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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "backpointers.h"
   5 #include "bkey_buf.h"
   6 #include "alloc_background.h"
   7 #include "btree_gc.h"
   8 #include "btree_update.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_io.h"
  11 #include "buckets.h"
  12 #include "dirent.h"
  13 #include "ec.h"
  14 #include "errcode.h"
  15 #include "error.h"
  16 #include "fs-common.h"
  17 #include "fsck.h"
  18 #include "journal_io.h"
  19 #include "journal_reclaim.h"
  20 #include "journal_seq_blacklist.h"
  21 #include "lru.h"
  22 #include "move.h"
  23 #include "quota.h"
  24 #include "recovery.h"
  25 #include "replicas.h"
  26 #include "subvolume.h"
  27 #include "super-io.h"
  28
  29 #include <linux/sort.h>
  30 #include <linux/stat.h>
  31
  32 #define QSTR(n) { { { .len = strlen(n) } }, .name = n }
  33
  34 /* for -o reconstruct_alloc: */
  35 static void drop_alloc_keys(struct journal_keys *keys)
  36 {
  37         size_t src, dst;
  38
  39         for (src = 0, dst = 0; src < keys->nr; src++)
  40                 if (keys->d[src].btree_id != BTREE_ID_alloc)
  41                         keys->d[dst++] = keys->d[src];
  42
  43         keys->nr = dst;
  44 }
  45
  46 /*
  47  * Btree node pointers have a field to stack a pointer to the in memory btree
  48  * node; we need to zero out this field when reading in btree nodes, or when
  49  * reading in keys from the journal:
  50  */
  51 static void zero_out_btree_mem_ptr(struct journal_keys *keys)
  52 {
  53         struct journal_key *i;
  54
  55         for (i = keys->d; i < keys->d + keys->nr; i++)
  56                 if (i->k->k.type == KEY_TYPE_btree_ptr_v2)
  57                         bkey_i_to_btree_ptr_v2(i->k)->v.mem_ptr = 0;
  58 }
  59
  60 /* iterate over keys read from the journal: */
  61
  62 static int __journal_key_cmp(enum btree_id      l_btree_id,
  63                              unsigned           l_level,
  64                              struct bpos        l_pos,
  65                              const struct journal_key *r)
  66 {
  67         return (cmp_int(l_btree_id,     r->btree_id) ?:
  68                 cmp_int(l_level,        r->level) ?:
  69                 bpos_cmp(l_pos, r->k->k.p));
  70 }
  71
  72 static int journal_key_cmp(const struct journal_key *l, const struct journal_key *r)
  73 {
  74         return __journal_key_cmp(l->btree_id, l->level, l->k->k.p, r);
  75 }
  76
  77 static inline size_t idx_to_pos(struct journal_keys *keys, size_t idx)
  78 {
  79         size_t gap_size = keys->size - keys->nr;
  80
  81         if (idx >= keys->gap)
  82                 idx += gap_size;
  83         return idx;
  84 }
  85
  86 static inline struct journal_key *idx_to_key(struct journal_keys *keys, size_t idx)
  87 {
  88         return keys->d + idx_to_pos(keys, idx);
  89 }
  90
  91 static size_t __bch2_journal_key_search(struct journal_keys *keys,
  92                                         enum btree_id id, unsigned level,
  93                                         struct bpos pos)
  94 {
  95         size_t l = 0, r = keys->nr, m;
  96
  97         while (l < r) {
  98                 m = l + ((r - l) >> 1);
  99                 if (__journal_key_cmp(id, level, pos, idx_to_key(keys, m)) > 0)
 100                         l = m + 1;
 101                 else
 102                         r = m;
 103         }
 104
 105         BUG_ON(l < keys->nr &&
 106                __journal_key_cmp(id, level, pos, idx_to_key(keys, l)) > 0);
 107
 108         BUG_ON(l &&
 109                __journal_key_cmp(id, level, pos, idx_to_key(keys, l - 1)) <= 0);
 110
 111         return l;
 112 }
 113
 114 static size_t bch2_journal_key_search(struct journal_keys *keys,
 115                                       enum btree_id id, unsigned level,
 116                                       struct bpos pos)
 117 {
 118         return idx_to_pos(keys, __bch2_journal_key_search(keys, id, level, pos));
 119 }
 120
 121 struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *c, enum btree_id btree_id,
 122                                            unsigned level, struct bpos pos,
 123                                            struct bpos end_pos, size_t *idx)
 124 {
 125         struct journal_keys *keys = &c->journal_keys;
 126         unsigned iters = 0;
 127         struct journal_key *k;
 128 search:
 129         if (!*idx)
 130                 *idx = __bch2_journal_key_search(keys, btree_id, level, pos);
 131
 132         while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
 133                 if (__journal_key_cmp(btree_id, level, end_pos, k) < 0)
 134                         return NULL;
 135
 136                 if (__journal_key_cmp(btree_id, level, pos, k) <= 0 &&
 137                     !k->overwritten)
 138                         return k->k;
 139
 140                 (*idx)++;
 141                 iters++;
 142                 if (iters == 10) {
 143                         *idx = 0;
 144                         goto search;
 145                 }
 146         }
 147
 148         return NULL;
 149 }
 150
 151 struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree_id,
 152                                            unsigned level, struct bpos pos)
 153 {
 154         size_t idx = 0;
 155
 156         return bch2_journal_keys_peek_upto(c, btree_id, level, pos, pos, &idx);
 157 }
 158
 159 static void journal_iters_fix(struct bch_fs *c)
 160 {
 161         struct journal_keys *keys = &c->journal_keys;
 162         /* The key we just inserted is immediately before the gap: */
 163         size_t gap_end = keys->gap + (keys->size - keys->nr);
 164         struct btree_and_journal_iter *iter;
 165
 166         /*
 167          * If an iterator points one after the key we just inserted, decrement
 168          * the iterator so it points at the key we just inserted - if the
 169          * decrement was unnecessary, bch2_btree_and_journal_iter_peek() will
 170          * handle that:
 171          */
 172         list_for_each_entry(iter, &c->journal_iters, journal.list)
 173                 if (iter->journal.idx == gap_end)
 174                         iter->journal.idx = keys->gap - 1;
 175 }
 176
 177 static void journal_iters_move_gap(struct bch_fs *c, size_t old_gap, size_t new_gap)
 178 {
 179         struct journal_keys *keys = &c->journal_keys;
 180         struct journal_iter *iter;
 181         size_t gap_size = keys->size - keys->nr;
 182
 183         list_for_each_entry(iter, &c->journal_iters, list) {
 184                 if (iter->idx > old_gap)
 185                         iter->idx -= gap_size;
 186                 if (iter->idx >= new_gap)
 187                         iter->idx += gap_size;
 188         }
 189 }
 190
 191 int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
 192                                  unsigned level, struct bkey_i *k)
 193 {
 194         struct journal_key n = {
 195                 .btree_id       = id,
 196                 .level          = level,
 197                 .k              = k,
 198                 .allocated      = true,
 199                 /*
 200                  * Ensure these keys are done last by journal replay, to unblock
 201                  * journal reclaim:
 202                  */
 203                 .journal_seq    = U32_MAX,
 204         };
 205         struct journal_keys *keys = &c->journal_keys;
 206         size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
 207
 208         BUG_ON(test_bit(BCH_FS_RW, &c->flags));
 209
 210         if (idx < keys->size &&
 211             journal_key_cmp(&n, &keys->d[idx]) == 0) {
 212                 if (keys->d[idx].allocated)
 213                         kfree(keys->d[idx].k);
 214                 keys->d[idx] = n;
 215                 return 0;
 216         }
 217
 218         if (idx > keys->gap)
 219                 idx -= keys->size - keys->nr;
 220
 221         if (keys->nr == keys->size) {
 222                 struct journal_keys new_keys = {
 223                         .nr                     = keys->nr,
 224                         .size                   = max_t(size_t, keys->size, 8) * 2,
 225                 };
 226
 227                 new_keys.d = kvmalloc_array(new_keys.size, sizeof(new_keys.d[0]), GFP_KERNEL);
 228                 if (!new_keys.d) {
 229                         bch_err(c, "%s: error allocating new key array (size %zu)",
 230                                 __func__, new_keys.size);
 231                         return -BCH_ERR_ENOMEM_journal_key_insert;
 232                 }
 233
 234                 /* Since @keys was full, there was no gap: */
 235                 memcpy(new_keys.d, keys->d, sizeof(keys->d[0]) * keys->nr);
 236                 kvfree(keys->d);
 237                 *keys = new_keys;
 238
 239                 /* And now the gap is at the end: */
 240                 keys->gap = keys->nr;
 241         }
 242
 243         journal_iters_move_gap(c, keys->gap, idx);
 244
 245         move_gap(keys->d, keys->nr, keys->size, keys->gap, idx);
 246         keys->gap = idx;
 247
 248         keys->nr++;
 249         keys->d[keys->gap++] = n;
 250
 251         journal_iters_fix(c);
 252
 253         return 0;
 254 }
 255
 256 /*
 257  * Can only be used from the recovery thread while we're still RO - can't be
 258  * used once we've got RW, as journal_keys is at that point used by multiple
 259  * threads:
 260  */
 261 int bch2_journal_key_insert(struct bch_fs *c, enum btree_id id,
 262                             unsigned level, struct bkey_i *k)
 263 {
 264         struct bkey_i *n;
 265         int ret;
 266
 267         n = kmalloc(bkey_bytes(&k->k), GFP_KERNEL);
 268         if (!n)
 269                 return -BCH_ERR_ENOMEM_journal_key_insert;
 270
 271         bkey_copy(n, k);
 272         ret = bch2_journal_key_insert_take(c, id, level, n);
 273         if (ret)
 274                 kfree(n);
 275         return ret;
 276 }
 277
 278 int bch2_journal_key_delete(struct bch_fs *c, enum btree_id id,
 279                             unsigned level, struct bpos pos)
 280 {
 281         struct bkey_i whiteout;
 282
 283         bkey_init(&whiteout.k);
 284         whiteout.k.p = pos;
 285
 286         return bch2_journal_key_insert(c, id, level, &whiteout);
 287 }
 288
 289 void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
 290                                   unsigned level, struct bpos pos)
 291 {
 292         struct journal_keys *keys = &c->journal_keys;
 293         size_t idx = bch2_journal_key_search(keys, btree, level, pos);
 294
 295         if (idx < keys->size &&
 296             keys->d[idx].btree_id       == btree &&
 297             keys->d[idx].level          == level &&
 298             bpos_eq(keys->d[idx].k->k.p, pos))
 299                 keys->d[idx].overwritten = true;
 300 }
 301
 302 static void bch2_journal_iter_advance(struct journal_iter *iter)
 303 {
 304         if (iter->idx < iter->keys->size) {
 305                 iter->idx++;
 306                 if (iter->idx == iter->keys->gap)
 307                         iter->idx += iter->keys->size - iter->keys->nr;
 308         }
 309 }
 310
 311 static struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
 312 {
 313         struct journal_key *k = iter->keys->d + iter->idx;
 314
 315         while (k < iter->keys->d + iter->keys->size &&
 316                k->btree_id      == iter->btree_id &&
 317                k->level         == iter->level) {
 318                 if (!k->overwritten)
 319                         return bkey_i_to_s_c(k->k);
 320
 321                 bch2_journal_iter_advance(iter);
 322                 k = iter->keys->d + iter->idx;
 323         }
 324
 325         return bkey_s_c_null;
 326 }
 327
 328 static void bch2_journal_iter_exit(struct journal_iter *iter)
 329 {
 330         list_del(&iter->list);
 331 }
 332
 333 static void bch2_journal_iter_init(struct bch_fs *c,
 334                                    struct journal_iter *iter,
 335                                    enum btree_id id, unsigned level,
 336                                    struct bpos pos)
 337 {
 338         iter->btree_id  = id;
 339         iter->level     = level;
 340         iter->keys      = &c->journal_keys;
 341         iter->idx       = bch2_journal_key_search(&c->journal_keys, id, level, pos);
 342 }
 343
 344 static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
 345 {
 346         return bch2_btree_node_iter_peek_unpack(&iter->node_iter,
 347                                                 iter->b, &iter->unpacked);
 348 }
 349
 350 static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
 351 {
 352         bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
 353 }
 354
 355 void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
 356 {
 357         if (bpos_eq(iter->pos, SPOS_MAX))
 358                 iter->at_end = true;
 359         else
 360                 iter->pos = bpos_successor(iter->pos);
 361 }
 362
 363 struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
 364 {
 365         struct bkey_s_c btree_k, journal_k, ret;
 366 again:
 367         if (iter->at_end)
 368                 return bkey_s_c_null;
 369
 370         while ((btree_k = bch2_journal_iter_peek_btree(iter)).k &&
 371                bpos_lt(btree_k.k->p, iter->pos))
 372                 bch2_journal_iter_advance_btree(iter);
 373
 374         while ((journal_k = bch2_journal_iter_peek(&iter->journal)).k &&
 375                bpos_lt(journal_k.k->p, iter->pos))
 376                 bch2_journal_iter_advance(&iter->journal);
 377
 378         ret = journal_k.k &&
 379                 (!btree_k.k || bpos_le(journal_k.k->p, btree_k.k->p))
 380                 ? journal_k
 381                 : btree_k;
 382
 383         if (ret.k && iter->b && bpos_gt(ret.k->p, iter->b->data->max_key))
 384                 ret = bkey_s_c_null;
 385
 386         if (ret.k) {
 387                 iter->pos = ret.k->p;
 388                 if (bkey_deleted(ret.k)) {
 389                         bch2_btree_and_journal_iter_advance(iter);
 390                         goto again;
 391                 }
 392         } else {
 393                 iter->pos = SPOS_MAX;
 394                 iter->at_end = true;
 395         }
 396
 397         return ret;
 398 }
 399
 400 void bch2_btree_and_journal_iter_exit(struct btree_and_journal_iter *iter)
 401 {
 402         bch2_journal_iter_exit(&iter->journal);
 403 }
 404
 405 void __bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
 406                                                   struct bch_fs *c,
 407                                                   struct btree *b,
 408                                                   struct btree_node_iter node_iter,
 409                                                   struct bpos pos)
 410 {
 411         memset(iter, 0, sizeof(*iter));
 412
 413         iter->b = b;
 414         iter->node_iter = node_iter;
 415         bch2_journal_iter_init(c, &iter->journal, b->c.btree_id, b->c.level, pos);
 416         INIT_LIST_HEAD(&iter->journal.list);
 417         iter->pos = b->data->min_key;
 418         iter->at_end = false;
 419 }
 420
 421 /*
 422  * this version is used by btree_gc before filesystem has gone RW and
 423  * multithreaded, so uses the journal_iters list:
 424  */
 425 void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
 426                                                 struct bch_fs *c,
 427                                                 struct btree *b)
 428 {
 429         struct btree_node_iter node_iter;
 430
 431         bch2_btree_node_iter_init_from_start(&node_iter, b);
 432         __bch2_btree_and_journal_iter_init_node_iter(iter, c, b, node_iter, b->data->min_key);
 433         list_add(&iter->journal.list, &c->journal_iters);
 434 }
 435
 436 /* sort and dedup all keys in the journal: */
 437
 438 void bch2_journal_entries_free(struct bch_fs *c)
 439 {
 440         struct journal_replay **i;
 441         struct genradix_iter iter;
 442
 443         genradix_for_each(&c->journal_entries, iter, i)
 444                 if (*i)
 445                         kvpfree(*i, offsetof(struct journal_replay, j) +
 446                                 vstruct_bytes(&(*i)->j));
 447         genradix_free(&c->journal_entries);
 448 }
 449
 450 /*
 451  * When keys compare equal, oldest compares first:
 452  */
 453 static int journal_sort_key_cmp(const void *_l, const void *_r)
 454 {
 455         const struct journal_key *l = _l;
 456         const struct journal_key *r = _r;
 457
 458         return  journal_key_cmp(l, r) ?:
 459                 cmp_int(l->journal_seq, r->journal_seq) ?:
 460                 cmp_int(l->journal_offset, r->journal_offset);
 461 }
 462
 463 void bch2_journal_keys_free(struct journal_keys *keys)
 464 {
 465         struct journal_key *i;
 466
 467         move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
 468         keys->gap = keys->nr;
 469
 470         for (i = keys->d; i < keys->d + keys->nr; i++)
 471                 if (i->allocated)
 472                         kfree(i->k);
 473
 474         kvfree(keys->d);
 475         keys->d = NULL;
 476         keys->nr = keys->gap = keys->size = 0;
 477 }
 478
 479 static void __journal_keys_sort(struct journal_keys *keys)
 480 {
 481         struct journal_key *src, *dst;
 482
 483         sort(keys->d, keys->nr, sizeof(keys->d[0]), journal_sort_key_cmp, NULL);
 484
 485         src = dst = keys->d;
 486         while (src < keys->d + keys->nr) {
 487                 while (src + 1 < keys->d + keys->nr &&
 488                        src[0].btree_id  == src[1].btree_id &&
 489                        src[0].level     == src[1].level &&
 490                        bpos_eq(src[0].k->k.p, src[1].k->k.p))
 491                         src++;
 492
 493                 *dst++ = *src++;
 494         }
 495
 496         keys->nr = dst - keys->d;
 497 }
 498
 499 static int journal_keys_sort(struct bch_fs *c)
 500 {
 501         struct genradix_iter iter;
 502         struct journal_replay *i, **_i;
 503         struct jset_entry *entry;
 504         struct bkey_i *k;
 505         struct journal_keys *keys = &c->journal_keys;
 506         size_t nr_keys = 0, nr_read = 0;
 507
 508         genradix_for_each(&c->journal_entries, iter, _i) {
 509                 i = *_i;
 510
 511                 if (!i || i->ignore)
 512                         continue;
 513
 514                 for_each_jset_key(k, entry, &i->j)
 515                         nr_keys++;
 516         }
 517
 518         if (!nr_keys)
 519                 return 0;
 520
 521         keys->size = roundup_pow_of_two(nr_keys);
 522
 523         keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
 524         if (!keys->d) {
 525                 bch_err(c, "Failed to allocate buffer for sorted journal keys (%zu keys); trying slowpath",
 526                         nr_keys);
 527
 528                 do {
 529                         keys->size >>= 1;
 530                         keys->d = kvmalloc_array(keys->size, sizeof(keys->d[0]), GFP_KERNEL);
 531                 } while (!keys->d && keys->size > nr_keys / 8);
 532
 533                 if (!keys->d) {
 534                         bch_err(c, "Failed to allocate %zu size buffer for sorted journal keys; exiting",
 535                                 keys->size);
 536                         return -BCH_ERR_ENOMEM_journal_keys_sort;
 537                 }
 538         }
 539
 540         genradix_for_each(&c->journal_entries, iter, _i) {
 541                 i = *_i;
 542
 543                 if (!i || i->ignore)
 544                         continue;
 545
 546                 cond_resched();
 547
 548                 for_each_jset_key(k, entry, &i->j) {
 549                         if (keys->nr == keys->size) {
 550                                 __journal_keys_sort(keys);
 551
 552                                 if (keys->nr > keys->size * 7 / 8) {
 553                                         bch_err(c, "Too many journal keys for slowpath; have %zu compacted, buf size %zu, processed %zu/%zu",
 554                                                 keys->nr, keys->size, nr_read, nr_keys);
 555                                         return -BCH_ERR_ENOMEM_journal_keys_sort;
 556                                 }
 557                         }
 558
 559                         keys->d[keys->nr++] = (struct journal_key) {
 560                                 .btree_id       = entry->btree_id,
 561                                 .level          = entry->level,
 562                                 .k              = k,
 563                                 .journal_seq    = le64_to_cpu(i->j.seq),
 564                                 .journal_offset = k->_data - i->j._data,
 565                         };
 566
 567                         nr_read++;
 568                 }
 569         }
 570
 571         __journal_keys_sort(keys);
 572         keys->gap = keys->nr;
 573
 574         bch_verbose(c, "Journal keys: %zu read, %zu after sorting and compacting", nr_keys, keys->nr);
 575         return 0;
 576 }
 577
 578 /* journal replay: */
 579
 580 static void replay_now_at(struct journal *j, u64 seq)
 581 {
 582         BUG_ON(seq < j->replay_journal_seq);
 583
 584         seq = min(seq, j->replay_journal_seq_end);
 585
 586         while (j->replay_journal_seq < seq)
 587                 bch2_journal_pin_put(j, j->replay_journal_seq++);
 588 }
 589
 590 static int bch2_journal_replay_key(struct btree_trans *trans,
 591                                    struct journal_key *k)
 592 {
 593         struct btree_iter iter;
 594         unsigned iter_flags =
 595                 BTREE_ITER_INTENT|
 596                 BTREE_ITER_NOT_EXTENTS;
 597         unsigned update_flags = BTREE_TRIGGER_NORUN;
 598         int ret;
 599
 600         /*
 601          * BTREE_UPDATE_KEY_CACHE_RECLAIM disables key cache lookup/update to
 602          * keep the key cache coherent with the underlying btree. Nothing
 603          * besides the allocator is doing updates yet so we don't need key cache
 604          * coherency for non-alloc btrees, and key cache fills for snapshots
 605          * btrees use BTREE_ITER_FILTER_SNAPSHOTS, which isn't available until
 606          * the snapshots recovery pass runs.
 607          */
 608         if (!k->level && k->btree_id == BTREE_ID_alloc)
 609                 iter_flags |= BTREE_ITER_CACHED;
 610         else
 611                 update_flags |= BTREE_UPDATE_KEY_CACHE_RECLAIM;
 612
 613         bch2_trans_node_iter_init(trans, &iter, k->btree_id, k->k->k.p,
 614                                   BTREE_MAX_DEPTH, k->level,
 615                                   iter_flags);
 616         ret = bch2_btree_iter_traverse(&iter);
 617         if (ret)
 618                 goto out;
 619
 620         /* Must be checked with btree locked: */
 621         if (k->overwritten)
 622                 goto out;
 623
 624         ret = bch2_trans_update(trans, &iter, k->k, update_flags);
 625 out:
 626         bch2_trans_iter_exit(trans, &iter);
 627         return ret;
 628 }
 629
 630 static int journal_sort_seq_cmp(const void *_l, const void *_r)
 631 {
 632         const struct journal_key *l = *((const struct journal_key **)_l);
 633         const struct journal_key *r = *((const struct journal_key **)_r);
 634
 635         return cmp_int(l->journal_seq, r->journal_seq);
 636 }
 637
 638 static int bch2_journal_replay(struct bch_fs *c)
 639 {
 640         struct journal_keys *keys = &c->journal_keys;
 641         struct journal_key **keys_sorted, *k;
 642         struct journal *j = &c->journal;
 643         u64 start_seq   = c->journal_replay_seq_start;
 644         u64 end_seq     = c->journal_replay_seq_start;
 645         size_t i;
 646         int ret;
 647
 648         move_gap(keys->d, keys->nr, keys->size, keys->gap, keys->nr);
 649         keys->gap = keys->nr;
 650
 651         keys_sorted = kvmalloc_array(sizeof(*keys_sorted), keys->nr, GFP_KERNEL);
 652         if (!keys_sorted)
 653                 return -BCH_ERR_ENOMEM_journal_replay;
 654
 655         for (i = 0; i < keys->nr; i++)
 656                 keys_sorted[i] = &keys->d[i];
 657
 658         sort(keys_sorted, keys->nr,
 659              sizeof(keys_sorted[0]),
 660              journal_sort_seq_cmp, NULL);
 661
 662         if (keys->nr) {
 663                 ret = bch2_journal_log_msg(c, "Starting journal replay (%zu keys in entries %llu-%llu)",
 664                                            keys->nr, start_seq, end_seq);
 665                 if (ret)
 666                         goto err;
 667         }
 668
 669         for (i = 0; i < keys->nr; i++) {
 670                 k = keys_sorted[i];
 671
 672                 cond_resched();
 673
 674                 replay_now_at(j, k->journal_seq);
 675
 676                 ret = bch2_trans_do(c, NULL, NULL,
 677                                     BTREE_INSERT_LAZY_RW|
 678                                     BTREE_INSERT_NOFAIL|
 679                                     (!k->allocated
 680                                      ? BTREE_INSERT_JOURNAL_REPLAY|BCH_WATERMARK_reclaim
 681                                      : 0),
 682                              bch2_journal_replay_key(&trans, k));
 683                 if (ret) {
 684                         bch_err(c, "journal replay: error while replaying key at btree %s level %u: %s",
 685                                 bch2_btree_ids[k->btree_id], k->level, bch2_err_str(ret));
 686                         goto err;
 687                 }
 688         }
 689
 690         replay_now_at(j, j->replay_journal_seq_end);
 691         j->replay_journal_seq = 0;
 692
 693         bch2_journal_set_replay_done(j);
 694         bch2_journal_flush_all_pins(j);
 695         ret = bch2_journal_error(j);
 696
 697         if (keys->nr && !ret)
 698                 bch2_journal_log_msg(c, "journal replay finished");
 699 err:
 700         kvfree(keys_sorted);
 701
 702         if (ret)
 703                 bch_err_fn(c, ret);
 704         return ret;
 705 }
 706
 707 /* journal replay early: */
 708
 709 static int journal_replay_entry_early(struct bch_fs *c,
 710                                       struct jset_entry *entry)
 711 {
 712         int ret = 0;
 713
 714         switch (entry->type) {
 715         case BCH_JSET_ENTRY_btree_root: {
 716                 struct btree_root *r;
 717
 718                 while (entry->btree_id >= c->btree_roots_extra.nr + BTREE_ID_NR) {
 719                         ret = darray_push(&c->btree_roots_extra, (struct btree_root) { NULL });
 720                         if (ret)
 721                                 return ret;
 722                 }
 723
 724                 r = bch2_btree_id_root(c, entry->btree_id);
 725
 726                 if (entry->u64s) {
 727                         r->level = entry->level;
 728                         bkey_copy(&r->key, &entry->start[0]);
 729                         r->error = 0;
 730                 } else {
 731                         r->error = -EIO;
 732                 }
 733                 r->alive = true;
 734                 break;
 735         }
 736         case BCH_JSET_ENTRY_usage: {
 737                 struct jset_entry_usage *u =
 738                         container_of(entry, struct jset_entry_usage, entry);
 739
 740                 switch (entry->btree_id) {
 741                 case BCH_FS_USAGE_reserved:
 742                         if (entry->level < BCH_REPLICAS_MAX)
 743                                 c->usage_base->persistent_reserved[entry->level] =
 744                                         le64_to_cpu(u->v);
 745                         break;
 746                 case BCH_FS_USAGE_inodes:
 747                         c->usage_base->nr_inodes = le64_to_cpu(u->v);
 748                         break;
 749                 case BCH_FS_USAGE_key_version:
 750                         atomic64_set(&c->key_version,
 751                                      le64_to_cpu(u->v));
 752                         break;
 753                 }
 754
 755                 break;
 756         }
 757         case BCH_JSET_ENTRY_data_usage: {
 758                 struct jset_entry_data_usage *u =
 759                         container_of(entry, struct jset_entry_data_usage, entry);
 760
 761                 ret = bch2_replicas_set_usage(c, &u->r,
 762                                               le64_to_cpu(u->v));
 763                 break;
 764         }
 765         case BCH_JSET_ENTRY_dev_usage: {
 766                 struct jset_entry_dev_usage *u =
 767                         container_of(entry, struct jset_entry_dev_usage, entry);
 768                 struct bch_dev *ca = bch_dev_bkey_exists(c, le32_to_cpu(u->dev));
 769                 unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 770
 771                 ca->usage_base->buckets_ec              = le64_to_cpu(u->buckets_ec);
 772
 773                 for (i = 0; i < min_t(unsigned, nr_types, BCH_DATA_NR); i++) {
 774                         ca->usage_base->d[i].buckets    = le64_to_cpu(u->d[i].buckets);
 775                         ca->usage_base->d[i].sectors    = le64_to_cpu(u->d[i].sectors);
 776                         ca->usage_base->d[i].fragmented = le64_to_cpu(u->d[i].fragmented);
 777                 }
 778
 779                 break;
 780         }
 781         case BCH_JSET_ENTRY_blacklist: {
 782                 struct jset_entry_blacklist *bl_entry =
 783                         container_of(entry, struct jset_entry_blacklist, entry);
 784
 785                 ret = bch2_journal_seq_blacklist_add(c,
 786                                 le64_to_cpu(bl_entry->seq),
 787                                 le64_to_cpu(bl_entry->seq) + 1);
 788                 break;
 789         }
 790         case BCH_JSET_ENTRY_blacklist_v2: {
 791                 struct jset_entry_blacklist_v2 *bl_entry =
 792                         container_of(entry, struct jset_entry_blacklist_v2, entry);
 793
 794                 ret = bch2_journal_seq_blacklist_add(c,
 795                                 le64_to_cpu(bl_entry->start),
 796                                 le64_to_cpu(bl_entry->end) + 1);
 797                 break;
 798         }
 799         case BCH_JSET_ENTRY_clock: {
 800                 struct jset_entry_clock *clock =
 801                         container_of(entry, struct jset_entry_clock, entry);
 802
 803                 atomic64_set(&c->io_clock[clock->rw].now, le64_to_cpu(clock->time));
 804         }
 805         }
 806
 807         return ret;
 808 }
 809
 810 static int journal_replay_early(struct bch_fs *c,
 811                                 struct bch_sb_field_clean *clean)
 812 {
 813         struct jset_entry *entry;
 814         int ret;
 815
 816         if (clean) {
 817                 for (entry = clean->start;
 818                      entry != vstruct_end(&clean->field);
 819                      entry = vstruct_next(entry)) {
 820                         ret = journal_replay_entry_early(c, entry);
 821                         if (ret)
 822                                 return ret;
 823                 }
 824         } else {
 825                 struct genradix_iter iter;
 826                 struct journal_replay *i, **_i;
 827
 828                 genradix_for_each(&c->journal_entries, iter, _i) {
 829                         i = *_i;
 830
 831                         if (!i || i->ignore)
 832                                 continue;
 833
 834                         vstruct_for_each(&i->j, entry) {
 835                                 ret = journal_replay_entry_early(c, entry);
 836                                 if (ret)
 837                                         return ret;
 838                         }
 839                 }
 840         }
 841
 842         bch2_fs_usage_initialize(c);
 843
 844         return 0;
 845 }
 846
 847 /* sb clean section: */
 848
 849 static struct bkey_i *btree_root_find(struct bch_fs *c,
 850                                       struct bch_sb_field_clean *clean,
 851                                       struct jset *j,
 852                                       enum btree_id id, unsigned *level)
 853 {
 854         struct bkey_i *k;
 855         struct jset_entry *entry, *start, *end;
 856
 857         if (clean) {
 858                 start = clean->start;
 859                 end = vstruct_end(&clean->field);
 860         } else {
 861                 start = j->start;
 862                 end = vstruct_last(j);
 863         }
 864
 865         for (entry = start; entry < end; entry = vstruct_next(entry))
 866                 if (entry->type == BCH_JSET_ENTRY_btree_root &&
 867                     entry->btree_id == id)
 868                         goto found;
 869
 870         return NULL;
 871 found:
 872         if (!entry->u64s)
 873                 return ERR_PTR(-EINVAL);
 874
 875         k = entry->start;
 876         *level = entry->level;
 877         return k;
 878 }
 879
 880 static int verify_superblock_clean(struct bch_fs *c,
 881                                    struct bch_sb_field_clean **cleanp,
 882                                    struct jset *j)
 883 {
 884         unsigned i;
 885         struct bch_sb_field_clean *clean = *cleanp;
 886         struct printbuf buf1 = PRINTBUF;
 887         struct printbuf buf2 = PRINTBUF;
 888         int ret = 0;
 889
 890         if (mustfix_fsck_err_on(j->seq != clean->journal_seq, c,
 891                         "superblock journal seq (%llu) doesn't match journal (%llu) after clean shutdown",
 892                         le64_to_cpu(clean->journal_seq),
 893                         le64_to_cpu(j->seq))) {
 894                 kfree(clean);
 895                 *cleanp = NULL;
 896                 return 0;
 897         }
 898
 899         for (i = 0; i < BTREE_ID_NR; i++) {
 900                 struct bkey_i *k1, *k2;
 901                 unsigned l1 = 0, l2 = 0;
 902
 903                 k1 = btree_root_find(c, clean, NULL, i, &l1);
 904                 k2 = btree_root_find(c, NULL, j, i, &l2);
 905
 906                 if (!k1 && !k2)
 907                         continue;
 908
 909                 printbuf_reset(&buf1);
 910                 printbuf_reset(&buf2);
 911
 912                 if (k1)
 913                         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(k1));
 914                 else
 915                         prt_printf(&buf1, "(none)");
 916
 917                 if (k2)
 918                         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(k2));
 919                 else
 920                         prt_printf(&buf2, "(none)");
 921
 922                 mustfix_fsck_err_on(!k1 || !k2 ||
 923                                     IS_ERR(k1) ||
 924                                     IS_ERR(k2) ||
 925                                     k1->k.u64s != k2->k.u64s ||
 926                                     memcmp(k1, k2, bkey_bytes(&k1->k)) ||
 927                                     l1 != l2, c,
 928                         "superblock btree root %u doesn't match journal after clean shutdown\n"
 929                         "sb:      l=%u %s\n"
 930                         "journal: l=%u %s\n", i,
 931                         l1, buf1.buf,
 932                         l2, buf2.buf);
 933         }
 934 fsck_err:
 935         printbuf_exit(&buf2);
 936         printbuf_exit(&buf1);
 937         return ret;
 938 }
 939
 940 static struct bch_sb_field_clean *read_superblock_clean(struct bch_fs *c)
 941 {
 942         struct bch_sb_field_clean *clean, *sb_clean;
 943         int ret;
 944
 945         mutex_lock(&c->sb_lock);
 946         sb_clean = bch2_sb_get_clean(c->disk_sb.sb);
 947
 948         if (fsck_err_on(!sb_clean, c,
 949                         "superblock marked clean but clean section not present")) {
 950                 SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
 951                 c->sb.clean = false;
 952                 mutex_unlock(&c->sb_lock);
 953                 return NULL;
 954         }
 955
 956         clean = kmemdup(sb_clean, vstruct_bytes(&sb_clean->field),
 957                         GFP_KERNEL);
 958         if (!clean) {
 959                 mutex_unlock(&c->sb_lock);
 960                 return ERR_PTR(-BCH_ERR_ENOMEM_read_superblock_clean);
 961         }
 962
 963         ret = bch2_sb_clean_validate_late(c, clean, READ);
 964         if (ret) {
 965                 mutex_unlock(&c->sb_lock);
 966                 return ERR_PTR(ret);
 967         }
 968
 969         mutex_unlock(&c->sb_lock);
 970
 971         return clean;
 972 fsck_err:
 973         mutex_unlock(&c->sb_lock);
 974         return ERR_PTR(ret);
 975 }
 976
 977 static bool btree_id_is_alloc(enum btree_id id)
 978 {
 979         switch (id) {
 980         case BTREE_ID_alloc:
 981         case BTREE_ID_backpointers:
 982         case BTREE_ID_need_discard:
 983         case BTREE_ID_freespace:
 984         case BTREE_ID_bucket_gens:
 985                 return true;
 986         default:
 987                 return false;
 988         }
 989 }
 990
 991 static int read_btree_roots(struct bch_fs *c)
 992 {
 993         unsigned i;
 994         int ret = 0;
 995
 996         for (i = 0; i < btree_id_nr_alive(c); i++) {
 997                 struct btree_root *r = bch2_btree_id_root(c, i);
 998
 999                 if (!r->alive)
1000                         continue;
1001
1002                 if (btree_id_is_alloc(i) &&
1003                     c->opts.reconstruct_alloc) {
1004                         c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
1005                         continue;
1006                 }
1007
1008                 if (r->error) {
1009                         __fsck_err(c, btree_id_is_alloc(i)
1010                                    ? FSCK_CAN_IGNORE : 0,
1011                                    "invalid btree root %s",
1012                                    bch2_btree_ids[i]);
1013                         if (i == BTREE_ID_alloc)
1014                                 c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
1015                 }
1016
1017                 ret = bch2_btree_root_read(c, i, &r->key, r->level);
1018                 if (ret) {
1019                         __fsck_err(c,
1020                                    btree_id_is_alloc(i)
1021                                    ? FSCK_CAN_IGNORE : 0,
1022                                    "error reading btree root %s",
1023                                    bch2_btree_ids[i]);
1024                         if (btree_id_is_alloc(i))
1025                                 c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
1026                 }
1027         }
1028
1029         for (i = 0; i < BTREE_ID_NR; i++) {
1030                 struct btree_root *r = bch2_btree_id_root(c, i);
1031
1032                 if (!r->b) {
1033                         r->alive = false;
1034                         r->level = 0;
1035                         bch2_btree_root_alloc(c, i);
1036                 }
1037         }
1038 fsck_err:
1039         return ret;
1040 }
1041
1042 static int bch2_initialize_subvolumes(struct bch_fs *c)
1043 {
1044         struct bkey_i_snapshot_tree     root_tree;
1045         struct bkey_i_snapshot          root_snapshot;
1046         struct bkey_i_subvolume         root_volume;
1047         int ret;
1048
1049         bkey_snapshot_tree_init(&root_tree.k_i);
1050         root_tree.k.p.offset            = 1;
1051         root_tree.v.master_subvol       = cpu_to_le32(1);
1052         root_tree.v.root_snapshot       = cpu_to_le32(U32_MAX);
1053
1054         bkey_snapshot_init(&root_snapshot.k_i);
1055         root_snapshot.k.p.offset = U32_MAX;
1056         root_snapshot.v.flags   = 0;
1057         root_snapshot.v.parent  = 0;
1058         root_snapshot.v.subvol  = cpu_to_le32(BCACHEFS_ROOT_SUBVOL);
1059         root_snapshot.v.tree    = cpu_to_le32(1);
1060         SET_BCH_SNAPSHOT_SUBVOL(&root_snapshot.v, true);
1061
1062         bkey_subvolume_init(&root_volume.k_i);
1063         root_volume.k.p.offset = BCACHEFS_ROOT_SUBVOL;
1064         root_volume.v.flags     = 0;
1065         root_volume.v.snapshot  = cpu_to_le32(U32_MAX);
1066         root_volume.v.inode     = cpu_to_le64(BCACHEFS_ROOT_INO);
1067
1068         ret =   bch2_btree_insert(c, BTREE_ID_snapshot_trees,
1069                                   &root_tree.k_i,
1070                                   NULL, NULL, 0) ?:
1071                 bch2_btree_insert(c, BTREE_ID_snapshots,
1072                                   &root_snapshot.k_i,
1073                                   NULL, NULL, 0) ?:
1074                 bch2_btree_insert(c, BTREE_ID_subvolumes,
1075                                   &root_volume.k_i,
1076                                   NULL, NULL, 0);
1077         if (ret)
1078                 bch_err_fn(c, ret);
1079         return ret;
1080 }
1081
1082 static int __bch2_fs_upgrade_for_subvolumes(struct btree_trans *trans)
1083 {
1084         struct btree_iter iter;
1085         struct bkey_s_c k;
1086         struct bch_inode_unpacked inode;
1087         int ret;
1088
1089         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
1090                                SPOS(0, BCACHEFS_ROOT_INO, U32_MAX), 0);
1091         ret = bkey_err(k);
1092         if (ret)
1093                 return ret;
1094
1095         if (!bkey_is_inode(k.k)) {
1096                 bch_err(trans->c, "root inode not found");
1097                 ret = -BCH_ERR_ENOENT_inode;
1098                 goto err;
1099         }
1100
1101         ret = bch2_inode_unpack(k, &inode);
1102         BUG_ON(ret);
1103
1104         inode.bi_subvol = BCACHEFS_ROOT_SUBVOL;
1105
1106         ret = bch2_inode_write(trans, &iter, &inode);
1107 err:
1108         bch2_trans_iter_exit(trans, &iter);
1109         return ret;
1110 }
1111
1112 /* set bi_subvol on root inode */
1113 noinline_for_stack
1114 static int bch2_fs_upgrade_for_subvolumes(struct bch_fs *c)
1115 {
1116         int ret = bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
1117                                 __bch2_fs_upgrade_for_subvolumes(&trans));
1118         if (ret)
1119                 bch_err_fn(c, ret);
1120         return ret;
1121 }
1122
1123 static void check_version_upgrade(struct bch_fs *c)
1124 {
1125         unsigned latest_compatible = bch2_version_compatible(c->sb.version);
1126         unsigned latest_version = bcachefs_metadata_version_current;
1127         unsigned old_version = c->sb.version_upgrade_complete ?: c->sb.version;
1128         unsigned new_version = 0;
1129         u64 recovery_passes;
1130
1131         if (old_version < bcachefs_metadata_required_upgrade_below) {
1132                 if (c->opts.version_upgrade == BCH_VERSION_UPGRADE_incompatible ||
1133                     latest_compatible < bcachefs_metadata_required_upgrade_below)
1134                         new_version = latest_version;
1135                 else
1136                         new_version = latest_compatible;
1137         } else {
1138                 switch (c->opts.version_upgrade) {
1139                 case BCH_VERSION_UPGRADE_compatible:
1140                         new_version = latest_compatible;
1141                         break;
1142                 case BCH_VERSION_UPGRADE_incompatible:
1143                         new_version = latest_version;
1144                         break;
1145                 case BCH_VERSION_UPGRADE_none:
1146                         new_version = old_version;
1147                         break;
1148                 }
1149         }
1150
1151         if (new_version > old_version) {
1152                 struct printbuf buf = PRINTBUF;
1153
1154                 if (old_version < bcachefs_metadata_required_upgrade_below)
1155                         prt_str(&buf, "Version upgrade required:\n");
1156
1157                 if (old_version != c->sb.version) {
1158                         prt_str(&buf, "Version upgrade from ");
1159                         bch2_version_to_text(&buf, c->sb.version_upgrade_complete);
1160                         prt_str(&buf, " to ");
1161                         bch2_version_to_text(&buf, c->sb.version);
1162                         prt_str(&buf, " incomplete\n");
1163                 }
1164
1165                 prt_printf(&buf, "Doing %s version upgrade from ",
1166                            BCH_VERSION_MAJOR(old_version) != BCH_VERSION_MAJOR(new_version)
1167                            ? "incompatible" : "compatible");
1168                 bch2_version_to_text(&buf, old_version);
1169                 prt_str(&buf, " to ");
1170                 bch2_version_to_text(&buf, new_version);
1171                 prt_newline(&buf);
1172
1173                 recovery_passes = bch2_upgrade_recovery_passes(c, old_version, new_version);
1174                 if (recovery_passes) {
1175                         prt_str(&buf, "fsck required");
1176
1177                         c->recovery_passes_explicit |= recovery_passes;
1178                         c->opts.fix_errors = FSCK_FIX_yes;
1179                 }
1180
1181                 bch_info(c, "%s", buf.buf);
1182
1183                 mutex_lock(&c->sb_lock);
1184                 bch2_sb_upgrade(c, new_version);
1185                 mutex_unlock(&c->sb_lock);
1186
1187                 printbuf_exit(&buf);
1188         }
1189 }
1190
1191 static int bch2_check_allocations(struct bch_fs *c)
1192 {
1193         return bch2_gc(c, true, c->opts.norecovery);
1194 }
1195
1196 static int bch2_set_may_go_rw(struct bch_fs *c)
1197 {
1198         set_bit(BCH_FS_MAY_GO_RW, &c->flags);
1199         return 0;
1200 }
1201
1202 struct recovery_pass_fn {
1203         int             (*fn)(struct bch_fs *);
1204         const char      *name;
1205         unsigned        when;
1206 };
1207
1208 static struct recovery_pass_fn recovery_passes[] = {
1209 #define x(_fn, _when)   { .fn = bch2_##_fn, .name = #_fn, .when = _when },
1210         BCH_RECOVERY_PASSES()
1211 #undef x
1212 };
1213
1214 u64 bch2_fsck_recovery_passes(void)
1215 {
1216         u64 ret = 0;
1217
1218         for (unsigned i = 0; i < ARRAY_SIZE(recovery_passes); i++)
1219                 if (recovery_passes[i].when & PASS_FSCK)
1220                         ret |= BIT_ULL(i);
1221         return ret;
1222 }
1223
1224 static bool should_run_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pass)
1225 {
1226         struct recovery_pass_fn *p = recovery_passes + c->curr_recovery_pass;
1227
1228         if (c->opts.norecovery && pass > BCH_RECOVERY_PASS_snapshots_read)
1229                 return false;
1230         if (c->recovery_passes_explicit & BIT_ULL(pass))
1231                 return true;
1232         if ((p->when & PASS_FSCK) && c->opts.fsck)
1233                 return true;
1234         if ((p->when & PASS_UNCLEAN) && !c->sb.clean)
1235                 return true;
1236         if (p->when & PASS_ALWAYS)
1237                 return true;
1238         return false;
1239 }
1240
1241 static int bch2_run_recovery_pass(struct bch_fs *c, enum bch_recovery_pass pass)
1242 {
1243         int ret;
1244
1245         c->curr_recovery_pass = pass;
1246
1247         if (should_run_recovery_pass(c, pass)) {
1248                 struct recovery_pass_fn *p = recovery_passes + pass;
1249
1250                 if (!(p->when & PASS_SILENT))
1251                         printk(KERN_INFO bch2_log_msg(c, "%s..."), p->name);
1252                 ret = p->fn(c);
1253                 if (ret)
1254                         return ret;
1255                 if (!(p->when & PASS_SILENT))
1256                         printk(KERN_CONT " done\n");
1257         }
1258
1259         return 0;
1260 }
1261
1262 static int bch2_run_recovery_passes(struct bch_fs *c)
1263 {
1264         int ret = 0;
1265
1266         while (c->curr_recovery_pass < ARRAY_SIZE(recovery_passes)) {
1267                 ret = bch2_run_recovery_pass(c, c->curr_recovery_pass);
1268                 if (bch2_err_matches(ret, BCH_ERR_restart_recovery))
1269                         continue;
1270                 if (ret)
1271                         break;
1272                 c->curr_recovery_pass++;
1273         }
1274
1275         return ret;
1276 }
1277
1278 int bch2_fs_recovery(struct bch_fs *c)
1279 {
1280         struct bch_sb_field_clean *clean = NULL;
1281         struct jset *last_journal_entry = NULL;
1282         u64 last_seq, blacklist_seq, journal_seq;
1283         bool write_sb = false;
1284         int ret = 0;
1285
1286         if (c->sb.clean)
1287                 clean = read_superblock_clean(c);
1288         ret = PTR_ERR_OR_ZERO(clean);
1289         if (ret)
1290                 goto err;
1291
1292         if (c->sb.clean)
1293                 bch_info(c, "recovering from clean shutdown, journal seq %llu",
1294                          le64_to_cpu(clean->journal_seq));
1295         else
1296                 bch_info(c, "recovering from unclean shutdown");
1297
1298         if (!(c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))) {
1299                 bch_err(c, "feature new_extent_overwrite not set, filesystem no longer supported");
1300                 ret = -EINVAL;
1301                 goto err;
1302         }
1303
1304         if (!c->sb.clean &&
1305             !(c->sb.features & (1ULL << BCH_FEATURE_extents_above_btree_updates))) {
1306                 bch_err(c, "filesystem needs recovery from older version; run fsck from older bcachefs-tools to fix");
1307                 ret = -EINVAL;
1308                 goto err;
1309         }
1310
1311         if (!(c->sb.compat & (1ULL << BCH_COMPAT_bformat_overflow_done))) {
1312                 bch_err(c, "filesystem may have incompatible bkey formats; run fsck from the compat branch to fix");
1313                 ret = -EINVAL;
1314                 goto err;
1315         }
1316
1317         if (c->opts.fsck || !(c->opts.nochanges && c->opts.norecovery))
1318                 check_version_upgrade(c);
1319
1320         if (c->opts.fsck && c->opts.norecovery) {
1321                 bch_err(c, "cannot select both norecovery and fsck");
1322                 ret = -EINVAL;
1323                 goto err;
1324         }
1325
1326         ret = bch2_blacklist_table_initialize(c);
1327         if (ret) {
1328                 bch_err(c, "error initializing blacklist table");
1329                 goto err;
1330         }
1331
1332         if (!c->sb.clean || c->opts.fsck || c->opts.keep_journal) {
1333                 struct genradix_iter iter;
1334                 struct journal_replay **i;
1335
1336                 bch_verbose(c, "starting journal read");
1337                 ret = bch2_journal_read(c, &last_seq, &blacklist_seq, &journal_seq);
1338                 if (ret)
1339                         goto err;
1340
1341                 /*
1342                  * note: cmd_list_journal needs the blacklist table fully up to date so
1343                  * it can asterisk ignored journal entries:
1344                  */
1345                 if (c->opts.read_journal_only)
1346                         goto out;
1347
1348                 genradix_for_each_reverse(&c->journal_entries, iter, i)
1349                         if (*i && !(*i)->ignore) {
1350                                 last_journal_entry = &(*i)->j;
1351                                 break;
1352                         }
1353
1354                 if (mustfix_fsck_err_on(c->sb.clean &&
1355                                         last_journal_entry &&
1356                                         !journal_entry_empty(last_journal_entry), c,
1357                                 "filesystem marked clean but journal not empty")) {
1358                         c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
1359                         SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
1360                         c->sb.clean = false;
1361                 }
1362
1363                 if (!last_journal_entry) {
1364                         fsck_err_on(!c->sb.clean, c, "no journal entries found");
1365                         if (clean)
1366                                 goto use_clean;
1367
1368                         genradix_for_each_reverse(&c->journal_entries, iter, i)
1369                                 if (*i) {
1370                                         last_journal_entry = &(*i)->j;
1371                                         (*i)->ignore = false;
1372                                         break;
1373                                 }
1374                 }
1375
1376                 ret = journal_keys_sort(c);
1377                 if (ret)
1378                         goto err;
1379
1380                 if (c->sb.clean && last_journal_entry) {
1381                         ret = verify_superblock_clean(c, &clean,
1382                                                       last_journal_entry);
1383                         if (ret)
1384                                 goto err;
1385                 }
1386         } else {
1387 use_clean:
1388                 if (!clean) {
1389                         bch_err(c, "no superblock clean section found");
1390                         ret = -BCH_ERR_fsck_repair_impossible;
1391                         goto err;
1392
1393                 }
1394                 blacklist_seq = journal_seq = le64_to_cpu(clean->journal_seq) + 1;
1395         }
1396
1397         c->journal_replay_seq_start     = last_seq;
1398         c->journal_replay_seq_end       = blacklist_seq - 1;;
1399
1400         if (c->opts.reconstruct_alloc) {
1401                 c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
1402                 drop_alloc_keys(&c->journal_keys);
1403         }
1404
1405         zero_out_btree_mem_ptr(&c->journal_keys);
1406
1407         ret = journal_replay_early(c, clean);
1408         if (ret)
1409                 goto err;
1410
1411         /*
1412          * After an unclean shutdown, skip then next few journal sequence
1413          * numbers as they may have been referenced by btree writes that
1414          * happened before their corresponding journal writes - those btree
1415          * writes need to be ignored, by skipping and blacklisting the next few
1416          * journal sequence numbers:
1417          */
1418         if (!c->sb.clean)
1419                 journal_seq += 8;
1420
1421         if (blacklist_seq != journal_seq) {
1422                 ret =   bch2_journal_log_msg(c, "blacklisting entries %llu-%llu",
1423                                              blacklist_seq, journal_seq) ?:
1424                         bch2_journal_seq_blacklist_add(c,
1425                                         blacklist_seq, journal_seq);
1426                 if (ret) {
1427                         bch_err(c, "error creating new journal seq blacklist entry");
1428                         goto err;
1429                 }
1430         }
1431
1432         ret =   bch2_journal_log_msg(c, "starting journal at entry %llu, replaying %llu-%llu",
1433                                      journal_seq, last_seq, blacklist_seq - 1) ?:
1434                 bch2_fs_journal_start(&c->journal, journal_seq);
1435         if (ret)
1436                 goto err;
1437
1438         if (c->opts.reconstruct_alloc)
1439                 bch2_journal_log_msg(c, "dropping alloc info");
1440
1441         /*
1442          * Skip past versions that might have possibly been used (as nonces),
1443          * but hadn't had their pointers written:
1444          */
1445         if (c->sb.encryption_type && !c->sb.clean)
1446                 atomic64_add(1 << 16, &c->key_version);
1447
1448         ret = read_btree_roots(c);
1449         if (ret)
1450                 goto err;
1451
1452         if (c->opts.fsck &&
1453             (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) ||
1454              BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb)))
1455                 c->recovery_passes_explicit |= BIT_ULL(BCH_RECOVERY_PASS_check_topology);
1456
1457         ret = bch2_run_recovery_passes(c);
1458         if (ret)
1459                 goto err;
1460
1461         if (enabled_qtypes(c)) {
1462                 bch_verbose(c, "reading quotas");
1463                 ret = bch2_fs_quota_read(c);
1464                 if (ret)
1465                         goto err;
1466                 bch_verbose(c, "quotas done");
1467         }
1468
1469         mutex_lock(&c->sb_lock);
1470         if (BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb) != c->sb.version) {
1471                 SET_BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb, c->sb.version);
1472                 write_sb = true;
1473         }
1474
1475         if (!test_bit(BCH_FS_ERROR, &c->flags)) {
1476                 c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_alloc_info);
1477                 write_sb = true;
1478         }
1479
1480         if (c->opts.fsck &&
1481             !test_bit(BCH_FS_ERROR, &c->flags) &&
1482             !test_bit(BCH_FS_ERRORS_NOT_FIXED, &c->flags)) {
1483                 SET_BCH_SB_HAS_ERRORS(c->disk_sb.sb, 0);
1484                 SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, 0);
1485                 write_sb = true;
1486         }
1487
1488         if (write_sb)
1489                 bch2_write_super(c);
1490         mutex_unlock(&c->sb_lock);
1491
1492         if (!(c->sb.compat & (1ULL << BCH_COMPAT_extents_above_btree_updates_done)) ||
1493             !(c->sb.compat & (1ULL << BCH_COMPAT_bformat_overflow_done)) ||
1494             c->sb.version_min < bcachefs_metadata_version_btree_ptr_sectors_written) {
1495                 struct bch_move_stats stats;
1496
1497                 bch2_move_stats_init(&stats, "recovery");
1498
1499                 bch_info(c, "scanning for old btree nodes");
1500                 ret =   bch2_fs_read_write(c) ?:
1501                         bch2_scan_old_btree_nodes(c, &stats);
1502                 if (ret)
1503                         goto err;
1504                 bch_info(c, "scanning for old btree nodes done");
1505         }
1506
1507         if (c->journal_seq_blacklist_table &&
1508             c->journal_seq_blacklist_table->nr > 128)
1509                 queue_work(system_long_wq, &c->journal_seq_blacklist_gc_work);
1510
1511         ret = 0;
1512 out:
1513         set_bit(BCH_FS_FSCK_DONE, &c->flags);
1514         bch2_flush_fsck_errs(c);
1515
1516         if (!c->opts.keep_journal &&
1517             test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) {
1518                 bch2_journal_keys_free(&c->journal_keys);
1519                 bch2_journal_entries_free(c);
1520         }
1521         kfree(clean);
1522
1523         if (!ret && test_bit(BCH_FS_HAVE_DELETED_SNAPSHOTS, &c->flags)) {
1524                 bch2_fs_read_write_early(c);
1525                 bch2_delete_dead_snapshots_async(c);
1526         }
1527
1528         if (ret)
1529                 bch_err_fn(c, ret);
1530         return ret;
1531 err:
1532 fsck_err:
1533         bch2_fs_emergency_read_only(c);
1534         goto out;
1535 }
1536
1537 int bch2_fs_initialize(struct bch_fs *c)
1538 {
1539         struct bch_inode_unpacked root_inode, lostfound_inode;
1540         struct bkey_inode_buf packed_inode;
1541         struct qstr lostfound = QSTR("lost+found");
1542         struct bch_dev *ca;
1543         unsigned i;
1544         int ret;
1545
1546         bch_notice(c, "initializing new filesystem");
1547
1548         mutex_lock(&c->sb_lock);
1549         c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_extents_above_btree_updates_done);
1550         c->disk_sb.sb->compat[0] |= cpu_to_le64(1ULL << BCH_COMPAT_bformat_overflow_done);
1551
1552         bch2_sb_maybe_downgrade(c);
1553
1554         if (c->opts.version_upgrade != BCH_VERSION_UPGRADE_none) {
1555                 bch2_sb_upgrade(c, bcachefs_metadata_version_current);
1556                 SET_BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb, bcachefs_metadata_version_current);
1557                 bch2_write_super(c);
1558         }
1559         mutex_unlock(&c->sb_lock);
1560
1561         c->curr_recovery_pass = ARRAY_SIZE(recovery_passes);
1562         set_bit(BCH_FS_MAY_GO_RW, &c->flags);
1563         set_bit(BCH_FS_FSCK_DONE, &c->flags);
1564
1565         for (i = 0; i < BTREE_ID_NR; i++)
1566                 bch2_btree_root_alloc(c, i);
1567
1568         for_each_online_member(ca, c, i)
1569                 bch2_dev_usage_init(ca);
1570
1571         for_each_online_member(ca, c, i) {
1572                 ret = bch2_dev_journal_alloc(ca);
1573                 if (ret) {
1574                         percpu_ref_put(&ca->io_ref);
1575                         goto err;
1576                 }
1577         }
1578
1579         /*
1580          * journal_res_get() will crash if called before this has
1581          * set up the journal.pin FIFO and journal.cur pointer:
1582          */
1583         bch2_fs_journal_start(&c->journal, 1);
1584         bch2_journal_set_replay_done(&c->journal);
1585
1586         ret = bch2_fs_read_write_early(c);
1587         if (ret)
1588                 goto err;
1589
1590         /*
1591          * Write out the superblock and journal buckets, now that we can do
1592          * btree updates
1593          */
1594         bch_verbose(c, "marking superblocks");
1595         for_each_member_device(ca, c, i) {
1596                 ret = bch2_trans_mark_dev_sb(c, ca);
1597                 if (ret) {
1598                         percpu_ref_put(&ca->ref);
1599                         goto err;
1600                 }
1601
1602                 ca->new_fs_bucket_idx = 0;
1603         }
1604
1605         ret = bch2_fs_freespace_init(c);
1606         if (ret)
1607                 goto err;
1608
1609         ret = bch2_initialize_subvolumes(c);
1610         if (ret)
1611                 goto err;
1612
1613         bch_verbose(c, "reading snapshots table");
1614         ret = bch2_snapshots_read(c);
1615         if (ret)
1616                 goto err;
1617         bch_verbose(c, "reading snapshots done");
1618
1619         bch2_inode_init(c, &root_inode, 0, 0, S_IFDIR|0755, 0, NULL);
1620         root_inode.bi_inum      = BCACHEFS_ROOT_INO;
1621         root_inode.bi_subvol    = BCACHEFS_ROOT_SUBVOL;
1622         bch2_inode_pack(&packed_inode, &root_inode);
1623         packed_inode.inode.k.p.snapshot = U32_MAX;
1624
1625         ret = bch2_btree_insert(c, BTREE_ID_inodes,
1626                                 &packed_inode.inode.k_i,
1627                                 NULL, NULL, 0);
1628         if (ret) {
1629                 bch_err_msg(c, ret, "creating root directory");
1630                 goto err;
1631         }
1632
1633         bch2_inode_init_early(c, &lostfound_inode);
1634
1635         ret = bch2_trans_do(c, NULL, NULL, 0,
1636                 bch2_create_trans(&trans,
1637                                   BCACHEFS_ROOT_SUBVOL_INUM,
1638                                   &root_inode, &lostfound_inode,
1639                                   &lostfound,
1640                                   0, 0, S_IFDIR|0700, 0,
1641                                   NULL, NULL, (subvol_inum) { 0 }, 0));
1642         if (ret) {
1643                 bch_err_msg(c, ret, "creating lost+found");
1644                 goto err;
1645         }
1646
1647         if (enabled_qtypes(c)) {
1648                 ret = bch2_fs_quota_read(c);
1649                 if (ret)
1650                         goto err;
1651         }
1652
1653         ret = bch2_journal_flush(&c->journal);
1654         if (ret) {
1655                 bch_err_msg(c, ret, "writing first journal entry");
1656                 goto err;
1657         }
1658
1659         mutex_lock(&c->sb_lock);
1660         SET_BCH_SB_INITIALIZED(c->disk_sb.sb, true);
1661         SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
1662
1663         bch2_write_super(c);
1664         mutex_unlock(&c->sb_lock);
1665
1666         return 0;
1667 err:
1668         bch_err_fn(ca, ret);
1669         return ret;
1670 }