fs/bcachefs/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                 bch2_prt_data_type(out, i);
 687                 prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 688                        le64_to_cpu(u->d[i].buckets),
 689                        le64_to_cpu(u->d[i].sectors),
 690                        le64_to_cpu(u->d[i].fragmented));
 691         }
 692 }
 693
 694 static int journal_entry_log_validate(struct bch_fs *c,
 695                                 struct jset *jset,
 696                                 struct jset_entry *entry,
 697                                 unsigned version, int big_endian,
 698                                 enum bkey_invalid_flags flags)
 699 {
 700         return 0;
 701 }
 702
 703 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 704                                       struct jset_entry *entry)
 705 {
 706         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 707         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 708
 709         prt_printf(out, "%.*s", bytes, l->d);
 710 }
 711
 712 static int journal_entry_overwrite_validate(struct bch_fs *c,
 713                                 struct jset *jset,
 714                                 struct jset_entry *entry,
 715                                 unsigned version, int big_endian,
 716                                 enum bkey_invalid_flags flags)
 717 {
 718         return journal_entry_btree_keys_validate(c, jset, entry,
 719                                 version, big_endian, READ);
 720 }
 721
 722 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
 723                                             struct jset_entry *entry)
 724 {
 725         journal_entry_btree_keys_to_text(out, c, entry);
 726 }
 727
 728 static int journal_entry_write_buffer_keys_validate(struct bch_fs *c,
 729                                 struct jset *jset,
 730                                 struct jset_entry *entry,
 731                                 unsigned version, int big_endian,
 732                                 enum bkey_invalid_flags flags)
 733 {
 734         return journal_entry_btree_keys_validate(c, jset, entry,
 735                                 version, big_endian, READ);
 736 }
 737
 738 static void journal_entry_write_buffer_keys_to_text(struct printbuf *out, struct bch_fs *c,
 739                                             struct jset_entry *entry)
 740 {
 741         journal_entry_btree_keys_to_text(out, c, entry);
 742 }
 743
 744 struct jset_entry_ops {
 745         int (*validate)(struct bch_fs *, struct jset *,
 746                         struct jset_entry *, unsigned, int,
 747                         enum bkey_invalid_flags);
 748         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 749 };
 750
 751 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 752 #define x(f, nr)                                                \
 753         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 754                 .validate       = journal_entry_##f##_validate, \
 755                 .to_text        = journal_entry_##f##_to_text,  \
 756         },
 757         BCH_JSET_ENTRY_TYPES()
 758 #undef x
 759 };
 760
 761 int bch2_journal_entry_validate(struct bch_fs *c,
 762                                 struct jset *jset,
 763                                 struct jset_entry *entry,
 764                                 unsigned version, int big_endian,
 765                                 enum bkey_invalid_flags flags)
 766 {
 767         return entry->type < BCH_JSET_ENTRY_NR
 768                 ? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
 769                                 version, big_endian, flags)
 770                 : 0;
 771 }
 772
 773 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 774                                 struct jset_entry *entry)
 775 {
 776         if (entry->type < BCH_JSET_ENTRY_NR) {
 777                 prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 778                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 779         } else {
 780                 prt_printf(out, "(unknown type %u)", entry->type);
 781         }
 782 }
 783
 784 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 785                                  enum bkey_invalid_flags flags)
 786 {
 787         unsigned version = le32_to_cpu(jset->version);
 788         int ret = 0;
 789
 790         vstruct_for_each(jset, entry) {
 791                 if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
 792                                 c, version, jset, entry,
 793                                 journal_entry_past_jset_end,
 794                                 "journal entry extends past end of jset")) {
 795                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 796                         break;
 797                 }
 798
 799                 ret = bch2_journal_entry_validate(c, jset, entry,
 800                                         version, JSET_BIG_ENDIAN(jset), flags);
 801                 if (ret)
 802                         break;
 803         }
 804 fsck_err:
 805         return ret;
 806 }
 807
 808 static int jset_validate(struct bch_fs *c,
 809                          struct bch_dev *ca,
 810                          struct jset *jset, u64 sector,
 811                          enum bkey_invalid_flags flags)
 812 {
 813         unsigned version;
 814         int ret = 0;
 815
 816         if (le64_to_cpu(jset->magic) != jset_magic(c))
 817                 return JOURNAL_ENTRY_NONE;
 818
 819         version = le32_to_cpu(jset->version);
 820         if (journal_entry_err_on(!bch2_version_compatible(version),
 821                         c, version, jset, NULL,
 822                         jset_unsupported_version,
 823                         "%s sector %llu seq %llu: incompatible journal entry version %u.%u",
 824                         ca ? ca->name : c->name,
 825                         sector, le64_to_cpu(jset->seq),
 826                         BCH_VERSION_MAJOR(version),
 827                         BCH_VERSION_MINOR(version))) {
 828                 /* don't try to continue: */
 829                 return -EINVAL;
 830         }
 831
 832         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
 833                         c, version, jset, NULL,
 834                         jset_unknown_csum,
 835                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 836                         ca ? ca->name : c->name,
 837                         sector, le64_to_cpu(jset->seq),
 838                         JSET_CSUM_TYPE(jset)))
 839                 ret = JOURNAL_ENTRY_BAD;
 840
 841         /* last_seq is ignored when JSET_NO_FLUSH is true */
 842         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 843                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
 844                                  c, version, jset, NULL,
 845                                  jset_last_seq_newer_than_seq,
 846                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 847                                  le64_to_cpu(jset->last_seq),
 848                                  le64_to_cpu(jset->seq))) {
 849                 jset->last_seq = jset->seq;
 850                 return JOURNAL_ENTRY_BAD;
 851         }
 852
 853         ret = jset_validate_entries(c, jset, flags);
 854 fsck_err:
 855         return ret;
 856 }
 857
 858 static int jset_validate_early(struct bch_fs *c,
 859                          struct bch_dev *ca,
 860                          struct jset *jset, u64 sector,
 861                          unsigned bucket_sectors_left,
 862                          unsigned sectors_read)
 863 {
 864         size_t bytes = vstruct_bytes(jset);
 865         unsigned version;
 866         enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
 867         int ret = 0;
 868
 869         if (le64_to_cpu(jset->magic) != jset_magic(c))
 870                 return JOURNAL_ENTRY_NONE;
 871
 872         version = le32_to_cpu(jset->version);
 873         if (journal_entry_err_on(!bch2_version_compatible(version),
 874                         c, version, jset, NULL,
 875                         jset_unsupported_version,
 876                         "%s sector %llu seq %llu: unknown journal entry version %u.%u",
 877                         ca ? ca->name : c->name,
 878                         sector, le64_to_cpu(jset->seq),
 879                         BCH_VERSION_MAJOR(version),
 880                         BCH_VERSION_MINOR(version))) {
 881                 /* don't try to continue: */
 882                 return -EINVAL;
 883         }
 884
 885         if (bytes > (sectors_read << 9) &&
 886             sectors_read < bucket_sectors_left)
 887                 return JOURNAL_ENTRY_REREAD;
 888
 889         if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
 890                         c, version, jset, NULL,
 891                         jset_past_bucket_end,
 892                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 893                         ca ? ca->name : c->name,
 894                         sector, le64_to_cpu(jset->seq), bytes))
 895                 le32_add_cpu(&jset->u64s,
 896                              -((bytes - (bucket_sectors_left << 9)) / 8));
 897 fsck_err:
 898         return ret;
 899 }
 900
 901 struct journal_read_buf {
 902         void            *data;
 903         size_t          size;
 904 };
 905
 906 static int journal_read_buf_realloc(struct journal_read_buf *b,
 907                                     size_t new_size)
 908 {
 909         void *n;
 910
 911         /* the bios are sized for this many pages, max: */
 912         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 913                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 914
 915         new_size = roundup_pow_of_two(new_size);
 916         n = kvpmalloc(new_size, GFP_KERNEL);
 917         if (!n)
 918                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 919
 920         kvpfree(b->data, b->size);
 921         b->data = n;
 922         b->size = new_size;
 923         return 0;
 924 }
 925
 926 static int journal_read_bucket(struct bch_dev *ca,
 927                                struct journal_read_buf *buf,
 928                                struct journal_list *jlist,
 929                                unsigned bucket)
 930 {
 931         struct bch_fs *c = ca->fs;
 932         struct journal_device *ja = &ca->journal;
 933         struct jset *j = NULL;
 934         unsigned sectors, sectors_read = 0;
 935         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 936             end = offset + ca->mi.bucket_size;
 937         bool saw_bad = false, csum_good;
 938         struct printbuf err = PRINTBUF;
 939         int ret = 0;
 940
 941         pr_debug("reading %u", bucket);
 942
 943         while (offset < end) {
 944                 if (!sectors_read) {
 945                         struct bio *bio;
 946                         unsigned nr_bvecs;
 947 reread:
 948                         sectors_read = min_t(unsigned,
 949                                 end - offset, buf->size >> 9);
 950                         nr_bvecs = buf_pages(buf->data, sectors_read << 9);
 951
 952                         bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
 953                         bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
 954
 955                         bio->bi_iter.bi_sector = offset;
 956                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 957
 958                         ret = submit_bio_wait(bio);
 959                         kfree(bio);
 960
 961                         if (bch2_dev_io_err_on(ret, ca, BCH_MEMBER_ERROR_read,
 962                                                "journal read error: sector %llu",
 963                                                offset) ||
 964                             bch2_meta_read_fault("journal")) {
 965                                 /*
 966                                  * We don't error out of the recovery process
 967                                  * here, since the relevant journal entry may be
 968                                  * found on a different device, and missing or
 969                                  * no journal entries will be handled later
 970                                  */
 971                                 goto out;
 972                         }
 973
 974                         j = buf->data;
 975                 }
 976
 977                 ret = jset_validate_early(c, ca, j, offset,
 978                                     end - offset, sectors_read);
 979                 switch (ret) {
 980                 case 0:
 981                         sectors = vstruct_sectors(j, c->block_bits);
 982                         break;
 983                 case JOURNAL_ENTRY_REREAD:
 984                         if (vstruct_bytes(j) > buf->size) {
 985                                 ret = journal_read_buf_realloc(buf,
 986                                                         vstruct_bytes(j));
 987                                 if (ret)
 988                                         goto err;
 989                         }
 990                         goto reread;
 991                 case JOURNAL_ENTRY_NONE:
 992                         if (!saw_bad)
 993                                 goto out;
 994                         /*
 995                          * On checksum error we don't really trust the size
 996                          * field of the journal entry we read, so try reading
 997                          * again at next block boundary:
 998                          */
 999                         sectors = block_sectors(c);
1000                         goto next_block;
1001                 default:
1002                         goto err;
1003                 }
1004
1005                 /*
1006                  * This happens sometimes if we don't have discards on -
1007                  * when we've partially overwritten a bucket with new
1008                  * journal entries. We don't need the rest of the
1009                  * bucket:
1010                  */
1011                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
1012                         goto out;
1013
1014                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
1015
1016                 enum bch_csum_type csum_type = JSET_CSUM_TYPE(j);
1017                 struct bch_csum csum;
1018                 csum_good = jset_csum_good(c, j, &csum);
1019
1020                 if (bch2_dev_io_err_on(!csum_good, ca, BCH_MEMBER_ERROR_checksum,
1021                                        "%s",
1022                                        (printbuf_reset(&err),
1023                                         prt_str(&err, "journal "),
1024                                         bch2_csum_err_msg(&err, csum_type, j->csum, csum),
1025                                         err.buf)))
1026                         saw_bad = true;
1027
1028                 ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
1029                              j->encrypted_start,
1030                              vstruct_end(j) - (void *) j->encrypted_start);
1031                 bch2_fs_fatal_err_on(ret, c,
1032                                 "error decrypting journal entry: %s",
1033                                 bch2_err_str(ret));
1034
1035                 mutex_lock(&jlist->lock);
1036                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
1037                                         .csum_good      = csum_good,
1038                                         .dev            = ca->dev_idx,
1039                                         .bucket         = bucket,
1040                                         .bucket_offset  = offset -
1041                                                 bucket_to_sector(ca, ja->buckets[bucket]),
1042                                         .sector         = offset,
1043                                         }, jlist, j);
1044                 mutex_unlock(&jlist->lock);
1045
1046                 switch (ret) {
1047                 case JOURNAL_ENTRY_ADD_OK:
1048                         break;
1049                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
1050                         break;
1051                 default:
1052                         goto err;
1053                 }
1054 next_block:
1055                 pr_debug("next");
1056                 offset          += sectors;
1057                 sectors_read    -= sectors;
1058                 j = ((void *) j) + (sectors << 9);
1059         }
1060
1061 out:
1062         ret = 0;
1063 err:
1064         printbuf_exit(&err);
1065         return ret;
1066 }
1067
1068 static CLOSURE_CALLBACK(bch2_journal_read_device)
1069 {
1070         closure_type(ja, struct journal_device, read);
1071         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
1072         struct bch_fs *c = ca->fs;
1073         struct journal_list *jlist =
1074                 container_of(cl->parent, struct journal_list, cl);
1075         struct journal_replay *r, **_r;
1076         struct genradix_iter iter;
1077         struct journal_read_buf buf = { NULL, 0 };
1078         unsigned i;
1079         int ret = 0;
1080
1081         if (!ja->nr)
1082                 goto out;
1083
1084         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
1085         if (ret)
1086                 goto err;
1087
1088         pr_debug("%u journal buckets", ja->nr);
1089
1090         for (i = 0; i < ja->nr; i++) {
1091                 ret = journal_read_bucket(ca, &buf, jlist, i);
1092                 if (ret)
1093                         goto err;
1094         }
1095
1096         ja->sectors_free = ca->mi.bucket_size;
1097
1098         mutex_lock(&jlist->lock);
1099         genradix_for_each_reverse(&c->journal_entries, iter, _r) {
1100                 r = *_r;
1101
1102                 if (!r)
1103                         continue;
1104
1105                 for (i = 0; i < r->nr_ptrs; i++) {
1106                         if (r->ptrs[i].dev == ca->dev_idx) {
1107                                 unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
1108                                         vstruct_sectors(&r->j, c->block_bits);
1109
1110                                 ja->cur_idx = r->ptrs[i].bucket;
1111                                 ja->sectors_free = ca->mi.bucket_size - wrote;
1112                                 goto found;
1113                         }
1114                 }
1115         }
1116 found:
1117         mutex_unlock(&jlist->lock);
1118
1119         if (ja->bucket_seq[ja->cur_idx] &&
1120             ja->sectors_free == ca->mi.bucket_size) {
1121 #if 0
1122                 /*
1123                  * Debug code for ZNS support, where we (probably) want to be
1124                  * correlated where we stopped in the journal to the zone write
1125                  * points:
1126                  */
1127                 bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
1128                 bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
1129                 for (i = 0; i < 3; i++) {
1130                         unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
1131
1132                         bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
1133                 }
1134 #endif
1135                 ja->sectors_free = 0;
1136         }
1137
1138         /*
1139          * Set dirty_idx to indicate the entire journal is full and needs to be
1140          * reclaimed - journal reclaim will immediately reclaim whatever isn't
1141          * pinned when it first runs:
1142          */
1143         ja->discard_idx = ja->dirty_idx_ondisk =
1144                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
1145 out:
1146         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
1147         kvpfree(buf.data, buf.size);
1148         percpu_ref_put(&ca->io_ref);
1149         closure_return(cl);
1150         return;
1151 err:
1152         mutex_lock(&jlist->lock);
1153         jlist->ret = ret;
1154         mutex_unlock(&jlist->lock);
1155         goto out;
1156 }
1157
1158 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1159                                struct journal_replay *j)
1160 {
1161         unsigned i;
1162
1163         for (i = 0; i < j->nr_ptrs; i++) {
1164                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
1165                 u64 offset;
1166
1167                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
1168
1169                 if (i)
1170                         prt_printf(out, " ");
1171                 prt_printf(out, "%u:%u:%u (sector %llu)",
1172                        j->ptrs[i].dev,
1173                        j->ptrs[i].bucket,
1174                        j->ptrs[i].bucket_offset,
1175                        j->ptrs[i].sector);
1176         }
1177 }
1178
1179 int bch2_journal_read(struct bch_fs *c,
1180                       u64 *last_seq,
1181                       u64 *blacklist_seq,
1182                       u64 *start_seq)
1183 {
1184         struct journal_list jlist;
1185         struct journal_replay *i, **_i, *prev = NULL;
1186         struct genradix_iter radix_iter;
1187         struct printbuf buf = PRINTBUF;
1188         bool degraded = false, last_write_torn = false;
1189         u64 seq;
1190         int ret = 0;
1191
1192         closure_init_stack(&jlist.cl);
1193         mutex_init(&jlist.lock);
1194         jlist.last_seq = 0;
1195         jlist.ret = 0;
1196
1197         for_each_member_device(c, ca) {
1198                 if (!c->opts.fsck &&
1199                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1200                         continue;
1201
1202                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1203                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1204                     percpu_ref_tryget(&ca->io_ref))
1205                         closure_call(&ca->journal.read,
1206                                      bch2_journal_read_device,
1207                                      system_unbound_wq,
1208                                      &jlist.cl);
1209                 else
1210                         degraded = true;
1211         }
1212
1213         closure_sync(&jlist.cl);
1214
1215         if (jlist.ret)
1216                 return jlist.ret;
1217
1218         *last_seq       = 0;
1219         *start_seq      = 0;
1220         *blacklist_seq  = 0;
1221
1222         /*
1223          * Find most recent flush entry, and ignore newer non flush entries -
1224          * those entries will be blacklisted:
1225          */
1226         genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
1227                 enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
1228
1229                 i = *_i;
1230
1231                 if (!i || i->ignore)
1232                         continue;
1233
1234                 if (!*start_seq)
1235                         *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
1236
1237                 if (JSET_NO_FLUSH(&i->j)) {
1238                         i->ignore = true;
1239                         continue;
1240                 }
1241
1242                 if (!last_write_torn && !i->csum_good) {
1243                         last_write_torn = true;
1244                         i->ignore = true;
1245                         continue;
1246                 }
1247
1248                 if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
1249                                          c, le32_to_cpu(i->j.version), &i->j, NULL,
1250                                          jset_last_seq_newer_than_seq,
1251                                          "invalid journal entry: last_seq > seq (%llu > %llu)",
1252                                          le64_to_cpu(i->j.last_seq),
1253                                          le64_to_cpu(i->j.seq)))
1254                         i->j.last_seq = i->j.seq;
1255
1256                 *last_seq       = le64_to_cpu(i->j.last_seq);
1257                 *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1258                 break;
1259         }
1260
1261         if (!*start_seq) {
1262                 bch_info(c, "journal read done, but no entries found");
1263                 return 0;
1264         }
1265
1266         if (!*last_seq) {
1267                 fsck_err(c, dirty_but_no_journal_entries_post_drop_nonflushes,
1268                          "journal read done, but no entries found after dropping non-flushes");
1269                 return 0;
1270         }
1271
1272         bch_info(c, "journal read done, replaying entries %llu-%llu",
1273                  *last_seq, *blacklist_seq - 1);
1274
1275         if (*start_seq != *blacklist_seq)
1276                 bch_info(c, "dropped unflushed entries %llu-%llu",
1277                          *blacklist_seq, *start_seq - 1);
1278
1279         /* Drop blacklisted entries and entries older than last_seq: */
1280         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1281                 i = *_i;
1282
1283                 if (!i || i->ignore)
1284                         continue;
1285
1286                 seq = le64_to_cpu(i->j.seq);
1287                 if (seq < *last_seq) {
1288                         journal_replay_free(c, i);
1289                         continue;
1290                 }
1291
1292                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1293                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1294                                     jset_seq_blacklisted,
1295                                     "found blacklisted journal entry %llu", seq);
1296                         i->ignore = true;
1297                 }
1298         }
1299
1300         /* Check for missing entries: */
1301         seq = *last_seq;
1302         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1303                 i = *_i;
1304
1305                 if (!i || i->ignore)
1306                         continue;
1307
1308                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1309
1310                 while (seq < le64_to_cpu(i->j.seq)) {
1311                         u64 missing_start, missing_end;
1312                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1313
1314                         while (seq < le64_to_cpu(i->j.seq) &&
1315                                bch2_journal_seq_is_blacklisted(c, seq, false))
1316                                 seq++;
1317
1318                         if (seq == le64_to_cpu(i->j.seq))
1319                                 break;
1320
1321                         missing_start = seq;
1322
1323                         while (seq < le64_to_cpu(i->j.seq) &&
1324                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1325                                 seq++;
1326
1327                         if (prev) {
1328                                 bch2_journal_ptrs_to_text(&buf1, c, prev);
1329                                 prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits));
1330                         } else
1331                                 prt_printf(&buf1, "(none)");
1332                         bch2_journal_ptrs_to_text(&buf2, c, i);
1333
1334                         missing_end = seq - 1;
1335                         fsck_err(c, journal_entries_missing,
1336                                  "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1337                                  "  prev at %s\n"
1338                                  "  next at %s",
1339                                  missing_start, missing_end,
1340                                  *last_seq, *blacklist_seq - 1,
1341                                  buf1.buf, buf2.buf);
1342
1343                         printbuf_exit(&buf1);
1344                         printbuf_exit(&buf2);
1345                 }
1346
1347                 prev = i;
1348                 seq++;
1349         }
1350
1351         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1352                 struct bch_replicas_padded replicas = {
1353                         .e.data_type = BCH_DATA_journal,
1354                         .e.nr_required = 1,
1355                 };
1356                 unsigned ptr;
1357
1358                 i = *_i;
1359                 if (!i || i->ignore)
1360                         continue;
1361
1362                 for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
1363                         struct bch_dev *ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
1364
1365                         if (!i->ptrs[ptr].csum_good)
1366                                 bch_err_dev_offset(ca, i->ptrs[ptr].sector,
1367                                                    "invalid journal checksum, seq %llu%s",
1368                                                    le64_to_cpu(i->j.seq),
1369                                                    i->csum_good ? " (had good copy on another device)" : "");
1370                 }
1371
1372                 ret = jset_validate(c,
1373                                     bch_dev_bkey_exists(c, i->ptrs[0].dev),
1374                                     &i->j,
1375                                     i->ptrs[0].sector,
1376                                     READ);
1377                 if (ret)
1378                         goto err;
1379
1380                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1381                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1382
1383                 bch2_replicas_entry_sort(&replicas.e);
1384
1385                 printbuf_reset(&buf);
1386                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1387
1388                 if (!degraded &&
1389                     !bch2_replicas_marked(c, &replicas.e) &&
1390                     (le64_to_cpu(i->j.seq) == *last_seq ||
1391                      fsck_err(c, journal_entry_replicas_not_marked,
1392                               "superblock not marked as containing replicas for journal entry %llu\n  %s",
1393                               le64_to_cpu(i->j.seq), buf.buf))) {
1394                         ret = bch2_mark_replicas(c, &replicas.e);
1395                         if (ret)
1396                                 goto err;
1397                 }
1398         }
1399 err:
1400 fsck_err:
1401         printbuf_exit(&buf);
1402         return ret;
1403 }
1404
1405 /* journal write: */
1406
1407 static void __journal_write_alloc(struct journal *j,
1408                                   struct journal_buf *w,
1409                                   struct dev_alloc_list *devs_sorted,
1410                                   unsigned sectors,
1411                                   unsigned *replicas,
1412                                   unsigned replicas_want)
1413 {
1414         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1415         struct journal_device *ja;
1416         struct bch_dev *ca;
1417         unsigned i;
1418
1419         if (*replicas >= replicas_want)
1420                 return;
1421
1422         for (i = 0; i < devs_sorted->nr; i++) {
1423                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1424                 if (!ca)
1425                         continue;
1426
1427                 ja = &ca->journal;
1428
1429                 /*
1430                  * Check that we can use this device, and aren't already using
1431                  * it:
1432                  */
1433                 if (!ca->mi.durability ||
1434                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1435                     !ja->nr ||
1436                     bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) ||
1437                     sectors > ja->sectors_free)
1438                         continue;
1439
1440                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1441
1442                 bch2_bkey_append_ptr(&w->key,
1443                         (struct bch_extent_ptr) {
1444                                   .offset = bucket_to_sector(ca,
1445                                         ja->buckets[ja->cur_idx]) +
1446                                         ca->mi.bucket_size -
1447                                         ja->sectors_free,
1448                                   .dev = ca->dev_idx,
1449                 });
1450
1451                 ja->sectors_free -= sectors;
1452                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1453
1454                 *replicas += ca->mi.durability;
1455
1456                 if (*replicas >= replicas_want)
1457                         break;
1458         }
1459 }
1460
1461 /**
1462  * journal_write_alloc - decide where to write next journal entry
1463  *
1464  * @j:          journal object
1465  * @w:          journal buf (entry to be written)
1466  *
1467  * Returns: 0 on success, or -EROFS on failure
1468  */
1469 static int journal_write_alloc(struct journal *j, struct journal_buf *w)
1470 {
1471         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1472         struct bch_devs_mask devs;
1473         struct journal_device *ja;
1474         struct bch_dev *ca;
1475         struct dev_alloc_list devs_sorted;
1476         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1477         unsigned target = c->opts.metadata_target ?:
1478                 c->opts.foreground_target;
1479         unsigned i, replicas = 0, replicas_want =
1480                 READ_ONCE(c->opts.metadata_replicas);
1481         unsigned replicas_need = min_t(unsigned, replicas_want,
1482                                        READ_ONCE(c->opts.metadata_replicas_required));
1483
1484         rcu_read_lock();
1485 retry:
1486         devs = target_rw_devs(c, BCH_DATA_journal, target);
1487
1488         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1489
1490         __journal_write_alloc(j, w, &devs_sorted,
1491                               sectors, &replicas, replicas_want);
1492
1493         if (replicas >= replicas_want)
1494                 goto done;
1495
1496         for (i = 0; i < devs_sorted.nr; i++) {
1497                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1498                 if (!ca)
1499                         continue;
1500
1501                 ja = &ca->journal;
1502
1503                 if (sectors > ja->sectors_free &&
1504                     sectors <= ca->mi.bucket_size &&
1505                     bch2_journal_dev_buckets_available(j, ja,
1506                                         journal_space_discarded)) {
1507                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1508                         ja->sectors_free = ca->mi.bucket_size;
1509
1510                         /*
1511                          * ja->bucket_seq[ja->cur_idx] must always have
1512                          * something sensible:
1513                          */
1514                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1515                 }
1516         }
1517
1518         __journal_write_alloc(j, w, &devs_sorted,
1519                               sectors, &replicas, replicas_want);
1520
1521         if (replicas < replicas_want && target) {
1522                 /* Retry from all devices: */
1523                 target = 0;
1524                 goto retry;
1525         }
1526 done:
1527         rcu_read_unlock();
1528
1529         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1530
1531         return replicas >= replicas_need ? 0 : -EROFS;
1532 }
1533
1534 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1535 {
1536         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1537
1538         /* we aren't holding j->lock: */
1539         unsigned new_size = READ_ONCE(j->buf_size_want);
1540         void *new_buf;
1541
1542         if (buf->buf_size >= new_size)
1543                 return;
1544
1545         size_t btree_write_buffer_size = new_size / 64;
1546
1547         if (bch2_btree_write_buffer_resize(c, btree_write_buffer_size))
1548                 return;
1549
1550         new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
1551         if (!new_buf)
1552                 return;
1553
1554         memcpy(new_buf, buf->data, buf->buf_size);
1555
1556         spin_lock(&j->lock);
1557         swap(buf->data,         new_buf);
1558         swap(buf->buf_size,     new_size);
1559         spin_unlock(&j->lock);
1560
1561         kvpfree(new_buf, new_size);
1562 }
1563
1564 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1565 {
1566         return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
1567 }
1568
1569 static CLOSURE_CALLBACK(journal_write_done)
1570 {
1571         closure_type(j, struct journal, io);
1572         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1573         struct journal_buf *w = journal_last_unwritten_buf(j);
1574         struct bch_replicas_padded replicas;
1575         union journal_res_state old, new;
1576         u64 v, seq;
1577         int err = 0;
1578
1579         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1580                                ? j->flush_write_time
1581                                : j->noflush_write_time, j->write_start_time);
1582
1583         if (!w->devs_written.nr) {
1584                 bch_err(c, "unable to write journal to sufficient devices");
1585                 err = -EIO;
1586         } else {
1587                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1588                                          w->devs_written);
1589                 if (bch2_mark_replicas(c, &replicas.e))
1590                         err = -EIO;
1591         }
1592
1593         if (err)
1594                 bch2_fatal_error(c);
1595
1596         spin_lock(&j->lock);
1597         seq = le64_to_cpu(w->data->seq);
1598
1599         if (seq >= j->pin.front)
1600                 journal_seq_pin(j, seq)->devs = w->devs_written;
1601
1602         if (!err) {
1603                 if (!JSET_NO_FLUSH(w->data)) {
1604                         j->flushed_seq_ondisk = seq;
1605                         j->last_seq_ondisk = w->last_seq;
1606
1607                         bch2_do_discards(c);
1608                         closure_wake_up(&c->freelist_wait);
1609
1610                         bch2_reset_alloc_cursors(c);
1611                 }
1612         } else if (!j->err_seq || seq < j->err_seq)
1613                 j->err_seq      = seq;
1614
1615         j->seq_ondisk           = seq;
1616
1617         /*
1618          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1619          * more buckets:
1620          *
1621          * Must come before signaling write completion, for
1622          * bch2_fs_journal_stop():
1623          */
1624         if (j->watermark != BCH_WATERMARK_stripe)
1625                 journal_reclaim_kick(&c->journal);
1626
1627         /* also must come before signalling write completion: */
1628         closure_debug_destroy(cl);
1629
1630         v = atomic64_read(&j->reservations.counter);
1631         do {
1632                 old.v = new.v = v;
1633                 BUG_ON(journal_state_count(new, new.unwritten_idx));
1634
1635                 new.unwritten_idx++;
1636         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1637                                        old.v, new.v)) != old.v);
1638
1639         bch2_journal_reclaim_fast(j);
1640         bch2_journal_space_available(j);
1641
1642         track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
1643                            &j->max_in_flight_start, false);
1644
1645         closure_wake_up(&w->wait);
1646         journal_wake(j);
1647
1648         if (!journal_state_count(new, new.unwritten_idx) &&
1649             journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
1650                 spin_unlock(&j->lock);
1651                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1652         } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
1653                    new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
1654                 struct journal_buf *buf = journal_cur_buf(j);
1655                 long delta = buf->expires - jiffies;
1656
1657                 /*
1658                  * We don't close a journal entry to write it while there's
1659                  * previous entries still in flight - the current journal entry
1660                  * might want to be written now:
1661                  */
1662
1663                 spin_unlock(&j->lock);
1664                 mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
1665         } else {
1666                 spin_unlock(&j->lock);
1667         }
1668 }
1669
1670 static void journal_write_endio(struct bio *bio)
1671 {
1672         struct bch_dev *ca = bio->bi_private;
1673         struct journal *j = &ca->fs->journal;
1674         struct journal_buf *w = journal_last_unwritten_buf(j);
1675         unsigned long flags;
1676
1677         if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1678                                "error writing journal entry %llu: %s",
1679                                le64_to_cpu(w->data->seq),
1680                                bch2_blk_status_to_str(bio->bi_status)) ||
1681             bch2_meta_write_fault("journal")) {
1682                 spin_lock_irqsave(&j->err_lock, flags);
1683                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1684                 spin_unlock_irqrestore(&j->err_lock, flags);
1685         }
1686
1687         closure_put(&j->io);
1688         percpu_ref_put(&ca->io_ref);
1689 }
1690
1691 static CLOSURE_CALLBACK(do_journal_write)
1692 {
1693         closure_type(j, struct journal, io);
1694         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1695         struct bch_dev *ca;
1696         struct journal_buf *w = journal_last_unwritten_buf(j);
1697         struct bio *bio;
1698         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1699
1700         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1701                 ca = bch_dev_bkey_exists(c, ptr->dev);
1702                 if (!percpu_ref_tryget(&ca->io_ref)) {
1703                         /* XXX: fix this */
1704                         bch_err(c, "missing device for journal write\n");
1705                         continue;
1706                 }
1707
1708                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1709                              sectors);
1710
1711                 bio = ca->journal.bio;
1712                 bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
1713                 bio->bi_iter.bi_sector  = ptr->offset;
1714                 bio->bi_end_io          = journal_write_endio;
1715                 bio->bi_private         = ca;
1716
1717                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1718                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1719
1720                 if (!JSET_NO_FLUSH(w->data))
1721                         bio->bi_opf    |= REQ_FUA;
1722                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1723                         bio->bi_opf    |= REQ_PREFLUSH;
1724
1725                 bch2_bio_map(bio, w->data, sectors << 9);
1726
1727                 trace_and_count(c, journal_write, bio);
1728                 closure_bio_submit(bio, cl);
1729
1730                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1731                         le64_to_cpu(w->data->seq);
1732         }
1733
1734         continue_at(cl, journal_write_done, c->io_complete_wq);
1735 }
1736
1737 static int bch2_journal_write_prep(struct journal *j, struct journal_buf *w)
1738 {
1739         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1740         struct jset_entry *start, *end;
1741         struct jset *jset = w->data;
1742         struct journal_keys_to_wb wb = { NULL };
1743         unsigned sectors, bytes, u64s;
1744         unsigned long btree_roots_have = 0;
1745         bool validate_before_checksum = false;
1746         u64 seq = le64_to_cpu(jset->seq);
1747         int ret;
1748
1749         /*
1750          * Simple compaction, dropping empty jset_entries (from journal
1751          * reservations that weren't fully used) and merging jset_entries that
1752          * can be.
1753          *
1754          * If we wanted to be really fancy here, we could sort all the keys in
1755          * the jset and drop keys that were overwritten - probably not worth it:
1756          */
1757         vstruct_for_each(jset, i) {
1758                 unsigned u64s = le16_to_cpu(i->u64s);
1759
1760                 /* Empty entry: */
1761                 if (!u64s)
1762                         continue;
1763
1764                 /*
1765                  * New btree roots are set by journalling them; when the journal
1766                  * entry gets written we have to propagate them to
1767                  * c->btree_roots
1768                  *
1769                  * But, every journal entry we write has to contain all the
1770                  * btree roots (at least for now); so after we copy btree roots
1771                  * to c->btree_roots we have to get any missing btree roots and
1772                  * add them to this journal entry:
1773                  */
1774                 switch (i->type) {
1775                 case BCH_JSET_ENTRY_btree_root:
1776                         bch2_journal_entry_to_btree_root(c, i);
1777                         __set_bit(i->btree_id, &btree_roots_have);
1778                         break;
1779                 case BCH_JSET_ENTRY_write_buffer_keys:
1780                         EBUG_ON(!w->need_flush_to_write_buffer);
1781
1782                         if (!wb.wb)
1783                                 bch2_journal_keys_to_write_buffer_start(c, &wb, seq);
1784
1785                         struct bkey_i *k;
1786                         jset_entry_for_each_key(i, k) {
1787                                 ret = bch2_journal_key_to_wb(c, &wb, i->btree_id, k);
1788                                 if (ret) {
1789                                         bch2_fs_fatal_error(c, "-ENOMEM flushing journal keys to btree write buffer");
1790                                         bch2_journal_keys_to_write_buffer_end(c, &wb);
1791                                         return ret;
1792                                 }
1793                         }
1794                         i->type = BCH_JSET_ENTRY_btree_keys;
1795                         break;
1796                 }
1797         }
1798
1799         if (wb.wb)
1800                 bch2_journal_keys_to_write_buffer_end(c, &wb);
1801         w->need_flush_to_write_buffer = false;
1802
1803         start = end = vstruct_last(jset);
1804
1805         end     = bch2_btree_roots_to_journal_entries(c, end, btree_roots_have);
1806
1807         bch2_journal_super_entries_add_common(c, &end, seq);
1808         u64s    = (u64 *) end - (u64 *) start;
1809         BUG_ON(u64s > j->entry_u64s_reserved);
1810
1811         le32_add_cpu(&jset->u64s, u64s);
1812
1813         sectors = vstruct_sectors(jset, c->block_bits);
1814         bytes   = vstruct_bytes(jset);
1815
1816         if (sectors > w->sectors) {
1817                 bch2_fs_fatal_error(c, "aieeee! journal write overran available space, %zu > %u (extra %u reserved %u/%u)",
1818                                     vstruct_bytes(jset), w->sectors << 9,
1819                                     u64s, w->u64s_reserved, j->entry_u64s_reserved);
1820                 return -EINVAL;
1821         }
1822
1823         jset->magic             = cpu_to_le64(jset_magic(c));
1824         jset->version           = cpu_to_le32(c->sb.version);
1825
1826         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1827         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1828
1829         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1830                 j->last_empty_seq = seq;
1831
1832         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1833                 validate_before_checksum = true;
1834
1835         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1836                 validate_before_checksum = true;
1837
1838         if (validate_before_checksum &&
1839             (ret = jset_validate(c, NULL, jset, 0, WRITE)))
1840                 return ret;
1841
1842         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1843                     jset->encrypted_start,
1844                     vstruct_end(jset) - (void *) jset->encrypted_start);
1845         if (bch2_fs_fatal_err_on(ret, c,
1846                         "error decrypting journal entry: %i", ret))
1847                 return ret;
1848
1849         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1850                                   journal_nonce(jset), jset);
1851
1852         if (!validate_before_checksum &&
1853             (ret = jset_validate(c, NULL, jset, 0, WRITE)))
1854                 return ret;
1855
1856         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1857         return 0;
1858 }
1859
1860 static int bch2_journal_write_pick_flush(struct journal *j, struct journal_buf *w)
1861 {
1862         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1863         int error = bch2_journal_error(j);
1864
1865         /*
1866          * If the journal is in an error state - we did an emergency shutdown -
1867          * we prefer to continue doing journal writes. We just mark them as
1868          * noflush so they'll never be used, but they'll still be visible by the
1869          * list_journal tool - this helps in debugging.
1870          *
1871          * There's a caveat: the first journal write after marking the
1872          * superblock dirty must always be a flush write, because on startup
1873          * from a clean shutdown we didn't necessarily read the journal and the
1874          * new journal write might overwrite whatever was in the journal
1875          * previously - we can't leave the journal without any flush writes in
1876          * it.
1877          *
1878          * So if we're in an error state, and we're still starting up, we don't
1879          * write anything at all.
1880          */
1881         if (error && test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags))
1882                 return -EIO;
1883
1884         if (error ||
1885             w->noflush ||
1886             (!w->must_flush &&
1887              (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1888              test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags))) {
1889                 w->noflush = true;
1890                 SET_JSET_NO_FLUSH(w->data, true);
1891                 w->data->last_seq       = 0;
1892                 w->last_seq             = 0;
1893
1894                 j->nr_noflush_writes++;
1895         } else {
1896                 j->last_flush_write = jiffies;
1897                 j->nr_flush_writes++;
1898                 clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
1899         }
1900
1901         return 0;
1902 }
1903
1904 CLOSURE_CALLBACK(bch2_journal_write)
1905 {
1906         closure_type(j, struct journal, io);
1907         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1908         struct journal_buf *w = journal_last_unwritten_buf(j);
1909         struct bch_replicas_padded replicas;
1910         struct bio *bio;
1911         struct printbuf journal_debug_buf = PRINTBUF;
1912         unsigned nr_rw_members = 0;
1913         int ret;
1914
1915         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1916
1917         j->write_start_time = local_clock();
1918
1919         spin_lock(&j->lock);
1920         ret = bch2_journal_write_pick_flush(j, w);
1921         spin_unlock(&j->lock);
1922         if (ret)
1923                 goto err;
1924
1925         mutex_lock(&j->buf_lock);
1926         journal_buf_realloc(j, w);
1927
1928         ret = bch2_journal_write_prep(j, w);
1929         mutex_unlock(&j->buf_lock);
1930         if (ret)
1931                 goto err;
1932
1933         j->entry_bytes_written += vstruct_bytes(w->data);
1934
1935         while (1) {
1936                 spin_lock(&j->lock);
1937                 ret = journal_write_alloc(j, w);
1938                 if (!ret || !j->can_discard)
1939                         break;
1940
1941                 spin_unlock(&j->lock);
1942                 bch2_journal_do_discards(j);
1943         }
1944
1945         if (ret) {
1946                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1947                 spin_unlock(&j->lock);
1948                 bch_err(c, "Unable to allocate journal write:\n%s",
1949                         journal_debug_buf.buf);
1950                 printbuf_exit(&journal_debug_buf);
1951                 goto err;
1952         }
1953
1954         /*
1955          * write is allocated, no longer need to account for it in
1956          * bch2_journal_space_available():
1957          */
1958         w->sectors = 0;
1959
1960         /*
1961          * journal entry has been compacted and allocated, recalculate space
1962          * available:
1963          */
1964         bch2_journal_space_available(j);
1965         spin_unlock(&j->lock);
1966
1967         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1968
1969         if (c->opts.nochanges)
1970                 goto no_io;
1971
1972         for_each_rw_member(c, ca)
1973                 nr_rw_members++;
1974
1975         if (nr_rw_members > 1)
1976                 w->separate_flush = true;
1977
1978         /*
1979          * Mark journal replicas before we submit the write to guarantee
1980          * recovery will find the journal entries after a crash.
1981          */
1982         bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1983                                  w->devs_written);
1984         ret = bch2_mark_replicas(c, &replicas.e);
1985         if (ret)
1986                 goto err;
1987
1988         if (!JSET_NO_FLUSH(w->data) && w->separate_flush) {
1989                 for_each_rw_member(c, ca) {
1990                         percpu_ref_get(&ca->io_ref);
1991
1992                         bio = ca->journal.bio;
1993                         bio_reset(bio, ca->disk_sb.bdev,
1994                                   REQ_OP_WRITE|REQ_PREFLUSH);
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 }