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, struct bch_csum *csum)
  31 {
  32         if (!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(j))) {
  33                 *csum = (struct bch_csum) {};
  34                 return false;
  35         }
  36
  37         *csum = csum_vstruct(c, JSET_CSUM_TYPE(j), journal_nonce(j), j);
  38         return !bch2_crc_cmp(j->csum, *csum);
  39 }
  40
  41 static inline u32 journal_entry_radix_idx(struct bch_fs *c, u64 seq)
  42 {
  43         return (seq - c->journal_entries_base_seq) & (~0U >> 1);
  44 }
  45
  46 static void __journal_replay_free(struct bch_fs *c,
  47                                   struct journal_replay *i)
  48 {
  49         struct journal_replay **p =
  50                 genradix_ptr(&c->journal_entries,
  51                              journal_entry_radix_idx(c, le64_to_cpu(i->j.seq)));
  52
  53         BUG_ON(*p != i);
  54         *p = NULL;
  55         kvpfree(i, offsetof(struct journal_replay, j) +
  56                 vstruct_bytes(&i->j));
  57 }
  58
  59 static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
  60 {
  61         i->ignore = true;
  62
  63         if (!c->opts.read_entire_journal)
  64                 __journal_replay_free(c, i);
  65 }
  66
  67 struct journal_list {
  68         struct closure          cl;
  69         u64                     last_seq;
  70         struct mutex            lock;
  71         int                     ret;
  72 };
  73
  74 #define JOURNAL_ENTRY_ADD_OK            0
  75 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  76
  77 /*
  78  * Given a journal entry we just read, add it to the list of journal entries to
  79  * be replayed:
  80  */
  81 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  82                              struct journal_ptr entry_ptr,
  83                              struct journal_list *jlist, struct jset *j)
  84 {
  85         struct genradix_iter iter;
  86         struct journal_replay **_i, *i, *dup;
  87         struct journal_ptr *ptr;
  88         size_t bytes = vstruct_bytes(j);
  89         u64 last_seq = !JSET_NO_FLUSH(j) ? le64_to_cpu(j->last_seq) : 0;
  90         int ret = JOURNAL_ENTRY_ADD_OK;
  91
  92         /* Is this entry older than the range we need? */
  93         if (!c->opts.read_entire_journal &&
  94             le64_to_cpu(j->seq) < jlist->last_seq)
  95                 return JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  96
  97         /*
  98          * genradixes are indexed by a ulong, not a u64, so we can't index them
  99          * by sequence number directly: Assume instead that they will all fall
 100          * within the range of +-2billion of the filrst one we find.
 101          */
 102         if (!c->journal_entries_base_seq)
 103                 c->journal_entries_base_seq = max_t(s64, 1, le64_to_cpu(j->seq) - S32_MAX);
 104
 105         /* Drop entries we don't need anymore */
 106         if (last_seq > jlist->last_seq && !c->opts.read_entire_journal) {
 107                 genradix_for_each_from(&c->journal_entries, iter, _i,
 108                                        journal_entry_radix_idx(c, jlist->last_seq)) {
 109                         i = *_i;
 110
 111                         if (!i || i->ignore)
 112                                 continue;
 113
 114                         if (le64_to_cpu(i->j.seq) >= last_seq)
 115                                 break;
 116                         journal_replay_free(c, i);
 117                 }
 118         }
 119
 120         jlist->last_seq = max(jlist->last_seq, last_seq);
 121
 122         _i = genradix_ptr_alloc(&c->journal_entries,
 123                                 journal_entry_radix_idx(c, le64_to_cpu(j->seq)),
 124                                 GFP_KERNEL);
 125         if (!_i)
 126                 return -BCH_ERR_ENOMEM_journal_entry_add;
 127
 128         /*
 129          * Duplicate journal entries? If so we want the one that didn't have a
 130          * checksum error:
 131          */
 132         dup = *_i;
 133         if (dup) {
 134                 if (bytes == vstruct_bytes(&dup->j) &&
 135                     !memcmp(j, &dup->j, bytes)) {
 136                         i = dup;
 137                         goto found;
 138                 }
 139
 140                 if (!entry_ptr.csum_good) {
 141                         i = dup;
 142                         goto found;
 143                 }
 144
 145                 if (!dup->csum_good)
 146                         goto replace;
 147
 148                 fsck_err(c, journal_entry_replicas_data_mismatch,
 149                          "found duplicate but non identical journal entries (seq %llu)",
 150                          le64_to_cpu(j->seq));
 151                 i = dup;
 152                 goto found;
 153         }
 154 replace:
 155         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 156         if (!i)
 157                 return -BCH_ERR_ENOMEM_journal_entry_add;
 158
 159         i->nr_ptrs      = 0;
 160         i->csum_good    = entry_ptr.csum_good;
 161         i->ignore       = false;
 162         unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
 163         i->ptrs[i->nr_ptrs++] = entry_ptr;
 164
 165         if (dup) {
 166                 if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) {
 167                         bch_err(c, "found too many copies of journal entry %llu",
 168                                 le64_to_cpu(i->j.seq));
 169                         dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1;
 170                 }
 171
 172                 /* The first ptr should represent the jset we kept: */
 173                 memcpy(i->ptrs + i->nr_ptrs,
 174                        dup->ptrs,
 175                        sizeof(dup->ptrs[0]) * dup->nr_ptrs);
 176                 i->nr_ptrs += dup->nr_ptrs;
 177                 __journal_replay_free(c, dup);
 178         }
 179
 180         *_i = i;
 181         return 0;
 182 found:
 183         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 184                 if (ptr->dev == ca->dev_idx) {
 185                         bch_err(c, "duplicate journal entry %llu on same device",
 186                                 le64_to_cpu(i->j.seq));
 187                         goto out;
 188                 }
 189         }
 190
 191         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 192                 bch_err(c, "found too many copies of journal entry %llu",
 193                         le64_to_cpu(i->j.seq));
 194                 goto out;
 195         }
 196
 197         i->ptrs[i->nr_ptrs++] = entry_ptr;
 198 out:
 199 fsck_err:
 200         return ret;
 201 }
 202
 203 /* this fills in a range with empty jset_entries: */
 204 static void journal_entry_null_range(void *start, void *end)
 205 {
 206         struct jset_entry *entry;
 207
 208         for (entry = start; entry != end; entry = vstruct_next(entry))
 209                 memset(entry, 0, sizeof(*entry));
 210 }
 211
 212 #define JOURNAL_ENTRY_REREAD    5
 213 #define JOURNAL_ENTRY_NONE      6
 214 #define JOURNAL_ENTRY_BAD       7
 215
 216 static void journal_entry_err_msg(struct printbuf *out,
 217                                   u32 version,
 218                                   struct jset *jset,
 219                                   struct jset_entry *entry)
 220 {
 221         prt_str(out, "invalid journal entry, version=");
 222         bch2_version_to_text(out, version);
 223
 224         if (entry) {
 225                 prt_str(out, " type=");
 226                 prt_str(out, bch2_jset_entry_types[entry->type]);
 227         }
 228
 229         if (!jset) {
 230                 prt_printf(out, " in superblock");
 231         } else {
 232
 233                 prt_printf(out, " seq=%llu", le64_to_cpu(jset->seq));
 234
 235                 if (entry)
 236                         prt_printf(out, " offset=%zi/%u",
 237                                    (u64 *) entry - jset->_data,
 238                                    le32_to_cpu(jset->u64s));
 239         }
 240
 241         prt_str(out, ": ");
 242 }
 243
 244 #define journal_entry_err(c, version, jset, entry, _err, msg, ...)      \
 245 ({                                                                      \
 246         struct printbuf _buf = PRINTBUF;                                \
 247                                                                         \
 248         journal_entry_err_msg(&_buf, version, jset, entry);             \
 249         prt_printf(&_buf, msg, ##__VA_ARGS__);                          \
 250                                                                         \
 251         switch (flags & BKEY_INVALID_WRITE) {                           \
 252         case READ:                                                      \
 253                 mustfix_fsck_err(c, _err, "%s", _buf.buf);              \
 254                 break;                                                  \
 255         case WRITE:                                                     \
 256                 bch2_sb_error_count(c, BCH_FSCK_ERR_##_err);            \
 257                 bch_err(c, "corrupt metadata before write: %s\n", _buf.buf);\
 258                 if (bch2_fs_inconsistent(c)) {                          \
 259                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 260                         goto fsck_err;                                  \
 261                 }                                                       \
 262                 break;                                                  \
 263         }                                                               \
 264                                                                         \
 265         printbuf_exit(&_buf);                                           \
 266         true;                                                           \
 267 })
 268
 269 #define journal_entry_err_on(cond, ...)                                 \
 270         ((cond) ? journal_entry_err(__VA_ARGS__) : false)
 271
 272 #define FSCK_DELETED_KEY        5
 273
 274 static int journal_validate_key(struct bch_fs *c,
 275                                 struct jset *jset,
 276                                 struct jset_entry *entry,
 277                                 unsigned level, enum btree_id btree_id,
 278                                 struct bkey_i *k,
 279                                 unsigned version, int big_endian,
 280                                 enum bkey_invalid_flags flags)
 281 {
 282         int write = flags & BKEY_INVALID_WRITE;
 283         void *next = vstruct_next(entry);
 284         struct printbuf buf = PRINTBUF;
 285         int ret = 0;
 286
 287         if (journal_entry_err_on(!k->k.u64s,
 288                                  c, version, jset, entry,
 289                                  journal_entry_bkey_u64s_0,
 290                                  "k->u64s 0")) {
 291                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 292                 journal_entry_null_range(vstruct_next(entry), next);
 293                 return FSCK_DELETED_KEY;
 294         }
 295
 296         if (journal_entry_err_on((void *) bkey_next(k) >
 297                                  (void *) vstruct_next(entry),
 298                                  c, version, jset, entry,
 299                                  journal_entry_bkey_past_end,
 300                                  "extends past end of journal entry")) {
 301                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 302                 journal_entry_null_range(vstruct_next(entry), next);
 303                 return FSCK_DELETED_KEY;
 304         }
 305
 306         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
 307                                  c, version, jset, entry,
 308                                  journal_entry_bkey_bad_format,
 309                                  "bad format %u", k->k.format)) {
 310                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 311                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 312                 journal_entry_null_range(vstruct_next(entry), next);
 313                 return FSCK_DELETED_KEY;
 314         }
 315
 316         if (!write)
 317                 bch2_bkey_compat(level, btree_id, version, big_endian,
 318                                  write, NULL, bkey_to_packed(k));
 319
 320         if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 321                               __btree_node_type(level, btree_id), write, &buf)) {
 322                 printbuf_reset(&buf);
 323                 journal_entry_err_msg(&buf, version, jset, entry);
 324                 prt_newline(&buf);
 325                 printbuf_indent_add(&buf, 2);
 326
 327                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
 328                 prt_newline(&buf);
 329                 bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 330                                   __btree_node_type(level, btree_id), write, &buf);
 331
 332                 mustfix_fsck_err(c, journal_entry_bkey_invalid,
 333                                  "%s", buf.buf);
 334
 335                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 336                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 337                 journal_entry_null_range(vstruct_next(entry), next);
 338
 339                 printbuf_exit(&buf);
 340                 return FSCK_DELETED_KEY;
 341         }
 342
 343         if (write)
 344                 bch2_bkey_compat(level, btree_id, version, big_endian,
 345                                  write, NULL, bkey_to_packed(k));
 346 fsck_err:
 347         printbuf_exit(&buf);
 348         return ret;
 349 }
 350
 351 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 352                                 struct jset *jset,
 353                                 struct jset_entry *entry,
 354                                 unsigned version, int big_endian,
 355                                 enum bkey_invalid_flags flags)
 356 {
 357         struct bkey_i *k = entry->start;
 358
 359         while (k != vstruct_last(entry)) {
 360                 int ret = journal_validate_key(c, jset, entry,
 361                                                entry->level,
 362                                                entry->btree_id,
 363                                                k, version, big_endian,
 364                                                flags|BKEY_INVALID_JOURNAL);
 365                 if (ret == FSCK_DELETED_KEY)
 366                         continue;
 367
 368                 k = bkey_next(k);
 369         }
 370
 371         return 0;
 372 }
 373
 374 static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
 375                                              struct jset_entry *entry)
 376 {
 377         struct bkey_i *k;
 378         bool first = true;
 379
 380         jset_entry_for_each_key(entry, k) {
 381                 if (!first) {
 382                         prt_newline(out);
 383                         prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 384                 }
 385                 prt_printf(out, "btree=%s l=%u ", bch2_btree_id_str(entry->btree_id), entry->level);
 386                 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 387                 first = false;
 388         }
 389 }
 390
 391 static int journal_entry_btree_root_validate(struct bch_fs *c,
 392                                 struct jset *jset,
 393                                 struct jset_entry *entry,
 394                                 unsigned version, int big_endian,
 395                                 enum bkey_invalid_flags flags)
 396 {
 397         struct bkey_i *k = entry->start;
 398         int ret = 0;
 399
 400         if (journal_entry_err_on(!entry->u64s ||
 401                                  le16_to_cpu(entry->u64s) != k->k.u64s,
 402                                  c, version, jset, entry,
 403                                  journal_entry_btree_root_bad_size,
 404                                  "invalid btree root journal entry: wrong number of keys")) {
 405                 void *next = vstruct_next(entry);
 406                 /*
 407                  * we don't want to null out this jset_entry,
 408                  * just the contents, so that later we can tell
 409                  * we were _supposed_ to have a btree root
 410                  */
 411                 entry->u64s = 0;
 412                 journal_entry_null_range(vstruct_next(entry), next);
 413                 return 0;
 414         }
 415
 416         ret = journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
 417                                    version, big_endian, flags);
 418         if (ret == FSCK_DELETED_KEY)
 419                 ret = 0;
 420 fsck_err:
 421         return ret;
 422 }
 423
 424 static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
 425                                              struct jset_entry *entry)
 426 {
 427         journal_entry_btree_keys_to_text(out, c, entry);
 428 }
 429
 430 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 431                                 struct jset *jset,
 432                                 struct jset_entry *entry,
 433                                 unsigned version, int big_endian,
 434                                 enum bkey_invalid_flags flags)
 435 {
 436         /* obsolete, don't care: */
 437         return 0;
 438 }
 439
 440 static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 441                                             struct jset_entry *entry)
 442 {
 443 }
 444
 445 static int journal_entry_blacklist_validate(struct bch_fs *c,
 446                                 struct jset *jset,
 447                                 struct jset_entry *entry,
 448                                 unsigned version, int big_endian,
 449                                 enum bkey_invalid_flags flags)
 450 {
 451         int ret = 0;
 452
 453         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1,
 454                                  c, version, jset, entry,
 455                                  journal_entry_blacklist_bad_size,
 456                 "invalid journal seq blacklist entry: bad size")) {
 457                 journal_entry_null_range(entry, vstruct_next(entry));
 458         }
 459 fsck_err:
 460         return ret;
 461 }
 462
 463 static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
 464                                             struct jset_entry *entry)
 465 {
 466         struct jset_entry_blacklist *bl =
 467                 container_of(entry, struct jset_entry_blacklist, entry);
 468
 469         prt_printf(out, "seq=%llu", le64_to_cpu(bl->seq));
 470 }
 471
 472 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 473                                 struct jset *jset,
 474                                 struct jset_entry *entry,
 475                                 unsigned version, int big_endian,
 476                                 enum bkey_invalid_flags flags)
 477 {
 478         struct jset_entry_blacklist_v2 *bl_entry;
 479         int ret = 0;
 480
 481         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2,
 482                                  c, version, jset, entry,
 483                                  journal_entry_blacklist_v2_bad_size,
 484                 "invalid journal seq blacklist entry: bad size")) {
 485                 journal_entry_null_range(entry, vstruct_next(entry));
 486                 goto out;
 487         }
 488
 489         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 490
 491         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 492                                  le64_to_cpu(bl_entry->end),
 493                                  c, version, jset, entry,
 494                                  journal_entry_blacklist_v2_start_past_end,
 495                 "invalid journal seq blacklist entry: start > end")) {
 496                 journal_entry_null_range(entry, vstruct_next(entry));
 497         }
 498 out:
 499 fsck_err:
 500         return ret;
 501 }
 502
 503 static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
 504                                                struct jset_entry *entry)
 505 {
 506         struct jset_entry_blacklist_v2 *bl =
 507                 container_of(entry, struct jset_entry_blacklist_v2, entry);
 508
 509         prt_printf(out, "start=%llu end=%llu",
 510                le64_to_cpu(bl->start),
 511                le64_to_cpu(bl->end));
 512 }
 513
 514 static int journal_entry_usage_validate(struct bch_fs *c,
 515                                 struct jset *jset,
 516                                 struct jset_entry *entry,
 517                                 unsigned version, int big_endian,
 518                                 enum bkey_invalid_flags flags)
 519 {
 520         struct jset_entry_usage *u =
 521                 container_of(entry, struct jset_entry_usage, entry);
 522         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 523         int ret = 0;
 524
 525         if (journal_entry_err_on(bytes < sizeof(*u),
 526                                  c, version, jset, entry,
 527                                  journal_entry_usage_bad_size,
 528                                  "invalid journal entry usage: bad size")) {
 529                 journal_entry_null_range(entry, vstruct_next(entry));
 530                 return ret;
 531         }
 532
 533 fsck_err:
 534         return ret;
 535 }
 536
 537 static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 538                                         struct jset_entry *entry)
 539 {
 540         struct jset_entry_usage *u =
 541                 container_of(entry, struct jset_entry_usage, entry);
 542
 543         prt_printf(out, "type=%s v=%llu",
 544                bch2_fs_usage_types[u->entry.btree_id],
 545                le64_to_cpu(u->v));
 546 }
 547
 548 static int journal_entry_data_usage_validate(struct bch_fs *c,
 549                                 struct jset *jset,
 550                                 struct jset_entry *entry,
 551                                 unsigned version, int big_endian,
 552                                 enum bkey_invalid_flags flags)
 553 {
 554         struct jset_entry_data_usage *u =
 555                 container_of(entry, struct jset_entry_data_usage, entry);
 556         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 557         struct printbuf err = PRINTBUF;
 558         int ret = 0;
 559
 560         if (journal_entry_err_on(bytes < sizeof(*u) ||
 561                                  bytes < sizeof(*u) + u->r.nr_devs,
 562                                  c, version, jset, entry,
 563                                  journal_entry_data_usage_bad_size,
 564                                  "invalid journal entry usage: bad size")) {
 565                 journal_entry_null_range(entry, vstruct_next(entry));
 566                 goto out;
 567         }
 568
 569         if (journal_entry_err_on(bch2_replicas_entry_validate(&u->r, c->disk_sb.sb, &err),
 570                                  c, version, jset, entry,
 571                                  journal_entry_data_usage_bad_size,
 572                                  "invalid journal entry usage: %s", err.buf)) {
 573                 journal_entry_null_range(entry, vstruct_next(entry));
 574                 goto out;
 575         }
 576 out:
 577 fsck_err:
 578         printbuf_exit(&err);
 579         return ret;
 580 }
 581
 582 static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
 583                                              struct jset_entry *entry)
 584 {
 585         struct jset_entry_data_usage *u =
 586                 container_of(entry, struct jset_entry_data_usage, entry);
 587
 588         bch2_replicas_entry_to_text(out, &u->r);
 589         prt_printf(out, "=%llu", le64_to_cpu(u->v));
 590 }
 591
 592 static int journal_entry_clock_validate(struct bch_fs *c,
 593                                 struct jset *jset,
 594                                 struct jset_entry *entry,
 595                                 unsigned version, int big_endian,
 596                                 enum bkey_invalid_flags flags)
 597 {
 598         struct jset_entry_clock *clock =
 599                 container_of(entry, struct jset_entry_clock, entry);
 600         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 601         int ret = 0;
 602
 603         if (journal_entry_err_on(bytes != sizeof(*clock),
 604                                  c, version, jset, entry,
 605                                  journal_entry_clock_bad_size,
 606                                  "bad size")) {
 607                 journal_entry_null_range(entry, vstruct_next(entry));
 608                 return ret;
 609         }
 610
 611         if (journal_entry_err_on(clock->rw > 1,
 612                                  c, version, jset, entry,
 613                                  journal_entry_clock_bad_rw,
 614                                  "bad rw")) {
 615                 journal_entry_null_range(entry, vstruct_next(entry));
 616                 return ret;
 617         }
 618
 619 fsck_err:
 620         return ret;
 621 }
 622
 623 static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 624                                         struct jset_entry *entry)
 625 {
 626         struct jset_entry_clock *clock =
 627                 container_of(entry, struct jset_entry_clock, entry);
 628
 629         prt_printf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
 630 }
 631
 632 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 633                                 struct jset *jset,
 634                                 struct jset_entry *entry,
 635                                 unsigned version, int big_endian,
 636                                 enum bkey_invalid_flags flags)
 637 {
 638         struct jset_entry_dev_usage *u =
 639                 container_of(entry, struct jset_entry_dev_usage, entry);
 640         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 641         unsigned expected = sizeof(*u);
 642         unsigned dev;
 643         int ret = 0;
 644
 645         if (journal_entry_err_on(bytes < expected,
 646                                  c, version, jset, entry,
 647                                  journal_entry_dev_usage_bad_size,
 648                                  "bad size (%u < %u)",
 649                                  bytes, expected)) {
 650                 journal_entry_null_range(entry, vstruct_next(entry));
 651                 return ret;
 652         }
 653
 654         dev = le32_to_cpu(u->dev);
 655
 656         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 657                                  c, version, jset, entry,
 658                                  journal_entry_dev_usage_bad_dev,
 659                                  "bad dev")) {
 660                 journal_entry_null_range(entry, vstruct_next(entry));
 661                 return ret;
 662         }
 663
 664         if (journal_entry_err_on(u->pad,
 665                                  c, version, jset, entry,
 666                                  journal_entry_dev_usage_bad_pad,
 667                                  "bad pad")) {
 668                 journal_entry_null_range(entry, vstruct_next(entry));
 669                 return ret;
 670         }
 671
 672 fsck_err:
 673         return ret;
 674 }
 675
 676 static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
 677                                             struct jset_entry *entry)
 678 {
 679         struct jset_entry_dev_usage *u =
 680                 container_of(entry, struct jset_entry_dev_usage, entry);
 681         unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 682
 683         prt_printf(out, "dev=%u", le32_to_cpu(u->dev));
 684
 685         for (i = 0; i < nr_types; i++) {
 686                 if (i < BCH_DATA_NR)
 687                         prt_printf(out, " %s", bch2_data_types[i]);
 688                 else
 689                         prt_printf(out, " (unknown data type %u)", i);
 690                 prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 691                        le64_to_cpu(u->d[i].buckets),
 692                        le64_to_cpu(u->d[i].sectors),
 693                        le64_to_cpu(u->d[i].fragmented));
 694         }
 695 }
 696
 697 static int journal_entry_log_validate(struct bch_fs *c,
 698                                 struct jset *jset,
 699                                 struct jset_entry *entry,
 700                                 unsigned version, int big_endian,
 701                                 enum bkey_invalid_flags flags)
 702 {
 703         return 0;
 704 }
 705
 706 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 707                                       struct jset_entry *entry)
 708 {
 709         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 710         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 711
 712         prt_printf(out, "%.*s", bytes, l->d);
 713 }
 714
 715 static int journal_entry_overwrite_validate(struct bch_fs *c,
 716                                 struct jset *jset,
 717                                 struct jset_entry *entry,
 718                                 unsigned version, int big_endian,
 719                                 enum bkey_invalid_flags flags)
 720 {
 721         return journal_entry_btree_keys_validate(c, jset, entry,
 722                                 version, big_endian, READ);
 723 }
 724
 725 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
 726                                             struct jset_entry *entry)
 727 {
 728         journal_entry_btree_keys_to_text(out, c, entry);
 729 }
 730
 731 static int journal_entry_write_buffer_keys_validate(struct bch_fs *c,
 732                                 struct jset *jset,
 733                                 struct jset_entry *entry,
 734                                 unsigned version, int big_endian,
 735                                 enum bkey_invalid_flags flags)
 736 {
 737         return journal_entry_btree_keys_validate(c, jset, entry,
 738                                 version, big_endian, READ);
 739 }
 740
 741 static void journal_entry_write_buffer_keys_to_text(struct printbuf *out, struct bch_fs *c,
 742                                             struct jset_entry *entry)
 743 {
 744         journal_entry_btree_keys_to_text(out, c, entry);
 745 }
 746
 747 struct jset_entry_ops {
 748         int (*validate)(struct bch_fs *, struct jset *,
 749                         struct jset_entry *, unsigned, int,
 750                         enum bkey_invalid_flags);
 751         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 752 };
 753
 754 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 755 #define x(f, nr)                                                \
 756         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 757                 .validate       = journal_entry_##f##_validate, \
 758                 .to_text        = journal_entry_##f##_to_text,  \
 759         },
 760         BCH_JSET_ENTRY_TYPES()
 761 #undef x
 762 };
 763
 764 int bch2_journal_entry_validate(struct bch_fs *c,
 765                                 struct jset *jset,
 766                                 struct jset_entry *entry,
 767                                 unsigned version, int big_endian,
 768                                 enum bkey_invalid_flags flags)
 769 {
 770         return entry->type < BCH_JSET_ENTRY_NR
 771                 ? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
 772                                 version, big_endian, flags)
 773                 : 0;
 774 }
 775
 776 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 777                                 struct jset_entry *entry)
 778 {
 779         if (entry->type < BCH_JSET_ENTRY_NR) {
 780                 prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 781                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 782         } else {
 783                 prt_printf(out, "(unknown type %u)", entry->type);
 784         }
 785 }
 786
 787 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 788                                  enum bkey_invalid_flags flags)
 789 {
 790         unsigned version = le32_to_cpu(jset->version);
 791         int ret = 0;
 792
 793         vstruct_for_each(jset, entry) {
 794                 if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
 795                                 c, version, jset, entry,
 796                                 journal_entry_past_jset_end,
 797                                 "journal entry extends past end of jset")) {
 798                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 799                         break;
 800                 }
 801
 802                 ret = bch2_journal_entry_validate(c, jset, entry,
 803                                         version, JSET_BIG_ENDIAN(jset), flags);
 804                 if (ret)
 805                         break;
 806         }
 807 fsck_err:
 808         return ret;
 809 }
 810
 811 static int jset_validate(struct bch_fs *c,
 812                          struct bch_dev *ca,
 813                          struct jset *jset, u64 sector,
 814                          enum bkey_invalid_flags flags)
 815 {
 816         unsigned version;
 817         int ret = 0;
 818
 819         if (le64_to_cpu(jset->magic) != jset_magic(c))
 820                 return JOURNAL_ENTRY_NONE;
 821
 822         version = le32_to_cpu(jset->version);
 823         if (journal_entry_err_on(!bch2_version_compatible(version),
 824                         c, version, jset, NULL,
 825                         jset_unsupported_version,
 826                         "%s sector %llu seq %llu: incompatible journal entry version %u.%u",
 827                         ca ? ca->name : c->name,
 828                         sector, le64_to_cpu(jset->seq),
 829                         BCH_VERSION_MAJOR(version),
 830                         BCH_VERSION_MINOR(version))) {
 831                 /* don't try to continue: */
 832                 return -EINVAL;
 833         }
 834
 835         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
 836                         c, version, jset, NULL,
 837                         jset_unknown_csum,
 838                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 839                         ca ? ca->name : c->name,
 840                         sector, le64_to_cpu(jset->seq),
 841                         JSET_CSUM_TYPE(jset)))
 842                 ret = JOURNAL_ENTRY_BAD;
 843
 844         /* last_seq is ignored when JSET_NO_FLUSH is true */
 845         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 846                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
 847                                  c, version, jset, NULL,
 848                                  jset_last_seq_newer_than_seq,
 849                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 850                                  le64_to_cpu(jset->last_seq),
 851                                  le64_to_cpu(jset->seq))) {
 852                 jset->last_seq = jset->seq;
 853                 return JOURNAL_ENTRY_BAD;
 854         }
 855
 856         ret = jset_validate_entries(c, jset, flags);
 857 fsck_err:
 858         return ret;
 859 }
 860
 861 static int jset_validate_early(struct bch_fs *c,
 862                          struct bch_dev *ca,
 863                          struct jset *jset, u64 sector,
 864                          unsigned bucket_sectors_left,
 865                          unsigned sectors_read)
 866 {
 867         size_t bytes = vstruct_bytes(jset);
 868         unsigned version;
 869         enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
 870         int ret = 0;
 871
 872         if (le64_to_cpu(jset->magic) != jset_magic(c))
 873                 return JOURNAL_ENTRY_NONE;
 874
 875         version = le32_to_cpu(jset->version);
 876         if (journal_entry_err_on(!bch2_version_compatible(version),
 877                         c, version, jset, NULL,
 878                         jset_unsupported_version,
 879                         "%s sector %llu seq %llu: unknown journal entry version %u.%u",
 880                         ca ? ca->name : c->name,
 881                         sector, le64_to_cpu(jset->seq),
 882                         BCH_VERSION_MAJOR(version),
 883                         BCH_VERSION_MINOR(version))) {
 884                 /* don't try to continue: */
 885                 return -EINVAL;
 886         }
 887
 888         if (bytes > (sectors_read << 9) &&
 889             sectors_read < bucket_sectors_left)
 890                 return JOURNAL_ENTRY_REREAD;
 891
 892         if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
 893                         c, version, jset, NULL,
 894                         jset_past_bucket_end,
 895                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 896                         ca ? ca->name : c->name,
 897                         sector, le64_to_cpu(jset->seq), bytes))
 898                 le32_add_cpu(&jset->u64s,
 899                              -((bytes - (bucket_sectors_left << 9)) / 8));
 900 fsck_err:
 901         return ret;
 902 }
 903
 904 struct journal_read_buf {
 905         void            *data;
 906         size_t          size;
 907 };
 908
 909 static int journal_read_buf_realloc(struct journal_read_buf *b,
 910                                     size_t new_size)
 911 {
 912         void *n;
 913
 914         /* the bios are sized for this many pages, max: */
 915         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 916                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 917
 918         new_size = roundup_pow_of_two(new_size);
 919         n = kvpmalloc(new_size, GFP_KERNEL);
 920         if (!n)
 921                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 922
 923         kvpfree(b->data, b->size);
 924         b->data = n;
 925         b->size = new_size;
 926         return 0;
 927 }
 928
 929 static int journal_read_bucket(struct bch_dev *ca,
 930                                struct journal_read_buf *buf,
 931                                struct journal_list *jlist,
 932                                unsigned bucket)
 933 {
 934         struct bch_fs *c = ca->fs;
 935         struct journal_device *ja = &ca->journal;
 936         struct jset *j = NULL;
 937         unsigned sectors, sectors_read = 0;
 938         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 939             end = offset + ca->mi.bucket_size;
 940         bool saw_bad = false, csum_good;
 941         struct printbuf err = PRINTBUF;
 942         int ret = 0;
 943
 944         pr_debug("reading %u", bucket);
 945
 946         while (offset < end) {
 947                 if (!sectors_read) {
 948                         struct bio *bio;
 949                         unsigned nr_bvecs;
 950 reread:
 951                         sectors_read = min_t(unsigned,
 952                                 end - offset, buf->size >> 9);
 953                         nr_bvecs = buf_pages(buf->data, sectors_read << 9);
 954
 955                         bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
 956                         bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
 957
 958                         bio->bi_iter.bi_sector = offset;
 959                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 960
 961                         ret = submit_bio_wait(bio);
 962                         kfree(bio);
 963
 964                         if (bch2_dev_io_err_on(ret, ca, BCH_MEMBER_ERROR_read,
 965                                                "journal read error: sector %llu",
 966                                                offset) ||
 967                             bch2_meta_read_fault("journal")) {
 968                                 /*
 969                                  * We don't error out of the recovery process
 970                                  * here, since the relevant journal entry may be
 971                                  * found on a different device, and missing or
 972                                  * no journal entries will be handled later
 973                                  */
 974                                 goto out;
 975                         }
 976
 977                         j = buf->data;
 978                 }
 979
 980                 ret = jset_validate_early(c, ca, j, offset,
 981                                     end - offset, sectors_read);
 982                 switch (ret) {
 983                 case 0:
 984                         sectors = vstruct_sectors(j, c->block_bits);
 985                         break;
 986                 case JOURNAL_ENTRY_REREAD:
 987                         if (vstruct_bytes(j) > buf->size) {
 988                                 ret = journal_read_buf_realloc(buf,
 989                                                         vstruct_bytes(j));
 990                                 if (ret)
 991                                         goto err;
 992                         }
 993                         goto reread;
 994                 case JOURNAL_ENTRY_NONE:
 995                         if (!saw_bad)
 996                                 goto out;
 997                         /*
 998                          * On checksum error we don't really trust the size
 999                          * field of the journal entry we read, so try reading
1000                          * again at next block boundary:
1001                          */
1002                         sectors = block_sectors(c);
1003                         goto next_block;
1004                 default:
1005                         goto err;
1006                 }
1007
1008                 /*
1009                  * This happens sometimes if we don't have discards on -
1010                  * when we've partially overwritten a bucket with new
1011                  * journal entries. We don't need the rest of the
1012                  * bucket:
1013                  */
1014                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
1015                         goto out;
1016
1017                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
1018
1019                 enum bch_csum_type csum_type = JSET_CSUM_TYPE(j);
1020                 struct bch_csum csum;
1021                 csum_good = jset_csum_good(c, j, &csum);
1022
1023                 if (bch2_dev_io_err_on(!csum_good, ca, BCH_MEMBER_ERROR_checksum,
1024                                        "%s",
1025                                        (printbuf_reset(&err),
1026                                         prt_str(&err, "journal "),
1027                                         bch2_csum_err_msg(&err, csum_type, j->csum, csum),
1028                                         err.buf)))
1029                         saw_bad = true;
1030
1031                 ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
1032                              j->encrypted_start,
1033                              vstruct_end(j) - (void *) j->encrypted_start);
1034                 bch2_fs_fatal_err_on(ret, c,
1035                                 "error decrypting journal entry: %s",
1036                                 bch2_err_str(ret));
1037
1038                 mutex_lock(&jlist->lock);
1039                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
1040                                         .csum_good      = csum_good,
1041                                         .dev            = ca->dev_idx,
1042                                         .bucket         = bucket,
1043                                         .bucket_offset  = offset -
1044                                                 bucket_to_sector(ca, ja->buckets[bucket]),
1045                                         .sector         = offset,
1046                                         }, jlist, j);
1047                 mutex_unlock(&jlist->lock);
1048
1049                 switch (ret) {
1050                 case JOURNAL_ENTRY_ADD_OK:
1051                         break;
1052                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
1053                         break;
1054                 default:
1055                         goto err;
1056                 }
1057 next_block:
1058                 pr_debug("next");
1059                 offset          += sectors;
1060                 sectors_read    -= sectors;
1061                 j = ((void *) j) + (sectors << 9);
1062         }
1063
1064 out:
1065         ret = 0;
1066 err:
1067         printbuf_exit(&err);
1068         return ret;
1069 }
1070
1071 static CLOSURE_CALLBACK(bch2_journal_read_device)
1072 {
1073         closure_type(ja, struct journal_device, read);
1074         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
1075         struct bch_fs *c = ca->fs;
1076         struct journal_list *jlist =
1077                 container_of(cl->parent, struct journal_list, cl);
1078         struct journal_replay *r, **_r;
1079         struct genradix_iter iter;
1080         struct journal_read_buf buf = { NULL, 0 };
1081         unsigned i;
1082         int ret = 0;
1083
1084         if (!ja->nr)
1085                 goto out;
1086
1087         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
1088         if (ret)
1089                 goto err;
1090
1091         pr_debug("%u journal buckets", ja->nr);
1092
1093         for (i = 0; i < ja->nr; i++) {
1094                 ret = journal_read_bucket(ca, &buf, jlist, i);
1095                 if (ret)
1096                         goto err;
1097         }
1098
1099         ja->sectors_free = ca->mi.bucket_size;
1100
1101         mutex_lock(&jlist->lock);
1102         genradix_for_each_reverse(&c->journal_entries, iter, _r) {
1103                 r = *_r;
1104
1105                 if (!r)
1106                         continue;
1107
1108                 for (i = 0; i < r->nr_ptrs; i++) {
1109                         if (r->ptrs[i].dev == ca->dev_idx) {
1110                                 unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
1111                                         vstruct_sectors(&r->j, c->block_bits);
1112
1113                                 ja->cur_idx = r->ptrs[i].bucket;
1114                                 ja->sectors_free = ca->mi.bucket_size - wrote;
1115                                 goto found;
1116                         }
1117                 }
1118         }
1119 found:
1120         mutex_unlock(&jlist->lock);
1121
1122         if (ja->bucket_seq[ja->cur_idx] &&
1123             ja->sectors_free == ca->mi.bucket_size) {
1124 #if 0
1125                 /*
1126                  * Debug code for ZNS support, where we (probably) want to be
1127                  * correlated where we stopped in the journal to the zone write
1128                  * points:
1129                  */
1130                 bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
1131                 bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
1132                 for (i = 0; i < 3; i++) {
1133                         unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
1134
1135                         bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
1136                 }
1137 #endif
1138                 ja->sectors_free = 0;
1139         }
1140
1141         /*
1142          * Set dirty_idx to indicate the entire journal is full and needs to be
1143          * reclaimed - journal reclaim will immediately reclaim whatever isn't
1144          * pinned when it first runs:
1145          */
1146         ja->discard_idx = ja->dirty_idx_ondisk =
1147                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
1148 out:
1149         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
1150         kvpfree(buf.data, buf.size);
1151         percpu_ref_put(&ca->io_ref);
1152         closure_return(cl);
1153         return;
1154 err:
1155         mutex_lock(&jlist->lock);
1156         jlist->ret = ret;
1157         mutex_unlock(&jlist->lock);
1158         goto out;
1159 }
1160
1161 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1162                                struct journal_replay *j)
1163 {
1164         unsigned i;
1165
1166         for (i = 0; i < j->nr_ptrs; i++) {
1167                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
1168                 u64 offset;
1169
1170                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
1171
1172                 if (i)
1173                         prt_printf(out, " ");
1174                 prt_printf(out, "%u:%u:%u (sector %llu)",
1175                        j->ptrs[i].dev,
1176                        j->ptrs[i].bucket,
1177                        j->ptrs[i].bucket_offset,
1178                        j->ptrs[i].sector);
1179         }
1180 }
1181
1182 int bch2_journal_read(struct bch_fs *c,
1183                       u64 *last_seq,
1184                       u64 *blacklist_seq,
1185                       u64 *start_seq)
1186 {
1187         struct journal_list jlist;
1188         struct journal_replay *i, **_i, *prev = NULL;
1189         struct genradix_iter radix_iter;
1190         struct printbuf buf = PRINTBUF;
1191         bool degraded = false, last_write_torn = false;
1192         u64 seq;
1193         int ret = 0;
1194
1195         closure_init_stack(&jlist.cl);
1196         mutex_init(&jlist.lock);
1197         jlist.last_seq = 0;
1198         jlist.ret = 0;
1199
1200         for_each_member_device(c, ca) {
1201                 if (!c->opts.fsck &&
1202                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1203                         continue;
1204
1205                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1206                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1207                     percpu_ref_tryget(&ca->io_ref))
1208                         closure_call(&ca->journal.read,
1209                                      bch2_journal_read_device,
1210                                      system_unbound_wq,
1211                                      &jlist.cl);
1212                 else
1213                         degraded = true;
1214         }
1215
1216         closure_sync(&jlist.cl);
1217
1218         if (jlist.ret)
1219                 return jlist.ret;
1220
1221         *last_seq       = 0;
1222         *start_seq      = 0;
1223         *blacklist_seq  = 0;
1224
1225         /*
1226          * Find most recent flush entry, and ignore newer non flush entries -
1227          * those entries will be blacklisted:
1228          */
1229         genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
1230                 enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
1231
1232                 i = *_i;
1233
1234                 if (!i || i->ignore)
1235                         continue;
1236
1237                 if (!*start_seq)
1238                         *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
1239
1240                 if (JSET_NO_FLUSH(&i->j)) {
1241                         i->ignore = true;
1242                         continue;
1243                 }
1244
1245                 if (!last_write_torn && !i->csum_good) {
1246                         last_write_torn = true;
1247                         i->ignore = true;
1248                         continue;
1249                 }
1250
1251                 if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
1252                                          c, le32_to_cpu(i->j.version), &i->j, NULL,
1253                                          jset_last_seq_newer_than_seq,
1254                                          "invalid journal entry: last_seq > seq (%llu > %llu)",
1255                                          le64_to_cpu(i->j.last_seq),
1256                                          le64_to_cpu(i->j.seq)))
1257                         i->j.last_seq = i->j.seq;
1258
1259                 *last_seq       = le64_to_cpu(i->j.last_seq);
1260                 *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1261                 break;
1262         }
1263
1264         if (!*start_seq) {
1265                 bch_info(c, "journal read done, but no entries found");
1266                 return 0;
1267         }
1268
1269         if (!*last_seq) {
1270                 fsck_err(c, dirty_but_no_journal_entries_post_drop_nonflushes,
1271                          "journal read done, but no entries found after dropping non-flushes");
1272                 return 0;
1273         }
1274
1275         bch_info(c, "journal read done, replaying entries %llu-%llu",
1276                  *last_seq, *blacklist_seq - 1);
1277
1278         if (*start_seq != *blacklist_seq)
1279                 bch_info(c, "dropped unflushed entries %llu-%llu",
1280                          *blacklist_seq, *start_seq - 1);
1281
1282         /* Drop blacklisted entries and entries older than last_seq: */
1283         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1284                 i = *_i;
1285
1286                 if (!i || i->ignore)
1287                         continue;
1288
1289                 seq = le64_to_cpu(i->j.seq);
1290                 if (seq < *last_seq) {
1291                         journal_replay_free(c, i);
1292                         continue;
1293                 }
1294
1295                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1296                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1297                                     jset_seq_blacklisted,
1298                                     "found blacklisted journal entry %llu", seq);
1299                         i->ignore = true;
1300                 }
1301         }
1302
1303         /* Check for missing entries: */
1304         seq = *last_seq;
1305         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1306                 i = *_i;
1307
1308                 if (!i || i->ignore)
1309                         continue;
1310
1311                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1312
1313                 while (seq < le64_to_cpu(i->j.seq)) {
1314                         u64 missing_start, missing_end;
1315                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1316
1317                         while (seq < le64_to_cpu(i->j.seq) &&
1318                                bch2_journal_seq_is_blacklisted(c, seq, false))
1319                                 seq++;
1320
1321                         if (seq == le64_to_cpu(i->j.seq))
1322                                 break;
1323
1324                         missing_start = seq;
1325
1326                         while (seq < le64_to_cpu(i->j.seq) &&
1327                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1328                                 seq++;
1329
1330                         if (prev) {
1331                                 bch2_journal_ptrs_to_text(&buf1, c, prev);
1332                                 prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits));
1333                         } else
1334                                 prt_printf(&buf1, "(none)");
1335                         bch2_journal_ptrs_to_text(&buf2, c, i);
1336
1337                         missing_end = seq - 1;
1338                         fsck_err(c, journal_entries_missing,
1339                                  "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1340                                  "  prev at %s\n"
1341                                  "  next at %s",
1342                                  missing_start, missing_end,
1343                                  *last_seq, *blacklist_seq - 1,
1344                                  buf1.buf, buf2.buf);
1345
1346                         printbuf_exit(&buf1);
1347                         printbuf_exit(&buf2);
1348                 }
1349
1350                 prev = i;
1351                 seq++;
1352         }
1353
1354         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1355                 struct bch_replicas_padded replicas = {
1356                         .e.data_type = BCH_DATA_journal,
1357                         .e.nr_required = 1,
1358                 };
1359                 unsigned ptr;
1360
1361                 i = *_i;
1362                 if (!i || i->ignore)
1363                         continue;
1364
1365                 for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
1366                         struct bch_dev *ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
1367
1368                         if (!i->ptrs[ptr].csum_good)
1369                                 bch_err_dev_offset(ca, i->ptrs[ptr].sector,
1370                                                    "invalid journal checksum, seq %llu%s",
1371                                                    le64_to_cpu(i->j.seq),
1372                                                    i->csum_good ? " (had good copy on another device)" : "");
1373                 }
1374
1375                 ret = jset_validate(c,
1376                                     bch_dev_bkey_exists(c, i->ptrs[0].dev),
1377                                     &i->j,
1378                                     i->ptrs[0].sector,
1379                                     READ);
1380                 if (ret)
1381                         goto err;
1382
1383                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1384                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1385
1386                 bch2_replicas_entry_sort(&replicas.e);
1387
1388                 printbuf_reset(&buf);
1389                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1390
1391                 if (!degraded &&
1392                     !bch2_replicas_marked(c, &replicas.e) &&
1393                     (le64_to_cpu(i->j.seq) == *last_seq ||
1394                      fsck_err(c, journal_entry_replicas_not_marked,
1395                               "superblock not marked as containing replicas for journal entry %llu\n  %s",
1396                               le64_to_cpu(i->j.seq), buf.buf))) {
1397                         ret = bch2_mark_replicas(c, &replicas.e);
1398                         if (ret)
1399                                 goto err;
1400                 }
1401         }
1402 err:
1403 fsck_err:
1404         printbuf_exit(&buf);
1405         return ret;
1406 }
1407
1408 /* journal write: */
1409
1410 static void __journal_write_alloc(struct journal *j,
1411                                   struct journal_buf *w,
1412                                   struct dev_alloc_list *devs_sorted,
1413                                   unsigned sectors,
1414                                   unsigned *replicas,
1415                                   unsigned replicas_want)
1416 {
1417         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1418         struct journal_device *ja;
1419         struct bch_dev *ca;
1420         unsigned i;
1421
1422         if (*replicas >= replicas_want)
1423                 return;
1424
1425         for (i = 0; i < devs_sorted->nr; i++) {
1426                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1427                 if (!ca)
1428                         continue;
1429
1430                 ja = &ca->journal;
1431
1432                 /*
1433                  * Check that we can use this device, and aren't already using
1434                  * it:
1435                  */
1436                 if (!ca->mi.durability ||
1437                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1438                     !ja->nr ||
1439                     bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) ||
1440                     sectors > ja->sectors_free)
1441                         continue;
1442
1443                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1444
1445                 bch2_bkey_append_ptr(&w->key,
1446                         (struct bch_extent_ptr) {
1447                                   .offset = bucket_to_sector(ca,
1448                                         ja->buckets[ja->cur_idx]) +
1449                                         ca->mi.bucket_size -
1450                                         ja->sectors_free,
1451                                   .dev = ca->dev_idx,
1452                 });
1453
1454                 ja->sectors_free -= sectors;
1455                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1456
1457                 *replicas += ca->mi.durability;
1458
1459                 if (*replicas >= replicas_want)
1460                         break;
1461         }
1462 }
1463
1464 /**
1465  * journal_write_alloc - decide where to write next journal entry
1466  *
1467  * @j:          journal object
1468  * @w:          journal buf (entry to be written)
1469  *
1470  * Returns: 0 on success, or -EROFS on failure
1471  */
1472 static int journal_write_alloc(struct journal *j, struct journal_buf *w)
1473 {
1474         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1475         struct bch_devs_mask devs;
1476         struct journal_device *ja;
1477         struct bch_dev *ca;
1478         struct dev_alloc_list devs_sorted;
1479         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1480         unsigned target = c->opts.metadata_target ?:
1481                 c->opts.foreground_target;
1482         unsigned i, replicas = 0, replicas_want =
1483                 READ_ONCE(c->opts.metadata_replicas);
1484
1485         rcu_read_lock();
1486 retry:
1487         devs = target_rw_devs(c, BCH_DATA_journal, target);
1488
1489         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1490
1491         __journal_write_alloc(j, w, &devs_sorted,
1492                               sectors, &replicas, replicas_want);
1493
1494         if (replicas >= replicas_want)
1495                 goto done;
1496
1497         for (i = 0; i < devs_sorted.nr; i++) {
1498                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1499                 if (!ca)
1500                         continue;
1501
1502                 ja = &ca->journal;
1503
1504                 if (sectors > ja->sectors_free &&
1505                     sectors <= ca->mi.bucket_size &&
1506                     bch2_journal_dev_buckets_available(j, ja,
1507                                         journal_space_discarded)) {
1508                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1509                         ja->sectors_free = ca->mi.bucket_size;
1510
1511                         /*
1512                          * ja->bucket_seq[ja->cur_idx] must always have
1513                          * something sensible:
1514                          */
1515                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1516                 }
1517         }
1518
1519         __journal_write_alloc(j, w, &devs_sorted,
1520                               sectors, &replicas, replicas_want);
1521
1522         if (replicas < replicas_want && target) {
1523                 /* Retry from all devices: */
1524                 target = 0;
1525                 goto retry;
1526         }
1527 done:
1528         rcu_read_unlock();
1529
1530         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1531
1532         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1533 }
1534
1535 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1536 {
1537         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1538
1539         /* we aren't holding j->lock: */
1540         unsigned new_size = READ_ONCE(j->buf_size_want);
1541         void *new_buf;
1542
1543         if (buf->buf_size >= new_size)
1544                 return;
1545
1546         size_t btree_write_buffer_size = new_size / 64;
1547
1548         if (bch2_btree_write_buffer_resize(c, btree_write_buffer_size))
1549                 return;
1550
1551         new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
1552         if (!new_buf)
1553                 return;
1554
1555         memcpy(new_buf, buf->data, buf->buf_size);
1556
1557         spin_lock(&j->lock);
1558         swap(buf->data,         new_buf);
1559         swap(buf->buf_size,     new_size);
1560         spin_unlock(&j->lock);
1561
1562         kvpfree(new_buf, new_size);
1563 }
1564
1565 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1566 {
1567         return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
1568 }
1569
1570 static CLOSURE_CALLBACK(journal_write_done)
1571 {
1572         closure_type(j, struct journal, io);
1573         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1574         struct journal_buf *w = journal_last_unwritten_buf(j);
1575         struct bch_replicas_padded replicas;
1576         union journal_res_state old, new;
1577         u64 v, seq;
1578         int err = 0;
1579
1580         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1581                                ? j->flush_write_time
1582                                : j->noflush_write_time, j->write_start_time);
1583
1584         if (!w->devs_written.nr) {
1585                 bch_err(c, "unable to write journal to sufficient devices");
1586                 err = -EIO;
1587         } else {
1588                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1589                                          w->devs_written);
1590                 if (bch2_mark_replicas(c, &replicas.e))
1591                         err = -EIO;
1592         }
1593
1594         if (err)
1595                 bch2_fatal_error(c);
1596
1597         spin_lock(&j->lock);
1598         seq = le64_to_cpu(w->data->seq);
1599
1600         if (seq >= j->pin.front)
1601                 journal_seq_pin(j, seq)->devs = w->devs_written;
1602
1603         if (!err) {
1604                 if (!JSET_NO_FLUSH(w->data)) {
1605                         j->flushed_seq_ondisk = seq;
1606                         j->last_seq_ondisk = w->last_seq;
1607
1608                         bch2_do_discards(c);
1609                         closure_wake_up(&c->freelist_wait);
1610
1611                         bch2_reset_alloc_cursors(c);
1612                 }
1613         } else if (!j->err_seq || seq < j->err_seq)
1614                 j->err_seq      = seq;
1615
1616         j->seq_ondisk           = seq;
1617
1618         /*
1619          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1620          * more buckets:
1621          *
1622          * Must come before signaling write completion, for
1623          * bch2_fs_journal_stop():
1624          */
1625         if (j->watermark != BCH_WATERMARK_stripe)
1626                 journal_reclaim_kick(&c->journal);
1627
1628         /* also must come before signalling write completion: */
1629         closure_debug_destroy(cl);
1630
1631         v = atomic64_read(&j->reservations.counter);
1632         do {
1633                 old.v = new.v = v;
1634                 BUG_ON(journal_state_count(new, new.unwritten_idx));
1635
1636                 new.unwritten_idx++;
1637         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1638                                        old.v, new.v)) != old.v);
1639
1640         bch2_journal_reclaim_fast(j);
1641         bch2_journal_space_available(j);
1642
1643         track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
1644                            &j->max_in_flight_start, false);
1645
1646         closure_wake_up(&w->wait);
1647         journal_wake(j);
1648
1649         if (!journal_state_count(new, new.unwritten_idx) &&
1650             journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
1651                 spin_unlock(&j->lock);
1652                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1653         } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
1654                    new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
1655                 struct journal_buf *buf = journal_cur_buf(j);
1656                 long delta = buf->expires - jiffies;
1657
1658                 /*
1659                  * We don't close a journal entry to write it while there's
1660                  * previous entries still in flight - the current journal entry
1661                  * might want to be written now:
1662                  */
1663
1664                 spin_unlock(&j->lock);
1665                 mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
1666         } else {
1667                 spin_unlock(&j->lock);
1668         }
1669 }
1670
1671 static void journal_write_endio(struct bio *bio)
1672 {
1673         struct bch_dev *ca = bio->bi_private;
1674         struct journal *j = &ca->fs->journal;
1675         struct journal_buf *w = journal_last_unwritten_buf(j);
1676         unsigned long flags;
1677
1678         if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1679                                "error writing journal entry %llu: %s",
1680                                le64_to_cpu(w->data->seq),
1681                                bch2_blk_status_to_str(bio->bi_status)) ||
1682             bch2_meta_write_fault("journal")) {
1683                 spin_lock_irqsave(&j->err_lock, flags);
1684                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1685                 spin_unlock_irqrestore(&j->err_lock, flags);
1686         }
1687
1688         closure_put(&j->io);
1689         percpu_ref_put(&ca->io_ref);
1690 }
1691
1692 static CLOSURE_CALLBACK(do_journal_write)
1693 {
1694         closure_type(j, struct journal, io);
1695         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1696         struct bch_dev *ca;
1697         struct journal_buf *w = journal_last_unwritten_buf(j);
1698         struct bio *bio;
1699         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1700
1701         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1702                 ca = bch_dev_bkey_exists(c, ptr->dev);
1703                 if (!percpu_ref_tryget(&ca->io_ref)) {
1704                         /* XXX: fix this */
1705                         bch_err(c, "missing device for journal write\n");
1706                         continue;
1707                 }
1708
1709                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1710                              sectors);
1711
1712                 bio = ca->journal.bio;
1713                 bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
1714                 bio->bi_iter.bi_sector  = ptr->offset;
1715                 bio->bi_end_io          = journal_write_endio;
1716                 bio->bi_private         = ca;
1717
1718                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1719                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1720
1721                 if (!JSET_NO_FLUSH(w->data))
1722                         bio->bi_opf    |= REQ_FUA;
1723                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1724                         bio->bi_opf    |= REQ_PREFLUSH;
1725
1726                 bch2_bio_map(bio, w->data, sectors << 9);
1727
1728                 trace_and_count(c, journal_write, bio);
1729                 closure_bio_submit(bio, cl);
1730
1731                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1732                         le64_to_cpu(w->data->seq);
1733         }
1734
1735         continue_at(cl, journal_write_done, c->io_complete_wq);
1736 }
1737
1738 static int bch2_journal_write_prep(struct journal *j, struct journal_buf *w)
1739 {
1740         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1741         struct jset_entry *start, *end;
1742         struct jset *jset = w->data;
1743         struct journal_keys_to_wb wb = { NULL };
1744         unsigned sectors, bytes, u64s;
1745         unsigned long btree_roots_have = 0;
1746         bool validate_before_checksum = false;
1747         u64 seq = le64_to_cpu(jset->seq);
1748         int ret;
1749
1750         /*
1751          * Simple compaction, dropping empty jset_entries (from journal
1752          * reservations that weren't fully used) and merging jset_entries that
1753          * can be.
1754          *
1755          * If we wanted to be really fancy here, we could sort all the keys in
1756          * the jset and drop keys that were overwritten - probably not worth it:
1757          */
1758         vstruct_for_each(jset, i) {
1759                 unsigned u64s = le16_to_cpu(i->u64s);
1760
1761                 /* Empty entry: */
1762                 if (!u64s)
1763                         continue;
1764
1765                 /*
1766                  * New btree roots are set by journalling them; when the journal
1767                  * entry gets written we have to propagate them to
1768                  * c->btree_roots
1769                  *
1770                  * But, every journal entry we write has to contain all the
1771                  * btree roots (at least for now); so after we copy btree roots
1772                  * to c->btree_roots we have to get any missing btree roots and
1773                  * add them to this journal entry:
1774                  */
1775                 switch (i->type) {
1776                 case BCH_JSET_ENTRY_btree_root:
1777                         bch2_journal_entry_to_btree_root(c, i);
1778                         __set_bit(i->btree_id, &btree_roots_have);
1779                         break;
1780                 case BCH_JSET_ENTRY_write_buffer_keys:
1781                         EBUG_ON(!w->need_flush_to_write_buffer);
1782
1783                         if (!wb.wb)
1784                                 bch2_journal_keys_to_write_buffer_start(c, &wb, seq);
1785
1786                         struct bkey_i *k;
1787                         jset_entry_for_each_key(i, k) {
1788                                 ret = bch2_journal_key_to_wb(c, &wb, i->btree_id, k);
1789                                 if (ret) {
1790                                         bch2_fs_fatal_error(c, "-ENOMEM flushing journal keys to btree write buffer");
1791                                         bch2_journal_keys_to_write_buffer_end(c, &wb);
1792                                         return ret;
1793                                 }
1794                         }
1795                         i->type = BCH_JSET_ENTRY_btree_keys;
1796                         break;
1797                 }
1798         }
1799
1800         if (wb.wb)
1801                 bch2_journal_keys_to_write_buffer_end(c, &wb);
1802         w->need_flush_to_write_buffer = false;
1803
1804         start = end = vstruct_last(jset);
1805
1806         end     = bch2_btree_roots_to_journal_entries(c, end, btree_roots_have);
1807
1808         bch2_journal_super_entries_add_common(c, &end, seq);
1809         u64s    = (u64 *) end - (u64 *) start;
1810         BUG_ON(u64s > j->entry_u64s_reserved);
1811
1812         le32_add_cpu(&jset->u64s, u64s);
1813
1814         sectors = vstruct_sectors(jset, c->block_bits);
1815         bytes   = vstruct_bytes(jset);
1816
1817         if (sectors > w->sectors) {
1818                 bch2_fs_fatal_error(c, "aieeee! journal write overran available space, %zu > %u (extra %u reserved %u/%u)",
1819                                     vstruct_bytes(jset), w->sectors << 9,
1820                                     u64s, w->u64s_reserved, j->entry_u64s_reserved);
1821                 return -EINVAL;
1822         }
1823
1824         jset->magic             = cpu_to_le64(jset_magic(c));
1825         jset->version           = cpu_to_le32(c->sb.version);
1826
1827         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1828         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1829
1830         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1831                 j->last_empty_seq = seq;
1832
1833         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1834                 validate_before_checksum = true;
1835
1836         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1837                 validate_before_checksum = true;
1838
1839         if (validate_before_checksum &&
1840             (ret = jset_validate(c, NULL, jset, 0, WRITE)))
1841                 return ret;
1842
1843         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1844                     jset->encrypted_start,
1845                     vstruct_end(jset) - (void *) jset->encrypted_start);
1846         if (bch2_fs_fatal_err_on(ret, c,
1847                         "error decrypting journal entry: %i", ret))
1848                 return ret;
1849
1850         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1851                                   journal_nonce(jset), jset);
1852
1853         if (!validate_before_checksum &&
1854             (ret = jset_validate(c, NULL, jset, 0, WRITE)))
1855                 return ret;
1856
1857         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1858         return 0;
1859 }
1860
1861 static int bch2_journal_write_pick_flush(struct journal *j, struct journal_buf *w)
1862 {
1863         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1864         int error = bch2_journal_error(j);
1865
1866         /*
1867          * If the journal is in an error state - we did an emergency shutdown -
1868          * we prefer to continue doing journal writes. We just mark them as
1869          * noflush so they'll never be used, but they'll still be visible by the
1870          * list_journal tool - this helps in debugging.
1871          *
1872          * There's a caveat: the first journal write after marking the
1873          * superblock dirty must always be a flush write, because on startup
1874          * from a clean shutdown we didn't necessarily read the journal and the
1875          * new journal write might overwrite whatever was in the journal
1876          * previously - we can't leave the journal without any flush writes in
1877          * it.
1878          *
1879          * So if we're in an error state, and we're still starting up, we don't
1880          * write anything at all.
1881          */
1882         if (error && test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags))
1883                 return -EIO;
1884
1885         if (error ||
1886             w->noflush ||
1887             (!w->must_flush &&
1888              (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1889              test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags))) {
1890                 w->noflush = true;
1891                 SET_JSET_NO_FLUSH(w->data, true);
1892                 w->data->last_seq       = 0;
1893                 w->last_seq             = 0;
1894
1895                 j->nr_noflush_writes++;
1896         } else {
1897                 j->last_flush_write = jiffies;
1898                 j->nr_flush_writes++;
1899                 clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
1900         }
1901
1902         return 0;
1903 }
1904
1905 CLOSURE_CALLBACK(bch2_journal_write)
1906 {
1907         closure_type(j, struct journal, io);
1908         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1909         struct journal_buf *w = journal_last_unwritten_buf(j);
1910         struct bch_replicas_padded replicas;
1911         struct bio *bio;
1912         struct printbuf journal_debug_buf = PRINTBUF;
1913         unsigned nr_rw_members = 0;
1914         int ret;
1915
1916         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1917
1918         j->write_start_time = local_clock();
1919
1920         spin_lock(&j->lock);
1921         ret = bch2_journal_write_pick_flush(j, w);
1922         spin_unlock(&j->lock);
1923         if (ret)
1924                 goto err;
1925
1926         mutex_lock(&j->buf_lock);
1927         journal_buf_realloc(j, w);
1928
1929         ret = bch2_journal_write_prep(j, w);
1930         mutex_unlock(&j->buf_lock);
1931         if (ret)
1932                 goto err;
1933
1934         j->entry_bytes_written += vstruct_bytes(w->data);
1935
1936         while (1) {
1937                 spin_lock(&j->lock);
1938                 ret = journal_write_alloc(j, w);
1939                 if (!ret || !j->can_discard)
1940                         break;
1941
1942                 spin_unlock(&j->lock);
1943                 bch2_journal_do_discards(j);
1944         }
1945
1946         if (ret) {
1947                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1948                 spin_unlock(&j->lock);
1949                 bch_err(c, "Unable to allocate journal write:\n%s",
1950                         journal_debug_buf.buf);
1951                 printbuf_exit(&journal_debug_buf);
1952                 goto err;
1953         }
1954
1955         /*
1956          * write is allocated, no longer need to account for it in
1957          * bch2_journal_space_available():
1958          */
1959         w->sectors = 0;
1960
1961         /*
1962          * journal entry has been compacted and allocated, recalculate space
1963          * available:
1964          */
1965         bch2_journal_space_available(j);
1966         spin_unlock(&j->lock);
1967
1968         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1969
1970         if (c->opts.nochanges)
1971                 goto no_io;
1972
1973         for_each_rw_member(c, ca)
1974                 nr_rw_members++;
1975
1976         if (nr_rw_members > 1)
1977                 w->separate_flush = true;
1978
1979         /*
1980          * Mark journal replicas before we submit the write to guarantee
1981          * recovery will find the journal entries after a crash.
1982          */
1983         bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1984                                  w->devs_written);
1985         ret = bch2_mark_replicas(c, &replicas.e);
1986         if (ret)
1987                 goto err;
1988
1989         if (!JSET_NO_FLUSH(w->data) && w->separate_flush) {
1990                 for_each_rw_member(c, ca) {
1991                         percpu_ref_get(&ca->io_ref);
1992
1993                         bio = ca->journal.bio;
1994                         bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
1995                         bio->bi_end_io          = journal_write_endio;
1996                         bio->bi_private         = ca;
1997                         closure_bio_submit(bio, cl);
1998                 }
1999         }
2000
2001         continue_at(cl, do_journal_write, c->io_complete_wq);
2002         return;
2003 no_io:
2004         continue_at(cl, journal_write_done, c->io_complete_wq);
2005         return;
2006 err:
2007         bch2_fatal_error(c);
2008         continue_at(cl, journal_write_done, c->io_complete_wq);
2009 }