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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_foreground.h"
   4 #include "btree_io.h"
   5 #include "btree_update_interior.h"
   6 #include "buckets.h"
   7 #include "checksum.h"
   8 #include "error.h"
   9 #include "io.h"
  10 #include "journal.h"
  11 #include "journal_io.h"
  12 #include "journal_reclaim.h"
  13 #include "journal_seq_blacklist.h"
  14 #include "replicas.h"
  15
  16 #include <trace/events/bcachefs.h>
  17
  18 static void __journal_replay_free(struct journal_replay *i)
  19 {
  20         list_del(&i->list);
  21         kvpfree(i, offsetof(struct journal_replay, j) +
  22                 vstruct_bytes(&i->j));
  23
  24 }
  25
  26 static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
  27 {
  28         i->ignore = true;
  29
  30         if (!c->opts.read_entire_journal)
  31                 __journal_replay_free(i);
  32 }
  33
  34 struct journal_list {
  35         struct closure          cl;
  36         struct mutex            lock;
  37         struct list_head        *head;
  38         int                     ret;
  39 };
  40
  41 #define JOURNAL_ENTRY_ADD_OK            0
  42 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  43
  44 /*
  45  * Given a journal entry we just read, add it to the list of journal entries to
  46  * be replayed:
  47  */
  48 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  49                              struct journal_list *jlist, struct jset *j,
  50                              bool bad)
  51 {
  52         struct journal_replay *i, *pos;
  53         struct bch_devs_list devs = { .nr = 0 };
  54         struct list_head *where;
  55         size_t bytes = vstruct_bytes(j);
  56         u64 last_seq = 0;
  57         int ret;
  58
  59         list_for_each_entry_reverse(i, jlist->head, list) {
  60                 if (!JSET_NO_FLUSH(&i->j)) {
  61                         last_seq = le64_to_cpu(i->j.last_seq);
  62                         break;
  63                 }
  64         }
  65
  66         /* Is this entry older than the range we need? */
  67         if (!c->opts.read_entire_journal &&
  68             le64_to_cpu(j->seq) < last_seq) {
  69                 ret = JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  70                 goto out;
  71         }
  72
  73         /* Drop entries we don't need anymore */
  74         if (!JSET_NO_FLUSH(j)) {
  75                 list_for_each_entry_safe(i, pos, jlist->head, list) {
  76                         if (le64_to_cpu(i->j.seq) >= le64_to_cpu(j->last_seq))
  77                                 break;
  78                         journal_replay_free(c, i);
  79                 }
  80         }
  81
  82         list_for_each_entry_reverse(i, jlist->head, list) {
  83                 if (le64_to_cpu(j->seq) > le64_to_cpu(i->j.seq)) {
  84                         where = &i->list;
  85                         goto add;
  86                 }
  87         }
  88
  89         where = jlist->head;
  90 add:
  91         i = where->next != jlist->head
  92                 ? container_of(where->next, struct journal_replay, list)
  93                 : NULL;
  94
  95         /*
  96          * Duplicate journal entries? If so we want the one that didn't have a
  97          * checksum error:
  98          */
  99         if (i && le64_to_cpu(j->seq) == le64_to_cpu(i->j.seq)) {
 100                 if (i->bad) {
 101                         devs = i->devs;
 102                         __journal_replay_free(i);
 103                 } else if (bad) {
 104                         goto found;
 105                 } else {
 106                         fsck_err_on(bytes != vstruct_bytes(&i->j) ||
 107                                     memcmp(j, &i->j, bytes), c,
 108                                     "found duplicate but non identical journal entries (seq %llu)",
 109                                     le64_to_cpu(j->seq));
 110                         goto found;
 111                 }
 112
 113         }
 114
 115         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 116         if (!i) {
 117                 ret = -ENOMEM;
 118                 goto out;
 119         }
 120
 121         list_add(&i->list, where);
 122         i->devs = devs;
 123         i->bad  = bad;
 124         i->ignore = false;
 125         memcpy(&i->j, j, bytes);
 126 found:
 127         if (!bch2_dev_list_has_dev(i->devs, ca->dev_idx))
 128                 bch2_dev_list_add_dev(&i->devs, ca->dev_idx);
 129         else
 130                 fsck_err_on(1, c, "duplicate journal entries on same device");
 131         ret = JOURNAL_ENTRY_ADD_OK;
 132 out:
 133 fsck_err:
 134         return ret;
 135 }
 136
 137 static struct nonce journal_nonce(const struct jset *jset)
 138 {
 139         return (struct nonce) {{
 140                 [0] = 0,
 141                 [1] = ((__le32 *) &jset->seq)[0],
 142                 [2] = ((__le32 *) &jset->seq)[1],
 143                 [3] = BCH_NONCE_JOURNAL,
 144         }};
 145 }
 146
 147 /* this fills in a range with empty jset_entries: */
 148 static void journal_entry_null_range(void *start, void *end)
 149 {
 150         struct jset_entry *entry;
 151
 152         for (entry = start; entry != end; entry = vstruct_next(entry))
 153                 memset(entry, 0, sizeof(*entry));
 154 }
 155
 156 #define JOURNAL_ENTRY_REREAD    5
 157 #define JOURNAL_ENTRY_NONE      6
 158 #define JOURNAL_ENTRY_BAD       7
 159
 160 #define journal_entry_err(c, msg, ...)                                  \
 161 ({                                                                      \
 162         switch (write) {                                                \
 163         case READ:                                                      \
 164                 mustfix_fsck_err(c, msg, ##__VA_ARGS__);                \
 165                 break;                                                  \
 166         case WRITE:                                                     \
 167                 bch_err(c, "corrupt metadata before write:\n"           \
 168                         msg, ##__VA_ARGS__);                            \
 169                 if (bch2_fs_inconsistent(c)) {                          \
 170                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 171                         goto fsck_err;                                  \
 172                 }                                                       \
 173                 break;                                                  \
 174         }                                                               \
 175         true;                                                           \
 176 })
 177
 178 #define journal_entry_err_on(cond, c, msg, ...)                         \
 179         ((cond) ? journal_entry_err(c, msg, ##__VA_ARGS__) : false)
 180
 181 #define FSCK_DELETED_KEY        5
 182
 183 static int journal_validate_key(struct bch_fs *c, struct jset *jset,
 184                                 struct jset_entry *entry,
 185                                 unsigned level, enum btree_id btree_id,
 186                                 struct bkey_i *k,
 187                                 const char *type, int write)
 188 {
 189         void *next = vstruct_next(entry);
 190         const char *invalid;
 191         unsigned version = le32_to_cpu(jset->version);
 192         int ret = 0;
 193
 194         if (journal_entry_err_on(!k->k.u64s, c,
 195                         "invalid %s in jset %llu offset %zi/%u entry offset %zi/%u: k->u64s 0",
 196                         type, le64_to_cpu(jset->seq),
 197                         (u64 *) entry - jset->_data,
 198                         le32_to_cpu(jset->u64s),
 199                         (u64 *) k - entry->_data,
 200                         le16_to_cpu(entry->u64s))) {
 201                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 202                 journal_entry_null_range(vstruct_next(entry), next);
 203                 return FSCK_DELETED_KEY;
 204         }
 205
 206         if (journal_entry_err_on((void *) bkey_next(k) >
 207                                 (void *) vstruct_next(entry), c,
 208                         "invalid %s in jset %llu offset %zi/%u entry offset %zi/%u: extends past end of journal entry",
 209                         type, le64_to_cpu(jset->seq),
 210                         (u64 *) entry - jset->_data,
 211                         le32_to_cpu(jset->u64s),
 212                         (u64 *) k - entry->_data,
 213                         le16_to_cpu(entry->u64s))) {
 214                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 215                 journal_entry_null_range(vstruct_next(entry), next);
 216                 return FSCK_DELETED_KEY;
 217         }
 218
 219         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT, c,
 220                         "invalid %s in jset %llu offset %zi/%u entry offset %zi/%u: bad format %u",
 221                         type, le64_to_cpu(jset->seq),
 222                         (u64 *) entry - jset->_data,
 223                         le32_to_cpu(jset->u64s),
 224                         (u64 *) k - entry->_data,
 225                         le16_to_cpu(entry->u64s),
 226                         k->k.format)) {
 227                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 228                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 229                 journal_entry_null_range(vstruct_next(entry), next);
 230                 return FSCK_DELETED_KEY;
 231         }
 232
 233         if (!write)
 234                 bch2_bkey_compat(level, btree_id, version,
 235                             JSET_BIG_ENDIAN(jset), write,
 236                             NULL, bkey_to_packed(k));
 237
 238         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 239                                     __btree_node_type(level, btree_id));
 240         if (invalid) {
 241                 char buf[160];
 242
 243                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(k));
 244                 mustfix_fsck_err(c, "invalid %s in jset %llu offset %zi/%u entry offset %zi/%u: %s\n%s",
 245                                  type, le64_to_cpu(jset->seq),
 246                                  (u64 *) entry - jset->_data,
 247                                  le32_to_cpu(jset->u64s),
 248                                  (u64 *) k - entry->_data,
 249                                  le16_to_cpu(entry->u64s),
 250                                  invalid, buf);
 251
 252                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 253                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 254                 journal_entry_null_range(vstruct_next(entry), next);
 255                 return FSCK_DELETED_KEY;
 256         }
 257
 258         if (write)
 259                 bch2_bkey_compat(level, btree_id, version,
 260                             JSET_BIG_ENDIAN(jset), write,
 261                             NULL, bkey_to_packed(k));
 262 fsck_err:
 263         return ret;
 264 }
 265
 266 static int journal_entry_validate_btree_keys(struct bch_fs *c,
 267                                              struct jset *jset,
 268                                              struct jset_entry *entry,
 269                                              int write)
 270 {
 271         struct bkey_i *k = entry->start;
 272
 273         while (k != vstruct_last(entry)) {
 274                 int ret = journal_validate_key(c, jset, entry,
 275                                                entry->level,
 276                                                entry->btree_id,
 277                                                k, "key", write);
 278                 if (ret == FSCK_DELETED_KEY)
 279                         continue;
 280
 281                 k = bkey_next(k);
 282         }
 283
 284         return 0;
 285 }
 286
 287 static int journal_entry_validate_btree_root(struct bch_fs *c,
 288                                              struct jset *jset,
 289                                              struct jset_entry *entry,
 290                                              int write)
 291 {
 292         struct bkey_i *k = entry->start;
 293         int ret = 0;
 294
 295         if (journal_entry_err_on(!entry->u64s ||
 296                                  le16_to_cpu(entry->u64s) != k->k.u64s, c,
 297                                  "invalid btree root journal entry: wrong number of keys")) {
 298                 void *next = vstruct_next(entry);
 299                 /*
 300                  * we don't want to null out this jset_entry,
 301                  * just the contents, so that later we can tell
 302                  * we were _supposed_ to have a btree root
 303                  */
 304                 entry->u64s = 0;
 305                 journal_entry_null_range(vstruct_next(entry), next);
 306                 return 0;
 307         }
 308
 309         return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
 310                                     "btree root", write);
 311 fsck_err:
 312         return ret;
 313 }
 314
 315 static int journal_entry_validate_prio_ptrs(struct bch_fs *c,
 316                                             struct jset *jset,
 317                                             struct jset_entry *entry,
 318                                             int write)
 319 {
 320         /* obsolete, don't care: */
 321         return 0;
 322 }
 323
 324 static int journal_entry_validate_blacklist(struct bch_fs *c,
 325                                             struct jset *jset,
 326                                             struct jset_entry *entry,
 327                                             int write)
 328 {
 329         int ret = 0;
 330
 331         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1, c,
 332                 "invalid journal seq blacklist entry: bad size")) {
 333                 journal_entry_null_range(entry, vstruct_next(entry));
 334         }
 335 fsck_err:
 336         return ret;
 337 }
 338
 339 static int journal_entry_validate_blacklist_v2(struct bch_fs *c,
 340                                                struct jset *jset,
 341                                                struct jset_entry *entry,
 342                                                int write)
 343 {
 344         struct jset_entry_blacklist_v2 *bl_entry;
 345         int ret = 0;
 346
 347         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2, c,
 348                 "invalid journal seq blacklist entry: bad size")) {
 349                 journal_entry_null_range(entry, vstruct_next(entry));
 350                 goto out;
 351         }
 352
 353         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 354
 355         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 356                                  le64_to_cpu(bl_entry->end), c,
 357                 "invalid journal seq blacklist entry: start > end")) {
 358                 journal_entry_null_range(entry, vstruct_next(entry));
 359         }
 360 out:
 361 fsck_err:
 362         return ret;
 363 }
 364
 365 static int journal_entry_validate_usage(struct bch_fs *c,
 366                                         struct jset *jset,
 367                                         struct jset_entry *entry,
 368                                         int write)
 369 {
 370         struct jset_entry_usage *u =
 371                 container_of(entry, struct jset_entry_usage, entry);
 372         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 373         int ret = 0;
 374
 375         if (journal_entry_err_on(bytes < sizeof(*u),
 376                                  c,
 377                                  "invalid journal entry usage: bad size")) {
 378                 journal_entry_null_range(entry, vstruct_next(entry));
 379                 return ret;
 380         }
 381
 382 fsck_err:
 383         return ret;
 384 }
 385
 386 static int journal_entry_validate_data_usage(struct bch_fs *c,
 387                                         struct jset *jset,
 388                                         struct jset_entry *entry,
 389                                         int write)
 390 {
 391         struct jset_entry_data_usage *u =
 392                 container_of(entry, struct jset_entry_data_usage, entry);
 393         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 394         int ret = 0;
 395
 396         if (journal_entry_err_on(bytes < sizeof(*u) ||
 397                                  bytes < sizeof(*u) + u->r.nr_devs,
 398                                  c,
 399                                  "invalid journal entry usage: bad size")) {
 400                 journal_entry_null_range(entry, vstruct_next(entry));
 401                 return ret;
 402         }
 403
 404 fsck_err:
 405         return ret;
 406 }
 407
 408 struct jset_entry_ops {
 409         int (*validate)(struct bch_fs *, struct jset *,
 410                         struct jset_entry *, int);
 411 };
 412
 413 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 414 #define x(f, nr)                                                \
 415         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 416                 .validate       = journal_entry_validate_##f,   \
 417         },
 418         BCH_JSET_ENTRY_TYPES()
 419 #undef x
 420 };
 421
 422 static int journal_entry_validate(struct bch_fs *c, struct jset *jset,
 423                                   struct jset_entry *entry, int write)
 424 {
 425         return entry->type < BCH_JSET_ENTRY_NR
 426                 ? bch2_jset_entry_ops[entry->type].validate(c, jset,
 427                                                             entry, write)
 428                 : 0;
 429 }
 430
 431 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 432                                  int write)
 433 {
 434         struct jset_entry *entry;
 435         int ret = 0;
 436
 437         vstruct_for_each(jset, entry) {
 438                 if (journal_entry_err_on(vstruct_next(entry) >
 439                                          vstruct_last(jset), c,
 440                                 "journal entry extends past end of jset")) {
 441                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 442                         break;
 443                 }
 444
 445                 ret = journal_entry_validate(c, jset, entry, write);
 446                 if (ret)
 447                         break;
 448         }
 449 fsck_err:
 450         return ret;
 451 }
 452
 453 static int jset_validate(struct bch_fs *c,
 454                          struct bch_dev *ca,
 455                          struct jset *jset, u64 sector,
 456                          unsigned bucket_sectors_left,
 457                          unsigned sectors_read,
 458                          int write)
 459 {
 460         size_t bytes = vstruct_bytes(jset);
 461         struct bch_csum csum;
 462         unsigned version;
 463         int ret = 0;
 464
 465         if (le64_to_cpu(jset->magic) != jset_magic(c))
 466                 return JOURNAL_ENTRY_NONE;
 467
 468         version = le32_to_cpu(jset->version);
 469         if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
 470                                   version < bcachefs_metadata_version_min) ||
 471                                  version >= bcachefs_metadata_version_max, c,
 472                         "%s sector %llu seq %llu: unknown journal entry version %u",
 473                         ca->name, sector, le64_to_cpu(jset->seq),
 474                         version)) {
 475                 /* don't try to continue: */
 476                 return EINVAL;
 477         }
 478
 479         if (bytes > (sectors_read << 9) &&
 480             sectors_read < bucket_sectors_left)
 481                 return JOURNAL_ENTRY_REREAD;
 482
 483         if (journal_entry_err_on(bytes > bucket_sectors_left << 9, c,
 484                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 485                         ca->name, sector, le64_to_cpu(jset->seq), bytes)) {
 486                 ret = JOURNAL_ENTRY_BAD;
 487                 le32_add_cpu(&jset->u64s,
 488                              -((bytes - (bucket_sectors_left << 9)) / 8));
 489         }
 490
 491         if (fsck_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)), c,
 492                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 493                         ca->name, sector, le64_to_cpu(jset->seq),
 494                         JSET_CSUM_TYPE(jset))) {
 495                 ret = JOURNAL_ENTRY_BAD;
 496                 goto bad_csum_type;
 497         }
 498
 499         csum = csum_vstruct(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), jset);
 500         if (journal_entry_err_on(bch2_crc_cmp(csum, jset->csum), c,
 501                                  "%s sector %llu seq %llu: journal checksum bad",
 502                                  ca->name, sector, le64_to_cpu(jset->seq)))
 503                 ret = JOURNAL_ENTRY_BAD;
 504
 505         bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
 506                      jset->encrypted_start,
 507                      vstruct_end(jset) - (void *) jset->encrypted_start);
 508 bad_csum_type:
 509         if (journal_entry_err_on(le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq), c,
 510                                  "invalid journal entry: last_seq > seq")) {
 511                 jset->last_seq = jset->seq;
 512                 return JOURNAL_ENTRY_BAD;
 513         }
 514 fsck_err:
 515         return ret;
 516 }
 517
 518 struct journal_read_buf {
 519         void            *data;
 520         size_t          size;
 521 };
 522
 523 static int journal_read_buf_realloc(struct journal_read_buf *b,
 524                                     size_t new_size)
 525 {
 526         void *n;
 527
 528         /* the bios are sized for this many pages, max: */
 529         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 530                 return -ENOMEM;
 531
 532         new_size = roundup_pow_of_two(new_size);
 533         n = kvpmalloc(new_size, GFP_KERNEL);
 534         if (!n)
 535                 return -ENOMEM;
 536
 537         kvpfree(b->data, b->size);
 538         b->data = n;
 539         b->size = new_size;
 540         return 0;
 541 }
 542
 543 static int journal_read_bucket(struct bch_dev *ca,
 544                                struct journal_read_buf *buf,
 545                                struct journal_list *jlist,
 546                                unsigned bucket)
 547 {
 548         struct bch_fs *c = ca->fs;
 549         struct journal_device *ja = &ca->journal;
 550         struct jset *j = NULL;
 551         unsigned sectors, sectors_read = 0;
 552         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 553             end = offset + ca->mi.bucket_size;
 554         bool saw_bad = false;
 555         int ret = 0;
 556
 557         pr_debug("reading %u", bucket);
 558
 559         while (offset < end) {
 560                 if (!sectors_read) {
 561                         struct bio *bio;
 562 reread:
 563                         sectors_read = min_t(unsigned,
 564                                 end - offset, buf->size >> 9);
 565
 566                         bio = bio_kmalloc(GFP_KERNEL,
 567                                           buf_pages(buf->data,
 568                                                     sectors_read << 9));
 569                         bio_set_dev(bio, ca->disk_sb.bdev);
 570                         bio->bi_iter.bi_sector  = offset;
 571                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 572                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 573
 574                         ret = submit_bio_wait(bio);
 575                         bio_put(bio);
 576
 577                         if (bch2_dev_io_err_on(ret, ca,
 578                                                "journal read error: sector %llu",
 579                                                offset) ||
 580                             bch2_meta_read_fault("journal")) {
 581                                 /*
 582                                  * We don't error out of the recovery process
 583                                  * here, since the relevant journal entry may be
 584                                  * found on a different device, and missing or
 585                                  * no journal entries will be handled later
 586                                  */
 587                                 return 0;
 588                         }
 589
 590                         j = buf->data;
 591                 }
 592
 593                 ret = jset_validate(c, ca, j, offset,
 594                                     end - offset, sectors_read,
 595                                     READ);
 596                 switch (ret) {
 597                 case BCH_FSCK_OK:
 598                         sectors = vstruct_sectors(j, c->block_bits);
 599                         break;
 600                 case JOURNAL_ENTRY_REREAD:
 601                         if (vstruct_bytes(j) > buf->size) {
 602                                 ret = journal_read_buf_realloc(buf,
 603                                                         vstruct_bytes(j));
 604                                 if (ret)
 605                                         return ret;
 606                         }
 607                         goto reread;
 608                 case JOURNAL_ENTRY_NONE:
 609                         if (!saw_bad)
 610                                 return 0;
 611                         sectors = c->opts.block_size;
 612                         goto next_block;
 613                 case JOURNAL_ENTRY_BAD:
 614                         saw_bad = true;
 615                         /*
 616                          * On checksum error we don't really trust the size
 617                          * field of the journal entry we read, so try reading
 618                          * again at next block boundary:
 619                          */
 620                         sectors = c->opts.block_size;
 621                         break;
 622                 default:
 623                         return ret;
 624                 }
 625
 626                 /*
 627                  * This happens sometimes if we don't have discards on -
 628                  * when we've partially overwritten a bucket with new
 629                  * journal entries. We don't need the rest of the
 630                  * bucket:
 631                  */
 632                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 633                         return 0;
 634
 635                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 636
 637                 mutex_lock(&jlist->lock);
 638                 ret = journal_entry_add(c, ca, jlist, j, ret != 0);
 639                 mutex_unlock(&jlist->lock);
 640
 641                 switch (ret) {
 642                 case JOURNAL_ENTRY_ADD_OK:
 643                         break;
 644                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
 645                         break;
 646                 default:
 647                         return ret;
 648                 }
 649 next_block:
 650                 pr_debug("next");
 651                 offset          += sectors;
 652                 sectors_read    -= sectors;
 653                 j = ((void *) j) + (sectors << 9);
 654         }
 655
 656         return 0;
 657 }
 658
 659 static void bch2_journal_read_device(struct closure *cl)
 660 {
 661         struct journal_device *ja =
 662                 container_of(cl, struct journal_device, read);
 663         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
 664         struct journal_list *jlist =
 665                 container_of(cl->parent, struct journal_list, cl);
 666         struct journal_read_buf buf = { NULL, 0 };
 667         u64 min_seq = U64_MAX;
 668         unsigned i;
 669         int ret;
 670
 671         if (!ja->nr)
 672                 goto out;
 673
 674         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
 675         if (ret)
 676                 goto err;
 677
 678         pr_debug("%u journal buckets", ja->nr);
 679
 680         for (i = 0; i < ja->nr; i++) {
 681                 ret = journal_read_bucket(ca, &buf, jlist, i);
 682                 if (ret)
 683                         goto err;
 684         }
 685
 686         /* Find the journal bucket with the highest sequence number: */
 687         for (i = 0; i < ja->nr; i++) {
 688                 if (ja->bucket_seq[i] > ja->bucket_seq[ja->cur_idx])
 689                         ja->cur_idx = i;
 690
 691                 min_seq = min(ja->bucket_seq[i], min_seq);
 692         }
 693
 694         /*
 695          * If there's duplicate journal entries in multiple buckets (which
 696          * definitely isn't supposed to happen, but...) - make sure to start
 697          * cur_idx at the last of those buckets, so we don't deadlock trying to
 698          * allocate
 699          */
 700         while (ja->bucket_seq[ja->cur_idx] > min_seq &&
 701                ja->bucket_seq[ja->cur_idx] >
 702                ja->bucket_seq[(ja->cur_idx + 1) % ja->nr])
 703                 ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 704
 705         ja->sectors_free = 0;
 706
 707         /*
 708          * Set dirty_idx to indicate the entire journal is full and needs to be
 709          * reclaimed - journal reclaim will immediately reclaim whatever isn't
 710          * pinned when it first runs:
 711          */
 712         ja->discard_idx = ja->dirty_idx_ondisk =
 713                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
 714 out:
 715         kvpfree(buf.data, buf.size);
 716         percpu_ref_put(&ca->io_ref);
 717         closure_return(cl);
 718         return;
 719 err:
 720         mutex_lock(&jlist->lock);
 721         jlist->ret = ret;
 722         mutex_unlock(&jlist->lock);
 723         goto out;
 724 }
 725
 726 int bch2_journal_read(struct bch_fs *c, struct list_head *list,
 727                       u64 *blacklist_seq, u64 *start_seq)
 728 {
 729         struct journal_list jlist;
 730         struct journal_replay *i, *t;
 731         struct bch_dev *ca;
 732         unsigned iter;
 733         size_t keys = 0, entries = 0;
 734         bool degraded = false;
 735         u64 seq, last_seq = 0;
 736         int ret = 0;
 737
 738         closure_init_stack(&jlist.cl);
 739         mutex_init(&jlist.lock);
 740         jlist.head = list;
 741         jlist.ret = 0;
 742
 743         for_each_member_device(ca, c, iter) {
 744                 if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
 745                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
 746                         continue;
 747
 748                 if ((ca->mi.state == BCH_MEMBER_STATE_RW ||
 749                      ca->mi.state == BCH_MEMBER_STATE_RO) &&
 750                     percpu_ref_tryget(&ca->io_ref))
 751                         closure_call(&ca->journal.read,
 752                                      bch2_journal_read_device,
 753                                      system_unbound_wq,
 754                                      &jlist.cl);
 755                 else
 756                         degraded = true;
 757         }
 758
 759         closure_sync(&jlist.cl);
 760
 761         if (jlist.ret)
 762                 return jlist.ret;
 763
 764         if (list_empty(list)) {
 765                 bch_info(c, "journal read done, but no entries found");
 766                 return 0;
 767         }
 768
 769         i = list_last_entry(list, struct journal_replay, list);
 770         *start_seq = le64_to_cpu(i->j.seq) + 1;
 771
 772         /*
 773          * Find most recent flush entry, and ignore newer non flush entries -
 774          * those entries will be blacklisted:
 775          */
 776         list_for_each_entry_safe_reverse(i, t, list, list) {
 777                 if (i->ignore)
 778                         continue;
 779
 780                 if (!JSET_NO_FLUSH(&i->j)) {
 781                         last_seq        = le64_to_cpu(i->j.last_seq);
 782                         *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
 783                         break;
 784                 }
 785
 786                 journal_replay_free(c, i);
 787         }
 788
 789         if (!last_seq) {
 790                 fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
 791                 return -1;
 792         }
 793
 794         /* Drop blacklisted entries and entries older than last_seq: */
 795         list_for_each_entry_safe(i, t, list, list) {
 796                 if (i->ignore)
 797                         continue;
 798
 799                 seq = le64_to_cpu(i->j.seq);
 800                 if (seq < last_seq) {
 801                         journal_replay_free(c, i);
 802                         continue;
 803                 }
 804
 805                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
 806                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
 807                                     "found blacklisted journal entry %llu", seq);
 808
 809                         journal_replay_free(c, i);
 810                 }
 811         }
 812
 813         /* Check for missing entries: */
 814         seq = last_seq;
 815         list_for_each_entry(i, list, list) {
 816                 if (i->ignore)
 817                         continue;
 818
 819                 BUG_ON(seq > le64_to_cpu(i->j.seq));
 820
 821                 while (seq < le64_to_cpu(i->j.seq)) {
 822                         u64 missing_start, missing_end;
 823
 824                         while (seq < le64_to_cpu(i->j.seq) &&
 825                                bch2_journal_seq_is_blacklisted(c, seq, false))
 826                                 seq++;
 827
 828                         if (seq == le64_to_cpu(i->j.seq))
 829                                 break;
 830
 831                         missing_start = seq;
 832
 833                         while (seq < le64_to_cpu(i->j.seq) &&
 834                                !bch2_journal_seq_is_blacklisted(c, seq, false))
 835                                 seq++;
 836
 837                         missing_end = seq - 1;
 838                         fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)",
 839                                  missing_start, missing_end,
 840                                  last_seq, *blacklist_seq - 1);
 841                 }
 842
 843                 seq++;
 844         }
 845
 846         list_for_each_entry(i, list, list) {
 847                 struct jset_entry *entry;
 848                 struct bkey_i *k, *_n;
 849                 struct bch_replicas_padded replicas;
 850                 char buf[80];
 851
 852                 if (i->ignore)
 853                         continue;
 854
 855                 ret = jset_validate_entries(c, &i->j, READ);
 856                 if (ret)
 857                         goto fsck_err;
 858
 859                 /*
 860                  * If we're mounting in degraded mode - if we didn't read all
 861                  * the devices - this is wrong:
 862                  */
 863
 864                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal, i->devs);
 865
 866                 if (!degraded &&
 867                     (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 868                      fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
 869                                  "superblock not marked as containing replicas %s",
 870                                  (bch2_replicas_entry_to_text(&PBUF(buf),
 871                                                               &replicas.e), buf)))) {
 872                         ret = bch2_mark_replicas(c, &replicas.e);
 873                         if (ret)
 874                                 return ret;
 875                 }
 876
 877                 for_each_jset_key(k, _n, entry, &i->j)
 878                         keys++;
 879                 entries++;
 880         }
 881
 882         bch_info(c, "journal read done, %zu keys in %zu entries, seq %llu",
 883                  keys, entries, *start_seq);
 884
 885         if (*start_seq != *blacklist_seq)
 886                 bch_info(c, "dropped unflushed entries %llu-%llu",
 887                          *blacklist_seq, *start_seq - 1);
 888 fsck_err:
 889         return ret;
 890 }
 891
 892 /* journal write: */
 893
 894 static void __journal_write_alloc(struct journal *j,
 895                                   struct journal_buf *w,
 896                                   struct dev_alloc_list *devs_sorted,
 897                                   unsigned sectors,
 898                                   unsigned *replicas,
 899                                   unsigned replicas_want)
 900 {
 901         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 902         struct journal_device *ja;
 903         struct bch_dev *ca;
 904         unsigned i;
 905
 906         if (*replicas >= replicas_want)
 907                 return;
 908
 909         for (i = 0; i < devs_sorted->nr; i++) {
 910                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
 911                 if (!ca)
 912                         continue;
 913
 914                 ja = &ca->journal;
 915
 916                 /*
 917                  * Check that we can use this device, and aren't already using
 918                  * it:
 919                  */
 920                 if (!ca->mi.durability ||
 921                     ca->mi.state != BCH_MEMBER_STATE_RW ||
 922                     !ja->nr ||
 923                     bch2_bkey_has_device(bkey_i_to_s_c(&w->key),
 924                                          ca->dev_idx) ||
 925                     sectors > ja->sectors_free)
 926                         continue;
 927
 928                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
 929
 930                 bch2_bkey_append_ptr(&w->key,
 931                         (struct bch_extent_ptr) {
 932                                   .offset = bucket_to_sector(ca,
 933                                         ja->buckets[ja->cur_idx]) +
 934                                         ca->mi.bucket_size -
 935                                         ja->sectors_free,
 936                                   .dev = ca->dev_idx,
 937                 });
 938
 939                 ja->sectors_free -= sectors;
 940                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
 941
 942                 *replicas += ca->mi.durability;
 943
 944                 if (*replicas >= replicas_want)
 945                         break;
 946         }
 947 }
 948
 949 /**
 950  * journal_next_bucket - move on to the next journal bucket if possible
 951  */
 952 static int journal_write_alloc(struct journal *j, struct journal_buf *w,
 953                                unsigned sectors)
 954 {
 955         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 956         struct journal_device *ja;
 957         struct bch_dev *ca;
 958         struct dev_alloc_list devs_sorted;
 959         unsigned i, replicas = 0, replicas_want =
 960                 READ_ONCE(c->opts.metadata_replicas);
 961
 962         rcu_read_lock();
 963
 964         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe,
 965                                           &c->rw_devs[BCH_DATA_journal]);
 966
 967         __journal_write_alloc(j, w, &devs_sorted,
 968                               sectors, &replicas, replicas_want);
 969
 970         if (replicas >= replicas_want)
 971                 goto done;
 972
 973         for (i = 0; i < devs_sorted.nr; i++) {
 974                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 975                 if (!ca)
 976                         continue;
 977
 978                 ja = &ca->journal;
 979
 980                 if (sectors > ja->sectors_free &&
 981                     sectors <= ca->mi.bucket_size &&
 982                     bch2_journal_dev_buckets_available(j, ja,
 983                                         journal_space_discarded)) {
 984                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 985                         ja->sectors_free = ca->mi.bucket_size;
 986
 987                         /*
 988                          * ja->bucket_seq[ja->cur_idx] must always have
 989                          * something sensible:
 990                          */
 991                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
 992                 }
 993         }
 994
 995         __journal_write_alloc(j, w, &devs_sorted,
 996                               sectors, &replicas, replicas_want);
 997 done:
 998         rcu_read_unlock();
 999
1000         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1001
1002         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1003 }
1004
1005 static void journal_write_compact(struct jset *jset)
1006 {
1007         struct jset_entry *i, *next, *prev = NULL;
1008
1009         /*
1010          * Simple compaction, dropping empty jset_entries (from journal
1011          * reservations that weren't fully used) and merging jset_entries that
1012          * can be.
1013          *
1014          * If we wanted to be really fancy here, we could sort all the keys in
1015          * the jset and drop keys that were overwritten - probably not worth it:
1016          */
1017         vstruct_for_each_safe(jset, i, next) {
1018                 unsigned u64s = le16_to_cpu(i->u64s);
1019
1020                 /* Empty entry: */
1021                 if (!u64s)
1022                         continue;
1023
1024                 /* Can we merge with previous entry? */
1025                 if (prev &&
1026                     i->btree_id == prev->btree_id &&
1027                     i->level    == prev->level &&
1028                     i->type     == prev->type &&
1029                     i->type     == BCH_JSET_ENTRY_btree_keys &&
1030                     le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
1031                         memmove_u64s_down(vstruct_next(prev),
1032                                           i->_data,
1033                                           u64s);
1034                         le16_add_cpu(&prev->u64s, u64s);
1035                         continue;
1036                 }
1037
1038                 /* Couldn't merge, move i into new position (after prev): */
1039                 prev = prev ? vstruct_next(prev) : jset->start;
1040                 if (i != prev)
1041                         memmove_u64s_down(prev, i, jset_u64s(u64s));
1042         }
1043
1044         prev = prev ? vstruct_next(prev) : jset->start;
1045         jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
1046 }
1047
1048 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1049 {
1050         /* we aren't holding j->lock: */
1051         unsigned new_size = READ_ONCE(j->buf_size_want);
1052         void *new_buf;
1053
1054         if (buf->buf_size >= new_size)
1055                 return;
1056
1057         new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN);
1058         if (!new_buf)
1059                 return;
1060
1061         memcpy(new_buf, buf->data, buf->buf_size);
1062
1063         spin_lock(&j->lock);
1064         swap(buf->data,         new_buf);
1065         swap(buf->buf_size,     new_size);
1066         spin_unlock(&j->lock);
1067
1068         kvpfree(new_buf, new_size);
1069 }
1070
1071 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1072 {
1073         return j->buf + j->reservations.unwritten_idx;
1074 }
1075
1076 static void journal_write_done(struct closure *cl)
1077 {
1078         struct journal *j = container_of(cl, struct journal, io);
1079         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1080         struct journal_buf *w = journal_last_unwritten_buf(j);
1081         struct bch_devs_list devs =
1082                 bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1083         struct bch_replicas_padded replicas;
1084         union journal_res_state old, new;
1085         u64 seq = le64_to_cpu(w->data->seq);
1086         u64 last_seq = le64_to_cpu(w->data->last_seq);
1087         u64 v;
1088         int err = 0;
1089
1090         bch2_time_stats_update(j->write_time, j->write_start_time);
1091
1092         if (!devs.nr) {
1093                 bch_err(c, "unable to write journal to sufficient devices");
1094                 err = -EIO;
1095         } else {
1096                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal, devs);
1097                 if (bch2_mark_replicas(c, &replicas.e))
1098                         err = -EIO;
1099         }
1100
1101         if (err)
1102                 bch2_fatal_error(c);
1103
1104         spin_lock(&j->lock);
1105         if (seq >= j->pin.front)
1106                 journal_seq_pin(j, seq)->devs = devs;
1107
1108         j->seq_ondisk           = seq;
1109         if (err && (!j->err_seq || seq < j->err_seq))
1110                 j->err_seq      = seq;
1111
1112         if (!w->noflush) {
1113                 j->flushed_seq_ondisk = seq;
1114                 j->last_seq_ondisk = last_seq;
1115         }
1116
1117         /*
1118          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1119          * more buckets:
1120          *
1121          * Must come before signaling write completion, for
1122          * bch2_fs_journal_stop():
1123          */
1124         journal_reclaim_kick(&c->journal);
1125
1126         /* also must come before signalling write completion: */
1127         closure_debug_destroy(cl);
1128
1129         v = atomic64_read(&j->reservations.counter);
1130         do {
1131                 old.v = new.v = v;
1132                 BUG_ON(new.idx == new.unwritten_idx);
1133
1134                 new.unwritten_idx++;
1135         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1136                                        old.v, new.v)) != old.v);
1137
1138         bch2_journal_space_available(j);
1139
1140         closure_wake_up(&w->wait);
1141         journal_wake(j);
1142
1143         if (test_bit(JOURNAL_NEED_WRITE, &j->flags))
1144                 mod_delayed_work(system_freezable_wq, &j->write_work, 0);
1145         spin_unlock(&j->lock);
1146
1147         if (new.unwritten_idx != new.idx &&
1148             !journal_state_count(new, new.unwritten_idx))
1149                 closure_call(&j->io, bch2_journal_write, system_highpri_wq, NULL);
1150 }
1151
1152 static void journal_write_endio(struct bio *bio)
1153 {
1154         struct bch_dev *ca = bio->bi_private;
1155         struct journal *j = &ca->fs->journal;
1156
1157         if (bch2_dev_io_err_on(bio->bi_status, ca, "journal write error: %s",
1158                                bch2_blk_status_to_str(bio->bi_status)) ||
1159             bch2_meta_write_fault("journal")) {
1160                 struct journal_buf *w = journal_last_unwritten_buf(j);
1161                 unsigned long flags;
1162
1163                 spin_lock_irqsave(&j->err_lock, flags);
1164                 bch2_bkey_drop_device(bkey_i_to_s(&w->key), ca->dev_idx);
1165                 spin_unlock_irqrestore(&j->err_lock, flags);
1166         }
1167
1168         closure_put(&j->io);
1169         percpu_ref_put(&ca->io_ref);
1170 }
1171
1172 void bch2_journal_write(struct closure *cl)
1173 {
1174         struct journal *j = container_of(cl, struct journal, io);
1175         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1176         struct bch_dev *ca;
1177         struct journal_buf *w = journal_last_unwritten_buf(j);
1178         struct jset_entry *start, *end;
1179         struct jset *jset;
1180         struct bio *bio;
1181         struct bch_extent_ptr *ptr;
1182         bool validate_before_checksum = false;
1183         unsigned i, sectors, bytes, u64s;
1184         int ret;
1185
1186         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1187
1188         journal_buf_realloc(j, w);
1189         jset = w->data;
1190
1191         j->write_start_time = local_clock();
1192
1193         spin_lock(&j->lock);
1194         if (c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush) &&
1195             !w->must_flush &&
1196             (jiffies - j->last_flush_write) < msecs_to_jiffies(j->write_delay_ms) &&
1197             test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)) {
1198                 w->noflush = true;
1199                 SET_JSET_NO_FLUSH(jset, true);
1200                 jset->last_seq = cpu_to_le64(j->last_seq_ondisk);
1201
1202                 j->nr_noflush_writes++;
1203         } else {
1204                 j->last_flush_write = jiffies;
1205                 j->nr_flush_writes++;
1206         }
1207         spin_unlock(&j->lock);
1208
1209         /*
1210          * New btree roots are set by journalling them; when the journal entry
1211          * gets written we have to propagate them to c->btree_roots
1212          *
1213          * But, every journal entry we write has to contain all the btree roots
1214          * (at least for now); so after we copy btree roots to c->btree_roots we
1215          * have to get any missing btree roots and add them to this journal
1216          * entry:
1217          */
1218
1219         bch2_journal_entries_to_btree_roots(c, jset);
1220
1221         start = end = vstruct_last(jset);
1222
1223         end     = bch2_btree_roots_to_journal_entries(c, jset->start, end);
1224
1225         end     = bch2_journal_super_entries_add_common(c, end,
1226                                                 le64_to_cpu(jset->seq));
1227         u64s    = (u64 *) end - (u64 *) start;
1228         BUG_ON(u64s > j->entry_u64s_reserved);
1229
1230         le32_add_cpu(&jset->u64s, u64s);
1231         BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors);
1232
1233         journal_write_compact(jset);
1234
1235         jset->read_clock        = cpu_to_le16(c->bucket_clock[READ].hand);
1236         jset->write_clock       = cpu_to_le16(c->bucket_clock[WRITE].hand);
1237         jset->magic             = cpu_to_le64(jset_magic(c));
1238
1239         jset->version           = c->sb.version < bcachefs_metadata_version_new_versioning
1240                 ? cpu_to_le32(BCH_JSET_VERSION_OLD)
1241                 : cpu_to_le32(c->sb.version);
1242
1243         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1244         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1245
1246         if (journal_entry_empty(jset))
1247                 j->last_empty_seq = le64_to_cpu(jset->seq);
1248
1249         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1250                 validate_before_checksum = true;
1251
1252         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_max)
1253                 validate_before_checksum = true;
1254
1255         if (validate_before_checksum &&
1256             jset_validate_entries(c, jset, WRITE))
1257                 goto err;
1258
1259         bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1260                     jset->encrypted_start,
1261                     vstruct_end(jset) - (void *) jset->encrypted_start);
1262
1263         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1264                                   journal_nonce(jset), jset);
1265
1266         if (!validate_before_checksum &&
1267             jset_validate_entries(c, jset, WRITE))
1268                 goto err;
1269
1270         sectors = vstruct_sectors(jset, c->block_bits);
1271         BUG_ON(sectors > w->sectors);
1272
1273         bytes = vstruct_bytes(jset);
1274         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1275
1276 retry_alloc:
1277         spin_lock(&j->lock);
1278         ret = journal_write_alloc(j, w, sectors);
1279
1280         if (ret && j->can_discard) {
1281                 spin_unlock(&j->lock);
1282                 bch2_journal_do_discards(j);
1283                 goto retry_alloc;
1284         }
1285
1286         /*
1287          * write is allocated, no longer need to account for it in
1288          * bch2_journal_space_available():
1289          */
1290         w->sectors = 0;
1291
1292         /*
1293          * journal entry has been compacted and allocated, recalculate space
1294          * available:
1295          */
1296         bch2_journal_space_available(j);
1297         spin_unlock(&j->lock);
1298
1299         if (ret) {
1300                 bch_err(c, "Unable to allocate journal write");
1301                 bch2_fatal_error(c);
1302                 continue_at(cl, journal_write_done, system_highpri_wq);
1303                 return;
1304         }
1305
1306         /*
1307          * XXX: we really should just disable the entire journal in nochanges
1308          * mode
1309          */
1310         if (c->opts.nochanges)
1311                 goto no_io;
1312
1313         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1314                 ca = bch_dev_bkey_exists(c, ptr->dev);
1315                 if (!percpu_ref_tryget(&ca->io_ref)) {
1316                         /* XXX: fix this */
1317                         bch_err(c, "missing device for journal write\n");
1318                         continue;
1319                 }
1320
1321                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1322                              sectors);
1323
1324                 bio = ca->journal.bio;
1325                 bio_reset(bio);
1326                 bio_set_dev(bio, ca->disk_sb.bdev);
1327                 bio->bi_iter.bi_sector  = ptr->offset;
1328                 bio->bi_end_io          = journal_write_endio;
1329                 bio->bi_private         = ca;
1330                 bio->bi_opf             = REQ_OP_WRITE|REQ_SYNC|REQ_META;
1331                 if (!JSET_NO_FLUSH(jset))
1332                         bio->bi_opf    |= REQ_PREFLUSH|REQ_FUA;
1333                 bch2_bio_map(bio, jset, sectors << 9);
1334
1335                 trace_journal_write(bio);
1336                 closure_bio_submit(bio, cl);
1337
1338                 ca->journal.bucket_seq[ca->journal.cur_idx] = le64_to_cpu(jset->seq);
1339         }
1340
1341         if (!JSET_NO_FLUSH(jset)) {
1342                 for_each_rw_member(ca, c, i)
1343                         if (journal_flushes_device(ca) &&
1344                             !bch2_bkey_has_device(bkey_i_to_s_c(&w->key), i)) {
1345                                 percpu_ref_get(&ca->io_ref);
1346
1347                                 bio = ca->journal.bio;
1348                                 bio_reset(bio);
1349                                 bio_set_dev(bio, ca->disk_sb.bdev);
1350                                 bio->bi_opf             = REQ_OP_FLUSH;
1351                                 bio->bi_end_io          = journal_write_endio;
1352                                 bio->bi_private         = ca;
1353                                 closure_bio_submit(bio, cl);
1354                         }
1355         }
1356 no_io:
1357         bch2_bucket_seq_cleanup(c);
1358
1359         continue_at(cl, journal_write_done, system_highpri_wq);
1360         return;
1361 err:
1362         bch2_inconsistent_error(c);
1363         continue_at(cl, journal_write_done, system_highpri_wq);
1364 }