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 "disk_groups.h"
   9 #include "error.h"
  10 #include "io.h"
  11 #include "journal.h"
  12 #include "journal_io.h"
  13 #include "journal_reclaim.h"
  14 #include "journal_seq_blacklist.h"
  15 #include "replicas.h"
  16
  17 #include <trace/events/bcachefs.h>
  18
  19 static void __journal_replay_free(struct journal_replay *i)
  20 {
  21         list_del(&i->list);
  22         kvpfree(i, offsetof(struct journal_replay, j) +
  23                 vstruct_bytes(&i->j));
  24
  25 }
  26
  27 static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
  28 {
  29         i->ignore = true;
  30
  31         if (!c->opts.read_entire_journal)
  32                 __journal_replay_free(i);
  33 }
  34
  35 struct journal_list {
  36         struct closure          cl;
  37         struct mutex            lock;
  38         struct list_head        *head;
  39         int                     ret;
  40 };
  41
  42 #define JOURNAL_ENTRY_ADD_OK            0
  43 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  44
  45 /*
  46  * Given a journal entry we just read, add it to the list of journal entries to
  47  * be replayed:
  48  */
  49 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  50                              struct bch_extent_ptr entry_ptr,
  51                              struct journal_list *jlist, struct jset *j,
  52                              bool bad)
  53 {
  54         struct journal_replay *i, *pos, *dup = NULL;
  55         struct bch_extent_ptr *ptr;
  56         struct list_head *where;
  57         size_t bytes = vstruct_bytes(j);
  58         u64 last_seq = 0;
  59         int ret = JOURNAL_ENTRY_ADD_OK;
  60
  61         list_for_each_entry_reverse(i, jlist->head, list) {
  62                 if (!JSET_NO_FLUSH(&i->j)) {
  63                         last_seq = le64_to_cpu(i->j.last_seq);
  64                         break;
  65                 }
  66         }
  67
  68         /* Is this entry older than the range we need? */
  69         if (!c->opts.read_entire_journal &&
  70             le64_to_cpu(j->seq) < last_seq) {
  71                 ret = JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  72                 goto out;
  73         }
  74
  75         /* Drop entries we don't need anymore */
  76         if (!JSET_NO_FLUSH(j)) {
  77                 list_for_each_entry_safe(i, pos, jlist->head, list) {
  78                         if (le64_to_cpu(i->j.seq) >= le64_to_cpu(j->last_seq))
  79                                 break;
  80                         journal_replay_free(c, i);
  81                 }
  82         }
  83
  84         list_for_each_entry_reverse(i, jlist->head, list) {
  85                 if (le64_to_cpu(j->seq) > le64_to_cpu(i->j.seq)) {
  86                         where = &i->list;
  87                         goto add;
  88                 }
  89         }
  90
  91         where = jlist->head;
  92 add:
  93         dup = where->next != jlist->head
  94                 ? container_of(where->next, struct journal_replay, list)
  95                 : NULL;
  96
  97         if (dup && le64_to_cpu(j->seq) != le64_to_cpu(dup->j.seq))
  98                 dup = NULL;
  99
 100         /*
 101          * Duplicate journal entries? If so we want the one that didn't have a
 102          * checksum error:
 103          */
 104         if (dup) {
 105                 if (dup->bad) {
 106                         /* we'll replace @dup: */
 107                 } else if (bad) {
 108                         i = dup;
 109                         goto found;
 110                 } else {
 111                         fsck_err_on(bytes != vstruct_bytes(&dup->j) ||
 112                                     memcmp(j, &dup->j, bytes), c,
 113                                     "found duplicate but non identical journal entries (seq %llu)",
 114                                     le64_to_cpu(j->seq));
 115                         i = dup;
 116                         goto found;
 117                 }
 118         }
 119
 120         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 121         if (!i) {
 122                 ret = -ENOMEM;
 123                 goto out;
 124         }
 125
 126         i->nr_ptrs       = 0;
 127         i->bad          = bad;
 128         i->ignore       = false;
 129         memcpy(&i->j, j, bytes);
 130
 131         if (dup) {
 132                 i->nr_ptrs = dup->nr_ptrs;
 133                 memcpy(i->ptrs, dup->ptrs, sizeof(dup->ptrs));
 134                 __journal_replay_free(dup);
 135         }
 136
 137         list_add(&i->list, where);
 138 found:
 139         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 140                 if (ptr->dev == ca->dev_idx) {
 141                         bch_err(c, "duplicate journal entry %llu on same device",
 142                                 le64_to_cpu(i->j.seq));
 143                         goto out;
 144                 }
 145         }
 146
 147         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 148                 bch_err(c, "found too many copies of journal entry %llu",
 149                         le64_to_cpu(i->j.seq));
 150                 goto out;
 151         }
 152
 153         i->ptrs[i->nr_ptrs++] = entry_ptr;
 154 out:
 155 fsck_err:
 156         return ret;
 157 }
 158
 159 static struct nonce journal_nonce(const struct jset *jset)
 160 {
 161         return (struct nonce) {{
 162                 [0] = 0,
 163                 [1] = ((__le32 *) &jset->seq)[0],
 164                 [2] = ((__le32 *) &jset->seq)[1],
 165                 [3] = BCH_NONCE_JOURNAL,
 166         }};
 167 }
 168
 169 /* this fills in a range with empty jset_entries: */
 170 static void journal_entry_null_range(void *start, void *end)
 171 {
 172         struct jset_entry *entry;
 173
 174         for (entry = start; entry != end; entry = vstruct_next(entry))
 175                 memset(entry, 0, sizeof(*entry));
 176 }
 177
 178 #define JOURNAL_ENTRY_REREAD    5
 179 #define JOURNAL_ENTRY_NONE      6
 180 #define JOURNAL_ENTRY_BAD       7
 181
 182 #define journal_entry_err(c, msg, ...)                                  \
 183 ({                                                                      \
 184         switch (write) {                                                \
 185         case READ:                                                      \
 186                 mustfix_fsck_err(c, msg, ##__VA_ARGS__);                \
 187                 break;                                                  \
 188         case WRITE:                                                     \
 189                 bch_err(c, "corrupt metadata before write:\n"           \
 190                         msg, ##__VA_ARGS__);                            \
 191                 if (bch2_fs_inconsistent(c)) {                          \
 192                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 193                         goto fsck_err;                                  \
 194                 }                                                       \
 195                 break;                                                  \
 196         }                                                               \
 197         true;                                                           \
 198 })
 199
 200 #define journal_entry_err_on(cond, c, msg, ...)                         \
 201         ((cond) ? journal_entry_err(c, msg, ##__VA_ARGS__) : false)
 202
 203 #define FSCK_DELETED_KEY        5
 204
 205 static int journal_validate_key(struct bch_fs *c, const char *where,
 206                                 struct jset_entry *entry,
 207                                 unsigned level, enum btree_id btree_id,
 208                                 struct bkey_i *k, const char *type,
 209                                 unsigned version, int big_endian, int write)
 210 {
 211         void *next = vstruct_next(entry);
 212         const char *invalid;
 213         int ret = 0;
 214
 215         if (journal_entry_err_on(!k->k.u64s, c,
 216                         "invalid %s in %s entry offset %zi/%u: k->u64s 0",
 217                         type, where,
 218                         (u64 *) k - entry->_data,
 219                         le16_to_cpu(entry->u64s))) {
 220                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 221                 journal_entry_null_range(vstruct_next(entry), next);
 222                 return FSCK_DELETED_KEY;
 223         }
 224
 225         if (journal_entry_err_on((void *) bkey_next(k) >
 226                                 (void *) vstruct_next(entry), c,
 227                         "invalid %s in %s entry offset %zi/%u: extends past end of journal entry",
 228                         type, where,
 229                         (u64 *) k - entry->_data,
 230                         le16_to_cpu(entry->u64s))) {
 231                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 232                 journal_entry_null_range(vstruct_next(entry), next);
 233                 return FSCK_DELETED_KEY;
 234         }
 235
 236         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT, c,
 237                         "invalid %s in %s entry offset %zi/%u: bad format %u",
 238                         type, where,
 239                         (u64 *) k - entry->_data,
 240                         le16_to_cpu(entry->u64s),
 241                         k->k.format)) {
 242                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 243                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 244                 journal_entry_null_range(vstruct_next(entry), next);
 245                 return FSCK_DELETED_KEY;
 246         }
 247
 248         if (!write)
 249                 bch2_bkey_compat(level, btree_id, version, big_endian,
 250                                  write, NULL, bkey_to_packed(k));
 251
 252         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 253                                     __btree_node_type(level, btree_id));
 254         if (invalid) {
 255                 char buf[160];
 256
 257                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(k));
 258                 mustfix_fsck_err(c, "invalid %s in %s entry offset %zi/%u: %s\n%s",
 259                                  type, where,
 260                                  (u64 *) k - entry->_data,
 261                                  le16_to_cpu(entry->u64s),
 262                                  invalid, buf);
 263
 264                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 265                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 266                 journal_entry_null_range(vstruct_next(entry), next);
 267                 return FSCK_DELETED_KEY;
 268         }
 269
 270         if (write)
 271                 bch2_bkey_compat(level, btree_id, version, big_endian,
 272                                  write, NULL, bkey_to_packed(k));
 273 fsck_err:
 274         return ret;
 275 }
 276
 277 static int journal_entry_validate_btree_keys(struct bch_fs *c,
 278                                              const char *where,
 279                                              struct jset_entry *entry,
 280                                              unsigned version, int big_endian, int write)
 281 {
 282         struct bkey_i *k = entry->start;
 283
 284         while (k != vstruct_last(entry)) {
 285                 int ret = journal_validate_key(c, where, entry,
 286                                                entry->level,
 287                                                entry->btree_id,
 288                                                k, "key", version, big_endian, write);
 289                 if (ret == FSCK_DELETED_KEY)
 290                         continue;
 291
 292                 k = bkey_next(k);
 293         }
 294
 295         return 0;
 296 }
 297
 298 static int journal_entry_validate_btree_root(struct bch_fs *c,
 299                                              const char *where,
 300                                              struct jset_entry *entry,
 301                                              unsigned version, int big_endian, int write)
 302 {
 303         struct bkey_i *k = entry->start;
 304         int ret = 0;
 305
 306         if (journal_entry_err_on(!entry->u64s ||
 307                                  le16_to_cpu(entry->u64s) != k->k.u64s, c,
 308                                  "invalid btree root journal entry: wrong number of keys")) {
 309                 void *next = vstruct_next(entry);
 310                 /*
 311                  * we don't want to null out this jset_entry,
 312                  * just the contents, so that later we can tell
 313                  * we were _supposed_ to have a btree root
 314                  */
 315                 entry->u64s = 0;
 316                 journal_entry_null_range(vstruct_next(entry), next);
 317                 return 0;
 318         }
 319
 320         return journal_validate_key(c, where, entry, 1, entry->btree_id, k,
 321                                     "btree root", version, big_endian, write);
 322 fsck_err:
 323         return ret;
 324 }
 325
 326 static int journal_entry_validate_prio_ptrs(struct bch_fs *c,
 327                                             const char *where,
 328                                             struct jset_entry *entry,
 329                                             unsigned version, int big_endian, int write)
 330 {
 331         /* obsolete, don't care: */
 332         return 0;
 333 }
 334
 335 static int journal_entry_validate_blacklist(struct bch_fs *c,
 336                                             const char *where,
 337                                             struct jset_entry *entry,
 338                                             unsigned version, int big_endian, int write)
 339 {
 340         int ret = 0;
 341
 342         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1, c,
 343                 "invalid journal seq blacklist entry: bad size")) {
 344                 journal_entry_null_range(entry, vstruct_next(entry));
 345         }
 346 fsck_err:
 347         return ret;
 348 }
 349
 350 static int journal_entry_validate_blacklist_v2(struct bch_fs *c,
 351                                                const char *where,
 352                                                struct jset_entry *entry,
 353                                                unsigned version, int big_endian, int write)
 354 {
 355         struct jset_entry_blacklist_v2 *bl_entry;
 356         int ret = 0;
 357
 358         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2, c,
 359                 "invalid journal seq blacklist entry: bad size")) {
 360                 journal_entry_null_range(entry, vstruct_next(entry));
 361                 goto out;
 362         }
 363
 364         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 365
 366         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 367                                  le64_to_cpu(bl_entry->end), c,
 368                 "invalid journal seq blacklist entry: start > end")) {
 369                 journal_entry_null_range(entry, vstruct_next(entry));
 370         }
 371 out:
 372 fsck_err:
 373         return ret;
 374 }
 375
 376 static int journal_entry_validate_usage(struct bch_fs *c,
 377                                         const char *where,
 378                                         struct jset_entry *entry,
 379                                         unsigned version, int big_endian, int write)
 380 {
 381         struct jset_entry_usage *u =
 382                 container_of(entry, struct jset_entry_usage, entry);
 383         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 384         int ret = 0;
 385
 386         if (journal_entry_err_on(bytes < sizeof(*u),
 387                                  c,
 388                                  "invalid journal entry usage: bad size")) {
 389                 journal_entry_null_range(entry, vstruct_next(entry));
 390                 return ret;
 391         }
 392
 393 fsck_err:
 394         return ret;
 395 }
 396
 397 static int journal_entry_validate_data_usage(struct bch_fs *c,
 398                                         const char *where,
 399                                         struct jset_entry *entry,
 400                                         unsigned version, int big_endian, int write)
 401 {
 402         struct jset_entry_data_usage *u =
 403                 container_of(entry, struct jset_entry_data_usage, entry);
 404         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 405         int ret = 0;
 406
 407         if (journal_entry_err_on(bytes < sizeof(*u) ||
 408                                  bytes < sizeof(*u) + u->r.nr_devs,
 409                                  c,
 410                                  "invalid journal entry usage: bad size")) {
 411                 journal_entry_null_range(entry, vstruct_next(entry));
 412                 return ret;
 413         }
 414
 415 fsck_err:
 416         return ret;
 417 }
 418
 419 static int journal_entry_validate_clock(struct bch_fs *c,
 420                                         const char *where,
 421                                         struct jset_entry *entry,
 422                                         unsigned version, int big_endian, int write)
 423 {
 424         struct jset_entry_clock *clock =
 425                 container_of(entry, struct jset_entry_clock, entry);
 426         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 427         int ret = 0;
 428
 429         if (journal_entry_err_on(bytes != sizeof(*clock),
 430                                  c, "invalid journal entry clock: bad size")) {
 431                 journal_entry_null_range(entry, vstruct_next(entry));
 432                 return ret;
 433         }
 434
 435         if (journal_entry_err_on(clock->rw > 1,
 436                                  c, "invalid journal entry clock: bad rw")) {
 437                 journal_entry_null_range(entry, vstruct_next(entry));
 438                 return ret;
 439         }
 440
 441 fsck_err:
 442         return ret;
 443 }
 444
 445 static int journal_entry_validate_dev_usage(struct bch_fs *c,
 446                                             const char *where,
 447                                             struct jset_entry *entry,
 448                                             unsigned version, int big_endian, int write)
 449 {
 450         struct jset_entry_dev_usage *u =
 451                 container_of(entry, struct jset_entry_dev_usage, entry);
 452         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 453         unsigned expected = sizeof(*u) + sizeof(u->d[0]) * 7; /* Current value of BCH_DATA_NR */
 454         unsigned dev;
 455         int ret = 0;
 456
 457         if (journal_entry_err_on(bytes < expected,
 458                                  c, "invalid journal entry dev usage: bad size (%u < %u)",
 459                                  bytes, expected)) {
 460                 journal_entry_null_range(entry, vstruct_next(entry));
 461                 return ret;
 462         }
 463
 464         dev = le32_to_cpu(u->dev);
 465
 466         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 467                                  c, "invalid journal entry dev usage: bad dev")) {
 468                 journal_entry_null_range(entry, vstruct_next(entry));
 469                 return ret;
 470         }
 471
 472         if (journal_entry_err_on(u->pad,
 473                                  c, "invalid journal entry dev usage: bad pad")) {
 474                 journal_entry_null_range(entry, vstruct_next(entry));
 475                 return ret;
 476         }
 477
 478 fsck_err:
 479         return ret;
 480 }
 481
 482 struct jset_entry_ops {
 483         int (*validate)(struct bch_fs *, const char *,
 484                         struct jset_entry *, unsigned, int, int);
 485 };
 486
 487 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 488 #define x(f, nr)                                                \
 489         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 490                 .validate       = journal_entry_validate_##f,   \
 491         },
 492         BCH_JSET_ENTRY_TYPES()
 493 #undef x
 494 };
 495
 496 int bch2_journal_entry_validate(struct bch_fs *c, const char *where,
 497                                 struct jset_entry *entry,
 498                                 unsigned version, int big_endian, int write)
 499 {
 500         return entry->type < BCH_JSET_ENTRY_NR
 501                 ? bch2_jset_entry_ops[entry->type].validate(c, where, entry,
 502                                 version, big_endian, write)
 503                 : 0;
 504 }
 505
 506 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 507                                  int write)
 508 {
 509         char buf[100];
 510         struct jset_entry *entry;
 511         int ret = 0;
 512
 513         vstruct_for_each(jset, entry) {
 514                 scnprintf(buf, sizeof(buf), "jset %llu entry offset %zi/%u",
 515                           le64_to_cpu(jset->seq),
 516                           (u64 *) entry - jset->_data,
 517                           le32_to_cpu(jset->u64s));
 518
 519                 if (journal_entry_err_on(vstruct_next(entry) >
 520                                          vstruct_last(jset), c,
 521                                 "journal entry extends past end of jset")) {
 522                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 523                         break;
 524                 }
 525
 526                 ret = bch2_journal_entry_validate(c, buf, entry,
 527                                         le32_to_cpu(jset->version),
 528                                         JSET_BIG_ENDIAN(jset), write);
 529                 if (ret)
 530                         break;
 531         }
 532 fsck_err:
 533         return ret;
 534 }
 535
 536 static int jset_validate(struct bch_fs *c,
 537                          struct bch_dev *ca,
 538                          struct jset *jset, u64 sector,
 539                          unsigned bucket_sectors_left,
 540                          unsigned sectors_read,
 541                          int write)
 542 {
 543         size_t bytes = vstruct_bytes(jset);
 544         struct bch_csum csum;
 545         unsigned version;
 546         int ret = 0;
 547
 548         if (le64_to_cpu(jset->magic) != jset_magic(c))
 549                 return JOURNAL_ENTRY_NONE;
 550
 551         version = le32_to_cpu(jset->version);
 552         if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
 553                                   version < bcachefs_metadata_version_min) ||
 554                                  version >= bcachefs_metadata_version_max, c,
 555                         "%s sector %llu seq %llu: unknown journal entry version %u",
 556                         ca ? ca->name : c->name,
 557                         sector, le64_to_cpu(jset->seq),
 558                         version)) {
 559                 /* don't try to continue: */
 560                 return EINVAL;
 561         }
 562
 563         if (bytes > (sectors_read << 9) &&
 564             sectors_read < bucket_sectors_left)
 565                 return JOURNAL_ENTRY_REREAD;
 566
 567         if (journal_entry_err_on(bytes > bucket_sectors_left << 9, c,
 568                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 569                         ca ? ca->name : c->name,
 570                         sector, le64_to_cpu(jset->seq), bytes)) {
 571                 ret = JOURNAL_ENTRY_BAD;
 572                 le32_add_cpu(&jset->u64s,
 573                              -((bytes - (bucket_sectors_left << 9)) / 8));
 574         }
 575
 576         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)), c,
 577                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 578                         ca ? ca->name : c->name,
 579                         sector, le64_to_cpu(jset->seq),
 580                         JSET_CSUM_TYPE(jset))) {
 581                 ret = JOURNAL_ENTRY_BAD;
 582                 goto csum_done;
 583         }
 584
 585         if (write)
 586                 goto csum_done;
 587
 588         csum = csum_vstruct(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), jset);
 589         if (journal_entry_err_on(bch2_crc_cmp(csum, jset->csum), c,
 590                                  "%s sector %llu seq %llu: journal checksum bad",
 591                                  ca ? ca->name : c->name,
 592                                  sector, le64_to_cpu(jset->seq)))
 593                 ret = JOURNAL_ENTRY_BAD;
 594
 595         bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
 596                      jset->encrypted_start,
 597                      vstruct_end(jset) - (void *) jset->encrypted_start);
 598 csum_done:
 599         /* last_seq is ignored when JSET_NO_FLUSH is true */
 600         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 601                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq), c,
 602                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 603                                  le64_to_cpu(jset->last_seq),
 604                                  le64_to_cpu(jset->seq))) {
 605                 jset->last_seq = jset->seq;
 606                 return JOURNAL_ENTRY_BAD;
 607         }
 608 fsck_err:
 609         return ret;
 610 }
 611
 612 static int jset_validate_for_write(struct bch_fs *c, struct jset *jset)
 613 {
 614         unsigned sectors = vstruct_sectors(jset, c->block_bits);
 615
 616         return jset_validate(c, NULL, jset, 0, sectors, sectors, WRITE) ?:
 617                 jset_validate_entries(c, jset, WRITE);
 618 }
 619
 620 struct journal_read_buf {
 621         void            *data;
 622         size_t          size;
 623 };
 624
 625 static int journal_read_buf_realloc(struct journal_read_buf *b,
 626                                     size_t new_size)
 627 {
 628         void *n;
 629
 630         /* the bios are sized for this many pages, max: */
 631         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 632                 return -ENOMEM;
 633
 634         new_size = roundup_pow_of_two(new_size);
 635         n = kvpmalloc(new_size, GFP_KERNEL);
 636         if (!n)
 637                 return -ENOMEM;
 638
 639         kvpfree(b->data, b->size);
 640         b->data = n;
 641         b->size = new_size;
 642         return 0;
 643 }
 644
 645 static int journal_read_bucket(struct bch_dev *ca,
 646                                struct journal_read_buf *buf,
 647                                struct journal_list *jlist,
 648                                unsigned bucket)
 649 {
 650         struct bch_fs *c = ca->fs;
 651         struct journal_device *ja = &ca->journal;
 652         struct jset *j = NULL;
 653         unsigned sectors, sectors_read = 0;
 654         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 655             end = offset + ca->mi.bucket_size;
 656         bool saw_bad = false;
 657         int ret = 0;
 658
 659         pr_debug("reading %u", bucket);
 660
 661         while (offset < end) {
 662                 if (!sectors_read) {
 663                         struct bio *bio;
 664 reread:
 665                         sectors_read = min_t(unsigned,
 666                                 end - offset, buf->size >> 9);
 667
 668                         bio = bio_kmalloc(GFP_KERNEL,
 669                                           buf_pages(buf->data,
 670                                                     sectors_read << 9));
 671                         bio_set_dev(bio, ca->disk_sb.bdev);
 672                         bio->bi_iter.bi_sector  = offset;
 673                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 674                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 675
 676                         ret = submit_bio_wait(bio);
 677                         bio_put(bio);
 678
 679                         if (bch2_dev_io_err_on(ret, ca,
 680                                                "journal read error: sector %llu",
 681                                                offset) ||
 682                             bch2_meta_read_fault("journal")) {
 683                                 /*
 684                                  * We don't error out of the recovery process
 685                                  * here, since the relevant journal entry may be
 686                                  * found on a different device, and missing or
 687                                  * no journal entries will be handled later
 688                                  */
 689                                 return 0;
 690                         }
 691
 692                         j = buf->data;
 693                 }
 694
 695                 ret = jset_validate(c, ca, j, offset,
 696                                     end - offset, sectors_read,
 697                                     READ);
 698                 switch (ret) {
 699                 case BCH_FSCK_OK:
 700                         sectors = vstruct_sectors(j, c->block_bits);
 701                         break;
 702                 case JOURNAL_ENTRY_REREAD:
 703                         if (vstruct_bytes(j) > buf->size) {
 704                                 ret = journal_read_buf_realloc(buf,
 705                                                         vstruct_bytes(j));
 706                                 if (ret)
 707                                         return ret;
 708                         }
 709                         goto reread;
 710                 case JOURNAL_ENTRY_NONE:
 711                         if (!saw_bad)
 712                                 return 0;
 713                         sectors = c->opts.block_size;
 714                         goto next_block;
 715                 case JOURNAL_ENTRY_BAD:
 716                         saw_bad = true;
 717                         /*
 718                          * On checksum error we don't really trust the size
 719                          * field of the journal entry we read, so try reading
 720                          * again at next block boundary:
 721                          */
 722                         sectors = c->opts.block_size;
 723                         break;
 724                 default:
 725                         return ret;
 726                 }
 727
 728                 /*
 729                  * This happens sometimes if we don't have discards on -
 730                  * when we've partially overwritten a bucket with new
 731                  * journal entries. We don't need the rest of the
 732                  * bucket:
 733                  */
 734                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 735                         return 0;
 736
 737                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 738
 739                 mutex_lock(&jlist->lock);
 740                 ret = journal_entry_add(c, ca, (struct bch_extent_ptr) {
 741                                         .dev = ca->dev_idx,
 742                                         .offset = offset,
 743                                         }, jlist, j, ret != 0);
 744                 mutex_unlock(&jlist->lock);
 745
 746                 switch (ret) {
 747                 case JOURNAL_ENTRY_ADD_OK:
 748                         break;
 749                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
 750                         break;
 751                 default:
 752                         return ret;
 753                 }
 754 next_block:
 755                 pr_debug("next");
 756                 offset          += sectors;
 757                 sectors_read    -= sectors;
 758                 j = ((void *) j) + (sectors << 9);
 759         }
 760
 761         return 0;
 762 }
 763
 764 static void bch2_journal_read_device(struct closure *cl)
 765 {
 766         struct journal_device *ja =
 767                 container_of(cl, struct journal_device, read);
 768         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
 769         struct journal_list *jlist =
 770                 container_of(cl->parent, struct journal_list, cl);
 771         struct journal_read_buf buf = { NULL, 0 };
 772         u64 min_seq = U64_MAX;
 773         unsigned i;
 774         int ret;
 775
 776         if (!ja->nr)
 777                 goto out;
 778
 779         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
 780         if (ret)
 781                 goto err;
 782
 783         pr_debug("%u journal buckets", ja->nr);
 784
 785         for (i = 0; i < ja->nr; i++) {
 786                 ret = journal_read_bucket(ca, &buf, jlist, i);
 787                 if (ret)
 788                         goto err;
 789         }
 790
 791         /* Find the journal bucket with the highest sequence number: */
 792         for (i = 0; i < ja->nr; i++) {
 793                 if (ja->bucket_seq[i] > ja->bucket_seq[ja->cur_idx])
 794                         ja->cur_idx = i;
 795
 796                 min_seq = min(ja->bucket_seq[i], min_seq);
 797         }
 798
 799         /*
 800          * If there's duplicate journal entries in multiple buckets (which
 801          * definitely isn't supposed to happen, but...) - make sure to start
 802          * cur_idx at the last of those buckets, so we don't deadlock trying to
 803          * allocate
 804          */
 805         while (ja->bucket_seq[ja->cur_idx] > min_seq &&
 806                ja->bucket_seq[ja->cur_idx] >
 807                ja->bucket_seq[(ja->cur_idx + 1) % ja->nr])
 808                 ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 809
 810         ja->sectors_free = 0;
 811
 812         /*
 813          * Set dirty_idx to indicate the entire journal is full and needs to be
 814          * reclaimed - journal reclaim will immediately reclaim whatever isn't
 815          * pinned when it first runs:
 816          */
 817         ja->discard_idx = ja->dirty_idx_ondisk =
 818                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
 819 out:
 820         kvpfree(buf.data, buf.size);
 821         percpu_ref_put(&ca->io_ref);
 822         closure_return(cl);
 823         return;
 824 err:
 825         mutex_lock(&jlist->lock);
 826         jlist->ret = ret;
 827         mutex_unlock(&jlist->lock);
 828         goto out;
 829 }
 830
 831 static void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 832                                       struct journal_replay *j)
 833 {
 834         unsigned i;
 835
 836         for (i = 0; i < j->nr_ptrs; i++) {
 837                 struct bch_dev *ca = c->devs[j->ptrs[i].dev];
 838                 u64 offset;
 839
 840                 div64_u64_rem(j->ptrs[i].offset, ca->mi.bucket_size, &offset);
 841
 842                 if (i)
 843                         pr_buf(out, " ");
 844                 pr_buf(out, "%u:%llu (offset %llu)",
 845                        j->ptrs[i].dev,
 846                        (u64) j->ptrs[i].offset, offset);
 847         }
 848 }
 849
 850 int bch2_journal_read(struct bch_fs *c, struct list_head *list,
 851                       u64 *blacklist_seq, u64 *start_seq)
 852 {
 853         struct journal_list jlist;
 854         struct journal_replay *i, *t;
 855         struct bch_dev *ca;
 856         unsigned iter;
 857         size_t keys = 0, entries = 0;
 858         bool degraded = false;
 859         u64 seq, last_seq = 0;
 860         int ret = 0;
 861
 862         closure_init_stack(&jlist.cl);
 863         mutex_init(&jlist.lock);
 864         jlist.head = list;
 865         jlist.ret = 0;
 866
 867         for_each_member_device(ca, c, iter) {
 868                 if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
 869                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
 870                         continue;
 871
 872                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
 873                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
 874                     percpu_ref_tryget(&ca->io_ref))
 875                         closure_call(&ca->journal.read,
 876                                      bch2_journal_read_device,
 877                                      system_unbound_wq,
 878                                      &jlist.cl);
 879                 else
 880                         degraded = true;
 881         }
 882
 883         closure_sync(&jlist.cl);
 884
 885         if (jlist.ret)
 886                 return jlist.ret;
 887
 888         if (list_empty(list)) {
 889                 bch_info(c, "journal read done, but no entries found");
 890                 return 0;
 891         }
 892
 893         i = list_last_entry(list, struct journal_replay, list);
 894         *start_seq = le64_to_cpu(i->j.seq) + 1;
 895
 896         /*
 897          * Find most recent flush entry, and ignore newer non flush entries -
 898          * those entries will be blacklisted:
 899          */
 900         list_for_each_entry_safe_reverse(i, t, list, list) {
 901                 if (i->ignore)
 902                         continue;
 903
 904                 if (!JSET_NO_FLUSH(&i->j)) {
 905                         last_seq        = le64_to_cpu(i->j.last_seq);
 906                         *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
 907                         break;
 908                 }
 909
 910                 journal_replay_free(c, i);
 911         }
 912
 913         if (!last_seq) {
 914                 fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
 915                 return -1;
 916         }
 917
 918         /* Drop blacklisted entries and entries older than last_seq: */
 919         list_for_each_entry_safe(i, t, list, list) {
 920                 if (i->ignore)
 921                         continue;
 922
 923                 seq = le64_to_cpu(i->j.seq);
 924                 if (seq < last_seq) {
 925                         journal_replay_free(c, i);
 926                         continue;
 927                 }
 928
 929                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
 930                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
 931                                     "found blacklisted journal entry %llu", seq);
 932
 933                         journal_replay_free(c, i);
 934                 }
 935         }
 936
 937         /* Check for missing entries: */
 938         seq = last_seq;
 939         list_for_each_entry(i, list, list) {
 940                 if (i->ignore)
 941                         continue;
 942
 943                 BUG_ON(seq > le64_to_cpu(i->j.seq));
 944
 945                 while (seq < le64_to_cpu(i->j.seq)) {
 946                         u64 missing_start, missing_end;
 947                         char buf1[200], buf2[200];
 948
 949                         while (seq < le64_to_cpu(i->j.seq) &&
 950                                bch2_journal_seq_is_blacklisted(c, seq, false))
 951                                 seq++;
 952
 953                         if (seq == le64_to_cpu(i->j.seq))
 954                                 break;
 955
 956                         missing_start = seq;
 957
 958                         while (seq < le64_to_cpu(i->j.seq) &&
 959                                !bch2_journal_seq_is_blacklisted(c, seq, false))
 960                                 seq++;
 961
 962                         if (i->list.prev != list) {
 963                                 struct printbuf out = PBUF(buf1);
 964                                 struct journal_replay *p = list_prev_entry(i, list);
 965
 966                                 bch2_journal_ptrs_to_text(&out, c, p);
 967                                 pr_buf(&out, " size %llu", vstruct_sectors(&p->j, c->block_bits));
 968                         } else
 969                                 sprintf(buf1, "(none)");
 970                         bch2_journal_ptrs_to_text(&PBUF(buf2), c, i);
 971
 972                         missing_end = seq - 1;
 973                         fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
 974                                  "  prev at %s\n"
 975                                  "  next at %s",
 976                                  missing_start, missing_end,
 977                                  last_seq, *blacklist_seq - 1,
 978                                  buf1, buf2);
 979                 }
 980
 981                 seq++;
 982         }
 983
 984         list_for_each_entry(i, list, list) {
 985                 struct jset_entry *entry;
 986                 struct bkey_i *k, *_n;
 987                 struct bch_replicas_padded replicas = {
 988                         .e.data_type = BCH_DATA_journal,
 989                         .e.nr_required = 1,
 990                 };
 991                 unsigned ptr;
 992                 char buf[80];
 993
 994                 if (i->ignore)
 995                         continue;
 996
 997                 ret = jset_validate_entries(c, &i->j, READ);
 998                 if (ret)
 999                         goto fsck_err;
1000
1001                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1002                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1003
1004                 bch2_replicas_entry_sort(&replicas.e);
1005
1006                 /*
1007                  * If we're mounting in degraded mode - if we didn't read all
1008                  * the devices - this is wrong:
1009                  */
1010
1011                 if (!degraded &&
1012                     (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
1013                      fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
1014                                  "superblock not marked as containing replicas %s",
1015                                  (bch2_replicas_entry_to_text(&PBUF(buf),
1016                                                               &replicas.e), buf)))) {
1017                         ret = bch2_mark_replicas(c, &replicas.e);
1018                         if (ret)
1019                                 return ret;
1020                 }
1021
1022                 for_each_jset_key(k, _n, entry, &i->j)
1023                         keys++;
1024                 entries++;
1025         }
1026
1027         bch_info(c, "journal read done, %zu keys in %zu entries, seq %llu",
1028                  keys, entries, *start_seq);
1029
1030         if (*start_seq != *blacklist_seq)
1031                 bch_info(c, "dropped unflushed entries %llu-%llu",
1032                          *blacklist_seq, *start_seq - 1);
1033 fsck_err:
1034         return ret;
1035 }
1036
1037 /* journal write: */
1038
1039 static void __journal_write_alloc(struct journal *j,
1040                                   struct journal_buf *w,
1041                                   struct dev_alloc_list *devs_sorted,
1042                                   unsigned sectors,
1043                                   unsigned *replicas,
1044                                   unsigned replicas_want)
1045 {
1046         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1047         struct journal_device *ja;
1048         struct bch_dev *ca;
1049         unsigned i;
1050
1051         if (*replicas >= replicas_want)
1052                 return;
1053
1054         for (i = 0; i < devs_sorted->nr; i++) {
1055                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1056                 if (!ca)
1057                         continue;
1058
1059                 ja = &ca->journal;
1060
1061                 /*
1062                  * Check that we can use this device, and aren't already using
1063                  * it:
1064                  */
1065                 if (!ca->mi.durability ||
1066                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1067                     !ja->nr ||
1068                     bch2_bkey_has_device(bkey_i_to_s_c(&w->key),
1069                                          ca->dev_idx) ||
1070                     sectors > ja->sectors_free)
1071                         continue;
1072
1073                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1074
1075                 bch2_bkey_append_ptr(&w->key,
1076                         (struct bch_extent_ptr) {
1077                                   .offset = bucket_to_sector(ca,
1078                                         ja->buckets[ja->cur_idx]) +
1079                                         ca->mi.bucket_size -
1080                                         ja->sectors_free,
1081                                   .dev = ca->dev_idx,
1082                 });
1083
1084                 ja->sectors_free -= sectors;
1085                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1086
1087                 *replicas += ca->mi.durability;
1088
1089                 if (*replicas >= replicas_want)
1090                         break;
1091         }
1092 }
1093
1094 /**
1095  * journal_next_bucket - move on to the next journal bucket if possible
1096  */
1097 static int journal_write_alloc(struct journal *j, struct journal_buf *w,
1098                                unsigned sectors)
1099 {
1100         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1101         struct bch_devs_mask devs;
1102         struct journal_device *ja;
1103         struct bch_dev *ca;
1104         struct dev_alloc_list devs_sorted;
1105         unsigned target = c->opts.metadata_target ?:
1106                 c->opts.foreground_target;
1107         unsigned i, replicas = 0, replicas_want =
1108                 READ_ONCE(c->opts.metadata_replicas);
1109
1110         rcu_read_lock();
1111 retry:
1112         devs = target_rw_devs(c, BCH_DATA_journal, target);
1113
1114         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1115
1116         __journal_write_alloc(j, w, &devs_sorted,
1117                               sectors, &replicas, replicas_want);
1118
1119         if (replicas >= replicas_want)
1120                 goto done;
1121
1122         for (i = 0; i < devs_sorted.nr; i++) {
1123                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1124                 if (!ca)
1125                         continue;
1126
1127                 ja = &ca->journal;
1128
1129                 if (sectors > ja->sectors_free &&
1130                     sectors <= ca->mi.bucket_size &&
1131                     bch2_journal_dev_buckets_available(j, ja,
1132                                         journal_space_discarded)) {
1133                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1134                         ja->sectors_free = ca->mi.bucket_size;
1135
1136                         /*
1137                          * ja->bucket_seq[ja->cur_idx] must always have
1138                          * something sensible:
1139                          */
1140                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1141                 }
1142         }
1143
1144         __journal_write_alloc(j, w, &devs_sorted,
1145                               sectors, &replicas, replicas_want);
1146
1147         if (replicas < replicas_want && target) {
1148                 /* Retry from all devices: */
1149                 target = 0;
1150                 goto retry;
1151         }
1152 done:
1153         rcu_read_unlock();
1154
1155         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1156
1157         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1158 }
1159
1160 static void journal_write_compact(struct jset *jset)
1161 {
1162         struct jset_entry *i, *next, *prev = NULL;
1163
1164         /*
1165          * Simple compaction, dropping empty jset_entries (from journal
1166          * reservations that weren't fully used) and merging jset_entries that
1167          * can be.
1168          *
1169          * If we wanted to be really fancy here, we could sort all the keys in
1170          * the jset and drop keys that were overwritten - probably not worth it:
1171          */
1172         vstruct_for_each_safe(jset, i, next) {
1173                 unsigned u64s = le16_to_cpu(i->u64s);
1174
1175                 /* Empty entry: */
1176                 if (!u64s)
1177                         continue;
1178
1179                 /* Can we merge with previous entry? */
1180                 if (prev &&
1181                     i->btree_id == prev->btree_id &&
1182                     i->level    == prev->level &&
1183                     i->type     == prev->type &&
1184                     i->type     == BCH_JSET_ENTRY_btree_keys &&
1185                     le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
1186                         memmove_u64s_down(vstruct_next(prev),
1187                                           i->_data,
1188                                           u64s);
1189                         le16_add_cpu(&prev->u64s, u64s);
1190                         continue;
1191                 }
1192
1193                 /* Couldn't merge, move i into new position (after prev): */
1194                 prev = prev ? vstruct_next(prev) : jset->start;
1195                 if (i != prev)
1196                         memmove_u64s_down(prev, i, jset_u64s(u64s));
1197         }
1198
1199         prev = prev ? vstruct_next(prev) : jset->start;
1200         jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
1201 }
1202
1203 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1204 {
1205         /* we aren't holding j->lock: */
1206         unsigned new_size = READ_ONCE(j->buf_size_want);
1207         void *new_buf;
1208
1209         if (buf->buf_size >= new_size)
1210                 return;
1211
1212         new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN);
1213         if (!new_buf)
1214                 return;
1215
1216         memcpy(new_buf, buf->data, buf->buf_size);
1217
1218         spin_lock(&j->lock);
1219         swap(buf->data,         new_buf);
1220         swap(buf->buf_size,     new_size);
1221         spin_unlock(&j->lock);
1222
1223         kvpfree(new_buf, new_size);
1224 }
1225
1226 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1227 {
1228         return j->buf + j->reservations.unwritten_idx;
1229 }
1230
1231 static void journal_write_done(struct closure *cl)
1232 {
1233         struct journal *j = container_of(cl, struct journal, io);
1234         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1235         struct journal_buf *w = journal_last_unwritten_buf(j);
1236         struct bch_devs_list devs =
1237                 bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1238         struct bch_replicas_padded replicas;
1239         union journal_res_state old, new;
1240         u64 v, seq, last_seq;
1241         int err = 0;
1242
1243         bch2_time_stats_update(j->write_time, j->write_start_time);
1244
1245         if (!devs.nr) {
1246                 bch_err(c, "unable to write journal to sufficient devices");
1247                 err = -EIO;
1248         } else {
1249                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal, devs);
1250                 if (bch2_mark_replicas(c, &replicas.e))
1251                         err = -EIO;
1252         }
1253
1254         if (err)
1255                 bch2_fatal_error(c);
1256
1257         spin_lock(&j->lock);
1258         seq = le64_to_cpu(w->data->seq);
1259         last_seq = le64_to_cpu(w->data->last_seq);
1260
1261         if (seq >= j->pin.front)
1262                 journal_seq_pin(j, seq)->devs = devs;
1263
1264         j->seq_ondisk           = seq;
1265         if (err && (!j->err_seq || seq < j->err_seq))
1266                 j->err_seq      = seq;
1267
1268         if (!JSET_NO_FLUSH(w->data)) {
1269                 j->flushed_seq_ondisk = seq;
1270                 j->last_seq_ondisk = last_seq;
1271         }
1272
1273         /*
1274          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1275          * more buckets:
1276          *
1277          * Must come before signaling write completion, for
1278          * bch2_fs_journal_stop():
1279          */
1280         journal_reclaim_kick(&c->journal);
1281
1282         /* also must come before signalling write completion: */
1283         closure_debug_destroy(cl);
1284
1285         v = atomic64_read(&j->reservations.counter);
1286         do {
1287                 old.v = new.v = v;
1288                 BUG_ON(new.idx == new.unwritten_idx);
1289
1290                 new.unwritten_idx++;
1291         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1292                                        old.v, new.v)) != old.v);
1293
1294         bch2_journal_space_available(j);
1295
1296         closure_wake_up(&w->wait);
1297         journal_wake(j);
1298
1299         if (test_bit(JOURNAL_NEED_WRITE, &j->flags))
1300                 mod_delayed_work(system_freezable_wq, &j->write_work, 0);
1301         spin_unlock(&j->lock);
1302
1303         if (new.unwritten_idx != new.idx &&
1304             !journal_state_count(new, new.unwritten_idx))
1305                 closure_call(&j->io, bch2_journal_write, system_highpri_wq, NULL);
1306 }
1307
1308 static void journal_write_endio(struct bio *bio)
1309 {
1310         struct bch_dev *ca = bio->bi_private;
1311         struct journal *j = &ca->fs->journal;
1312
1313         if (bch2_dev_io_err_on(bio->bi_status, ca, "journal write error: %s",
1314                                bch2_blk_status_to_str(bio->bi_status)) ||
1315             bch2_meta_write_fault("journal")) {
1316                 struct journal_buf *w = journal_last_unwritten_buf(j);
1317                 unsigned long flags;
1318
1319                 spin_lock_irqsave(&j->err_lock, flags);
1320                 bch2_bkey_drop_device(bkey_i_to_s(&w->key), ca->dev_idx);
1321                 spin_unlock_irqrestore(&j->err_lock, flags);
1322         }
1323
1324         closure_put(&j->io);
1325         percpu_ref_put(&ca->io_ref);
1326 }
1327
1328 static void do_journal_write(struct closure *cl)
1329 {
1330         struct journal *j = container_of(cl, struct journal, io);
1331         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1332         struct bch_dev *ca;
1333         struct journal_buf *w = journal_last_unwritten_buf(j);
1334         struct bch_extent_ptr *ptr;
1335         struct bio *bio;
1336         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1337
1338         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1339                 ca = bch_dev_bkey_exists(c, ptr->dev);
1340                 if (!percpu_ref_tryget(&ca->io_ref)) {
1341                         /* XXX: fix this */
1342                         bch_err(c, "missing device for journal write\n");
1343                         continue;
1344                 }
1345
1346                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1347                              sectors);
1348
1349                 bio = ca->journal.bio;
1350                 bio_reset(bio);
1351                 bio_set_dev(bio, ca->disk_sb.bdev);
1352                 bio->bi_iter.bi_sector  = ptr->offset;
1353                 bio->bi_end_io          = journal_write_endio;
1354                 bio->bi_private         = ca;
1355                 bio->bi_opf             = REQ_OP_WRITE|REQ_SYNC|REQ_META;
1356
1357                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1358                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1359
1360                 if (!JSET_NO_FLUSH(w->data))
1361                         bio->bi_opf    |= REQ_FUA;
1362                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1363                         bio->bi_opf    |= REQ_PREFLUSH;
1364
1365                 bch2_bio_map(bio, w->data, sectors << 9);
1366
1367                 trace_journal_write(bio);
1368                 closure_bio_submit(bio, cl);
1369
1370                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1371                         le64_to_cpu(w->data->seq);
1372         }
1373
1374         continue_at(cl, journal_write_done, system_highpri_wq);
1375         return;
1376 }
1377
1378 void bch2_journal_write(struct closure *cl)
1379 {
1380         struct journal *j = container_of(cl, struct journal, io);
1381         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1382         struct bch_dev *ca;
1383         struct journal_buf *w = journal_last_unwritten_buf(j);
1384         struct jset_entry *start, *end;
1385         struct jset *jset;
1386         struct bio *bio;
1387         char *journal_debug_buf = NULL;
1388         bool validate_before_checksum = false;
1389         unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
1390         int ret;
1391
1392         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1393
1394         journal_buf_realloc(j, w);
1395         jset = w->data;
1396
1397         j->write_start_time = local_clock();
1398
1399         spin_lock(&j->lock);
1400         if (c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush) &&
1401             !w->must_flush &&
1402             (jiffies - j->last_flush_write) < msecs_to_jiffies(j->write_delay_ms) &&
1403             test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)) {
1404                 w->noflush = true;
1405                 SET_JSET_NO_FLUSH(jset, true);
1406                 jset->last_seq = 0;
1407
1408                 j->nr_noflush_writes++;
1409         } else {
1410                 j->last_flush_write = jiffies;
1411                 j->nr_flush_writes++;
1412         }
1413         spin_unlock(&j->lock);
1414
1415         /*
1416          * New btree roots are set by journalling them; when the journal entry
1417          * gets written we have to propagate them to c->btree_roots
1418          *
1419          * But, every journal entry we write has to contain all the btree roots
1420          * (at least for now); so after we copy btree roots to c->btree_roots we
1421          * have to get any missing btree roots and add them to this journal
1422          * entry:
1423          */
1424
1425         bch2_journal_entries_to_btree_roots(c, jset);
1426
1427         start = end = vstruct_last(jset);
1428
1429         end     = bch2_btree_roots_to_journal_entries(c, jset->start, end);
1430
1431         bch2_journal_super_entries_add_common(c, &end,
1432                                 le64_to_cpu(jset->seq));
1433         u64s    = (u64 *) end - (u64 *) start;
1434         BUG_ON(u64s > j->entry_u64s_reserved);
1435
1436         le32_add_cpu(&jset->u64s, u64s);
1437         BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors);
1438
1439         journal_write_compact(jset);
1440
1441         jset->magic             = cpu_to_le64(jset_magic(c));
1442         jset->version           = c->sb.version < bcachefs_metadata_version_new_versioning
1443                 ? cpu_to_le32(BCH_JSET_VERSION_OLD)
1444                 : cpu_to_le32(c->sb.version);
1445
1446         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1447         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1448
1449         if (journal_entry_empty(jset))
1450                 j->last_empty_seq = le64_to_cpu(jset->seq);
1451
1452         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1453                 validate_before_checksum = true;
1454
1455         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1456                 validate_before_checksum = true;
1457
1458         if (validate_before_checksum &&
1459             jset_validate_for_write(c, jset))
1460                 goto err;
1461
1462         bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1463                     jset->encrypted_start,
1464                     vstruct_end(jset) - (void *) jset->encrypted_start);
1465
1466         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1467                                   journal_nonce(jset), jset);
1468
1469         if (!validate_before_checksum &&
1470             jset_validate_for_write(c, jset))
1471                 goto err;
1472
1473         sectors = vstruct_sectors(jset, c->block_bits);
1474         BUG_ON(sectors > w->sectors);
1475
1476         bytes = vstruct_bytes(jset);
1477         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1478
1479 retry_alloc:
1480         spin_lock(&j->lock);
1481         ret = journal_write_alloc(j, w, sectors);
1482
1483         if (ret && j->can_discard) {
1484                 spin_unlock(&j->lock);
1485                 bch2_journal_do_discards(j);
1486                 goto retry_alloc;
1487         }
1488
1489         if (ret) {
1490                 journal_debug_buf = kmalloc(4096, GFP_ATOMIC);
1491                 if (journal_debug_buf)
1492                         __bch2_journal_debug_to_text(&_PBUF(journal_debug_buf, 4096), j);
1493         }
1494
1495         /*
1496          * write is allocated, no longer need to account for it in
1497          * bch2_journal_space_available():
1498          */
1499         w->sectors = 0;
1500
1501         /*
1502          * journal entry has been compacted and allocated, recalculate space
1503          * available:
1504          */
1505         bch2_journal_space_available(j);
1506         spin_unlock(&j->lock);
1507
1508         if (ret) {
1509                 bch_err(c, "Unable to allocate journal write:\n%s",
1510                         journal_debug_buf);
1511                 kfree(journal_debug_buf);
1512                 bch2_fatal_error(c);
1513                 continue_at(cl, journal_write_done, system_highpri_wq);
1514                 return;
1515         }
1516
1517         /*
1518          * XXX: we really should just disable the entire journal in nochanges
1519          * mode
1520          */
1521         if (c->opts.nochanges)
1522                 goto no_io;
1523
1524         for_each_rw_member(ca, c, i)
1525                 nr_rw_members++;
1526
1527         if (nr_rw_members > 1)
1528                 w->separate_flush = true;
1529
1530         if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
1531                 for_each_rw_member(ca, c, i) {
1532                         percpu_ref_get(&ca->io_ref);
1533
1534                         bio = ca->journal.bio;
1535                         bio_reset(bio);
1536                         bio_set_dev(bio, ca->disk_sb.bdev);
1537                         bio->bi_opf             = REQ_OP_FLUSH;
1538                         bio->bi_end_io          = journal_write_endio;
1539                         bio->bi_private         = ca;
1540                         closure_bio_submit(bio, cl);
1541                 }
1542         }
1543
1544         bch2_bucket_seq_cleanup(c);
1545
1546         continue_at(cl, do_journal_write, system_highpri_wq);
1547         return;
1548 no_io:
1549         bch2_bucket_seq_cleanup(c);
1550
1551         continue_at(cl, journal_write_done, system_highpri_wq);
1552         return;
1553 err:
1554         bch2_inconsistent_error(c);
1555         continue_at(cl, journal_write_done, system_highpri_wq);
1556 }