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 "buckets.h"
   8 #include "checksum.h"
   9 #include "disk_groups.h"
  10 #include "error.h"
  11 #include "journal.h"
  12 #include "journal_io.h"
  13 #include "journal_reclaim.h"
  14 #include "journal_seq_blacklist.h"
  15 #include "replicas.h"
  16 #include "sb-clean.h"
  17 #include "trace.h"
  18
  19 static struct nonce journal_nonce(const struct jset *jset)
  20 {
  21         return (struct nonce) {{
  22                 [0] = 0,
  23                 [1] = ((__le32 *) &jset->seq)[0],
  24                 [2] = ((__le32 *) &jset->seq)[1],
  25                 [3] = BCH_NONCE_JOURNAL,
  26         }};
  27 }
  28
  29 static bool jset_csum_good(struct bch_fs *c, struct jset *j)
  30 {
  31         return bch2_checksum_type_valid(c, JSET_CSUM_TYPE(j)) &&
  32                 !bch2_crc_cmp(j->csum,
  33                               csum_vstruct(c, JSET_CSUM_TYPE(j), journal_nonce(j), j));
  34 }
  35
  36 static inline u32 journal_entry_radix_idx(struct bch_fs *c, u64 seq)
  37 {
  38         return (seq - c->journal_entries_base_seq) & (~0U >> 1);
  39 }
  40
  41 static void __journal_replay_free(struct bch_fs *c,
  42                                   struct journal_replay *i)
  43 {
  44         struct journal_replay **p =
  45                 genradix_ptr(&c->journal_entries,
  46                              journal_entry_radix_idx(c, le64_to_cpu(i->j.seq)));
  47
  48         BUG_ON(*p != i);
  49         *p = NULL;
  50         kvpfree(i, offsetof(struct journal_replay, j) +
  51                 vstruct_bytes(&i->j));
  52 }
  53
  54 static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
  55 {
  56         i->ignore = true;
  57
  58         if (!c->opts.read_entire_journal)
  59                 __journal_replay_free(c, i);
  60 }
  61
  62 struct journal_list {
  63         struct closure          cl;
  64         u64                     last_seq;
  65         struct mutex            lock;
  66         int                     ret;
  67 };
  68
  69 #define JOURNAL_ENTRY_ADD_OK            0
  70 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  71
  72 /*
  73  * Given a journal entry we just read, add it to the list of journal entries to
  74  * be replayed:
  75  */
  76 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  77                              struct journal_ptr entry_ptr,
  78                              struct journal_list *jlist, struct jset *j)
  79 {
  80         struct genradix_iter iter;
  81         struct journal_replay **_i, *i, *dup;
  82         struct journal_ptr *ptr;
  83         size_t bytes = vstruct_bytes(j);
  84         u64 last_seq = !JSET_NO_FLUSH(j) ? le64_to_cpu(j->last_seq) : 0;
  85         int ret = JOURNAL_ENTRY_ADD_OK;
  86
  87         /* Is this entry older than the range we need? */
  88         if (!c->opts.read_entire_journal &&
  89             le64_to_cpu(j->seq) < jlist->last_seq)
  90                 return JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  91
  92         /*
  93          * genradixes are indexed by a ulong, not a u64, so we can't index them
  94          * by sequence number directly: Assume instead that they will all fall
  95          * within the range of +-2billion of the filrst one we find.
  96          */
  97         if (!c->journal_entries_base_seq)
  98                 c->journal_entries_base_seq = max_t(s64, 1, le64_to_cpu(j->seq) - S32_MAX);
  99
 100         /* Drop entries we don't need anymore */
 101         if (last_seq > jlist->last_seq && !c->opts.read_entire_journal) {
 102                 genradix_for_each_from(&c->journal_entries, iter, _i,
 103                                        journal_entry_radix_idx(c, jlist->last_seq)) {
 104                         i = *_i;
 105
 106                         if (!i || i->ignore)
 107                                 continue;
 108
 109                         if (le64_to_cpu(i->j.seq) >= last_seq)
 110                                 break;
 111                         journal_replay_free(c, i);
 112                 }
 113         }
 114
 115         jlist->last_seq = max(jlist->last_seq, last_seq);
 116
 117         _i = genradix_ptr_alloc(&c->journal_entries,
 118                                 journal_entry_radix_idx(c, le64_to_cpu(j->seq)),
 119                                 GFP_KERNEL);
 120         if (!_i)
 121                 return -BCH_ERR_ENOMEM_journal_entry_add;
 122
 123         /*
 124          * Duplicate journal entries? If so we want the one that didn't have a
 125          * checksum error:
 126          */
 127         dup = *_i;
 128         if (dup) {
 129                 if (bytes == vstruct_bytes(&dup->j) &&
 130                     !memcmp(j, &dup->j, bytes)) {
 131                         i = dup;
 132                         goto found;
 133                 }
 134
 135                 if (!entry_ptr.csum_good) {
 136                         i = dup;
 137                         goto found;
 138                 }
 139
 140                 if (!dup->csum_good)
 141                         goto replace;
 142
 143                 fsck_err(c, journal_entry_replicas_data_mismatch,
 144                          "found duplicate but non identical journal entries (seq %llu)",
 145                          le64_to_cpu(j->seq));
 146                 i = dup;
 147                 goto found;
 148         }
 149 replace:
 150         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 151         if (!i)
 152                 return -BCH_ERR_ENOMEM_journal_entry_add;
 153
 154         i->nr_ptrs      = 0;
 155         i->csum_good    = entry_ptr.csum_good;
 156         i->ignore       = false;
 157         unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
 158         i->ptrs[i->nr_ptrs++] = entry_ptr;
 159
 160         if (dup) {
 161                 if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) {
 162                         bch_err(c, "found too many copies of journal entry %llu",
 163                                 le64_to_cpu(i->j.seq));
 164                         dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1;
 165                 }
 166
 167                 /* The first ptr should represent the jset we kept: */
 168                 memcpy(i->ptrs + i->nr_ptrs,
 169                        dup->ptrs,
 170                        sizeof(dup->ptrs[0]) * dup->nr_ptrs);
 171                 i->nr_ptrs += dup->nr_ptrs;
 172                 __journal_replay_free(c, dup);
 173         }
 174
 175         *_i = i;
 176         return 0;
 177 found:
 178         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 179                 if (ptr->dev == ca->dev_idx) {
 180                         bch_err(c, "duplicate journal entry %llu on same device",
 181                                 le64_to_cpu(i->j.seq));
 182                         goto out;
 183                 }
 184         }
 185
 186         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 187                 bch_err(c, "found too many copies of journal entry %llu",
 188                         le64_to_cpu(i->j.seq));
 189                 goto out;
 190         }
 191
 192         i->ptrs[i->nr_ptrs++] = entry_ptr;
 193 out:
 194 fsck_err:
 195         return ret;
 196 }
 197
 198 /* this fills in a range with empty jset_entries: */
 199 static void journal_entry_null_range(void *start, void *end)
 200 {
 201         struct jset_entry *entry;
 202
 203         for (entry = start; entry != end; entry = vstruct_next(entry))
 204                 memset(entry, 0, sizeof(*entry));
 205 }
 206
 207 #define JOURNAL_ENTRY_REREAD    5
 208 #define JOURNAL_ENTRY_NONE      6
 209 #define JOURNAL_ENTRY_BAD       7
 210
 211 static void journal_entry_err_msg(struct printbuf *out,
 212                                   u32 version,
 213                                   struct jset *jset,
 214                                   struct jset_entry *entry)
 215 {
 216         prt_str(out, "invalid journal entry, version=");
 217         bch2_version_to_text(out, version);
 218
 219         if (entry) {
 220                 prt_str(out, " type=");
 221                 prt_str(out, bch2_jset_entry_types[entry->type]);
 222         }
 223
 224         if (!jset) {
 225                 prt_printf(out, " in superblock");
 226         } else {
 227
 228                 prt_printf(out, " seq=%llu", le64_to_cpu(jset->seq));
 229
 230                 if (entry)
 231                         prt_printf(out, " offset=%zi/%u",
 232                                    (u64 *) entry - jset->_data,
 233                                    le32_to_cpu(jset->u64s));
 234         }
 235
 236         prt_str(out, ": ");
 237 }
 238
 239 #define journal_entry_err(c, version, jset, entry, _err, msg, ...)      \
 240 ({                                                                      \
 241         struct printbuf _buf = PRINTBUF;                                \
 242                                                                         \
 243         journal_entry_err_msg(&_buf, version, jset, entry);             \
 244         prt_printf(&_buf, msg, ##__VA_ARGS__);                          \
 245                                                                         \
 246         switch (flags & BKEY_INVALID_WRITE) {                           \
 247         case READ:                                                      \
 248                 mustfix_fsck_err(c, _err, "%s", _buf.buf);              \
 249                 break;                                                  \
 250         case WRITE:                                                     \
 251                 bch2_sb_error_count(c, BCH_FSCK_ERR_##_err);            \
 252                 bch_err(c, "corrupt metadata before write: %s\n", _buf.buf);\
 253                 if (bch2_fs_inconsistent(c)) {                          \
 254                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 255                         goto fsck_err;                                  \
 256                 }                                                       \
 257                 break;                                                  \
 258         }                                                               \
 259                                                                         \
 260         printbuf_exit(&_buf);                                           \
 261         true;                                                           \
 262 })
 263
 264 #define journal_entry_err_on(cond, ...)                                 \
 265         ((cond) ? journal_entry_err(__VA_ARGS__) : false)
 266
 267 #define FSCK_DELETED_KEY        5
 268
 269 static int journal_validate_key(struct bch_fs *c,
 270                                 struct jset *jset,
 271                                 struct jset_entry *entry,
 272                                 unsigned level, enum btree_id btree_id,
 273                                 struct bkey_i *k,
 274                                 unsigned version, int big_endian,
 275                                 enum bkey_invalid_flags flags)
 276 {
 277         int write = flags & BKEY_INVALID_WRITE;
 278         void *next = vstruct_next(entry);
 279         struct printbuf buf = PRINTBUF;
 280         int ret = 0;
 281
 282         if (journal_entry_err_on(!k->k.u64s,
 283                                  c, version, jset, entry,
 284                                  journal_entry_bkey_u64s_0,
 285                                  "k->u64s 0")) {
 286                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 287                 journal_entry_null_range(vstruct_next(entry), next);
 288                 return FSCK_DELETED_KEY;
 289         }
 290
 291         if (journal_entry_err_on((void *) bkey_next(k) >
 292                                  (void *) vstruct_next(entry),
 293                                  c, version, jset, entry,
 294                                  journal_entry_bkey_past_end,
 295                                  "extends past end of journal entry")) {
 296                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 297                 journal_entry_null_range(vstruct_next(entry), next);
 298                 return FSCK_DELETED_KEY;
 299         }
 300
 301         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
 302                                  c, version, jset, entry,
 303                                  journal_entry_bkey_bad_format,
 304                                  "bad format %u", k->k.format)) {
 305                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 306                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 307                 journal_entry_null_range(vstruct_next(entry), next);
 308                 return FSCK_DELETED_KEY;
 309         }
 310
 311         if (!write)
 312                 bch2_bkey_compat(level, btree_id, version, big_endian,
 313                                  write, NULL, bkey_to_packed(k));
 314
 315         if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 316                               __btree_node_type(level, btree_id), write, &buf)) {
 317                 printbuf_reset(&buf);
 318                 journal_entry_err_msg(&buf, version, jset, entry);
 319                 prt_newline(&buf);
 320                 printbuf_indent_add(&buf, 2);
 321
 322                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
 323                 prt_newline(&buf);
 324                 bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 325                                   __btree_node_type(level, btree_id), write, &buf);
 326
 327                 mustfix_fsck_err(c, journal_entry_bkey_invalid,
 328                                  "%s", buf.buf);
 329
 330                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 331                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 332                 journal_entry_null_range(vstruct_next(entry), next);
 333
 334                 printbuf_exit(&buf);
 335                 return FSCK_DELETED_KEY;
 336         }
 337
 338         if (write)
 339                 bch2_bkey_compat(level, btree_id, version, big_endian,
 340                                  write, NULL, bkey_to_packed(k));
 341 fsck_err:
 342         printbuf_exit(&buf);
 343         return ret;
 344 }
 345
 346 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 347                                 struct jset *jset,
 348                                 struct jset_entry *entry,
 349                                 unsigned version, int big_endian,
 350                                 enum bkey_invalid_flags flags)
 351 {
 352         struct bkey_i *k = entry->start;
 353
 354         while (k != vstruct_last(entry)) {
 355                 int ret = journal_validate_key(c, jset, entry,
 356                                                entry->level,
 357                                                entry->btree_id,
 358                                                k, version, big_endian,
 359                                                flags|BKEY_INVALID_JOURNAL);
 360                 if (ret == FSCK_DELETED_KEY)
 361                         continue;
 362
 363                 k = bkey_next(k);
 364         }
 365
 366         return 0;
 367 }
 368
 369 static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
 370                                              struct jset_entry *entry)
 371 {
 372         struct bkey_i *k;
 373         bool first = true;
 374
 375         jset_entry_for_each_key(entry, k) {
 376                 if (!first) {
 377                         prt_newline(out);
 378                         prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 379                 }
 380                 prt_printf(out, "btree=%s l=%u ", bch2_btree_id_str(entry->btree_id), entry->level);
 381                 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 382                 first = false;
 383         }
 384 }
 385
 386 static int journal_entry_btree_root_validate(struct bch_fs *c,
 387                                 struct jset *jset,
 388                                 struct jset_entry *entry,
 389                                 unsigned version, int big_endian,
 390                                 enum bkey_invalid_flags flags)
 391 {
 392         struct bkey_i *k = entry->start;
 393         int ret = 0;
 394
 395         if (journal_entry_err_on(!entry->u64s ||
 396                                  le16_to_cpu(entry->u64s) != k->k.u64s,
 397                                  c, version, jset, entry,
 398                                  journal_entry_btree_root_bad_size,
 399                                  "invalid btree root journal entry: wrong number of keys")) {
 400                 void *next = vstruct_next(entry);
 401                 /*
 402                  * we don't want to null out this jset_entry,
 403                  * just the contents, so that later we can tell
 404                  * we were _supposed_ to have a btree root
 405                  */
 406                 entry->u64s = 0;
 407                 journal_entry_null_range(vstruct_next(entry), next);
 408                 return 0;
 409         }
 410
 411         return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
 412                                     version, big_endian, flags);
 413 fsck_err:
 414         return ret;
 415 }
 416
 417 static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
 418                                              struct jset_entry *entry)
 419 {
 420         journal_entry_btree_keys_to_text(out, c, entry);
 421 }
 422
 423 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 424                                 struct jset *jset,
 425                                 struct jset_entry *entry,
 426                                 unsigned version, int big_endian,
 427                                 enum bkey_invalid_flags flags)
 428 {
 429         /* obsolete, don't care: */
 430         return 0;
 431 }
 432
 433 static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 434                                             struct jset_entry *entry)
 435 {
 436 }
 437
 438 static int journal_entry_blacklist_validate(struct bch_fs *c,
 439                                 struct jset *jset,
 440                                 struct jset_entry *entry,
 441                                 unsigned version, int big_endian,
 442                                 enum bkey_invalid_flags flags)
 443 {
 444         int ret = 0;
 445
 446         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1,
 447                                  c, version, jset, entry,
 448                                  journal_entry_blacklist_bad_size,
 449                 "invalid journal seq blacklist entry: bad size")) {
 450                 journal_entry_null_range(entry, vstruct_next(entry));
 451         }
 452 fsck_err:
 453         return ret;
 454 }
 455
 456 static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
 457                                             struct jset_entry *entry)
 458 {
 459         struct jset_entry_blacklist *bl =
 460                 container_of(entry, struct jset_entry_blacklist, entry);
 461
 462         prt_printf(out, "seq=%llu", le64_to_cpu(bl->seq));
 463 }
 464
 465 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 466                                 struct jset *jset,
 467                                 struct jset_entry *entry,
 468                                 unsigned version, int big_endian,
 469                                 enum bkey_invalid_flags flags)
 470 {
 471         struct jset_entry_blacklist_v2 *bl_entry;
 472         int ret = 0;
 473
 474         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2,
 475                                  c, version, jset, entry,
 476                                  journal_entry_blacklist_v2_bad_size,
 477                 "invalid journal seq blacklist entry: bad size")) {
 478                 journal_entry_null_range(entry, vstruct_next(entry));
 479                 goto out;
 480         }
 481
 482         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 483
 484         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 485                                  le64_to_cpu(bl_entry->end),
 486                                  c, version, jset, entry,
 487                                  journal_entry_blacklist_v2_start_past_end,
 488                 "invalid journal seq blacklist entry: start > end")) {
 489                 journal_entry_null_range(entry, vstruct_next(entry));
 490         }
 491 out:
 492 fsck_err:
 493         return ret;
 494 }
 495
 496 static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
 497                                                struct jset_entry *entry)
 498 {
 499         struct jset_entry_blacklist_v2 *bl =
 500                 container_of(entry, struct jset_entry_blacklist_v2, entry);
 501
 502         prt_printf(out, "start=%llu end=%llu",
 503                le64_to_cpu(bl->start),
 504                le64_to_cpu(bl->end));
 505 }
 506
 507 static int journal_entry_usage_validate(struct bch_fs *c,
 508                                 struct jset *jset,
 509                                 struct jset_entry *entry,
 510                                 unsigned version, int big_endian,
 511                                 enum bkey_invalid_flags flags)
 512 {
 513         struct jset_entry_usage *u =
 514                 container_of(entry, struct jset_entry_usage, entry);
 515         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 516         int ret = 0;
 517
 518         if (journal_entry_err_on(bytes < sizeof(*u),
 519                                  c, version, jset, entry,
 520                                  journal_entry_usage_bad_size,
 521                                  "invalid journal entry usage: bad size")) {
 522                 journal_entry_null_range(entry, vstruct_next(entry));
 523                 return ret;
 524         }
 525
 526 fsck_err:
 527         return ret;
 528 }
 529
 530 static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 531                                         struct jset_entry *entry)
 532 {
 533         struct jset_entry_usage *u =
 534                 container_of(entry, struct jset_entry_usage, entry);
 535
 536         prt_printf(out, "type=%s v=%llu",
 537                bch2_fs_usage_types[u->entry.btree_id],
 538                le64_to_cpu(u->v));
 539 }
 540
 541 static int journal_entry_data_usage_validate(struct bch_fs *c,
 542                                 struct jset *jset,
 543                                 struct jset_entry *entry,
 544                                 unsigned version, int big_endian,
 545                                 enum bkey_invalid_flags flags)
 546 {
 547         struct jset_entry_data_usage *u =
 548                 container_of(entry, struct jset_entry_data_usage, entry);
 549         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 550         struct printbuf err = PRINTBUF;
 551         int ret = 0;
 552
 553         if (journal_entry_err_on(bytes < sizeof(*u) ||
 554                                  bytes < sizeof(*u) + u->r.nr_devs,
 555                                  c, version, jset, entry,
 556                                  journal_entry_data_usage_bad_size,
 557                                  "invalid journal entry usage: bad size")) {
 558                 journal_entry_null_range(entry, vstruct_next(entry));
 559                 goto out;
 560         }
 561
 562         if (journal_entry_err_on(bch2_replicas_entry_validate(&u->r, c->disk_sb.sb, &err),
 563                                  c, version, jset, entry,
 564                                  journal_entry_data_usage_bad_size,
 565                                  "invalid journal entry usage: %s", err.buf)) {
 566                 journal_entry_null_range(entry, vstruct_next(entry));
 567                 goto out;
 568         }
 569 out:
 570 fsck_err:
 571         printbuf_exit(&err);
 572         return ret;
 573 }
 574
 575 static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
 576                                              struct jset_entry *entry)
 577 {
 578         struct jset_entry_data_usage *u =
 579                 container_of(entry, struct jset_entry_data_usage, entry);
 580
 581         bch2_replicas_entry_to_text(out, &u->r);
 582         prt_printf(out, "=%llu", le64_to_cpu(u->v));
 583 }
 584
 585 static int journal_entry_clock_validate(struct bch_fs *c,
 586                                 struct jset *jset,
 587                                 struct jset_entry *entry,
 588                                 unsigned version, int big_endian,
 589                                 enum bkey_invalid_flags flags)
 590 {
 591         struct jset_entry_clock *clock =
 592                 container_of(entry, struct jset_entry_clock, entry);
 593         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 594         int ret = 0;
 595
 596         if (journal_entry_err_on(bytes != sizeof(*clock),
 597                                  c, version, jset, entry,
 598                                  journal_entry_clock_bad_size,
 599                                  "bad size")) {
 600                 journal_entry_null_range(entry, vstruct_next(entry));
 601                 return ret;
 602         }
 603
 604         if (journal_entry_err_on(clock->rw > 1,
 605                                  c, version, jset, entry,
 606                                  journal_entry_clock_bad_rw,
 607                                  "bad rw")) {
 608                 journal_entry_null_range(entry, vstruct_next(entry));
 609                 return ret;
 610         }
 611
 612 fsck_err:
 613         return ret;
 614 }
 615
 616 static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 617                                         struct jset_entry *entry)
 618 {
 619         struct jset_entry_clock *clock =
 620                 container_of(entry, struct jset_entry_clock, entry);
 621
 622         prt_printf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
 623 }
 624
 625 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 626                                 struct jset *jset,
 627                                 struct jset_entry *entry,
 628                                 unsigned version, int big_endian,
 629                                 enum bkey_invalid_flags flags)
 630 {
 631         struct jset_entry_dev_usage *u =
 632                 container_of(entry, struct jset_entry_dev_usage, entry);
 633         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 634         unsigned expected = sizeof(*u);
 635         unsigned dev;
 636         int ret = 0;
 637
 638         if (journal_entry_err_on(bytes < expected,
 639                                  c, version, jset, entry,
 640                                  journal_entry_dev_usage_bad_size,
 641                                  "bad size (%u < %u)",
 642                                  bytes, expected)) {
 643                 journal_entry_null_range(entry, vstruct_next(entry));
 644                 return ret;
 645         }
 646
 647         dev = le32_to_cpu(u->dev);
 648
 649         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 650                                  c, version, jset, entry,
 651                                  journal_entry_dev_usage_bad_dev,
 652                                  "bad dev")) {
 653                 journal_entry_null_range(entry, vstruct_next(entry));
 654                 return ret;
 655         }
 656
 657         if (journal_entry_err_on(u->pad,
 658                                  c, version, jset, entry,
 659                                  journal_entry_dev_usage_bad_pad,
 660                                  "bad pad")) {
 661                 journal_entry_null_range(entry, vstruct_next(entry));
 662                 return ret;
 663         }
 664
 665 fsck_err:
 666         return ret;
 667 }
 668
 669 static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
 670                                             struct jset_entry *entry)
 671 {
 672         struct jset_entry_dev_usage *u =
 673                 container_of(entry, struct jset_entry_dev_usage, entry);
 674         unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 675
 676         prt_printf(out, "dev=%u", le32_to_cpu(u->dev));
 677
 678         for (i = 0; i < nr_types; i++) {
 679                 if (i < BCH_DATA_NR)
 680                         prt_printf(out, " %s", bch2_data_types[i]);
 681                 else
 682                         prt_printf(out, " (unknown data type %u)", i);
 683                 prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 684                        le64_to_cpu(u->d[i].buckets),
 685                        le64_to_cpu(u->d[i].sectors),
 686                        le64_to_cpu(u->d[i].fragmented));
 687         }
 688 }
 689
 690 static int journal_entry_log_validate(struct bch_fs *c,
 691                                 struct jset *jset,
 692                                 struct jset_entry *entry,
 693                                 unsigned version, int big_endian,
 694                                 enum bkey_invalid_flags flags)
 695 {
 696         return 0;
 697 }
 698
 699 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 700                                       struct jset_entry *entry)
 701 {
 702         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 703         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 704
 705         prt_printf(out, "%.*s", bytes, l->d);
 706 }
 707
 708 static int journal_entry_overwrite_validate(struct bch_fs *c,
 709                                 struct jset *jset,
 710                                 struct jset_entry *entry,
 711                                 unsigned version, int big_endian,
 712                                 enum bkey_invalid_flags flags)
 713 {
 714         return journal_entry_btree_keys_validate(c, jset, entry,
 715                                 version, big_endian, READ);
 716 }
 717
 718 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
 719                                             struct jset_entry *entry)
 720 {
 721         journal_entry_btree_keys_to_text(out, c, entry);
 722 }
 723
 724 struct jset_entry_ops {
 725         int (*validate)(struct bch_fs *, struct jset *,
 726                         struct jset_entry *, unsigned, int,
 727                         enum bkey_invalid_flags);
 728         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 729 };
 730
 731 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 732 #define x(f, nr)                                                \
 733         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 734                 .validate       = journal_entry_##f##_validate, \
 735                 .to_text        = journal_entry_##f##_to_text,  \
 736         },
 737         BCH_JSET_ENTRY_TYPES()
 738 #undef x
 739 };
 740
 741 int bch2_journal_entry_validate(struct bch_fs *c,
 742                                 struct jset *jset,
 743                                 struct jset_entry *entry,
 744                                 unsigned version, int big_endian,
 745                                 enum bkey_invalid_flags flags)
 746 {
 747         return entry->type < BCH_JSET_ENTRY_NR
 748                 ? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
 749                                 version, big_endian, flags)
 750                 : 0;
 751 }
 752
 753 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 754                                 struct jset_entry *entry)
 755 {
 756         if (entry->type < BCH_JSET_ENTRY_NR) {
 757                 prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 758                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 759         } else {
 760                 prt_printf(out, "(unknown type %u)", entry->type);
 761         }
 762 }
 763
 764 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 765                                  enum bkey_invalid_flags flags)
 766 {
 767         struct jset_entry *entry;
 768         unsigned version = le32_to_cpu(jset->version);
 769         int ret = 0;
 770
 771         vstruct_for_each(jset, entry) {
 772                 if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
 773                                 c, version, jset, entry,
 774                                 journal_entry_past_jset_end,
 775                                 "journal entry extends past end of jset")) {
 776                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 777                         break;
 778                 }
 779
 780                 ret = bch2_journal_entry_validate(c, jset, entry,
 781                                         version, JSET_BIG_ENDIAN(jset), flags);
 782                 if (ret)
 783                         break;
 784         }
 785 fsck_err:
 786         return ret;
 787 }
 788
 789 static int jset_validate(struct bch_fs *c,
 790                          struct bch_dev *ca,
 791                          struct jset *jset, u64 sector,
 792                          enum bkey_invalid_flags flags)
 793 {
 794         unsigned version;
 795         int ret = 0;
 796
 797         if (le64_to_cpu(jset->magic) != jset_magic(c))
 798                 return JOURNAL_ENTRY_NONE;
 799
 800         version = le32_to_cpu(jset->version);
 801         if (journal_entry_err_on(!bch2_version_compatible(version),
 802                         c, version, jset, NULL,
 803                         jset_unsupported_version,
 804                         "%s sector %llu seq %llu: incompatible journal entry version %u.%u",
 805                         ca ? ca->name : c->name,
 806                         sector, le64_to_cpu(jset->seq),
 807                         BCH_VERSION_MAJOR(version),
 808                         BCH_VERSION_MINOR(version))) {
 809                 /* don't try to continue: */
 810                 return -EINVAL;
 811         }
 812
 813         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
 814                         c, version, jset, NULL,
 815                         jset_unknown_csum,
 816                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 817                         ca ? ca->name : c->name,
 818                         sector, le64_to_cpu(jset->seq),
 819                         JSET_CSUM_TYPE(jset)))
 820                 ret = JOURNAL_ENTRY_BAD;
 821
 822         /* last_seq is ignored when JSET_NO_FLUSH is true */
 823         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 824                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
 825                                  c, version, jset, NULL,
 826                                  jset_last_seq_newer_than_seq,
 827                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 828                                  le64_to_cpu(jset->last_seq),
 829                                  le64_to_cpu(jset->seq))) {
 830                 jset->last_seq = jset->seq;
 831                 return JOURNAL_ENTRY_BAD;
 832         }
 833
 834         ret = jset_validate_entries(c, jset, flags);
 835 fsck_err:
 836         return ret;
 837 }
 838
 839 static int jset_validate_early(struct bch_fs *c,
 840                          struct bch_dev *ca,
 841                          struct jset *jset, u64 sector,
 842                          unsigned bucket_sectors_left,
 843                          unsigned sectors_read)
 844 {
 845         size_t bytes = vstruct_bytes(jset);
 846         unsigned version;
 847         enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
 848         int ret = 0;
 849
 850         if (le64_to_cpu(jset->magic) != jset_magic(c))
 851                 return JOURNAL_ENTRY_NONE;
 852
 853         version = le32_to_cpu(jset->version);
 854         if (journal_entry_err_on(!bch2_version_compatible(version),
 855                         c, version, jset, NULL,
 856                         jset_unsupported_version,
 857                         "%s sector %llu seq %llu: unknown journal entry version %u.%u",
 858                         ca ? ca->name : c->name,
 859                         sector, le64_to_cpu(jset->seq),
 860                         BCH_VERSION_MAJOR(version),
 861                         BCH_VERSION_MINOR(version))) {
 862                 /* don't try to continue: */
 863                 return -EINVAL;
 864         }
 865
 866         if (bytes > (sectors_read << 9) &&
 867             sectors_read < bucket_sectors_left)
 868                 return JOURNAL_ENTRY_REREAD;
 869
 870         if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
 871                         c, version, jset, NULL,
 872                         jset_past_bucket_end,
 873                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 874                         ca ? ca->name : c->name,
 875                         sector, le64_to_cpu(jset->seq), bytes))
 876                 le32_add_cpu(&jset->u64s,
 877                              -((bytes - (bucket_sectors_left << 9)) / 8));
 878 fsck_err:
 879         return ret;
 880 }
 881
 882 struct journal_read_buf {
 883         void            *data;
 884         size_t          size;
 885 };
 886
 887 static int journal_read_buf_realloc(struct journal_read_buf *b,
 888                                     size_t new_size)
 889 {
 890         void *n;
 891
 892         /* the bios are sized for this many pages, max: */
 893         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 894                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 895
 896         new_size = roundup_pow_of_two(new_size);
 897         n = kvpmalloc(new_size, GFP_KERNEL);
 898         if (!n)
 899                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 900
 901         kvpfree(b->data, b->size);
 902         b->data = n;
 903         b->size = new_size;
 904         return 0;
 905 }
 906
 907 static int journal_read_bucket(struct bch_dev *ca,
 908                                struct journal_read_buf *buf,
 909                                struct journal_list *jlist,
 910                                unsigned bucket)
 911 {
 912         struct bch_fs *c = ca->fs;
 913         struct journal_device *ja = &ca->journal;
 914         struct jset *j = NULL;
 915         unsigned sectors, sectors_read = 0;
 916         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 917             end = offset + ca->mi.bucket_size;
 918         bool saw_bad = false, csum_good;
 919         int ret = 0;
 920
 921         pr_debug("reading %u", bucket);
 922
 923         while (offset < end) {
 924                 if (!sectors_read) {
 925                         struct bio *bio;
 926                         unsigned nr_bvecs;
 927 reread:
 928                         sectors_read = min_t(unsigned,
 929                                 end - offset, buf->size >> 9);
 930                         nr_bvecs = buf_pages(buf->data, sectors_read << 9);
 931
 932                         bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
 933                         bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
 934
 935                         bio->bi_iter.bi_sector = offset;
 936                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 937
 938                         ret = submit_bio_wait(bio);
 939                         kfree(bio);
 940
 941                         if (bch2_dev_io_err_on(ret, ca, BCH_MEMBER_ERROR_read,
 942                                                "journal read error: sector %llu",
 943                                                offset) ||
 944                             bch2_meta_read_fault("journal")) {
 945                                 /*
 946                                  * We don't error out of the recovery process
 947                                  * here, since the relevant journal entry may be
 948                                  * found on a different device, and missing or
 949                                  * no journal entries will be handled later
 950                                  */
 951                                 return 0;
 952                         }
 953
 954                         j = buf->data;
 955                 }
 956
 957                 ret = jset_validate_early(c, ca, j, offset,
 958                                     end - offset, sectors_read);
 959                 switch (ret) {
 960                 case 0:
 961                         sectors = vstruct_sectors(j, c->block_bits);
 962                         break;
 963                 case JOURNAL_ENTRY_REREAD:
 964                         if (vstruct_bytes(j) > buf->size) {
 965                                 ret = journal_read_buf_realloc(buf,
 966                                                         vstruct_bytes(j));
 967                                 if (ret)
 968                                         return ret;
 969                         }
 970                         goto reread;
 971                 case JOURNAL_ENTRY_NONE:
 972                         if (!saw_bad)
 973                                 return 0;
 974                         /*
 975                          * On checksum error we don't really trust the size
 976                          * field of the journal entry we read, so try reading
 977                          * again at next block boundary:
 978                          */
 979                         sectors = block_sectors(c);
 980                         goto next_block;
 981                 default:
 982                         return ret;
 983                 }
 984
 985                 /*
 986                  * This happens sometimes if we don't have discards on -
 987                  * when we've partially overwritten a bucket with new
 988                  * journal entries. We don't need the rest of the
 989                  * bucket:
 990                  */
 991                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 992                         return 0;
 993
 994                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 995
 996                 csum_good = jset_csum_good(c, j);
 997                 if (bch2_dev_io_err_on(!csum_good, ca, BCH_MEMBER_ERROR_checksum,
 998                                        "journal checksum error"))
 999                         saw_bad = true;
1000
1001                 ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
1002                              j->encrypted_start,
1003                              vstruct_end(j) - (void *) j->encrypted_start);
1004                 bch2_fs_fatal_err_on(ret, c,
1005                                 "error decrypting journal entry: %i", ret);
1006
1007                 mutex_lock(&jlist->lock);
1008                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
1009                                         .csum_good      = csum_good,
1010                                         .dev            = ca->dev_idx,
1011                                         .bucket         = bucket,
1012                                         .bucket_offset  = offset -
1013                                                 bucket_to_sector(ca, ja->buckets[bucket]),
1014                                         .sector         = offset,
1015                                         }, jlist, j);
1016                 mutex_unlock(&jlist->lock);
1017
1018                 switch (ret) {
1019                 case JOURNAL_ENTRY_ADD_OK:
1020                         break;
1021                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
1022                         break;
1023                 default:
1024                         return ret;
1025                 }
1026 next_block:
1027                 pr_debug("next");
1028                 offset          += sectors;
1029                 sectors_read    -= sectors;
1030                 j = ((void *) j) + (sectors << 9);
1031         }
1032
1033         return 0;
1034 }
1035
1036 static CLOSURE_CALLBACK(bch2_journal_read_device)
1037 {
1038         closure_type(ja, struct journal_device, read);
1039         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
1040         struct bch_fs *c = ca->fs;
1041         struct journal_list *jlist =
1042                 container_of(cl->parent, struct journal_list, cl);
1043         struct journal_replay *r, **_r;
1044         struct genradix_iter iter;
1045         struct journal_read_buf buf = { NULL, 0 };
1046         unsigned i;
1047         int ret = 0;
1048
1049         if (!ja->nr)
1050                 goto out;
1051
1052         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
1053         if (ret)
1054                 goto err;
1055
1056         pr_debug("%u journal buckets", ja->nr);
1057
1058         for (i = 0; i < ja->nr; i++) {
1059                 ret = journal_read_bucket(ca, &buf, jlist, i);
1060                 if (ret)
1061                         goto err;
1062         }
1063
1064         ja->sectors_free = ca->mi.bucket_size;
1065
1066         mutex_lock(&jlist->lock);
1067         genradix_for_each_reverse(&c->journal_entries, iter, _r) {
1068                 r = *_r;
1069
1070                 if (!r)
1071                         continue;
1072
1073                 for (i = 0; i < r->nr_ptrs; i++) {
1074                         if (r->ptrs[i].dev == ca->dev_idx) {
1075                                 unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
1076                                         vstruct_sectors(&r->j, c->block_bits);
1077
1078                                 ja->cur_idx = r->ptrs[i].bucket;
1079                                 ja->sectors_free = ca->mi.bucket_size - wrote;
1080                                 goto found;
1081                         }
1082                 }
1083         }
1084 found:
1085         mutex_unlock(&jlist->lock);
1086
1087         if (ja->bucket_seq[ja->cur_idx] &&
1088             ja->sectors_free == ca->mi.bucket_size) {
1089 #if 0
1090                 /*
1091                  * Debug code for ZNS support, where we (probably) want to be
1092                  * correlated where we stopped in the journal to the zone write
1093                  * points:
1094                  */
1095                 bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
1096                 bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
1097                 for (i = 0; i < 3; i++) {
1098                         unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
1099
1100                         bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
1101                 }
1102 #endif
1103                 ja->sectors_free = 0;
1104         }
1105
1106         /*
1107          * Set dirty_idx to indicate the entire journal is full and needs to be
1108          * reclaimed - journal reclaim will immediately reclaim whatever isn't
1109          * pinned when it first runs:
1110          */
1111         ja->discard_idx = ja->dirty_idx_ondisk =
1112                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
1113 out:
1114         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
1115         kvpfree(buf.data, buf.size);
1116         percpu_ref_put(&ca->io_ref);
1117         closure_return(cl);
1118         return;
1119 err:
1120         mutex_lock(&jlist->lock);
1121         jlist->ret = ret;
1122         mutex_unlock(&jlist->lock);
1123         goto out;
1124 }
1125
1126 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1127                                struct journal_replay *j)
1128 {
1129         unsigned i;
1130
1131         for (i = 0; i < j->nr_ptrs; i++) {
1132                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
1133                 u64 offset;
1134
1135                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
1136
1137                 if (i)
1138                         prt_printf(out, " ");
1139                 prt_printf(out, "%u:%u:%u (sector %llu)",
1140                        j->ptrs[i].dev,
1141                        j->ptrs[i].bucket,
1142                        j->ptrs[i].bucket_offset,
1143                        j->ptrs[i].sector);
1144         }
1145 }
1146
1147 int bch2_journal_read(struct bch_fs *c,
1148                       u64 *last_seq,
1149                       u64 *blacklist_seq,
1150                       u64 *start_seq)
1151 {
1152         struct journal_list jlist;
1153         struct journal_replay *i, **_i, *prev = NULL;
1154         struct genradix_iter radix_iter;
1155         struct bch_dev *ca;
1156         unsigned iter;
1157         struct printbuf buf = PRINTBUF;
1158         bool degraded = false, last_write_torn = false;
1159         u64 seq;
1160         int ret = 0;
1161
1162         closure_init_stack(&jlist.cl);
1163         mutex_init(&jlist.lock);
1164         jlist.last_seq = 0;
1165         jlist.ret = 0;
1166
1167         for_each_member_device(ca, c, iter) {
1168                 if (!c->opts.fsck &&
1169                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1170                         continue;
1171
1172                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1173                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1174                     percpu_ref_tryget(&ca->io_ref))
1175                         closure_call(&ca->journal.read,
1176                                      bch2_journal_read_device,
1177                                      system_unbound_wq,
1178                                      &jlist.cl);
1179                 else
1180                         degraded = true;
1181         }
1182
1183         closure_sync(&jlist.cl);
1184
1185         if (jlist.ret)
1186                 return jlist.ret;
1187
1188         *last_seq       = 0;
1189         *start_seq      = 0;
1190         *blacklist_seq  = 0;
1191
1192         /*
1193          * Find most recent flush entry, and ignore newer non flush entries -
1194          * those entries will be blacklisted:
1195          */
1196         genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
1197                 enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
1198
1199                 i = *_i;
1200
1201                 if (!i || i->ignore)
1202                         continue;
1203
1204                 if (!*start_seq)
1205                         *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
1206
1207                 if (JSET_NO_FLUSH(&i->j)) {
1208                         i->ignore = true;
1209                         continue;
1210                 }
1211
1212                 if (!last_write_torn && !i->csum_good) {
1213                         last_write_torn = true;
1214                         i->ignore = true;
1215                         continue;
1216                 }
1217
1218                 if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
1219                                          c, le32_to_cpu(i->j.version), &i->j, NULL,
1220                                          jset_last_seq_newer_than_seq,
1221                                          "invalid journal entry: last_seq > seq (%llu > %llu)",
1222                                          le64_to_cpu(i->j.last_seq),
1223                                          le64_to_cpu(i->j.seq)))
1224                         i->j.last_seq = i->j.seq;
1225
1226                 *last_seq       = le64_to_cpu(i->j.last_seq);
1227                 *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1228                 break;
1229         }
1230
1231         if (!*start_seq) {
1232                 bch_info(c, "journal read done, but no entries found");
1233                 return 0;
1234         }
1235
1236         if (!*last_seq) {
1237                 fsck_err(c, dirty_but_no_journal_entries_post_drop_nonflushes,
1238                          "journal read done, but no entries found after dropping non-flushes");
1239                 return 0;
1240         }
1241
1242         bch_info(c, "journal read done, replaying entries %llu-%llu",
1243                  *last_seq, *blacklist_seq - 1);
1244
1245         if (*start_seq != *blacklist_seq)
1246                 bch_info(c, "dropped unflushed entries %llu-%llu",
1247                          *blacklist_seq, *start_seq - 1);
1248
1249         /* Drop blacklisted entries and entries older than last_seq: */
1250         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1251                 i = *_i;
1252
1253                 if (!i || i->ignore)
1254                         continue;
1255
1256                 seq = le64_to_cpu(i->j.seq);
1257                 if (seq < *last_seq) {
1258                         journal_replay_free(c, i);
1259                         continue;
1260                 }
1261
1262                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1263                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1264                                     jset_seq_blacklisted,
1265                                     "found blacklisted journal entry %llu", seq);
1266                         i->ignore = true;
1267                 }
1268         }
1269
1270         /* Check for missing entries: */
1271         seq = *last_seq;
1272         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1273                 i = *_i;
1274
1275                 if (!i || i->ignore)
1276                         continue;
1277
1278                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1279
1280                 while (seq < le64_to_cpu(i->j.seq)) {
1281                         u64 missing_start, missing_end;
1282                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1283
1284                         while (seq < le64_to_cpu(i->j.seq) &&
1285                                bch2_journal_seq_is_blacklisted(c, seq, false))
1286                                 seq++;
1287
1288                         if (seq == le64_to_cpu(i->j.seq))
1289                                 break;
1290
1291                         missing_start = seq;
1292
1293                         while (seq < le64_to_cpu(i->j.seq) &&
1294                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1295                                 seq++;
1296
1297                         if (prev) {
1298                                 bch2_journal_ptrs_to_text(&buf1, c, prev);
1299                                 prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits));
1300                         } else
1301                                 prt_printf(&buf1, "(none)");
1302                         bch2_journal_ptrs_to_text(&buf2, c, i);
1303
1304                         missing_end = seq - 1;
1305                         fsck_err(c, journal_entries_missing,
1306                                  "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1307                                  "  prev at %s\n"
1308                                  "  next at %s",
1309                                  missing_start, missing_end,
1310                                  *last_seq, *blacklist_seq - 1,
1311                                  buf1.buf, buf2.buf);
1312
1313                         printbuf_exit(&buf1);
1314                         printbuf_exit(&buf2);
1315                 }
1316
1317                 prev = i;
1318                 seq++;
1319         }
1320
1321         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1322                 struct bch_replicas_padded replicas = {
1323                         .e.data_type = BCH_DATA_journal,
1324                         .e.nr_required = 1,
1325                 };
1326                 unsigned ptr;
1327
1328                 i = *_i;
1329                 if (!i || i->ignore)
1330                         continue;
1331
1332                 for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
1333                         ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
1334
1335                         if (!i->ptrs[ptr].csum_good)
1336                                 bch_err_dev_offset(ca, i->ptrs[ptr].sector,
1337                                                    "invalid journal checksum, seq %llu%s",
1338                                                    le64_to_cpu(i->j.seq),
1339                                                    i->csum_good ? " (had good copy on another device)" : "");
1340                 }
1341
1342                 ret = jset_validate(c,
1343                                     bch_dev_bkey_exists(c, i->ptrs[0].dev),
1344                                     &i->j,
1345                                     i->ptrs[0].sector,
1346                                     READ);
1347                 if (ret)
1348                         goto err;
1349
1350                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1351                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1352
1353                 bch2_replicas_entry_sort(&replicas.e);
1354
1355                 printbuf_reset(&buf);
1356                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1357
1358                 if (!degraded &&
1359                     !bch2_replicas_marked(c, &replicas.e) &&
1360                     (le64_to_cpu(i->j.seq) == *last_seq ||
1361                      fsck_err(c, journal_entry_replicas_not_marked,
1362                               "superblock not marked as containing replicas for journal entry %llu\n  %s",
1363                               le64_to_cpu(i->j.seq), buf.buf))) {
1364                         ret = bch2_mark_replicas(c, &replicas.e);
1365                         if (ret)
1366                                 goto err;
1367                 }
1368         }
1369 err:
1370 fsck_err:
1371         printbuf_exit(&buf);
1372         return ret;
1373 }
1374
1375 /* journal write: */
1376
1377 static void __journal_write_alloc(struct journal *j,
1378                                   struct journal_buf *w,
1379                                   struct dev_alloc_list *devs_sorted,
1380                                   unsigned sectors,
1381                                   unsigned *replicas,
1382                                   unsigned replicas_want)
1383 {
1384         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1385         struct journal_device *ja;
1386         struct bch_dev *ca;
1387         unsigned i;
1388
1389         if (*replicas >= replicas_want)
1390                 return;
1391
1392         for (i = 0; i < devs_sorted->nr; i++) {
1393                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1394                 if (!ca)
1395                         continue;
1396
1397                 ja = &ca->journal;
1398
1399                 /*
1400                  * Check that we can use this device, and aren't already using
1401                  * it:
1402                  */
1403                 if (!ca->mi.durability ||
1404                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1405                     !ja->nr ||
1406                     bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) ||
1407                     sectors > ja->sectors_free)
1408                         continue;
1409
1410                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1411
1412                 bch2_bkey_append_ptr(&w->key,
1413                         (struct bch_extent_ptr) {
1414                                   .offset = bucket_to_sector(ca,
1415                                         ja->buckets[ja->cur_idx]) +
1416                                         ca->mi.bucket_size -
1417                                         ja->sectors_free,
1418                                   .dev = ca->dev_idx,
1419                 });
1420
1421                 ja->sectors_free -= sectors;
1422                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1423
1424                 *replicas += ca->mi.durability;
1425
1426                 if (*replicas >= replicas_want)
1427                         break;
1428         }
1429 }
1430
1431 /**
1432  * journal_write_alloc - decide where to write next journal entry
1433  *
1434  * @j:          journal object
1435  * @w:          journal buf (entry to be written)
1436  *
1437  * Returns: 0 on success, or -EROFS on failure
1438  */
1439 static int journal_write_alloc(struct journal *j, struct journal_buf *w)
1440 {
1441         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1442         struct bch_devs_mask devs;
1443         struct journal_device *ja;
1444         struct bch_dev *ca;
1445         struct dev_alloc_list devs_sorted;
1446         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1447         unsigned target = c->opts.metadata_target ?:
1448                 c->opts.foreground_target;
1449         unsigned i, replicas = 0, replicas_want =
1450                 READ_ONCE(c->opts.metadata_replicas);
1451
1452         rcu_read_lock();
1453 retry:
1454         devs = target_rw_devs(c, BCH_DATA_journal, target);
1455
1456         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1457
1458         __journal_write_alloc(j, w, &devs_sorted,
1459                               sectors, &replicas, replicas_want);
1460
1461         if (replicas >= replicas_want)
1462                 goto done;
1463
1464         for (i = 0; i < devs_sorted.nr; i++) {
1465                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1466                 if (!ca)
1467                         continue;
1468
1469                 ja = &ca->journal;
1470
1471                 if (sectors > ja->sectors_free &&
1472                     sectors <= ca->mi.bucket_size &&
1473                     bch2_journal_dev_buckets_available(j, ja,
1474                                         journal_space_discarded)) {
1475                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1476                         ja->sectors_free = ca->mi.bucket_size;
1477
1478                         /*
1479                          * ja->bucket_seq[ja->cur_idx] must always have
1480                          * something sensible:
1481                          */
1482                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1483                 }
1484         }
1485
1486         __journal_write_alloc(j, w, &devs_sorted,
1487                               sectors, &replicas, replicas_want);
1488
1489         if (replicas < replicas_want && target) {
1490                 /* Retry from all devices: */
1491                 target = 0;
1492                 goto retry;
1493         }
1494 done:
1495         rcu_read_unlock();
1496
1497         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1498
1499         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1500 }
1501
1502 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1503 {
1504         /* we aren't holding j->lock: */
1505         unsigned new_size = READ_ONCE(j->buf_size_want);
1506         void *new_buf;
1507
1508         if (buf->buf_size >= new_size)
1509                 return;
1510
1511         new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
1512         if (!new_buf)
1513                 return;
1514
1515         memcpy(new_buf, buf->data, buf->buf_size);
1516
1517         spin_lock(&j->lock);
1518         swap(buf->data,         new_buf);
1519         swap(buf->buf_size,     new_size);
1520         spin_unlock(&j->lock);
1521
1522         kvpfree(new_buf, new_size);
1523 }
1524
1525 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1526 {
1527         return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
1528 }
1529
1530 static CLOSURE_CALLBACK(journal_write_done)
1531 {
1532         closure_type(j, struct journal, io);
1533         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1534         struct journal_buf *w = journal_last_unwritten_buf(j);
1535         struct bch_replicas_padded replicas;
1536         union journal_res_state old, new;
1537         u64 v, seq;
1538         int err = 0;
1539
1540         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1541                                ? j->flush_write_time
1542                                : j->noflush_write_time, j->write_start_time);
1543
1544         if (!w->devs_written.nr) {
1545                 bch_err(c, "unable to write journal to sufficient devices");
1546                 err = -EIO;
1547         } else {
1548                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1549                                          w->devs_written);
1550                 if (bch2_mark_replicas(c, &replicas.e))
1551                         err = -EIO;
1552         }
1553
1554         if (err)
1555                 bch2_fatal_error(c);
1556
1557         spin_lock(&j->lock);
1558         seq = le64_to_cpu(w->data->seq);
1559
1560         if (seq >= j->pin.front)
1561                 journal_seq_pin(j, seq)->devs = w->devs_written;
1562
1563         if (!err) {
1564                 if (!JSET_NO_FLUSH(w->data)) {
1565                         j->flushed_seq_ondisk = seq;
1566                         j->last_seq_ondisk = w->last_seq;
1567
1568                         bch2_do_discards(c);
1569                         closure_wake_up(&c->freelist_wait);
1570
1571                         bch2_reset_alloc_cursors(c);
1572                 }
1573         } else if (!j->err_seq || seq < j->err_seq)
1574                 j->err_seq      = seq;
1575
1576         j->seq_ondisk           = seq;
1577
1578         /*
1579          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1580          * more buckets:
1581          *
1582          * Must come before signaling write completion, for
1583          * bch2_fs_journal_stop():
1584          */
1585         if (j->watermark != BCH_WATERMARK_stripe)
1586                 journal_reclaim_kick(&c->journal);
1587
1588         /* also must come before signalling write completion: */
1589         closure_debug_destroy(cl);
1590
1591         v = atomic64_read(&j->reservations.counter);
1592         do {
1593                 old.v = new.v = v;
1594                 BUG_ON(journal_state_count(new, new.unwritten_idx));
1595
1596                 new.unwritten_idx++;
1597         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1598                                        old.v, new.v)) != old.v);
1599
1600         bch2_journal_space_available(j);
1601
1602         track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
1603                            &j->max_in_flight_start, false);
1604
1605         closure_wake_up(&w->wait);
1606         journal_wake(j);
1607
1608         if (!journal_state_count(new, new.unwritten_idx) &&
1609             journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
1610                 spin_unlock(&j->lock);
1611                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1612         } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
1613                    new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
1614                 struct journal_buf *buf = journal_cur_buf(j);
1615                 long delta = buf->expires - jiffies;
1616
1617                 /*
1618                  * We don't close a journal entry to write it while there's
1619                  * previous entries still in flight - the current journal entry
1620                  * might want to be written now:
1621                  */
1622
1623                 spin_unlock(&j->lock);
1624                 mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
1625         } else {
1626                 spin_unlock(&j->lock);
1627         }
1628 }
1629
1630 static void journal_write_endio(struct bio *bio)
1631 {
1632         struct bch_dev *ca = bio->bi_private;
1633         struct journal *j = &ca->fs->journal;
1634         struct journal_buf *w = journal_last_unwritten_buf(j);
1635         unsigned long flags;
1636
1637         if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1638                                "error writing journal entry %llu: %s",
1639                                le64_to_cpu(w->data->seq),
1640                                bch2_blk_status_to_str(bio->bi_status)) ||
1641             bch2_meta_write_fault("journal")) {
1642                 spin_lock_irqsave(&j->err_lock, flags);
1643                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1644                 spin_unlock_irqrestore(&j->err_lock, flags);
1645         }
1646
1647         closure_put(&j->io);
1648         percpu_ref_put(&ca->io_ref);
1649 }
1650
1651 static CLOSURE_CALLBACK(do_journal_write)
1652 {
1653         closure_type(j, struct journal, io);
1654         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1655         struct bch_dev *ca;
1656         struct journal_buf *w = journal_last_unwritten_buf(j);
1657         struct bch_extent_ptr *ptr;
1658         struct bio *bio;
1659         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1660
1661         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1662                 ca = bch_dev_bkey_exists(c, ptr->dev);
1663                 if (!percpu_ref_tryget(&ca->io_ref)) {
1664                         /* XXX: fix this */
1665                         bch_err(c, "missing device for journal write\n");
1666                         continue;
1667                 }
1668
1669                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1670                              sectors);
1671
1672                 bio = ca->journal.bio;
1673                 bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
1674                 bio->bi_iter.bi_sector  = ptr->offset;
1675                 bio->bi_end_io          = journal_write_endio;
1676                 bio->bi_private         = ca;
1677
1678                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1679                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1680
1681                 if (!JSET_NO_FLUSH(w->data))
1682                         bio->bi_opf    |= REQ_FUA;
1683                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1684                         bio->bi_opf    |= REQ_PREFLUSH;
1685
1686                 bch2_bio_map(bio, w->data, sectors << 9);
1687
1688                 trace_and_count(c, journal_write, bio);
1689                 closure_bio_submit(bio, cl);
1690
1691                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1692                         le64_to_cpu(w->data->seq);
1693         }
1694
1695         continue_at(cl, journal_write_done, c->io_complete_wq);
1696 }
1697
1698 static int bch2_journal_write_prep(struct journal *j, struct journal_buf *w)
1699 {
1700         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1701         struct jset_entry *start, *end, *i, *next, *prev = NULL;
1702         struct jset *jset = w->data;
1703         unsigned sectors, bytes, u64s;
1704         bool validate_before_checksum = false;
1705         unsigned long btree_roots_have = 0;
1706         int ret;
1707
1708         /*
1709          * Simple compaction, dropping empty jset_entries (from journal
1710          * reservations that weren't fully used) and merging jset_entries that
1711          * can be.
1712          *
1713          * If we wanted to be really fancy here, we could sort all the keys in
1714          * the jset and drop keys that were overwritten - probably not worth it:
1715          */
1716         vstruct_for_each_safe(jset, i, next) {
1717                 unsigned u64s = le16_to_cpu(i->u64s);
1718
1719                 /* Empty entry: */
1720                 if (!u64s)
1721                         continue;
1722
1723                 /*
1724                  * New btree roots are set by journalling them; when the journal
1725                  * entry gets written we have to propagate them to
1726                  * c->btree_roots
1727                  *
1728                  * But, every journal entry we write has to contain all the
1729                  * btree roots (at least for now); so after we copy btree roots
1730                  * to c->btree_roots we have to get any missing btree roots and
1731                  * add them to this journal entry:
1732                  */
1733                 if (i->type == BCH_JSET_ENTRY_btree_root) {
1734                         bch2_journal_entry_to_btree_root(c, i);
1735                         __set_bit(i->btree_id, &btree_roots_have);
1736                 }
1737
1738                 /* Can we merge with previous entry? */
1739                 if (prev &&
1740                     i->btree_id == prev->btree_id &&
1741                     i->level    == prev->level &&
1742                     i->type     == prev->type &&
1743                     i->type     == BCH_JSET_ENTRY_btree_keys &&
1744                     le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
1745                         memmove_u64s_down(vstruct_next(prev),
1746                                           i->_data,
1747                                           u64s);
1748                         le16_add_cpu(&prev->u64s, u64s);
1749                         continue;
1750                 }
1751
1752                 /* Couldn't merge, move i into new position (after prev): */
1753                 prev = prev ? vstruct_next(prev) : jset->start;
1754                 if (i != prev)
1755                         memmove_u64s_down(prev, i, jset_u64s(u64s));
1756         }
1757
1758         prev = prev ? vstruct_next(prev) : jset->start;
1759         jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
1760
1761         start = end = vstruct_last(jset);
1762
1763         end     = bch2_btree_roots_to_journal_entries(c, end, btree_roots_have);
1764
1765         bch2_journal_super_entries_add_common(c, &end,
1766                                 le64_to_cpu(jset->seq));
1767         u64s    = (u64 *) end - (u64 *) start;
1768         BUG_ON(u64s > j->entry_u64s_reserved);
1769
1770         le32_add_cpu(&jset->u64s, u64s);
1771
1772         sectors = vstruct_sectors(jset, c->block_bits);
1773         bytes   = vstruct_bytes(jset);
1774
1775         if (sectors > w->sectors) {
1776                 bch2_fs_fatal_error(c, "aieeee! journal write overran available space, %zu > %u (extra %u reserved %u/%u)",
1777                                     vstruct_bytes(jset), w->sectors << 9,
1778                                     u64s, w->u64s_reserved, j->entry_u64s_reserved);
1779                 return -EINVAL;
1780         }
1781
1782         jset->magic             = cpu_to_le64(jset_magic(c));
1783         jset->version           = cpu_to_le32(c->sb.version);
1784
1785         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1786         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1787
1788         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1789                 j->last_empty_seq = le64_to_cpu(jset->seq);
1790
1791         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1792                 validate_before_checksum = true;
1793
1794         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1795                 validate_before_checksum = true;
1796
1797         if (validate_before_checksum &&
1798             (ret = jset_validate(c, NULL, jset, 0, WRITE)))
1799                 return ret;
1800
1801         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1802                     jset->encrypted_start,
1803                     vstruct_end(jset) - (void *) jset->encrypted_start);
1804         if (bch2_fs_fatal_err_on(ret, c,
1805                         "error decrypting journal entry: %i", ret))
1806                 return ret;
1807
1808         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1809                                   journal_nonce(jset), jset);
1810
1811         if (!validate_before_checksum &&
1812             (ret = jset_validate(c, NULL, jset, 0, WRITE)))
1813                 return ret;
1814
1815         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1816         return 0;
1817 }
1818
1819 static int bch2_journal_write_pick_flush(struct journal *j, struct journal_buf *w)
1820 {
1821         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1822         int error = bch2_journal_error(j);
1823
1824         /*
1825          * If the journal is in an error state - we did an emergency shutdown -
1826          * we prefer to continue doing journal writes. We just mark them as
1827          * noflush so they'll never be used, but they'll still be visible by the
1828          * list_journal tool - this helps in debugging.
1829          *
1830          * There's a caveat: the first journal write after marking the
1831          * superblock dirty must always be a flush write, because on startup
1832          * from a clean shutdown we didn't necessarily read the journal and the
1833          * new journal write might overwrite whatever was in the journal
1834          * previously - we can't leave the journal without any flush writes in
1835          * it.
1836          *
1837          * So if we're in an error state, and we're still starting up, we don't
1838          * write anything at all.
1839          */
1840         if (error && test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags))
1841                 return -EIO;
1842
1843         if (error ||
1844             w->noflush ||
1845             (!w->must_flush &&
1846              (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1847              test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags))) {
1848                 w->noflush = true;
1849                 SET_JSET_NO_FLUSH(w->data, true);
1850                 w->data->last_seq       = 0;
1851                 w->last_seq             = 0;
1852
1853                 j->nr_noflush_writes++;
1854         } else {
1855                 j->last_flush_write = jiffies;
1856                 j->nr_flush_writes++;
1857                 clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
1858         }
1859
1860         return 0;
1861 }
1862
1863 CLOSURE_CALLBACK(bch2_journal_write)
1864 {
1865         closure_type(j, struct journal, io);
1866         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1867         struct bch_dev *ca;
1868         struct journal_buf *w = journal_last_unwritten_buf(j);
1869         struct bch_replicas_padded replicas;
1870         struct bio *bio;
1871         struct printbuf journal_debug_buf = PRINTBUF;
1872         unsigned i, nr_rw_members = 0;
1873         int ret;
1874
1875         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1876
1877         j->write_start_time = local_clock();
1878
1879         spin_lock(&j->lock);
1880         ret = bch2_journal_write_pick_flush(j, w);
1881         spin_unlock(&j->lock);
1882         if (ret)
1883                 goto err;
1884
1885         journal_buf_realloc(j, w);
1886
1887         ret = bch2_journal_write_prep(j, w);
1888         if (ret)
1889                 goto err;
1890
1891         j->entry_bytes_written += vstruct_bytes(w->data);
1892
1893         while (1) {
1894                 spin_lock(&j->lock);
1895                 ret = journal_write_alloc(j, w);
1896                 if (!ret || !j->can_discard)
1897                         break;
1898
1899                 spin_unlock(&j->lock);
1900                 bch2_journal_do_discards(j);
1901         }
1902
1903         if (ret) {
1904                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1905                 spin_unlock(&j->lock);
1906                 bch_err(c, "Unable to allocate journal write:\n%s",
1907                         journal_debug_buf.buf);
1908                 printbuf_exit(&journal_debug_buf);
1909                 goto err;
1910         }
1911
1912         /*
1913          * write is allocated, no longer need to account for it in
1914          * bch2_journal_space_available():
1915          */
1916         w->sectors = 0;
1917
1918         /*
1919          * journal entry has been compacted and allocated, recalculate space
1920          * available:
1921          */
1922         bch2_journal_space_available(j);
1923         spin_unlock(&j->lock);
1924
1925         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1926
1927         if (c->opts.nochanges)
1928                 goto no_io;
1929
1930         for_each_rw_member(ca, c, i)
1931                 nr_rw_members++;
1932
1933         if (nr_rw_members > 1)
1934                 w->separate_flush = true;
1935
1936         /*
1937          * Mark journal replicas before we submit the write to guarantee
1938          * recovery will find the journal entries after a crash.
1939          */
1940         bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1941                                  w->devs_written);
1942         ret = bch2_mark_replicas(c, &replicas.e);
1943         if (ret)
1944                 goto err;
1945
1946         if (!JSET_NO_FLUSH(w->data) && w->separate_flush) {
1947                 for_each_rw_member(ca, c, i) {
1948                         percpu_ref_get(&ca->io_ref);
1949
1950                         bio = ca->journal.bio;
1951                         bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
1952                         bio->bi_end_io          = journal_write_endio;
1953                         bio->bi_private         = ca;
1954                         closure_bio_submit(bio, cl);
1955                 }
1956         }
1957
1958         continue_at(cl, do_journal_write, c->io_complete_wq);
1959         return;
1960 no_io:
1961         continue_at(cl, journal_write_done, c->io_complete_wq);
1962         return;
1963 err:
1964         bch2_fatal_error(c);
1965         continue_at(cl, journal_write_done, c->io_complete_wq);
1966 }