ret;
2430 }
2431
2432 struct pathbuf_entry {
2433         u64     inum;
2434         u32     snapshot;
2435 };
2436
2437 typedef DARRAY(struct pathbuf_entry) pathbuf;
2438
2439 static bool path_is_dup(pathbuf *p, u64 inum, u32 snapshot)
2440 {
2441         darray_for_each(*p, i)
2442                 if (i->inum     == inum &&
2443                     i->snapshot == snapshot)
2444                         return true;
2445         return false;
2446 }
2447
2448 /*
2449  * Check that a given inode is reachable from its subvolume root - we already
2450  * verified subvolume connectivity:
2451  *
2452  * XXX: we should also be verifying that inodes are in the right subvolumes
2453  */
2454 static int check_path(struct btree_trans *trans, pathbuf *p, struct bkey_s_c inode_k)
2455 {
2456         struct bch_fs *c = trans->c;
2457         struct btree_iter inode_iter = {};
2458         struct bch_inode_unpacked inode;
2459         struct printbuf buf = PRINTBUF;
2460         u32 snapshot = bch2_snapshot_equiv(c, inode_k.k->p.snapshot);
2461         int ret = 0;
2462
2463         p->nr = 0;
2464
2465         BUG_ON(bch2_inode_unpack(inode_k, &inode));
2466
2467         while (!inode.bi_subvol) {
2468                 struct btree_iter dirent_iter;
2469                 struct bkey_s_c_dirent d;
2470                 u32 parent_snapshot = snapshot;
2471
2472                 d = inode_get_dirent(trans, &dirent_iter, &inode, &parent_snapshot);
2473                 ret = bkey_err(d.s_c);
2474                 if (ret && !bch2_err_matches(ret, ENOENT))
2475                         break;
2476
2477                 if (!ret && !dirent_points_to_inode(d, &inode)) {
2478                         bch2_trans_iter_exit(trans, &dirent_iter);
2479                         ret = -BCH_ERR_ENOENT_dirent_doesnt_match_inode;
2480                 }
2481
2482                 if (bch2_err_matches(ret, ENOENT)) {
2483                         ret = 0;
2484                         if (fsck_err(c, inode_unreachable,
2485                                      "unreachable inode\n%s",
2486                                      (printbuf_reset(&buf),
2487                                       bch2_bkey_val_to_text(&buf, c, inode_k),
2488                                       buf.buf)))
2489                                 ret = reattach_inode(trans, &inode, snapshot);
2490                         goto out;
2491                 }
2492
2493                 bch2_trans_iter_exit(trans, &dirent_iter);
2494
2495                 if (!S_ISDIR(inode.bi_mode))
2496                         break;
2497
2498                 ret = darray_push(p, ((struct pathbuf_entry) {
2499                         .inum           = inode.bi_inum,
2500                         .snapshot       = snapshot,
2501                 }));
2502                 if (ret)
2503                         return ret;
2504
2505                 snapshot = parent_snapshot;
2506
2507                 bch2_trans_iter_exit(trans, &inode_iter);
2508                 inode_k = bch2_bkey_get_iter(trans, &inode_iter, BTREE_ID_inodes,
2509                                              SPOS(0, inode.bi_dir, snapshot), 0);
2510                 ret = bkey_err(inode_k) ?:
2511                         !bkey_is_inode(inode_k.k) ? -BCH_ERR_ENOENT_inode
2512                         : bch2_inode_unpack(inode_k, &inode);
2513                 if (ret) {
2514                         /* Should have been caught in dirents pass */
2515                         if (!bch2_err_matches(ret, BCH_ERR_transaction_restart))
2516                                 bch_err(c, "error looking up parent directory: %i", ret);
2517                         break;
2518                 }
2519
2520                 snapshot = inode_k.k->p.snapshot;
2521
2522                 if (path_is_dup(p, inode.bi_inum, snapshot)) {
2523                         /* XXX print path */
2524                         bch_err(c, "directory structure loop");
2525
2526                         darray_for_each(*p, i)
2527                                 pr_err("%llu:%u", i->inum, i->snapshot);
2528                         pr_err("%llu:%u", inode.bi_inum, snapshot);
2529
2530                         if (fsck_err(c, dir_loop, "directory structure loop")) {
2531                                 ret = remove_backpointer(trans, &inode);
2532                                 bch_err_msg(c, ret, "removing dirent");
2533                                 if (ret)
2534                                         break;
2535
2536                                 ret = reattach_inode(trans, &inode, snapshot);
2537                                 bch_err_msg(c, ret, "reattaching inode %llu", inode.bi_inum);
2538                         }
2539                         break;
2540                 }
2541         }
2542 out:
2543 fsck_err:
2544         bch2_trans_iter_exit(trans, &inode_iter);
2545         printbuf_exit(&buf);
2546         bch_err_fn(c, ret);
2547         return ret;
2548 }
2549
2550 /*
2551  * Check for unreachable inodes, as well as loops in the directory structure:
2552  * After bch2_check_dirents(), if an inode backpointer doesn't exist that means it's
2553  * unreachable:
2554  */
2555 int bch2_check_directory_structure(struct bch_fs *c)
2556 {
2557         pathbuf path = { 0, };
2558         int ret;
2559
2560         ret = bch2_trans_run(c,
2561                 for_each_btree_key_commit(trans, iter, BTREE_ID_inodes, POS_MIN,
2562                                           BTREE_ITER_INTENT|
2563                                           BTREE_ITER_PREFETCH|
2564                                           BTREE_ITER_ALL_SNAPSHOTS, k,
2565                                           NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
2566                         if (!bkey_is_inode(k.k))
2567                                 continue;
2568
2569                         if (bch2_inode_flags(k) & BCH_INODE_unlinked)
2570                                 continue;
2571
2572                         check_path(trans, &path, k);
2573                 })));
2574         darray_exit(&path);
2575
2576         bch_err_fn(c, ret);
2577         return ret;
2578 }
2579
2580 struct nlink_table {
2581         size_t          nr;
2582         size_t          size;
2583
2584         struct nlink {
2585                 u64     inum;
2586                 u32     snapshot;
2587                 u32     count;
2588         }               *d;
2589 };
2590
2591 static int add_nlink(struct bch_fs *c, struct nlink_table *t,
2592                      u64 inum, u32 snapshot)
2593 {
2594         if (t->nr == t->size) {
2595                 size_t new_size = max_t(size_t, 128UL, t->size * 2);
2596                 void *d = kvmalloc_array(new_size, sizeof(t->d[0]), GFP_KERNEL);
2597
2598                 if (!d) {
2599                         bch_err(c, "fsck: error allocating memory for nlink_table, size %zu",
2600                                 new_size);
2601                         return -BCH_ERR_ENOMEM_fsck_add_nlink;
2602                 }
2603
2604                 if (t->d)
2605                         memcpy(d, t->d, t->size * sizeof(t->d[0]));
2606                 kvfree(t->d);
2607
2608                 t->d = d;
2609                 t->size = new_size;
2610         }
2611
2612
2613         t->d[t->nr++] = (struct nlink) {
2614                 .inum           = inum,
2615                 .snapshot       = snapshot,
2616         };
2617
2618         return 0;
2619 }
2620
2621 static int nlink_cmp(const void *_l, const void *_r)
2622 {
2623         const struct nlink *l = _l;
2624         const struct nlink *r = _r;
2625
2626         return cmp_int(l->inum, r->inum);
2627 }
2628
2629 static void inc_link(struct bch_fs *c, struct snapshots_seen *s,
2630                      struct nlink_table *links,
2631                      u64 range_start, u64 range_end, u64 inum, u32 snapshot)
2632 {
2633         struct nlink *link, key = {
2634                 .inum = inum, .snapshot = U32_MAX,
2635         };
2636
2637         if (inum < range_start || inum >= range_end)
2638                 return;
2639
2640         link = __inline_bsearch(&key, links->d, links->nr,
2641                                 sizeof(links->d[0]), nlink_cmp);
2642         if (!link)
2643                 return;
2644
2645         while (link > links->d && link[0].inum == link[-1].inum)
2646                 --link;
2647
2648         for (; link < links->d + links->nr && link->inum == inum; link++)
2649                 if (ref_visible(c, s, snapshot, link->snapshot)) {
2650                         link->count++;
2651                         if (link->snapshot >= snapshot)
2652                                 break;
2653                 }
2654 }
2655
2656 noinline_for_stack
2657 static int check_nlinks_find_hardlinks(struct bch_fs *c,
2658                                        struct nlink_table *t,
2659                                        u64 start, u64 *end)
2660 {
2661         int ret = bch2_trans_run(c,
2662                 for_each_btree_key(trans, iter, BTREE_ID_inodes,
2663                                    POS(0, start),
2664                                    BTREE_ITER_INTENT|
2665                                    BTREE_ITER_PREFETCH|
2666                                    BTREE_ITER_ALL_SNAPSHOTS, k, ({
2667                         if (!bkey_is_inode(k.k))
2668                                 continue;
2669
2670                         /* Should never fail, checked by bch2_inode_invalid: */
2671                         struct bch_inode_unpacked u;
2672                         BUG_ON(bch2_inode_unpack(k, &u));
2673
2674                         /*
2675                          * Backpointer and directory structure checks are sufficient for
2676                          * directories, since they can't have hardlinks:
2677                          */
2678                         if (S_ISDIR(u.bi_mode))
2679                                 continue;
2680
2681                         if (!u.bi_nlink)
2682                                 continue;
2683
2684                         ret = add_nlink(c, t, k.k->p.offset, k.k->p.snapshot);
2685                         if (ret) {
2686                                 *end = k.k->p.offset;
2687                                 ret = 0;
2688                                 break;
2689                         }
2690                         0;
2691                 })));
2692
2693         bch_err_fn(c, ret);
2694         return ret;
2695 }
2696
2697 noinline_for_stack
2698 static int check_nlinks_walk_dirents(struct bch_fs *c, struct nlink_table *links,
2699                                      u64 range_start, u64 range_end)
2700 {
2701         struct snapshots_seen s;
2702
2703         snapshots_seen_init(&s);
2704
2705         int ret = bch2_trans_run(c,
2706                 for_each_btree_key(trans, iter, BTREE_ID_dirents, POS_MIN,
2707                                    BTREE_ITER_INTENT|
2708                                    BTREE_ITER_PREFETCH|
2709                                    BTREE_ITER_ALL_SNAPSHOTS, k, ({
2710                         ret = snapshots_seen_update(c, &s, iter.btree_id, k.k->p);
2711                         if (ret)
2712                                 break;
2713
2714                         if (k.k->type == KEY_TYPE_dirent) {
2715                                 struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
2716
2717                                 if (d.v->d_type != DT_DIR &&
2718                                     d.v->d_type != DT_SUBVOL)
2719                                         inc_link(c, &s, links, range_start, range_end,
2720                                                  le64_to_cpu(d.v->d_inum),
2721                                                  bch2_snapshot_equiv(c, d.k->p.snapshot));
2722                         }
2723                         0;
2724                 })));
2725
2726         snapshots_seen_exit(&s);
2727
2728         bch_err_fn(c, ret);
2729         return ret;
2730 }
2731
2732 static int check_nlinks_update_inode(struct btree_trans *trans, struct btree_iter *iter,
2733                                      struct bkey_s_c k,
2734                                      struct nlink_table *links,
2735                                      size_t *idx, u64 range_end)
2736 {
2737         struct bch_fs *c = trans->c;
2738         struct bch_inode_unpacked u;
2739         struct nlink *link = &links->d[*idx];
2740         int ret = 0;
2741
2742         if (k.k->p.offset >= range_end)
2743                 return 1;
2744
2745         if (!bkey_is_inode(k.k))
2746                 return 0;
2747
2748         BUG_ON(bch2_inode_unpack(k, &u));
2749
2750         if (S_ISDIR(u.bi_mode))
2751                 return 0;
2752
2753         if (!u.bi_nlink)
2754                 return 0;
2755
2756         while ((cmp_int(link->inum, k.k->p.offset) ?:
2757                 cmp_int(link->snapshot, k.k->p.snapshot)) < 0) {
2758                 BUG_ON(*idx == links->nr);
2759                 link = &links->d[++*idx];
2760         }
2761
2762         if (fsck_err_on(bch2_inode_nlink_get(&u) != link->count,
2763                         c, inode_wrong_nlink,
2764                         "inode %llu type %s has wrong i_nlink (%u, should be %u)",
2765                         u.bi_inum, bch2_d_types[mode_to_type(u.bi_mode)],
2766                         bch2_inode_nlink_get(&u), link->count)) {
2767                 bch2_inode_nlink_set(&u, link->count);
2768                 ret = __bch2_fsck_write_inode(trans, &u, k.k->p.snapshot);
2769         }
2770 fsck_err:
2771         return ret;
2772 }
2773
2774 noinline_for_stack
2775 static int check_nlinks_update_hardlinks(struct bch_fs *c,
2776                                struct nlink_table *links,
2777                                u64 range_start, u64 range_end)
2778 {
2779         size_t idx = 0;
2780
2781         int ret = bch2_trans_run(c,
2782                 for_each_btree_key_commit(trans, iter, BTREE_ID_inodes,
2783                                 POS(0, range_start),
2784                                 BTREE_ITER_INTENT|BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS, k,
2785                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
2786                         check_nlinks_update_inode(trans, &iter, k, links, &idx, range_end)));
2787         if (ret < 0) {
2788                 bch_err(c, "error in fsck walking inodes: %s", bch2_err_str(ret));
2789                 return ret;
2790         }
2791
2792         return 0;
2793 }
2794
2795 int bch2_check_nlinks(struct bch_fs *c)
2796 {
2797         struct nlink_table links = { 0 };
2798         u64 this_iter_range_start, next_iter_range_start = 0;
2799         int ret = 0;
2800
2801         do {
2802                 this_iter_range_start = next_iter_range_start;
2803                 next_iter_range_start = U64_MAX;
2804
2805                 ret = check_nlinks_find_hardlinks(c, &links,
2806                                                   this_iter_range_start,
2807                                                   &next_iter_range_start);
2808
2809                 ret = check_nlinks_walk_dirents(c, &links,
2810                                           this_iter_range_start,
2811                                           next_iter_range_start);
2812                 if (ret)
2813                         break;
2814
2815                 ret = check_nlinks_update_hardlinks(c, &links,
2816                                          this_iter_range_start,
2817                                          next_iter_range_start);
2818                 if (ret)
2819                         break;
2820
2821                 links.nr = 0;
2822         } while (next_iter_range_start != U64_MAX);
2823
2824         kvfree(links.d);
2825         bch_err_fn(c, ret);
2826         return ret;
2827 }
2828
2829 static int fix_reflink_p_key(struct btree_trans *trans, struct btree_iter *iter,
2830                              struct bkey_s_c k)
2831 {
2832         struct bkey_s_c_reflink_p p;
2833         struct bkey_i_reflink_p *u;
2834
2835         if (k.k->type != KEY_TYPE_reflink_p)
2836                 return 0;
2837
2838         p = bkey_s_c_to_reflink_p(k);
2839
2840         if (!p.v->front_pad && !p.v->back_pad)
2841                 return 0;
2842
2843         u = bch2_trans_kmalloc(trans, sizeof(*u));
2844         int ret = PTR_ERR_OR_ZERO(u);
2845         if (ret)
2846                 return ret;
2847
2848         bkey_reassemble(&u->k_i, k);
2849         u->v.front_pad  = 0;
2850         u->v.back_pad   = 0;
2851
2852         return bch2_trans_update(trans, iter, &u->k_i, BTREE_TRIGGER_NORUN);
2853 }
2854
2855 int bch2_fix_reflink_p(struct bch_fs *c)
2856 {
2857         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix)
2858                 return 0;
2859
2860         int ret = bch2_trans_run(c,
2861                 for_each_btree_key_commit(trans, iter,
2862                                 BTREE_ID_extents, POS_MIN,
2863                                 BTREE_ITER_INTENT|BTREE_ITER_PREFETCH|
2864                                 BTREE_ITER_ALL_SNAPSHOTS, k,
2865                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
2866                         fix_reflink_p_key(trans, &iter, k)));
2867         bch_err_fn(c, ret);
2868         return ret;
2869 }