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_background.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_io.h"
   6 #include "btree_update_interior.h"
   7 #include "btree_write_buffer.h"
   8 #include "buckets.h"
   9 #include "checksum.h"
  10 #include "disk_groups.h"
  11 #include "error.h"
  12 #include "journal.h"
  13 #include "journal_io.h"
  14 #include "journal_reclaim.h"
  15 #include "journal_seq_blacklist.h"
  16 #include "replicas.h"
  17 #include "sb-clean.h"
  18 #include "trace.h"
  19
  20 static struct nonce journal_nonce(const struct jset *jset)
  21 {
  22         return (struct nonce) {{
  23                 [0] = 0,
  24                 [1] = ((__le32 *) &jset->seq)[0],
  25                 [2] = ((__le32 *) &jset->seq)[1],
  26                 [3] = BCH_NONCE_JOURNAL,
  27         }};
  28 }
  29
  30 static bool jset_csum_good(struct bch_fs *c, struct jset *j)
  31 {
  32         return bch2_checksum_type_valid(c, JSET_CSUM_TYPE(j)) &&
  33                 !bch2_crc_cmp(j->csum,
  34                               csum_vstruct(c, JSET_CSUM_TYPE(j), journal_nonce(j), j));
  35 }
  36
  37 static inline u32 journal_entry_radix_idx(struct bch_fs *c, u64 seq)
  38 {
  39         return (seq - c->journal_entries_base_seq) & (~0U >> 1);
  40 }
  41
  42 static void __journal_replay_free(struct bch_fs *c,
  43                                   struct journal_replay *i)
  44 {
  45         struct journal_replay **p =
  46                 genradix_ptr(&c->journal_entries,
  47                              journal_entry_radix_idx(c, le64_to_cpu(i->j.seq)));
  48
  49         BUG_ON(*p != i);
  50         *p = NULL;
  51         kvpfree(i, offsetof(struct journal_replay, j) +
  52                 vstruct_bytes(&i->j));
  53 }
  54
  55 static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
  56 {
  57         i->ignore = true;
  58
  59         if (!c->opts.read_entire_journal)
  60                 __journal_replay_free(c, i);
  61 }
  62
  63 struct journal_list {
  64         struct closure          cl;
  65         u64                     last_seq;
  66         struct mutex            lock;
  67         int                     ret;
  68 };
  69
  70 #define JOURNAL_ENTRY_ADD_OK            0
  71 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  72
  73 /*
  74  * Given a journal entry we just read, add it to the list of journal entries to
  75  * be replayed:
  76  */
  77 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  78                              struct journal_ptr entry_ptr,
  79                              struct journal_list *jlist, struct jset *j)
  80 {
  81         struct genradix_iter iter;
  82         struct journal_replay **_i, *i, *dup;
  83         struct journal_ptr *ptr;
  84         size_t bytes = vstruct_bytes(j);
  85         u64 last_seq = !JSET_NO_FLUSH(j) ? le64_to_cpu(j->last_seq) : 0;
  86         int ret = JOURNAL_ENTRY_ADD_OK;
  87
  88         /* Is this entry older than the range we need? */
  89         if (!c->opts.read_entire_journal &&
  90             le64_to_cpu(j->seq) < jlist->last_seq)
  91                 return JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  92
  93         /*
  94          * genradixes are indexed by a ulong, not a u64, so we can't index them
  95          * by sequence number directly: Assume instead that they will all fall
  96          * within the range of +-2billion of the filrst one we find.
  97          */
  98         if (!c->journal_entries_base_seq)
  99                 c->journal_entries_base_seq = max_t(s64, 1, le64_to_cpu(j->seq) - S32_MAX);
 100
 101         /* Drop entries we don't need anymore */
 102         if (last_seq > jlist->last_seq && !c->opts.read_entire_journal) {
 103                 genradix_for_each_from(&c->journal_entries, iter, _i,
 104                                        journal_entry_radix_idx(c, jlist->last_seq)) {
 105                         i = *_i;
 106
 107                         if (!i || i->ignore)
 108                                 continue;
 109
 110                         if (le64_to_cpu(i->j.seq) >= last_seq)
 111                                 break;
 112                         journal_replay_free(c, i);
 113                 }
 114         }
 115
 116         jlist->last_seq = max(jlist->last_seq, last_seq);
 117
 118         _i = genradix_ptr_alloc(&c->journal_entries,
 119                                 journal_entry_radix_idx(c, le64_to_cpu(j->seq)),
 120                                 GFP_KERNEL);
 121         if (!_i)
 122                 return -BCH_ERR_ENOMEM_journal_entry_add;
 123
 124         /*
 125          * Duplicate journal entries? If so we want the one that didn't have a
 126          * checksum error:
 127          */
 128         dup = *_i;
 129         if (dup) {
 130                 if (bytes == vstruct_bytes(&dup->j) &&
 131                     !memcmp(j, &dup->j, bytes)) {
 132                         i = dup;
 133                         goto found;
 134                 }
 135
 136                 if (!entry_ptr.csum_good) {
 137                         i = dup;
 138                         goto found;
 139                 }
 140
 141                 if (!dup->csum_good)
 142                         goto replace;
 143
 144                 fsck_err(c, journal_entry_replicas_data_mismatch,
 145                          "found duplicate but non identical journal entries (seq %llu)",
 146                          le64_to_cpu(j->seq));
 147                 i = dup;
 148                 goto found;
 149         }
 150 replace:
 151         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 152         if (!i)
 153                 return -BCH_ERR_ENOMEM_journal_entry_add;
 154
 155         i->nr_ptrs      = 0;
 156         i->csum_good    = entry_ptr.csum_good;
 157         i->ignore       = false;
 158         unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
 159         i->ptrs[i->nr_ptrs++] = entry_ptr;
 160
 161         if (dup) {
 162                 if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) {
 163                         bch_err(c, "found too many copies of journal entry %llu",
 164                                 le64_to_cpu(i->j.seq));
 165                         dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1;
 166                 }
 167
 168                 /* The first ptr should represent the jset we kept: */
 169                 memcpy(i->ptrs + i->nr_ptrs,
 170                        dup->ptrs,
 171                        sizeof(dup->ptrs[0]) * dup->nr_ptrs);
 172                 i->nr_ptrs += dup->nr_ptrs;
 173                 __journal_replay_free(c, dup);
 174         }
 175
 176         *_i = i;
 177         return 0;
 178 found:
 179         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 180                 if (ptr->dev == ca->dev_idx) {
 181                         bch_err(c, "duplicate journal entry %llu on same device",
 182                                 le64_to_cpu(i->j.seq));
 183                         goto out;
 184                 }
 185         }
 186
 187         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 188                 bch_err(c, "found too many copies of journal entry %llu",
 189                         le64_to_cpu(i->j.seq));
 190                 goto out;
 191         }
 192
 193         i->ptrs[i->nr_ptrs++] = entry_ptr;
 194 out:
 195 fsck_err:
 196         return ret;
 197 }
 198
 199 /* this fills in a range with empty jset_entries: */
 200 static void journal_entry_null_range(void *start, void *end)
 201 {
 202         struct jset_entry *entry;
 203
 204         for (entry = start; entry != end; entry = vstruct_next(entry))
 205                 memset(entry, 0, sizeof(*entry));
 206 }
 207
 208 #define JOURNAL_ENTRY_REREAD    5
 209 #define JOURNAL_ENTRY_NONE      6
 210 #define JOURNAL_ENTRY_BAD       7
 211
 212 static void journal_entry_err_msg(struct printbuf *out,
 213                                   u32 version,
 214                                   struct jset *jset,
 215                                   struct jset_entry *entry)
 216 {
 217         prt_str(out, "invalid journal entry, version=");
 218         bch2_version_to_text(out, version);
 219
 220         if (entry) {
 221                 prt_str(out, " type=");
 222                 prt_str(out, bch2_jset_entry_types[entry->type]);
 223         }
 224
 225         if (!jset) {
 226                 prt_printf(out, " in superblock");
 227         } else {
 228
 229                 prt_printf(out, " seq=%llu", le64_to_cpu(jset->seq));
 230
 231                 if (entry)
 232                         prt_printf(out, " offset=%zi/%u",
 233                                    (u64 *) entry - jset->_data,
 234                                    le32_to_cpu(jset->u64s));
 235         }
 236
 237         prt_str(out, ": ");
 238 }
 239
 240 #define journal_entry_err(c, version, jset, entry, _err, msg, ...)      \
 241 ({                                                                      \
 242         struct printbuf _buf = PRINTBUF;                                \
 243                                                                         \
 244         journal_entry_err_msg(&_buf, version, jset, entry);             \
 245         prt_printf(&_buf, msg, ##__VA_ARGS__);                          \
 246                                                                         \
 247         switch (flags & BKEY_INVALID_WRITE) {                           \
 248         case READ:                                                      \
 249                 mustfix_fsck_err(c, _err, "%s", _buf.buf);              \
 250                 break;                                                  \
 251         case WRITE:                                                     \
 252                 bch2_sb_error_count(c, BCH_FSCK_ERR_##_err);            \
 253                 bch_err(c, "corrupt metadata before write: %s\n", _buf.buf);\
 254                 if (bch2_fs_inconsistent(c)) {                          \
 255                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 256                         goto fsck_err;                                  \
 257                 }                                                       \
 258                 break;                                                  \
 259         }                                                               \
 260                                                                         \
 261         printbuf_exit(&_buf);                                           \
 262         true;                                                           \
 263 })
 264
 265 #define journal_entry_err_on(cond, ...)                                 \
 266         ((cond) ? journal_entry_err(__VA_ARGS__) : false)
 267
 268 #define FSCK_DELETED_KEY        5
 269
 270 static int journal_validate_key(struct bch_fs *c,
 271                                 struct jset *jset,
 272                                 struct jset_entry *entry,
 273                                 unsigned level, enum btree_id btree_id,
 274                                 struct bkey_i *k,
 275                                 unsigned version, int big_endian,
 276                                 enum bkey_invalid_flags flags)
 277 {
 278         int write = flags & BKEY_INVALID_WRITE;
 279         void *next = vstruct_next(entry);
 280         struct printbuf buf = PRINTBUF;
 281         int ret = 0;
 282
 283         if (journal_entry_err_on(!k->k.u64s,
 284                                  c, version, jset, entry,
 285                                  journal_entry_bkey_u64s_0,
 286                                  "k->u64s 0")) {
 287                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 288                 journal_entry_null_range(vstruct_next(entry), next);
 289                 return FSCK_DELETED_KEY;
 290         }
 291
 292         if (journal_entry_err_on((void *) bkey_next(k) >
 293                                  (void *) vstruct_next(entry),
 294                                  c, version, jset, entry,
 295                                  journal_entry_bkey_past_end,
 296                                  "extends past end of journal entry")) {
 297                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 298                 journal_entry_null_range(vstruct_next(entry), next);
 299                 return FSCK_DELETED_KEY;
 300         }
 301
 302         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
 303                                  c, version, jset, entry,
 304                                  journal_entry_bkey_bad_format,
 305                                  "bad format %u", k->k.format)) {
 306                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 307                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 308                 journal_entry_null_range(vstruct_next(entry), next);
 309                 return FSCK_DELETED_KEY;
 310         }
 311
 312         if (!write)
 313                 bch2_bkey_compat(level, btree_id, version, big_endian,
 314                                  write, NULL, bkey_to_packed(k));
 315
 316         if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 317                               __btree_node_type(level, btree_id), write, &buf)) {
 318                 printbuf_reset(&buf);
 319                 journal_entry_err_msg(&buf, version, jset, entry);
 320                 prt_newline(&buf);
 321                 printbuf_indent_add(&buf, 2);
 322
 323                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
 324                 prt_newline(&buf);
 325                 bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 326                                   __btree_node_type(level, btree_id), write, &buf);
 327
 328                 mustfix_fsck_err(c, journal_entry_bkey_invalid,
 329                                  "%s", buf.buf);
 330
 331                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 332                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 333                 journal_entry_null_range(vstruct_next(entry), next);
 334
 335                 printbuf_exit(&buf);
 336                 return FSCK_DELETED_KEY;
 337         }
 338
 339         if (write)
 340                 bch2_bkey_compat(level, btree_id, version, big_endian,
 341                                  write, NULL, bkey_to_packed(k));
 342 fsck_err:
 343         printbuf_exit(&buf);
 344         return ret;
 345 }
 346
 347 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 348                                 struct jset *jset,
 349                                 struct jset_entry *entry,
 350                                 unsigned version, int big_endian,
 351                                 enum bkey_invalid_flags flags)
 352 {
 353         struct bkey_i *k = entry->start;
 354
 355         while (k != vstruct_last(entry)) {
 356                 int ret = journal_validate_key(c, jset, entry,
 357                                                entry->level,
 358                                                entry->btree_id,
 359                                                k, version, big_endian,
 360                                                flags|BKEY_INVALID_JOURNAL);
 361                 if (ret == FSCK_DELETED_KEY)
 362                         continue;
 363
 364                 k = bkey_next(k);
 365         }
 366
 367         return 0;
 368 }
 369
 370 static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
 371                                              struct jset_entry *entry)
 372 {
 373         struct bkey_i *k;
 374         bool first = true;
 375
 376         jset_entry_for_each_key(entry, k) {
 377                 if (!first) {
 378                         prt_newline(out);
 379                         prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 380                 }
 381                 prt_printf(out, "btree=%s l=%u ", bch2_btree_id_str(entry->btree_id), entry->level);
 382                 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 383                 first = false;
 384         }
 385 }
 386
 387 static int journal_entry_btree_root_validate(struct bch_fs *c,
 388                                 struct jset *jset,
 389                                 struct jset_entry *entry,
 390                                 unsigned version, int big_endian,
 391                                 enum bkey_invalid_flags flags)
 392 {
 393         struct bkey_i *k = entry->start;
 394         int ret = 0;
 395
 396         if (journal_entry_err_on(!entry->u64s ||
 397                                  le16_to_cpu(entry->u64s) != k->k.u64s,
 398                                  c, version, jset, entry,
 399                                  journal_entry_btree_root_bad_size,
 400                                  "invalid btree root journal entry: wrong number of keys")) {
 401                 void *next = vstruct_next(entry);
 402                 /*
 403                  * we don't want to null out this jset_entry,
 404                  * just the contents, so that later we can tell
 405                  * we were _supposed_ to have a btree root
 406                  */
 407                 entry->u64s = 0;
 408                 journal_entry_null_range(vstruct_next(entry), next);
 409                 return 0;
 410         }
 411
 412         return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
 413                                     version, big_endian, flags);
 414 fsck_err:
 415         return ret;
 416 }
 417
 418 static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
 419                                              struct jset_entry *entry)
 420 {
 421         journal_entry_btree_keys_to_text(out, c, entry);
 422 }
 423
 424 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 425                                 struct jset *jset,
 426                                 struct jset_entry *entry,
 427                                 unsigned version, int big_endian,
 428                                 enum bkey_invalid_flags flags)
 429 {
 430         /* obsolete, don't care: */
 431         return 0;
 432 }
 433
 434 static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 435                                             struct jset_entry *entry)
 436 {
 437 }
 438
 439 static int journal_entry_blacklist_validate(struct bch_fs *c,
 440                                 struct jset *jset,
 441                                 struct jset_entry *entry,
 442                                 unsigned version, int big_endian,
 443                                 enum bkey_invalid_flags flags)
 444 {
 445         int ret = 0;
 446
 447         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1,
 448                                  c, version, jset, entry,
 449                                  journal_entry_blacklist_bad_size,
 450                 "invalid journal seq blacklist entry: bad size")) {
 451                 journal_entry_null_range(entry, vstruct_next(entry));
 452         }
 453 fsck_err:
 454         return ret;
 455 }
 456
 457 static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
 458                                             struct jset_entry *entry)
 459 {
 460         struct jset_entry_blacklist *bl =
 461                 container_of(entry, struct jset_entry_blacklist, entry);
 462
 463         prt_printf(out, "seq=%llu", le64_to_cpu(bl->seq));
 464 }
 465
 466 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 467                                 struct jset *jset,
 468                                 struct jset_entry *entry,
 469                                 unsigned version, int big_endian,
 470                                 enum bkey_invalid_flags flags)
 471 {
 472         struct jset_entry_blacklist_v2 *bl_entry;
 473         int ret = 0;
 474
 475         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2,
 476                                  c, version, jset, entry,
 477                                  journal_entry_blacklist_v2_bad_size,
 478                 "invalid journal seq blacklist entry: bad size")) {
 479                 journal_entry_null_range(entry, vstruct_next(entry));
 480                 goto out;
 481         }
 482
 483         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 484
 485         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 486                                  le64_to_cpu(bl_entry->end),
 487                                  c, version, jset, entry,
 488                                  journal_entry_blacklist_v2_start_past_end,
 489                 "invalid journal seq blacklist entry: start > end")) {
 490                 journal_entry_null_range(entry, vstruct_next(entry));
 491         }
 492 out:
 493 fsck_err:
 494         return ret;
 495 }
 496
 497 static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
 498                                                struct jset_entry *entry)
 499 {
 500         struct jset_entry_blacklist_v2 *bl =
 501                 container_of(entry, struct jset_entry_blacklist_v2, entry);
 502
 503         prt_printf(out, "start=%llu end=%llu",
 504                le64_to_cpu(bl->start),
 505                le64_to_cpu(bl->end));
 506 }
 507
 508 static int journal_entry_usage_validate(struct bch_fs *c,
 509                                 struct jset *jset,
 510                                 struct jset_entry *entry,
 511                                 unsigned version, int big_endian,
 512                                 enum bkey_invalid_flags flags)
 513 {
 514         struct jset_entry_usage *u =
 515                 container_of(entry, struct jset_entry_usage, entry);
 516         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 517         int ret = 0;
 518
 519         if (journal_entry_err_on(bytes < sizeof(*u),
 520                                  c, version, jset, entry,
 521                                  journal_entry_usage_bad_size,
 522                                  "invalid journal entry usage: bad size")) {
 523                 journal_entry_null_range(entry, vstruct_next(entry));
 524                 return ret;
 525         }
 526
 527 fsck_err:
 528         return ret;
 529 }
 530
 531 static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 532                                         struct jset_entry *entry)
 533 {
 534         struct jset_entry_usage *u =
 535                 container_of(entry, struct jset_entry_usage, entry);
 536
 537         prt_printf(out, "type=%s v=%llu",
 538                bch2_fs_usage_types[u->entry.btree_id],
 539                le64_to_cpu(u->v));
 540 }
 541
 542 static int journal_entry_data_usage_validate(struct bch_fs *c,
 543                                 struct jset *jset,
 544                                 struct jset_entry *entry,
 545                                 unsigned version, int big_endian,
 546                                 enum bkey_invalid_flags flags)
 547 {
 548         struct jset_entry_data_usage *u =
 549                 container_of(entry, struct jset_entry_data_usage, entry);
 550         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 551         struct printbuf err = PRINTBUF;
 552         int ret = 0;
 553
 554         if (journal_entry_err_on(bytes < sizeof(*u) ||
 555                                  bytes < sizeof(*u) + u->r.nr_devs,
 556                                  c, version, jset, entry,
 557                                  journal_entry_data_usage_bad_size,
 558                                  "invalid journal entry usage: bad size")) {
 559                 journal_entry_null_range(entry, vstruct_next(entry));
 560                 goto out;
 561         }
 562
 563         if (journal_entry_err_on(bch2_replicas_entry_validate(&u->r, c->disk_sb.sb, &err),
 564                                  c, version, jset, entry,
 565                                  journal_entry_data_usage_bad_size,
 566                                  "invalid journal entry usage: %s", err.buf)) {
 567                 journal_entry_null_range(entry, vstruct_next(entry));
 568                 goto out;
 569         }
 570 out:
 571 fsck_err:
 572         printbuf_exit(&err);
 573         return ret;
 574 }
 575
 576 static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
 577                                              struct jset_entry *entry)
 578 {
 579         struct jset_entry_data_usage *u =
 580                 container_of(entry, struct jset_entry_data_usage, entry);
 581
 582         bch2_replicas_entry_to_text(out, &u->r);
 583         prt_printf(out, "=%llu", le64_to_cpu(u->v));
 584 }
 585
 586 static int journal_entry_clock_validate(struct bch_fs *c,
 587                                 struct jset *jset,
 588                                 struct jset_entry *entry,
 589                                 unsigned version, int big_endian,
 590                                 enum bkey_invalid_flags flags)
 591 {
 592         struct jset_entry_clock *clock =
 593                 container_of(entry, struct jset_entry_clock, entry);
 594         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 595         int ret = 0;
 596
 597         if (journal_entry_err_on(bytes != sizeof(*clock),
 598                                  c, version, jset, entry,
 599                                  journal_entry_clock_bad_size,
 600                                  "bad size")) {
 601                 journal_entry_null_range(entry, vstruct_next(entry));
 602                 return ret;
 603         }
 604
 605         if (journal_entry_err_on(clock->rw > 1,
 606                                  c, version, jset, entry,
 607                                  journal_entry_clock_bad_rw,
 608                                  "bad rw")) {
 609                 journal_entry_null_range(entry, vstruct_next(entry));
 610                 return ret;
 611         }
 612
 613 fsck_err:
 614         return ret;
 615 }
 616
 617 static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 618                                         struct jset_entry *entry)
 619 {
 620         struct jset_entry_clock *clock =
 621                 container_of(entry, struct jset_entry_clock, entry);
 622
 623         prt_printf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
 624 }
 625
 626 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 627                                 struct jset *jset,
 628                                 struct jset_entry *entry,
 629                                 unsigned version, int big_endian,
 630                                 enum bkey_invalid_flags flags)
 631 {
 632         struct jset_entry_dev_usage *u =
 633                 container_of(entry, struct jset_entry_dev_usage, entry);
 634         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 635         unsigned expected = sizeof(*u);
 636         unsigned dev;
 637         int ret = 0;
 638
 639         if (journal_entry_err_on(bytes < expected,
 640                                  c, version, jset, entry,
 641                                  journal_entry_dev_usage_bad_size,
 642                                  "bad size (%u < %u)",
 643                                  bytes, expected)) {
 644                 journal_entry_null_range(entry, vstruct_next(entry));
 645                 return ret;
 646         }
 647
 648         dev = le32_to_cpu(u->dev);
 649
 650         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 651                                  c, version, jset, entry,
 652                                  journal_entry_dev_usage_bad_dev,
 653                                  "bad dev")) {
 654                 journal_entry_null_range(entry, vstruct_next(entry));
 655                 return ret;
 656         }
 657
 658         if (journal_entry_err_on(u->pad,
 659                                  c, version, jset, entry,
 660                                  journal_entry_dev_usage_bad_pad,
 661                                  "bad pad")) {
 662                 journal_entry_null_range(entry, vstruct_next(entry));
 663                 return ret;
 664         }
 665
 666 fsck_err:
 667         return ret;
 668 }
 669
 670 static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
 671                                             struct jset_entry *entry)
 672 {
 673         struct jset_entry_dev_usage *u =
 674                 container_of(entry, struct jset_entry_dev_usage, entry);
 675         unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 676
 677         prt_printf(out, "dev=%u", le32_to_cpu(u->dev));
 678
 679         for (i = 0; i < nr_types; i++) {
 680                 if (i < BCH_DATA_NR)
 681                         prt_printf(out, " %s", bch2_data_types[i]);
 682                 else
 683                         prt_printf(out, " (unknown data type %u)", i);
 684                 prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 685                        le64_to_cpu(u->d[i].buckets),
 686                        le64_to_cpu(u->d[i].sectors),
 687                        le64_to_cpu(u->d[i].fragmented));
 688         }
 689 }
 690
 691 static int journal_entry_log_validate(struct bch_fs *c,
 692                                 struct jset *jset,
 693                                 struct jset_entry *entry,
 694                                 unsigned version, int big_endian,
 695                                 enum bkey_invalid_flags flags)
 696 {
 697         return 0;
 698 }
 699
 700 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 701                                       struct jset_entry *entry)
 702 {
 703         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 704         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 705
 706         prt_printf(out, "%.*s", bytes, l->d);
 707 }
 708
 709 static int journal_entry_overwrite_validate(struct bch_fs *c,
 710                                 struct jset *jset,
 711                                 struct jset_entry *entry,
 712                                 unsigned version, int big_endian,
 713                                 enum bkey_invalid_flags flags)
 714 {
 715         return journal_entry_btree_keys_validate(c, jset, entry,
 716                                 version, big_endian, READ);
 717 }
 718
 719 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
 720                                             struct jset_entry *entry)
 721 {
 722         journal_entry_btree_keys_to_text(out, c, entry);
 723 }
 724
 725 static int journal_entry_write_buffer_keys_validate(struct bch_fs *c,
 726                                 struct jset *jset,
 727                                 struct jset_entry *entry,
 728                                 unsigned version, int big_endian,
 729                                 enum bkey_invalid_flags flags)
 730 {
 731         return journal_entry_btree_keys_validate(c, jset, entry,
 732                                 version, big_endian, READ);
 733 }
 734
 735 static void journal_entry_write_buffer_keys_to_text(struct printbuf *out, struct bch_fs *c,
 736                                             struct jset_entry *entry)
 737 {
 738         journal_entry_btree_keys_to_text(out, c, entry);
 739 }
 740
 741 struct jset_entry_ops {
 742         int (*validate)(struct bch_fs *, struct jset *,
 743                         struct jset_entry *, unsigned, int,
 744                         enum bkey_invalid_flags);
 745         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 746 };
 747
 748 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 749 #define x(f, nr)                                                \
 750         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 751                 .validate       = journal_entry_##f##_validate, \
 752                 .to_text        = journal_entry_##f##_to_text,  \
 753         },
 754         BCH_JSET_ENTRY_TYPES()
 755 #undef x
 756 };
 757
 758 int bch2_journal_entry_validate(struct bch_fs *c,
 759                                 struct jset *jset,
 760                                 struct jset_entry *entry,
 761                                 unsigned version, int big_endian,
 762                                 enum bkey_invalid_flags flags)
 763 {
 764         return entry->type < BCH_JSET_ENTRY_NR
 765                 ? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
 766                                 version, big_endian, flags)
 767                 : 0;
 768 }
 769
 770 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 771                                 struct jset_entry *entry)
 772 {
 773         if (entry->type < BCH_JSET_ENTRY_NR) {
 774                 prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 775                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 776         } else {
 777                 prt_printf(out, "(unknown type %u)", entry->type);
 778         }
 779 }
 780
 781 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 782                                  enum bkey_invalid_flags flags)
 783 {
 784         struct jset_entry *entry;
 785         unsigned version = le32_to_cpu(jset->version);
 786         int ret = 0;
 787
 788         vstruct_for_each(jset, entry) {
 789                 if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
 790                                 c, version, jset, entry,
 791                                 journal_entry_past_jset_end,
 792                                 "journal entry extends past end of jset")) {
 793                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 794                         break;
 795                 }
 796
 797                 ret = bch2_journal_entry_validate(c, jset, entry,
 798                                         version, JSET_BIG_ENDIAN(jset), flags);
 799                 if (ret)
 800                         break;
 801         }
 802 fsck_err:
 803         return ret;
 804 }
 805
 806 static int jset_validate(struct bch_fs *c,
 807                          struct bch_dev *ca,
 808                          struct jset *jset, u64 sector,
 809                          enum bkey_invalid_flags flags)
 810 {
 811         unsigned version;
 812         int ret = 0;
 813
 814         if (le64_to_cpu(jset->magic) != jset_magic(c))
 815                 return JOURNAL_ENTRY_NONE;
 816
 817         version = le32_to_cpu(jset->version);
 818         if (journal_entry_err_on(!bch2_version_compatible(version),
 819                         c, version, jset, NULL,
 820                         jset_unsupported_version,
 821                         "%s sector %llu seq %llu: incompatible journal entry version %u.%u",
 822                         ca ? ca->name : c->name,
 823                         sector, le64_to_cpu(jset->seq),
 824                         BCH_VERSION_MAJOR(version),
 825                         BCH_VERSION_MINOR(version))) {
 826                 /* don't try to continue: */
 827                 return -EINVAL;
 828         }
 829
 830         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
 831                         c, version, jset, NULL,
 832                         jset_unknown_csum,
 833                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 834                         ca ? ca->name : c->name,
 835                         sector, le64_to_cpu(jset->seq),
 836                         JSET_CSUM_TYPE(jset)))
 837                 ret = JOURNAL_ENTRY_BAD;
 838
 839         /* last_seq is ignored when JSET_NO_FLUSH is true */
 840         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 841                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
 842                                  c, version, jset, NULL,
 843                                  jset_last_seq_newer_than_seq,
 844                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 845                                  le64_to_cpu(jset->last_seq),
 846                                  le64_to_cpu(jset->seq))) {
 847                 jset->last_seq = jset->seq;
 848                 return JOURNAL_ENTRY_BAD;
 849         }
 850
 851         ret = jset_validate_entries(c, jset, flags);
 852 fsck_err:
 853         return ret;
 854 }
 855
 856 static int jset_validate_early(struct bch_fs *c,
 857                          struct bch_dev *ca,
 858                          struct jset *jset, u64 sector,
 859                          unsigned bucket_sectors_left,
 860                          unsigned sectors_read)
 861 {
 862         size_t bytes = vstruct_bytes(jset);
 863         unsigned version;
 864         enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
 865         int ret = 0;
 866
 867         if (le64_to_cpu(jset->magic) != jset_magic(c))
 868                 return JOURNAL_ENTRY_NONE;
 869
 870         version = le32_to_cpu(jset->version);
 871         if (journal_entry_err_on(!bch2_version_compatible(version),
 872                         c, version, jset, NULL,
 873                         jset_unsupported_version,
 874                         "%s sector %llu seq %llu: unknown journal entry version %u.%u",
 875                         ca ? ca->name : c->name,
 876                         sector, le64_to_cpu(jset->seq),
 877                         BCH_VERSION_MAJOR(version),
 878                         BCH_VERSION_MINOR(version))) {
 879                 /* don't try to continue: */
 880                 return -EINVAL;
 881         }
 882
 883         if (bytes > (sectors_read << 9) &&
 884             sectors_read < bucket_sectors_left)
 885                 return JOURNAL_ENTRY_REREAD;
 886
 887         if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
 888                         c, version, jset, NULL,
 889                         jset_past_bucket_end,
 890                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 891                         ca ? ca->name : c->name,
 892                         sector, le64_to_cpu(jset->seq), bytes))
 893                 le32_add_cpu(&jset->u64s,
 894                              -((bytes - (bucket_sectors_left << 9)) / 8));
 895 fsck_err:
 896         return ret;
 897 }
 898
 899 struct journal_read_buf {
 900         void            *data;
 901         size_t          size;
 902 };
 903
 904 static int journal_read_buf_realloc(struct journal_read_buf *b,
 905                                     size_t new_size)
 906 {
 907         void *n;
 908
 909         /* the bios are sized for this many pages, max: */
 910         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 911                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 912
 913         new_size = roundup_pow_of_two(new_size);
 914         n = kvpmalloc(new_size, GFP_KERNEL);
 915         if (!n)
 916                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 917
 918         kvpfree(b->data, b->size);
 919         b->data = n;
 920         b->size = new_size;
 921         return 0;
 922 }
 923
 924 static int journal_read_bucket(struct bch_dev *ca,
 925                                struct journal_read_buf *buf,
 926                                struct journal_list *jlist,
 927                                unsigned bucket)
 928 {
 929         struct bch_fs *c = ca->fs;
 930         struct journal_device *ja = &ca->journal;
 931         struct jset *j = NULL;
 932         unsigned sectors, sectors_read = 0;
 933         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 934             end = offset + ca->mi.bucket_size;
 935         bool saw_bad = false, csum_good;
 936         int ret = 0;
 937
 938         pr_debug("reading %u", bucket);
 939
 940         while (offset < end) {
 941                 if (!sectors_read) {
 942                         struct bio *bio;
 943                         unsigned nr_bvecs;
 944 reread:
 945                         sectors_read = min_t(unsigned,
 946                                 end - offset, buf->size >> 9);
 947                         nr_bvecs = buf_pages(buf->data, sectors_read << 9);
 948
 949                         bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
 950                         bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
 951
 952                         bio->bi_iter.bi_sector = offset;
 953                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 954
 955                         ret = submit_bio_wait(bio);
 956                         kfree(bio);
 957
 958                         if (bch2_dev_io_err_on(ret, ca, BCH_MEMBER_ERROR_read,
 959                                                "journal read error: sector %llu",
 960                                                offset) ||
 961                             bch2_meta_read_fault("journal")) {
 962                                 /*
 963                                  * We don't error out of the recovery process
 964                                  * here, since the relevant journal entry may be
 965                                  * found on a different device, and missing or
 966                                  * no journal entries will be handled later
 967                                  */
 968                                 return 0;
 969                         }
 970
 971                         j = buf->data;
 972                 }
 973
 974                 ret = jset_validate_early(c, ca, j, offset,
 975                                     end - offset, sectors_read);
 976                 switch (ret) {
 977                 case 0:
 978                         sectors = vstruct_sectors(j, c->block_bits);
 979                         break;
 980                 case JOURNAL_ENTRY_REREAD:
 981                         if (vstruct_bytes(j) > buf->size) {
 982                                 ret = journal_read_buf_realloc(buf,
 983                                                         vstruct_bytes(j));
 984                                 if (ret)
 985                                         return ret;
 986                         }
 987                         goto reread;
 988                 case JOURNAL_ENTRY_NONE:
 989                         if (!saw_bad)
 990                                 return 0;
 991                         /*
 992                          * On checksum error we don't really trust the size
 993                          * field of the journal entry we read, so try reading
 994                          * again at next block boundary:
 995                          */
 996                         sectors = block_sectors(c);
 997                         goto next_block;
 998                 default:
 999                         return ret;
1000                 }
1001
1002                 /*
1003                  * This happens sometimes if we don't have discards on -
1004                  * when we've partially overwritten a bucket with new
1005                  * journal entries. We don't need the rest of the
1006                  * bucket:
1007                  */
1008                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
1009                         return 0;
1010
1011                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
1012
1013                 csum_good = jset_csum_good(c, j);
1014                 if (bch2_dev_io_err_on(!csum_good, ca, BCH_MEMBER_ERROR_checksum,
1015                                        "journal checksum error"))
1016                         saw_bad = true;
1017
1018                 ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
1019                              j->encrypted_start,
1020                              vstruct_end(j) - (void *) j->encrypted_start);
1021                 bch2_fs_fatal_err_on(ret, c,
1022                                 "error decrypting journal entry: %i", ret);
1023
1024                 mutex_lock(&jlist->lock);
1025                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
1026                                         .csum_good      = csum_good,
1027                                         .dev            = ca->dev_idx,
1028                                         .bucket         = bucket,
1029                                         .bucket_offset  = offset -
1030                                                 bucket_to_sector(ca, ja->buckets[bucket]),
1031                                         .sector         = offset,
1032                                         }, jlist, j);
1033                 mutex_unlock(&jlist->lock);
1034
1035                 switch (ret) {
1036                 case JOURNAL_ENTRY_ADD_OK:
1037                         break;
1038                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
1039                         break;
1040                 default:
1041                         return ret;
1042                 }
1043 next_block:
1044                 pr_debug("next");
1045                 offset          += sectors;
1046                 sectors_read    -= sectors;
1047                 j = ((void *) j) + (sectors << 9);
1048         }
1049
1050         return 0;
1051 }
1052
1053 static CLOSURE_CALLBACK(bch2_journal_read_device)
1054 {
1055         closure_type(ja, struct journal_device, read);
1056         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
1057         struct bch_fs *c = ca->fs;
1058         struct journal_list *jlist =
1059                 container_of(cl->parent, struct journal_list, cl);
1060         struct journal_replay *r, **_r;
1061         struct genradix_iter iter;
1062         struct journal_read_buf buf = { NULL, 0 };
1063         unsigned i;
1064         int ret = 0;
1065
1066         if (!ja->nr)
1067                 goto out;
1068
1069         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
1070         if (ret)
1071                 goto err;
1072
1073         pr_debug("%u journal buckets", ja->nr);
1074
1075         for (i = 0; i < ja->nr; i++) {
1076                 ret = journal_read_bucket(ca, &buf, jlist, i);
1077                 if (ret)
1078                         goto err;
1079         }
1080
1081         ja->sectors_free = ca->mi.bucket_size;
1082
1083         mutex_lock(&jlist->lock);
1084         genradix_for_each_reverse(&c->journal_entries, iter, _r) {
1085                 r = *_r;
1086
1087                 if (!r)
1088                         continue;
1089
1090                 for (i = 0; i < r->nr_ptrs; i++) {
1091                         if (r->ptrs[i].dev == ca->dev_idx) {
1092                                 unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
1093                                         vstruct_sectors(&r->j, c->block_bits);
1094
1095                                 ja->cur_idx = r->ptrs[i].bucket;
1096                                 ja->sectors_free = ca->mi.bucket_size - wrote;
1097                                 goto found;
1098                         }
1099                 }
1100         }
1101 found:
1102         mutex_unlock(&jlist->lock);
1103
1104         if (ja->bucket_seq[ja->cur_idx] &&
1105             ja->sectors_free == ca->mi.bucket_size) {
1106 #if 0
1107                 /*
1108                  * Debug code for ZNS support, where we (probably) want to be
1109                  * correlated where we stopped in the journal to the zone write
1110                  * points:
1111                  */
1112                 bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
1113                 bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
1114                 for (i = 0; i < 3; i++) {
1115                         unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
1116
1117                         bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
1118                 }
1119 #endif
1120                 ja->sectors_free = 0;
1121         }
1122
1123         /*
1124          * Set dirty_idx to indicate the entire journal is full and needs to be
1125          * reclaimed - journal reclaim will immediately reclaim whatever isn't
1126          * pinned when it first runs:
1127          */
1128         ja->discard_idx = ja->dirty_idx_ondisk =
1129                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
1130 out:
1131         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
1132         kvpfree(buf.data, buf.size);
1133         percpu_ref_put(&ca->io_ref);
1134         closure_return(cl);
1135         return;
1136 err:
1137         mutex_lock(&jlist->lock);
1138         jlist->ret = ret;
1139         mutex_unlock(&jlist->lock);
1140         goto out;
1141 }
1142
1143 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1144                                struct journal_replay *j)
1145 {
1146         unsigned i;
1147
1148         for (i = 0; i < j->nr_ptrs; i++) {
1149                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
1150                 u64 offset;
1151
1152                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
1153
1154                 if (i)
1155                         prt_printf(out, " ");
1156                 prt_printf(out, "%u:%u:%u (sector %llu)",
1157                        j->ptrs[i].dev,
1158                        j->ptrs[i].bucket,
1159                        j->ptrs[i].bucket_offset,
1160                        j->ptrs[i].sector);
1161         }
1162 }
1163
1164 int bch2_journal_read(struct bch_fs *c,
1165                       u64 *last_seq,
1166                       u64 *blacklist_seq,
1167                       u64 *start_seq)
1168 {
1169         struct journal_list jlist;
1170         struct journal_replay *i, **_i, *prev = NULL;
1171         struct genradix_iter radix_iter;
1172         struct bch_dev *ca;
1173         unsigned iter;
1174         struct printbuf buf = PRINTBUF;
1175         bool degraded = false, last_write_torn = false;
1176         u64 seq;
1177         int ret = 0;
1178
1179         closure_init_stack(&jlist.cl);
1180         mutex_init(&jlist.lock);
1181         jlist.last_seq = 0;
1182         jlist.ret = 0;
1183
1184         for_each_member_device(ca, c, iter) {
1185                 if (!c->opts.fsck &&
1186                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1187                         continue;
1188
1189                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1190                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1191                     percpu_ref_tryget(&ca->io_ref))
1192                         closure_call(&ca->journal.read,
1193                                      bch2_journal_read_device,
1194                                      system_unbound_wq,
1195                                      &jlist.cl);
1196                 else
1197                         degraded = true;
1198         }
1199
1200         closure_sync(&jlist.cl);
1201
1202         if (jlist.ret)
1203                 return jlist.ret;
1204
1205         *last_seq       = 0;
1206         *start_seq      = 0;
1207         *blacklist_seq  = 0;
1208
1209         /*
1210          * Find most recent flush entry, and ignore newer non flush entries -
1211          * those entries will be blacklisted:
1212          */
1213         genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
1214                 enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
1215
1216                 i = *_i;
1217
1218                 if (!i || i->ignore)
1219                         continue;
1220
1221                 if (!*start_seq)
1222                         *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
1223
1224                 if (JSET_NO_FLUSH(&i->j)) {
1225                         i->ignore = true;
1226                         continue;
1227                 }
1228
1229                 if (!last_write_torn && !i->csum_good) {
1230                         last_write_torn = true;
1231                         i->ignore = true;
1232                         continue;
1233                 }
1234
1235                 if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
1236                                          c, le32_to_cpu(i->j.version), &i->j, NULL,
1237                                          jset_last_seq_newer_than_seq,
1238                                          "invalid journal entry: last_seq > seq (%llu > %llu)",
1239                                          le64_to_cpu(i->j.last_seq),
1240                                          le64_to_cpu(i->j.seq)))
1241                         i->j.last_seq = i->j.seq;
1242
1243                 *last_seq       = le64_to_cpu(i->j.last_seq);
1244                 *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1245                 break;
1246         }
1247
1248         if (!*start_seq) {
1249                 bch_info(c, "journal read done, but no entries found");
1250                 return 0;
1251         }
1252
1253         if (!*last_seq) {
1254                 fsck_err(c, dirty_but_no_journal_entries_post_drop_nonflushes,
1255                          "journal read done, but no entries found after dropping non-flushes");
1256                 return 0;
1257         }
1258
1259         bch_info(c, "journal read done, replaying entries %llu-%llu",
1260                  *last_seq, *blacklist_seq - 1);
1261
1262         if (*start_seq != *blacklist_seq)
1263                 bch_info(c, "dropped unflushed entries %llu-%llu",
1264                          *blacklist_seq, *start_seq - 1);
1265
1266         /* Drop blacklisted entries and entries older than last_seq: */
1267         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1268                 i = *_i;
1269
1270                 if (!i || i->ignore)
1271                         continue;
1272
1273                 seq = le64_to_cpu(i->j.seq);
1274                 if (seq < *last_seq) {
1275                         journal_replay_free(c, i);
1276                         continue;
1277                 }
1278
1279                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1280                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1281                                     jset_seq_blacklisted,
1282                                     "found blacklisted journal entry %llu", seq);
1283                         i->ignore = true;
1284                 }
1285         }
1286
1287         /* Check for missing entries: */
1288         seq = *last_seq;
1289         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1290                 i = *_i;
1291
1292                 if (!i || i->ignore)
1293                         continue;
1294
1295                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1296
1297                 while (seq < le64_to_cpu(i->j.seq)) {
1298                         u64 missing_start, missing_end;
1299                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1300
1301                         while (seq < le64_to_cpu(i->j.seq) &&
1302                                bch2_journal_seq_is_blacklisted(c, seq, false))
1303                                 seq++;
1304
1305                         if (seq == le64_to_cpu(i->j.seq))
1306                                 break;
1307
1308                         missing_start = seq;
1309
1310                         while (seq < le64_to_cpu(i->j.seq) &&
1311                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1312                                 seq++;
1313
1314                         if (prev) {
1315                                 bch2_journal_ptrs_to_text(&buf1, c, prev);
1316                                 prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits));
1317                         } else
1318                                 prt_printf(&buf1, "(none)");
1319                         bch2_journal_ptrs_to_text(&buf2, c, i);
1320
1321                         missing_end = seq - 1;
1322                         fsck_err(c, journal_entries_missing,
1323                                  "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1324                                  "  prev at %s\n"
1325                                  "  next at %s",
1326                                  missing_start, missing_end,
1327                                  *last_seq, *blacklist_seq - 1,
1328                                  buf1.buf, buf2.buf);
1329
1330                         printbuf_exit(&buf1);
1331                         printbuf_exit(&buf2);
1332                 }
1333
1334                 prev = i;
1335                 seq++;
1336         }
1337
1338         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1339                 struct bch_replicas_padded replicas = {
1340                         .e.data_type = BCH_DATA_journal,
1341                         .e.nr_required = 1,
1342                 };
1343                 unsigned ptr;
1344
1345                 i = *_i;
1346                 if (!i || i->ignore)
1347                         continue;
1348
1349                 for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
1350                         ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
1351
1352                         if (!i->ptrs[ptr].csum_good)
1353                                 bch_err_dev_offset(ca, i->ptrs[ptr].sector,
1354                                                    "invalid journal checksum, seq %llu%s",
1355                                                    le64_to_cpu(i->j.seq),
1356                                                    i->csum_good ? " (had good copy on another device)" : "");
1357                 }
1358
1359                 ret = jset_validate(c,
1360                                     bch_dev_bkey_exists(c, i->ptrs[0].dev),
1361                                     &i->j,
1362                                     i->ptrs[0].sector,
1363                                     READ);
1364                 if (ret)
1365                         goto err;
1366
1367                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1368                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1369
1370                 bch2_replicas_entry_sort(&replicas.e);
1371
1372                 printbuf_reset(&buf);
1373                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1374
1375                 if (!degraded &&
1376                     !bch2_replicas_marked(c, &replicas.e) &&
1377                     (le64_to_cpu(i->j.seq) == *last_seq ||
1378                      fsck_err(c, journal_entry_replicas_not_marked,
1379                               "superblock not marked as containing replicas for journal entry %llu\n  %s",
1380                               le64_to_cpu(i->j.seq), buf.buf))) {
1381                         ret = bch2_mark_replicas(c, &replicas.e);
1382                         if (ret)
1383                                 goto err;
1384                 }
1385         }
1386 err:
1387 fsck_err:
1388         printbuf_exit(&buf);
1389         return ret;
1390 }
1391
1392 /* journal write: */
1393
1394 static void __journal_write_alloc(struct journal *j,
1395                                   struct journal_buf *w,
1396                                   struct dev_alloc_list *devs_sorted,
1397                                   unsigned sectors,
1398                                   unsigned *replicas,
1399                                   unsigned replicas_want)
1400 {
1401         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1402         struct journal_device *ja;
1403         struct bch_dev *ca;
1404         unsigned i;
1405
1406         if (*replicas >= replicas_want)
1407                 return;
1408
1409         for (i = 0; i < devs_sorted->nr; i++) {
1410                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1411                 if (!ca)
1412                         continue;
1413
1414                 ja = &ca->journal;
1415
1416                 /*
1417                  * Check that we can use this device, and aren't already using
1418                  * it:
1419                  */
1420                 if (!ca->mi.durability ||
1421                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1422                     !ja->nr ||
1423                     bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) ||
1424                     sectors > ja->sectors_free)
1425                         continue;
1426
1427                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1428
1429                 bch2_bkey_append_ptr(&w->key,
1430                         (struct bch_extent_ptr) {
1431                                   .offset = bucket_to_sector(ca,
1432                                         ja->buckets[ja->cur_idx]) +
1433                                         ca->mi.bucket_size -
1434                                         ja->sectors_free,
1435                                   .dev = ca->dev_idx,
1436                 });
1437
1438                 ja->sectors_free -= sectors;
1439                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1440
1441                 *replicas += ca->mi.durability;
1442
1443                 if (*replicas >= replicas_want)
1444                         break;
1445         }
1446 }
1447
1448 /**
1449  * journal_write_alloc - decide where to write next journal entry
1450  *
1451  * @j:          journal object
1452  * @w:          journal buf (entry to be written)
1453  *
1454  * Returns: 0 on success, or -EROFS on failure
1455  */
1456 static int journal_write_alloc(struct journal *j, struct journal_buf *w)
1457 {
1458         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1459         struct bch_devs_mask devs;
1460         struct journal_device *ja;
1461         struct bch_dev *ca;
1462         struct dev_alloc_list devs_sorted;
1463         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1464         unsigned target = c->opts.metadata_target ?:
1465                 c->opts.foreground_target;
1466         unsigned i, replicas = 0, replicas_want =
1467                 READ_ONCE(c->opts.metadata_replicas);
1468
1469         rcu_read_lock();
1470 retry:
1471         devs = target_rw_devs(c, BCH_DATA_journal, target);
1472
1473         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1474
1475         __journal_write_alloc(j, w, &devs_sorted,
1476                               sectors, &replicas, replicas_want);
1477
1478         if (replicas >= replicas_want)
1479                 goto done;
1480
1481         for (i = 0; i < devs_sorted.nr; i++) {
1482                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1483                 if (!ca)
1484                         continue;
1485
1486                 ja = &ca->journal;
1487
1488                 if (sectors > ja->sectors_free &&
1489                     sectors <= ca->mi.bucket_size &&
1490                     bch2_journal_dev_buckets_available(j, ja,
1491                                         journal_space_discarded)) {
1492                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1493                         ja->sectors_free = ca->mi.bucket_size;
1494
1495                         /*
1496                          * ja->bucket_seq[ja->cur_idx] must always have
1497                          * something sensible:
1498                          */
1499                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1500                 }
1501         }
1502
1503         __journal_write_alloc(j, w, &devs_sorted,
1504                               sectors, &replicas, replicas_want);
1505
1506         if (replicas < replicas_want && target) {
1507                 /* Retry from all devices: */
1508                 target = 0;
1509                 goto retry;
1510         }
1511 done:
1512         rcu_read_unlock();
1513
1514         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1515
1516         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1517 }
1518
1519 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1520 {
1521         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1522
1523         /* we aren't holding j->lock: */
1524         unsigned new_size = READ_ONCE(j->buf_size_want);
1525         void *new_buf;
1526
1527         if (buf->buf_size >= new_size)
1528                 return;
1529
1530         size_t btree_write_buffer_size = new_size / 64;
1531
1532         if (bch2_btree_write_buffer_resize(c, btree_write_buffer_size))
1533                 return;
1534
1535         new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
1536         if (!new_buf)
1537                 return;
1538
1539         memcpy(new_buf, buf->data, buf->buf_size);
1540
1541         spin_lock(&j->lock);
1542         swap(buf->data,         new_buf);
1543         swap(buf->buf_size,     new_size);
1544         spin_unlock(&j->lock);
1545
1546         kvpfree(new_buf, new_size);
1547 }
1548
1549 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1550 {
1551         return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
1552 }
1553
1554 static CLOSURE_CALLBACK(journal_write_done)
1555 {
1556         closure_type(j, struct journal, io);
1557         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1558         struct journal_buf *w = journal_last_unwritten_buf(j);
1559         struct bch_replicas_padded replicas;
1560         union journal_res_state old, new;
1561         u64 v, seq;
1562         int err = 0;
1563
1564         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1565                                ? j->flush_write_time
1566                                : j->noflush_write_time, j->write_start_time);
1567
1568         if (!w->devs_written.nr) {
1569                 bch_err(c, "unable to write journal to sufficient devices");
1570                 err = -EIO;
1571         } else {
1572                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1573                                          w->devs_written);
1574                 if (bch2_mark_replicas(c, &replicas.e))
1575                         err = -EIO;
1576         }
1577
1578         if (err)
1579                 bch2_fatal_error(c);
1580
1581         spin_lock(&j->lock);
1582         seq = le64_to_cpu(w->data->seq);
1583
1584         if (seq >= j->pin.front)
1585                 journal_seq_pin(j, seq)->devs = w->devs_written;
1586
1587         if (!err) {
1588                 if (!JSET_NO_FLUSH(w->data)) {
1589                         j->flushed_seq_ondisk = seq;
1590                         j->last_seq_ondisk = w->last_seq;
1591
1592                         bch2_do_discards(c);
1593                         closure_wake_up(&c->freelist_wait);
1594
1595                         bch2_reset_alloc_cursors(c);
1596                 }
1597         } else if (!j->err_seq || seq < j->err_seq)
1598                 j->err_seq      = seq;
1599
1600         j->seq_ondisk           = seq;
1601
1602         /*
1603          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1604          * more buckets:
1605          *
1606          * Must come before signaling write completion, for
1607          * bch2_fs_journal_stop():
1608          */
1609         if (j->watermark != BCH_WATERMARK_stripe)
1610                 journal_reclaim_kick(&c->journal);
1611
1612         /* also must come before signalling write completion: */
1613         closure_debug_destroy(cl);
1614
1615         v = atomic64_read(&j->reservations.counter);
1616         do {
1617                 old.v = new.v = v;
1618                 BUG_ON(journal_state_count(new, new.unwritten_idx));
1619
1620                 new.unwritten_idx++;
1621         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1622                                        old.v, new.v)) != old.v);
1623
1624         bch2_journal_reclaim_fast(j);
1625         bch2_journal_space_available(j);
1626
1627         track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
1628                            &j->max_in_flight_start, false);
1629
1630         closure_wake_up(&w->wait);
1631         journal_wake(j);
1632
1633         if (!journal_state_count(new, new.unwritten_idx) &&
1634             journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
1635                 spin_unlock(&j->lock);
1636                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1637         } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
1638                    new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
1639                 struct journal_buf *buf = journal_cur_buf(j);
1640                 long delta = buf->expires - jiffies;
1641
1642                 /*
1643                  * We don't close a journal entry to write it while there's
1644                  * previous entries still in flight - the current journal entry
1645                  * might want to be written now:
1646                  */
1647
1648                 spin_unlock(&j->lock);
1649                 mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
1650         } else {
1651                 spin_unlock(&j->lock);
1652         }
1653 }
1654
1655 static void journal_write_endio(struct bio *bio)
1656 {
1657         struct bch_dev *ca = bio->bi_private;
1658         struct journal *j = &ca->fs->journal;
1659         struct journal_buf *w = journal_last_unwritten_buf(j);
1660         unsigned long flags;
1661
1662         if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1663                                "error writing journal entry %llu: %s",
1664                                le64_to_cpu(w->data->seq),
1665                                bch2_blk_status_to_str(bio->bi_status)) ||
1666             bch2_meta_write_fault("journal")) {
1667                 spin_lock_irqsave(&j->err_lock, flags);
1668                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1669                 spin_unlock_irqrestore(&j->err_lock, flags);
1670         }
1671
1672         closure_put(&j->io);
1673         percpu_ref_put(&ca->io_ref);
1674 }
1675
1676 static CLOSURE_CALLBACK(do_journal_write)
1677 {
1678         closure_type(j, struct journal, io);
1679         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1680         struct bch_dev *ca;
1681         struct journal_buf *w = journal_last_unwritten_buf(j);
1682         struct bch_extent_ptr *ptr;
1683         struct bio *bio;
1684         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1685
1686         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1687                 ca = bch_dev_bkey_exists(c, ptr->dev);
1688                 if (!percpu_ref_tryget(&ca->io_ref)) {
1689                         /* XXX: fix this */
1690                         bch_err(c, "missing device for journal write\n");
1691                         continue;
1692                 }
1693
1694                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1695                              sectors);
1696
1697                 bio = ca->journal.bio;
1698                 bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
1699                 bio->bi_iter.bi_sector  = ptr->offset;
1700                 bio->bi_end_io          = journal_write_endio;
1701                 bio->bi_private         = ca;
1702
1703                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1704                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1705
1706                 if (!JSET_NO_FLUSH(w->data))
1707                         bio->bi_opf    |= REQ_FUA;
1708                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1709                         bio->bi_opf    |= REQ_PREFLUSH;
1710
1711                 bch2_bio_map(bio, w->data, sectors << 9);
1712
1713                 trace_and_count(c, journal_write, bio);
1714                 closure_bio_submit(bio, cl);
1715
1716                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1717                         le64_to_cpu(w->data->seq);
1718         }
1719
1720         continue_at(cl, journal_write_done, c->io_complete_wq);
1721 }
1722
1723 static int bch2_journal_write_prep(struct journal *j, struct journal_buf *w)
1724 {
1725         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1726         struct jset_entry *start, *end, *i;
1727         struct jset *jset = w->data;
1728         struct journal_keys_to_wb wb = { NULL };
1729         unsigned sectors, bytes, u64s;
1730         unsigned long btree_roots_have = 0;
1731         bool validate_before_checksum = false;
1732         u64 seq = le64_to_cpu(jset->seq);
1733         int ret;
1734
1735         /*
1736          * Simple compaction, dropping empty jset_entries (from journal
1737          * reservations that weren't fully used) and merging jset_entries that
1738          * can be.
1739          *
1740          * If we wanted to be really fancy here, we could sort all the keys in
1741          * the jset and drop keys that were overwritten - probably not worth it:
1742          */
1743         vstruct_for_each(jset, i) {
1744                 unsigned u64s = le16_to_cpu(i->u64s);
1745
1746                 /* Empty entry: */
1747                 if (!u64s)
1748                         continue;
1749
1750                 /*
1751                  * New btree roots are set by journalling them; when the journal
1752                  * entry gets written we have to propagate them to
1753                  * c->btree_roots
1754                  *
1755                  * But, every journal entry we write has to contain all the
1756                  * btree roots (at least for now); so after we copy btree roots
1757                  * to c->btree_roots we have to get any missing btree roots and
1758                  * add them to this journal entry:
1759                  */
1760                 switch (i->type) {
1761                 case BCH_JSET_ENTRY_btree_root:
1762                         bch2_journal_entry_to_btree_root(c, i);
1763                         __set_bit(i->btree_id, &btree_roots_have);
1764                         break;
1765                 case BCH_JSET_ENTRY_write_buffer_keys:
1766                         EBUG_ON(!w->need_flush_to_write_buffer);
1767
1768                         if (!wb.wb)
1769                                 bch2_journal_keys_to_write_buffer_start(c, &wb, seq);
1770
1771                         struct bkey_i *k;
1772                         jset_entry_for_each_key(i, k) {
1773                                 ret = bch2_journal_key_to_wb(c, &wb, i->btree_id, k);
1774                                 if (ret) {
1775                                         bch2_fs_fatal_error(c, "-ENOMEM flushing journal keys to btree write buffer");
1776                                         bch2_journal_keys_to_write_buffer_end(c, &wb);
1777                                         return ret;
1778                                 }
1779                         }
1780                         i->type = BCH_JSET_ENTRY_btree_keys;
1781                         break;
1782                 }
1783         }
1784
1785         if (wb.wb)
1786                 bch2_journal_keys_to_write_buffer_end(c, &wb);
1787         w->need_flush_to_write_buffer = false;
1788
1789         start = end = vstruct_last(jset);
1790
1791         end     = bch2_btree_roots_to_journal_entries(c, end, btree_roots_have);
1792
1793         bch2_journal_super_entries_add_common(c, &end, seq);
1794         u64s    = (u64 *) end - (u64 *) start;
1795         BUG_ON(u64s > j->entry_u64s_reserved);
1796
1797         le32_add_cpu(&jset->u64s, u64s);
1798
1799         sectors = vstruct_sectors(jset, c->block_bits);
1800         bytes   = vstruct_bytes(jset);
1801
1802         if (sectors > w->sectors) {
1803                 bch2_fs_fatal_error(c, "aieeee! journal write overran available space, %zu > %u (extra %u reserved %u/%u)",
1804                                     vstruct_bytes(jset), w->sectors << 9,
1805                                     u64s, w->u64s_reserved, j->entry_u64s_reserved);
1806                 return -EINVAL;
1807         }
1808
1809         jset->magic             = cpu_to_le64(jset_magic(c));
1810         jset->version           = cpu_to_le32(c->sb.version);
1811
1812         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1813         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1814
1815         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1816                 j->last_empty_seq = seq;
1817
1818         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1819                 validate_before_checksum = true;
1820
1821         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1822                 validate_before_checksum = true;
1823
1824         if (validate_before_checksum &&
1825             (ret = jset_validate(c, NULL, jset, 0, WRITE)))
1826                 return ret;
1827
1828         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1829                     jset->encrypted_start,
1830                     vstruct_end(jset) - (void *) jset->encrypted_start);
1831         if (bch2_fs_fatal_err_on(ret, c,
1832                         "error decrypting journal entry: %i", ret))
1833                 return ret;
1834
1835         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1836                                   journal_nonce(jset), jset);
1837
1838         if (!validate_before_checksum &&
1839             (ret = jset_validate(c, NULL, jset, 0, WRITE)))
1840                 return ret;
1841
1842         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1843         return 0;
1844 }
1845
1846 static int bch2_journal_write_pick_flush(struct journal *j, struct journal_buf *w)
1847 {
1848         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1849         int error = bch2_journal_error(j);
1850
1851         /*
1852          * If the journal is in an error state - we did an emergency shutdown -
1853          * we prefer to continue doing journal writes. We just mark them as
1854          * noflush so they'll never be used, but they'll still be visible by the
1855          * list_journal tool - this helps in debugging.
1856          *
1857          * There's a caveat: the first journal write after marking the
1858          * superblock dirty must always be a flush write, because on startup
1859          * from a clean shutdown we didn't necessarily read the journal and the
1860          * new journal write might overwrite whatever was in the journal
1861          * previously - we can't leave the journal without any flush writes in
1862          * it.
1863          *
1864          * So if we're in an error state, and we're still starting up, we don't
1865          * write anything at all.
1866          */
1867         if (error && test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags))
1868                 return -EIO;
1869
1870         if (error ||
1871             w->noflush ||
1872             (!w->must_flush &&
1873              (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1874              test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags))) {
1875                 w->noflush = true;
1876                 SET_JSET_NO_FLUSH(w->data, true);
1877                 w->data->last_seq       = 0;
1878                 w->last_seq             = 0;
1879
1880                 j->nr_noflush_writes++;
1881         } else {
1882                 j->last_flush_write = jiffies;
1883                 j->nr_flush_writes++;
1884                 clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
1885         }
1886
1887         return 0;
1888 }
1889
1890 CLOSURE_CALLBACK(bch2_journal_write)
1891 {
1892         closure_type(j, struct journal, io);
1893         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1894         struct bch_dev *ca;
1895         struct journal_buf *w = journal_last_unwritten_buf(j);
1896         struct bch_replicas_padded replicas;
1897         struct bio *bio;
1898         struct printbuf journal_debug_buf = PRINTBUF;
1899         unsigned i, nr_rw_members = 0;
1900         int ret;
1901
1902         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1903
1904         j->write_start_time = local_clock();
1905
1906         spin_lock(&j->lock);
1907         ret = bch2_journal_write_pick_flush(j, w);
1908         spin_unlock(&j->lock);
1909         if (ret)
1910                 goto err;
1911
1912         mutex_lock(&j->buf_lock);
1913         journal_buf_realloc(j, w);
1914
1915         ret = bch2_journal_write_prep(j, w);
1916         mutex_unlock(&j->buf_lock);
1917         if (ret)
1918                 goto err;
1919
1920         j->entry_bytes_written += vstruct_bytes(w->data);
1921
1922         while (1) {
1923                 spin_lock(&j->lock);
1924                 ret = journal_write_alloc(j, w);
1925                 if (!ret || !j->can_discard)
1926                         break;
1927
1928                 spin_unlock(&j->lock);
1929                 bch2_journal_do_discards(j);
1930         }
1931
1932         if (ret) {
1933                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1934                 spin_unlock(&j->lock);
1935                 bch_err(c, "Unable to allocate journal write:\n%s",
1936                         journal_debug_buf.buf);
1937                 printbuf_exit(&journal_debug_buf);
1938                 goto err;
1939         }
1940
1941         /*
1942          * write is allocated, no longer need to account for it in
1943          * bch2_journal_space_available():
1944          */
1945         w->sectors = 0;
1946
1947         /*
1948          * journal entry has been compacted and allocated, recalculate space
1949          * available:
1950          */
1951         bch2_journal_space_available(j);
1952         spin_unlock(&j->lock);
1953
1954         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1955
1956         if (c->opts.nochanges)
1957                 goto no_io;
1958
1959         for_each_rw_member(ca, c, i)
1960                 nr_rw_members++;
1961
1962         if (nr_rw_members > 1)
1963                 w->separate_flush = true;
1964
1965         /*
1966          * Mark journal replicas before we submit the write to guarantee
1967          * recovery will find the journal entries after a crash.
1968          */
1969         bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1970                                  w->devs_written);
1971         ret = bch2_mark_replicas(c, &replicas.e);
1972         if (ret)
1973                 goto err;
1974
1975         if (!JSET_NO_FLUSH(w->data) && w->separate_flush) {
1976                 for_each_rw_member(ca, c, i) {
1977                         percpu_ref_get(&ca->io_ref);
1978
1979                         bio = ca->journal.bio;
1980                         bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
1981                         bio->bi_end_io          = journal_write_endio;
1982                         bio->bi_private         = ca;
1983                         closure_bio_submit(bio, cl);
1984                 }
1985         }
1986
1987         continue_at(cl, do_journal_write, c->io_complete_wq);
1988         return;
1989 no_io:
1990         continue_at(cl, journal_write_done, c->io_complete_wq);
1991         return;
1992 err:
1993         bch2_fatal_error(c);
1994         continue_at(cl, journal_write_done, c->io_complete_wq);
1995 }