git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/journal_io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_io.h"
   6 #include "btree_update_interior.h"
   7 #include "btree_write_buffer.h"
   8 #include "buckets.h"
   9 #include "checksum.h"
  10 #include "disk_groups.h"
  11 #include "error.h"
  12 #include "journal.h"
  13 #include "journal_io.h"
  14 #include "journal_reclaim.h"
  15 #include "journal_seq_blacklist.h"
  16 #include "replicas.h"
  17 #include "sb-clean.h"
  18 #include "trace.h"
  19
  20 static struct nonce journal_nonce(const struct jset *jset)
  21 {
  22         return (struct nonce) {{
  23                 [0] = 0,
  24                 [1] = ((__le32 *) &jset->seq)[0],
  25                 [2] = ((__le32 *) &jset->seq)[1],
  26                 [3] = BCH_NONCE_JOURNAL,
  27         }};
  28 }
  29
  30 static bool jset_csum_good(struct bch_fs *c, struct jset *j)
  31 {
  32         return bch2_checksum_type_valid(c, JSET_CSUM_TYPE(j)) &&
  33                 !bch2_crc_cmp(j->csum,
  34                               csum_vstruct(c, JSET_CSUM_TYPE(j), journal_nonce(j), j));
  35 }
  36
  37 static inline u32 journal_entry_radix_idx(struct bch_fs *c, u64 seq)
  38 {
  39         return (seq - c->journal_entries_base_seq) & (~0U >> 1);
  40 }
  41
  42 static void __journal_replay_free(struct bch_fs *c,
  43                                   struct journal_replay *i)
  44 {
  45         struct journal_replay **p =
  46                 genradix_ptr(&c->journal_entries,
  47                              journal_entry_radix_idx(c, le64_to_cpu(i->j.seq)));
  48
  49         BUG_ON(*p != i);
  50         *p = NULL;
  51         kvpfree(i, offsetof(struct journal_replay, j) +
  52                 vstruct_bytes(&i->j));
  53 }
  54
  55 static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
  56 {
  57         i->ignore = true;
  58
  59         if (!c->opts.read_entire_journal)
  60                 __journal_replay_free(c, i);
  61 }
  62
  63 struct journal_list {
  64         struct closure          cl;
  65         u64                     last_seq;
  66         struct mutex            lock;
  67         int                     ret;
  68 };
  69
  70 #define JOURNAL_ENTRY_ADD_OK            0
  71 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  72
  73 /*
  74  * Given a journal entry we just read, add it to the list of journal entries to
  75  * be replayed:
  76  */
  77 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  78                              struct journal_ptr entry_ptr,
  79                              struct journal_list *jlist, struct jset *j)
  80 {
  81         struct genradix_iter iter;
  82         struct journal_replay **_i, *i, *dup;
  83         struct journal_ptr *ptr;
  84         size_t bytes = vstruct_bytes(j);
  85         u64 last_seq = !JSET_NO_FLUSH(j) ? le64_to_cpu(j->last_seq) : 0;
  86         int ret = JOURNAL_ENTRY_ADD_OK;
  87
  88         /* Is this entry older than the range we need? */
  89         if (!c->opts.read_entire_journal &&
  90             le64_to_cpu(j->seq) < jlist->last_seq)
  91                 return JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  92
  93         /*
  94          * genradixes are indexed by a ulong, not a u64, so we can't index them
  95          * by sequence number directly: Assume instead that they will all fall
  96          * within the range of +-2billion of the filrst one we find.
  97          */
  98         if (!c->journal_entries_base_seq)
  99                 c->journal_entries_base_seq = max_t(s64, 1, le64_to_cpu(j->seq) - S32_MAX);
 100
 101         /* Drop entries we don't need anymore */
 102         if (last_seq > jlist->last_seq && !c->opts.read_entire_journal) {
 103                 genradix_for_each_from(&c->journal_entries, iter, _i,
 104                                        journal_entry_radix_idx(c, jlist->last_seq)) {
 105                         i = *_i;
 106
 107                         if (!i || i->ignore)
 108                                 continue;
 109
 110                         if (le64_to_cpu(i->j.seq) >= last_seq)
 111                                 break;
 112                         journal_replay_free(c, i);
 113                 }
 114         }
 115
 116         jlist->last_seq = max(jlist->last_seq, last_seq);
 117
 118         _i = genradix_ptr_alloc(&c->journal_entries,
 119                                 journal_entry_radix_idx(c, le64_to_cpu(j->seq)),
 120                                 GFP_KERNEL);
 121         if (!_i)
 122                 return -BCH_ERR_ENOMEM_journal_entry_add;
 123
 124         /*
 125          * Duplicate journal entries? If so we want the one that didn't have a
 126          * checksum error:
 127          */
 128         dup = *_i;
 129         if (dup) {
 130                 if (bytes == vstruct_bytes(&dup->j) &&
 131                     !memcmp(j, &dup->j, bytes)) {
 132                         i = dup;
 133                         goto found;
 134                 }
 135
 136                 if (!entry_ptr.csum_good) {
 137                         i = dup;
 138                         goto found;
 139                 }
 140
 141                 if (!dup->csum_good)
 142                         goto replace;
 143
 144                 fsck_err(c, journal_entry_replicas_data_mismatch,
 145                          "found duplicate but non identical journal entries (seq %llu)",
 146                          le64_to_cpu(j->seq));
 147                 i = dup;
 148                 goto found;
 149         }
 150 replace:
 151         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 152         if (!i)
 153                 return -BCH_ERR_ENOMEM_journal_entry_add;
 154
 155         i->nr_ptrs      = 0;
 156         i->csum_good    = entry_ptr.csum_good;
 157         i->ignore       = false;
 158         unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
 159         i->ptrs[i->nr_ptrs++] = entry_ptr;
 160
 161         if (dup) {
 162                 if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) {
 163                         bch_err(c, "found too many copies of journal entry %llu",
 164                                 le64_to_cpu(i->j.seq));
 165                         dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1;
 166                 }
 167
 168                 /* The first ptr should represent the jset we kept: */
 169                 memcpy(i->ptrs + i->nr_ptrs,
 170                        dup->ptrs,
 171                        sizeof(dup->ptrs[0]) * dup->nr_ptrs);
 172                 i->nr_ptrs += dup->nr_ptrs;
 173                 __journal_replay_free(c, dup);
 174         }
 175
 176         *_i = i;
 177         return 0;
 178 found:
 179         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 180                 if (ptr->dev == ca->dev_idx) {
 181                         bch_err(c, "duplicate journal entry %llu on same device",
 182                                 le64_to_cpu(i->j.seq));
 183                         goto out;
 184                 }
 185         }
 186
 187         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 188                 bch_err(c, "found too many copies of journal entry %llu",
 189                         le64_to_cpu(i->j.seq));
 190                 goto out;
 191         }
 192
 193         i->ptrs[i->nr_ptrs++] = entry_ptr;
 194 out:
 195 fsck_err:
 196         return ret;
 197 }
 198
 199 /* this fills in a range with empty jset_entries: */
 200 static void journal_entry_null_range(void *start, void *end)
 201 {
 202         struct jset_entry *entry;
 203
 204         for (entry = start; entry != end; entry = vstruct_next(entry))
 205                 memset(entry, 0, sizeof(*entry));
 206 }
 207
 208 #define JOURNAL_ENTRY_REREAD    5
 209 #define JOURNAL_ENTRY_NONE      6
 210 #define JOURNAL_ENTRY_BAD       7
 211
 212 static void journal_entry_err_msg(struct printbuf *out,
 213                                   u32 version,
 214                                   struct jset *jset,
 215                                   struct jset_entry *entry)
 216 {
 217         prt_str(out, "invalid journal entry, version=");
 218         bch2_version_to_text(out, version);
 219
 220         if (entry) {
 221                 prt_str(out, " type=");
 222                 prt_str(out, bch2_jset_entry_types[entry->type]);
 223         }
 224
 225         if (!jset) {
 226                 prt_printf(out, " in superblock");
 227         } else {
 228
 229                 prt_printf(out, " seq=%llu", le64_to_cpu(jset->seq));
 230
 231                 if (entry)
 232                         prt_printf(out, " offset=%zi/%u",
 233                                    (u64 *) entry - jset->_data,
 234                                    le32_to_cpu(jset->u64s));
 235         }
 236
 237         prt_str(out, ": ");
 238 }
 239
 240 #define journal_entry_err(c, version, jset, entry, _err, msg, ...)      \
 241 ({                                                                      \
 242         struct printbuf _buf = PRINTBUF;                                \
 243                                                                         \
 244         journal_entry_err_msg(&_buf, version, jset, entry);             \
 245         prt_printf(&_buf, msg, ##__VA_ARGS__);                          \
 246                                                                         \
 247         switch (flags & BKEY_INVALID_WRITE) {                           \
 248         case READ:                                                      \
 249                 mustfix_fsck_err(c, _err, "%s", _buf.buf);              \
 250                 break;                                                  \
 251         case WRITE:                                                     \
 252                 bch2_sb_error_count(c, BCH_FSCK_ERR_##_err);            \
 253                 bch_err(c, "corrupt metadata before write: %s\n", _buf.buf);\
 254                 if (bch2_fs_inconsistent(c)) {                          \
 255                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 256                         goto fsck_err;                                  \
 257                 }                                                       \
 258                 break;                                                  \
 259         }                                                               \
 260                                                                         \
 261         printbuf_exit(&_buf);                                           \
 262         true;                                                           \
 263 })
 264
 265 #define journal_entry_err_on(cond, ...)                                 \
 266         ((cond) ? journal_entry_err(__VA_ARGS__) : false)
 267
 268 #define FSCK_DELETED_KEY        5
 269
 270 static int journal_validate_key(struct bch_fs *c,
 271                                 struct jset *jset,
 272                                 struct jset_entry *entry,
 273                                 unsigned level, enum btree_id btree_id,
 274                                 struct bkey_i *k,
 275                                 unsigned version, int big_endian,
 276                                 enum bkey_invalid_flags flags)
 277 {
 278         int write = flags & BKEY_INVALID_WRITE;
 279         void *next = vstruct_next(entry);
 280         struct printbuf buf = PRINTBUF;
 281         int ret = 0;
 282
 283         if (journal_entry_err_on(!k->k.u64s,
 284                                  c, version, jset, entry,
 285                                  journal_entry_bkey_u64s_0,
 286                                  "k->u64s 0")) {
 287                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 288                 journal_entry_null_range(vstruct_next(entry), next);
 289                 return FSCK_DELETED_KEY;
 290         }
 291
 292         if (journal_entry_err_on((void *) bkey_next(k) >
 293                                  (void *) vstruct_next(entry),
 294                                  c, version, jset, entry,
 295                                  journal_entry_bkey_past_end,
 296                                  "extends past end of journal entry")) {
 297                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 298                 journal_entry_null_range(vstruct_next(entry), next);
 299                 return FSCK_DELETED_KEY;
 300         }
 301
 302         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
 303                                  c, version, jset, entry,
 304                                  journal_entry_bkey_bad_format,
 305                                  "bad format %u", k->k.format)) {
 306                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 307                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 308                 journal_entry_null_range(vstruct_next(entry), next);
 309                 return FSCK_DELETED_KEY;
 310         }
 311
 312         if (!write)
 313                 bch2_bkey_compat(level, btree_id, version, big_endian,
 314                                  write, NULL, bkey_to_packed(k));
 315
 316         if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 317                               __btree_node_type(level, btree_id), write, &buf)) {
 318                 printbuf_reset(&buf);
 319                 journal_entry_err_msg(&buf, version, jset, entry);
 320                 prt_newline(&buf);
 321                 printbuf_indent_add(&buf, 2);
 322
 323                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
 324                 prt_newline(&buf);
 325                 bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 326                                   __btree_node_type(level, btree_id), write, &buf);
 327
 328                 mustfix_fsck_err(c, journal_entry_bkey_invalid,
 329                                  "%s", buf.buf);
 330
 331                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 332                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 333                 journal_entry_null_range(vstruct_next(entry), next);
 334
 335                 printbuf_exit(&buf);
 336                 return FSCK_DELETED_KEY;
 337         }
 338
 339         if (write)
 340                 bch2_bkey_compat(level, btree_id, version, big_endian,
 341                                  write, NULL, bkey_to_packed(k));
 342 fsck_err:
 343         printbuf_exit(&buf);
 344         return ret;
 345 }
 346
 347 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 348                                 struct jset *jset,
 349                                 struct jset_entry *entry,
 350                                 unsigned version, int big_endian,
 351                                 enum bkey_invalid_flags flags)
 352 {
 353         struct bkey_i *k = entry->start;
 354
 355         while (k != vstruct_last(entry)) {
 356                 int ret = journal_validate_key(c, jset, entry,
 357                                                entry->level,
 358                                                entry->btree_id,
 359                                                k, version, big_endian,
 360                                                flags|BKEY_INVALID_JOURNAL);
 361                 if (ret == FSCK_DELETED_KEY)
 362                         continue;
 363
 364                 k = bkey_next(k);
 365         }
 366
 367         return 0;
 368 }
 369
 370 static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
 371                                              struct jset_entry *entry)
 372 {
 373         struct bkey_i *k;
 374         bool first = true;
 375
 376         jset_entry_for_each_key(entry, k) {
 377                 if (!first) {
 378                         prt_newline(out);
 379                         prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 380                 }
 381                 prt_printf(out, "btree=%s l=%u ", bch2_btree_id_str(entry->btree_id), entry->level);
 382                 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 383                 first = false;
 384         }
 385 }
 386
 387 static int journal_entry_btree_root_validate(struct bch_fs *c,
 388                                 struct jset *jset,
 389                                 struct jset_entry *entry,
 390                                 unsigned version, int big_endian,
 391                                 enum bkey_invalid_flags flags)
 392 {
 393         struct bkey_i *k = entry->start;
 394         int ret = 0;
 395
 396         if (journal_entry_err_on(!entry->u64s ||
 397                                  le16_to_cpu(entry->u64s) != k->k.u64s,
 398                                  c, version, jset, entry,
 399                                  journal_entry_btree_root_bad_size,
 400                                  "invalid btree root journal entry: wrong number of keys")) {
 401                 void *next = vstruct_next(entry);
 402                 /*
 403                  * we don't want to null out this jset_entry,
 404                  * just the contents, so that later we can tell
 405                  * we were _supposed_ to have a btree root
 406                  */
 407                 entry->u64s = 0;
 408                 journal_entry_null_range(vstruct_next(entry), next);
 409                 return 0;
 410         }
 411
 412         return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
 413                                     version, big_endian, flags);
 414 fsck_err:
 415         return ret;
 416 }
 417
 418 static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
 419                                              struct jset_entry *entry)
 420 {
 421         journal_entry_btree_keys_to_text(out, c, entry);
 422 }
 423
 424 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 425                                 struct jset *jset,
 426                                 struct jset_entry *entry,
 427                                 unsigned version, int big_endian,
 428                                 enum bkey_invalid_flags flags)
 429 {
 430         /* obsolete, don't care: */
 431         return 0;
 432 }
 433
 434 static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 435                                             struct jset_entry *entry)
 436 {
 437 }
 438
 439 static int journal_entry_blacklist_validate(struct bch_fs *c,
 440                                 struct jset *jset,
 441                                 struct jset_entry *entry,
 442                                 unsigned version, int big_endian,
 443                                 enum bkey_invalid_flags flags)
 444 {
 445         int ret = 0;
 446
 447         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1,
 448                                  c, version, jset, entry,
 449                                  journal_entry_blacklist_bad_size,
 450                 "invalid journal seq blacklist entry: bad size")) {
 451                 journal_entry_null_range(entry, vstruct_next(entry));
 452         }
 453 fsck_err:
 454         return ret;
 455 }
 456
 457 static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
 458                                             struct jset_entry *entry)
 459 {
 460         struct jset_entry_blacklist *bl =
 461                 container_of(entry, struct jset_entry_blacklist, entry);
 462
 463         prt_printf(out, "seq=%llu", le64_to_cpu(bl->seq));
 464 }
 465
 466 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 467                                 struct jset *jset,
 468                                 struct jset_entry *entry,
 469                                 unsigned version, int big_endian,
 470                                 enum bkey_invalid_flags flags)
 471 {
 472         struct jset_entry_blacklist_v2 *bl_entry;
 473         int ret = 0;
 474
 475         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2,
 476                                  c, version, jset, entry,
 477                                  journal_entry_blacklist_v2_bad_size,
 478                 "invalid journal seq blacklist entry: bad size")) {
 479                 journal_entry_null_range(entry, vstruct_next(entry));
 480                 goto out;
 481         }
 482
 483         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 484
 485         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 486                                  le64_to_cpu(bl_entry->end),
 487                                  c, version, jset, entry,
 488                                  journal_entry_blacklist_v2_start_past_end,
 489                 "invalid journal seq blacklist entry: start > end")) {
 490                 journal_entry_null_range(entry, vstruct_next(entry));
 491         }
 492 out:
 493 fsck_err:
 494         return ret;
 495 }
 496
 497 static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
 498                                                struct jset_entry *entry)
 499 {
 500         struct jset_entry_blacklist_v2 *bl =
 501                 container_of(entry, struct jset_entry_blacklist_v2, entry);
 502
 503         prt_printf(out, "start=%llu end=%llu",
 504                le64_to_cpu(bl->start),
 505                le64_to_cpu(bl->end));
 506 }
 507
 508 static int journal_entry_usage_validate(struct bch_fs *c,
 509                                 struct jset *jset,
 510                                 struct jset_entry *entry,
 511                                 unsigned version, int big_endian,
 512                                 enum bkey_invalid_flags flags)
 513 {
 514         struct jset_entry_usage *u =
 515                 container_of(entry, struct jset_entry_usage, entry);
 516         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 517         int ret = 0;
 518
 519         if (journal_entry_err_on(bytes < sizeof(*u),
 520                                  c, version, jset, entry,
 521                                  journal_entry_usage_bad_size,
 522                                  "invalid journal entry usage: bad size")) {
 523                 journal_entry_null_range(entry, vstruct_next(entry));
 524                 return ret;
 525         }
 526
 527 fsck_err:
 528         return ret;
 529 }
 530
 531 static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 532                                         struct jset_entry *entry)
 533 {
 534         struct jset_entry_usage *u =
 535                 container_of(entry, struct jset_entry_usage, entry);
 536
 537         prt_printf(out, "type=%s v=%llu",
 538                bch2_fs_usage_types[u->entry.btree_id],
 539                le64_to_cpu(u->v));
 540 }
 541
 542 static int journal_entry_data_usage_validate(struct bch_fs *c,
 543                                 struct jset *jset,
 544                                 struct jset_entry *entry,
 545                                 unsigned version, int big_endian,
 546                                 enum bkey_invalid_flags flags)
 547 {
 548         struct jset_entry_data_usage *u =
 549                 container_of(entry, struct jset_entry_data_usage, entry);
 550         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 551         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                 return ret;
 560         }
 561
 562 fsck_err:
 563         return ret;
 564 }
 565
 566 static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
 567                                              struct jset_entry *entry)
 568 {
 569         struct jset_entry_data_usage *u =
 570                 container_of(entry, struct jset_entry_data_usage, entry);
 571
 572         bch2_replicas_entry_to_text(out, &u->r);
 573         prt_printf(out, "=%llu", le64_to_cpu(u->v));
 574 }
 575
 576 static int journal_entry_clock_validate(struct bch_fs *c,
 577                                 struct jset *jset,
 578                                 struct jset_entry *entry,
 579                                 unsigned version, int big_endian,
 580                                 enum bkey_invalid_flags flags)
 581 {
 582         struct jset_entry_clock *clock =
 583                 container_of(entry, struct jset_entry_clock, entry);
 584         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 585         int ret = 0;
 586
 587         if (journal_entry_err_on(bytes != sizeof(*clock),
 588                                  c, version, jset, entry,
 589                                  journal_entry_clock_bad_size,
 590                                  "bad size")) {
 591                 journal_entry_null_range(entry, vstruct_next(entry));
 592                 return ret;
 593         }
 594
 595         if (journal_entry_err_on(clock->rw > 1,
 596                                  c, version, jset, entry,
 597                                  journal_entry_clock_bad_rw,
 598                                  "bad rw")) {
 599                 journal_entry_null_range(entry, vstruct_next(entry));
 600                 return ret;
 601         }
 602
 603 fsck_err:
 604         return ret;
 605 }
 606
 607 static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 608                                         struct jset_entry *entry)
 609 {
 610         struct jset_entry_clock *clock =
 611                 container_of(entry, struct jset_entry_clock, entry);
 612
 613         prt_printf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
 614 }
 615
 616 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 617                                 struct jset *jset,
 618                                 struct jset_entry *entry,
 619                                 unsigned version, int big_endian,
 620                                 enum bkey_invalid_flags flags)
 621 {
 622         struct jset_entry_dev_usage *u =
 623                 container_of(entry, struct jset_entry_dev_usage, entry);
 624         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 625         unsigned expected = sizeof(*u);
 626         unsigned dev;
 627         int ret = 0;
 628
 629         if (journal_entry_err_on(bytes < expected,
 630                                  c, version, jset, entry,
 631                                  journal_entry_dev_usage_bad_size,
 632                                  "bad size (%u < %u)",
 633                                  bytes, expected)) {
 634                 journal_entry_null_range(entry, vstruct_next(entry));
 635                 return ret;
 636         }
 637
 638         dev = le32_to_cpu(u->dev);
 639
 640         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 641                                  c, version, jset, entry,
 642                                  journal_entry_dev_usage_bad_dev,
 643                                  "bad dev")) {
 644                 journal_entry_null_range(entry, vstruct_next(entry));
 645                 return ret;
 646         }
 647
 648         if (journal_entry_err_on(u->pad,
 649                                  c, version, jset, entry,
 650                                  journal_entry_dev_usage_bad_pad,
 651                                  "bad pad")) {
 652                 journal_entry_null_range(entry, vstruct_next(entry));
 653                 return ret;
 654         }
 655
 656 fsck_err:
 657         return ret;
 658 }
 659
 660 static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
 661                                             struct jset_entry *entry)
 662 {
 663         struct jset_entry_dev_usage *u =
 664                 container_of(entry, struct jset_entry_dev_usage, entry);
 665         unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 666
 667         prt_printf(out, "dev=%u", le32_to_cpu(u->dev));
 668
 669         for (i = 0; i < nr_types; i++) {
 670                 if (i < BCH_DATA_NR)
 671                         prt_printf(out, " %s", bch2_data_types[i]);
 672                 else
 673                         prt_printf(out, " (unknown data type %u)", i);
 674                 prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 675                        le64_to_cpu(u->d[i].buckets),
 676                        le64_to_cpu(u->d[i].sectors),
 677                        le64_to_cpu(u->d[i].fragmented));
 678         }
 679
 680         prt_printf(out, " buckets_ec: %llu", le64_to_cpu(u->buckets_ec));
 681 }
 682
 683 static int journal_entry_log_validate(struct bch_fs *c,
 684                                 struct jset *jset,
 685                                 struct jset_entry *entry,
 686                                 unsigned version, int big_endian,
 687                                 enum bkey_invalid_flags flags)
 688 {
 689         return 0;
 690 }
 691
 692 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 693                                       struct jset_entry *entry)
 694 {
 695         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 696         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 697
 698         prt_printf(out, "%.*s", bytes, l->d);
 699 }
 700
 701 static int journal_entry_overwrite_validate(struct bch_fs *c,
 702                                 struct jset *jset,
 703                                 struct jset_entry *entry,
 704                                 unsigned version, int big_endian,
 705                                 enum bkey_invalid_flags flags)
 706 {
 707         return journal_entry_btree_keys_validate(c, jset, entry,
 708                                 version, big_endian, READ);
 709 }
 710
 711 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
 712                                             struct jset_entry *entry)
 713 {
 714         journal_entry_btree_keys_to_text(out, c, entry);
 715 }
 716
 717 static int journal_entry_write_buffer_keys_validate(struct bch_fs *c,
 718                                 struct jset *jset,
 719                                 struct jset_entry *entry,
 720                                 unsigned version, int big_endian,
 721                                 enum bkey_invalid_flags flags)
 722 {
 723         return journal_entry_btree_keys_validate(c, jset, entry,
 724                                 version, big_endian, READ);
 725 }
 726
 727 static void journal_entry_write_buffer_keys_to_text(struct printbuf *out, struct bch_fs *c,
 728                                             struct jset_entry *entry)
 729 {
 730         journal_entry_btree_keys_to_text(out, c, entry);
 731 }
 732
 733 struct jset_entry_ops {
 734         int (*validate)(struct bch_fs *, struct jset *,
 735                         struct jset_entry *, unsigned, int,
 736                         enum bkey_invalid_flags);
 737         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 738 };
 739
 740 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 741 #define x(f, nr)                                                \
 742         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 743                 .validate       = journal_entry_##f##_validate, \
 744                 .to_text        = journal_entry_##f##_to_text,  \
 745         },
 746         BCH_JSET_ENTRY_TYPES()
 747 #undef x
 748 };
 749
 750 int bch2_journal_entry_validate(struct bch_fs *c,
 751                                 struct jset *jset,
 752                                 struct jset_entry *entry,
 753                                 unsigned version, int big_endian,
 754                                 enum bkey_invalid_flags flags)
 755 {
 756         return entry->type < BCH_JSET_ENTRY_NR
 757                 ? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
 758                                 version, big_endian, flags)
 759                 : 0;
 760 }
 761
 762 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 763                                 struct jset_entry *entry)
 764 {
 765         if (entry->type < BCH_JSET_ENTRY_NR) {
 766                 prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 767                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 768         } else {
 769                 prt_printf(out, "(unknown type %u)", entry->type);
 770         }
 771 }
 772
 773 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 774                                  enum bkey_invalid_flags flags)
 775 {
 776         struct jset_entry *entry;
 777         unsigned version = le32_to_cpu(jset->version);
 778         int ret = 0;
 779
 780         vstruct_for_each(jset, entry) {
 781                 if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
 782                                 c, version, jset, entry,
 783                                 journal_entry_past_jset_end,
 784                                 "journal entry extends past end of jset")) {
 785                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 786                         break;
 787                 }
 788
 789                 ret = bch2_journal_entry_validate(c, jset, entry,
 790                                         version, JSET_BIG_ENDIAN(jset), flags);
 791                 if (ret)
 792                         break;
 793         }
 794 fsck_err:
 795         return ret;
 796 }
 797
 798 static int jset_validate(struct bch_fs *c,
 799                          struct bch_dev *ca,
 800                          struct jset *jset, u64 sector,
 801                          enum bkey_invalid_flags flags)
 802 {
 803         unsigned version;
 804         int ret = 0;
 805
 806         if (le64_to_cpu(jset->magic) != jset_magic(c))
 807                 return JOURNAL_ENTRY_NONE;
 808
 809         version = le32_to_cpu(jset->version);
 810         if (journal_entry_err_on(!bch2_version_compatible(version),
 811                         c, version, jset, NULL,
 812                         jset_unsupported_version,
 813                         "%s sector %llu seq %llu: incompatible journal entry version %u.%u",
 814                         ca ? ca->name : c->name,
 815                         sector, le64_to_cpu(jset->seq),
 816                         BCH_VERSION_MAJOR(version),
 817                         BCH_VERSION_MINOR(version))) {
 818                 /* don't try to continue: */
 819                 return -EINVAL;
 820         }
 821
 822         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
 823                         c, version, jset, NULL,
 824                         jset_unknown_csum,
 825                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 826                         ca ? ca->name : c->name,
 827                         sector, le64_to_cpu(jset->seq),
 828                         JSET_CSUM_TYPE(jset)))
 829                 ret = JOURNAL_ENTRY_BAD;
 830
 831         /* last_seq is ignored when JSET_NO_FLUSH is true */
 832         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 833                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
 834                                  c, version, jset, NULL,
 835                                  jset_last_seq_newer_than_seq,
 836                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 837                                  le64_to_cpu(jset->last_seq),
 838                                  le64_to_cpu(jset->seq))) {
 839                 jset->last_seq = jset->seq;
 840                 return JOURNAL_ENTRY_BAD;
 841         }
 842
 843         ret = jset_validate_entries(c, jset, flags);
 844 fsck_err:
 845         return ret;
 846 }
 847
 848 static int jset_validate_early(struct bch_fs *c,
 849                          struct bch_dev *ca,
 850                          struct jset *jset, u64 sector,
 851                          unsigned bucket_sectors_left,
 852                          unsigned sectors_read)
 853 {
 854         size_t bytes = vstruct_bytes(jset);
 855         unsigned version;
 856         enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
 857         int ret = 0;
 858
 859         if (le64_to_cpu(jset->magic) != jset_magic(c))
 860                 return JOURNAL_ENTRY_NONE;
 861
 862         version = le32_to_cpu(jset->version);
 863         if (journal_entry_err_on(!bch2_version_compatible(version),
 864                         c, version, jset, NULL,
 865                         jset_unsupported_version,
 866                         "%s sector %llu seq %llu: unknown journal entry version %u.%u",
 867                         ca ? ca->name : c->name,
 868                         sector, le64_to_cpu(jset->seq),
 869                         BCH_VERSION_MAJOR(version),
 870                         BCH_VERSION_MINOR(version))) {
 871                 /* don't try to continue: */
 872                 return -EINVAL;
 873         }
 874
 875         if (bytes > (sectors_read << 9) &&
 876             sectors_read < bucket_sectors_left)
 877                 return JOURNAL_ENTRY_REREAD;
 878
 879         if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
 880                         c, version, jset, NULL,
 881                         jset_past_bucket_end,
 882                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 883                         ca ? ca->name : c->name,
 884                         sector, le64_to_cpu(jset->seq), bytes))
 885                 le32_add_cpu(&jset->u64s,
 886                              -((bytes - (bucket_sectors_left << 9)) / 8));
 887 fsck_err:
 888         return ret;
 889 }
 890
 891 struct journal_read_buf {
 892         void            *data;
 893         size_t          size;
 894 };
 895
 896 static int journal_read_buf_realloc(struct journal_read_buf *b,
 897                                     size_t new_size)
 898 {
 899         void *n;
 900
 901         /* the bios are sized for this many pages, max: */
 902         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 903                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 904
 905         new_size = roundup_pow_of_two(new_size);
 906         n = kvpmalloc(new_size, GFP_KERNEL);
 907         if (!n)
 908                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 909
 910         kvpfree(b->data, b->size);
 911         b->data = n;
 912         b->size = new_size;
 913         return 0;
 914 }
 915
 916 static int journal_read_bucket(struct bch_dev *ca,
 917                                struct journal_read_buf *buf,
 918                                struct journal_list *jlist,
 919                                unsigned bucket)
 920 {
 921         struct bch_fs *c = ca->fs;
 922         struct journal_device *ja = &ca->journal;
 923         struct jset *j = NULL;
 924         unsigned sectors, sectors_read = 0;
 925         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 926             end = offset + ca->mi.bucket_size;
 927         bool saw_bad = false, csum_good;
 928         int ret = 0;
 929
 930         pr_debug("reading %u", bucket);
 931
 932         while (offset < end) {
 933                 if (!sectors_read) {
 934                         struct bio *bio;
 935                         unsigned nr_bvecs;
 936 reread:
 937                         sectors_read = min_t(unsigned,
 938                                 end - offset, buf->size >> 9);
 939                         nr_bvecs = buf_pages(buf->data, sectors_read << 9);
 940
 941                         bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
 942                         bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
 943
 944                         bio->bi_iter.bi_sector = offset;
 945                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 946
 947                         ret = submit_bio_wait(bio);
 948                         kfree(bio);
 949
 950                         if (bch2_dev_io_err_on(ret, ca, BCH_MEMBER_ERROR_read,
 951                                                "journal read error: sector %llu",
 952                                                offset) ||
 953                             bch2_meta_read_fault("journal")) {
 954                                 /*
 955                                  * We don't error out of the recovery process
 956                                  * here, since the relevant journal entry may be
 957                                  * found on a different device, and missing or
 958                                  * no journal entries will be handled later
 959                                  */
 960                                 return 0;
 961                         }
 962
 963                         j = buf->data;
 964                 }
 965
 966                 ret = jset_validate_early(c, ca, j, offset,
 967                                     end - offset, sectors_read);
 968                 switch (ret) {
 969                 case 0:
 970                         sectors = vstruct_sectors(j, c->block_bits);
 971                         break;
 972                 case JOURNAL_ENTRY_REREAD:
 973                         if (vstruct_bytes(j) > buf->size) {
 974                                 ret = journal_read_buf_realloc(buf,
 975                                                         vstruct_bytes(j));
 976                                 if (ret)
 977                                         return ret;
 978                         }
 979                         goto reread;
 980                 case JOURNAL_ENTRY_NONE:
 981                         if (!saw_bad)
 982                                 return 0;
 983                         /*
 984                          * On checksum error we don't really trust the size
 985                          * field of the journal entry we read, so try reading
 986                          * again at next block boundary:
 987                          */
 988                         sectors = block_sectors(c);
 989                         goto next_block;
 990                 default:
 991                         return ret;
 992                 }
 993
 994                 /*
 995                  * This happens sometimes if we don't have discards on -
 996                  * when we've partially overwritten a bucket with new
 997                  * journal entries. We don't need the rest of the
 998                  * bucket:
 999                  */
1000                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
1001                         return 0;
1002
1003                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
1004
1005                 csum_good = jset_csum_good(c, j);
1006                 if (bch2_dev_io_err_on(!csum_good, ca, BCH_MEMBER_ERROR_checksum,
1007                                        "journal checksum error"))
1008                         saw_bad = true;
1009
1010                 ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
1011                              j->encrypted_start,
1012                              vstruct_end(j) - (void *) j->encrypted_start);
1013                 bch2_fs_fatal_err_on(ret, c,
1014                                 "error decrypting journal entry: %i", ret);
1015
1016                 mutex_lock(&jlist->lock);
1017                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
1018                                         .csum_good      = csum_good,
1019                                         .dev            = ca->dev_idx,
1020                                         .bucket         = bucket,
1021                                         .bucket_offset  = offset -
1022                                                 bucket_to_sector(ca, ja->buckets[bucket]),
1023                                         .sector         = offset,
1024                                         }, jlist, j);
1025                 mutex_unlock(&jlist->lock);
1026
1027                 switch (ret) {
1028                 case JOURNAL_ENTRY_ADD_OK:
1029                         break;
1030                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
1031                         break;
1032                 default:
1033                         return ret;
1034                 }
1035 next_block:
1036                 pr_debug("next");
1037                 offset          += sectors;
1038                 sectors_read    -= sectors;
1039                 j = ((void *) j) + (sectors << 9);
1040         }
1041
1042         return 0;
1043 }
1044
1045 static CLOSURE_CALLBACK(bch2_journal_read_device)
1046 {
1047         closure_type(ja, struct journal_device, read);
1048         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
1049         struct bch_fs *c = ca->fs;
1050         struct journal_list *jlist =
1051                 container_of(cl->parent, struct journal_list, cl);
1052         struct journal_replay *r, **_r;
1053         struct genradix_iter iter;
1054         struct journal_read_buf buf = { NULL, 0 };
1055         unsigned i;
1056         int ret = 0;
1057
1058         if (!ja->nr)
1059                 goto out;
1060
1061         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
1062         if (ret)
1063                 goto err;
1064
1065         pr_debug("%u journal buckets", ja->nr);
1066
1067         for (i = 0; i < ja->nr; i++) {
1068                 ret = journal_read_bucket(ca, &buf, jlist, i);
1069                 if (ret)
1070                         goto err;
1071         }
1072
1073         ja->sectors_free = ca->mi.bucket_size;
1074
1075         mutex_lock(&jlist->lock);
1076         genradix_for_each_reverse(&c->journal_entries, iter, _r) {
1077                 r = *_r;
1078
1079                 if (!r)
1080                         continue;
1081
1082                 for (i = 0; i < r->nr_ptrs; i++) {
1083                         if (r->ptrs[i].dev == ca->dev_idx) {
1084                                 unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
1085                                         vstruct_sectors(&r->j, c->block_bits);
1086
1087                                 ja->cur_idx = r->ptrs[i].bucket;
1088                                 ja->sectors_free = ca->mi.bucket_size - wrote;
1089                                 goto found;
1090                         }
1091                 }
1092         }
1093 found:
1094         mutex_unlock(&jlist->lock);
1095
1096         if (ja->bucket_seq[ja->cur_idx] &&
1097             ja->sectors_free == ca->mi.bucket_size) {
1098 #if 0
1099                 /*
1100                  * Debug code for ZNS support, where we (probably) want to be
1101                  * correlated where we stopped in the journal to the zone write
1102                  * points:
1103                  */
1104                 bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
1105                 bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
1106                 for (i = 0; i < 3; i++) {
1107                         unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
1108
1109                         bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
1110                 }
1111 #endif
1112                 ja->sectors_free = 0;
1113         }
1114
1115         /*
1116          * Set dirty_idx to indicate the entire journal is full and needs to be
1117          * reclaimed - journal reclaim will immediately reclaim whatever isn't
1118          * pinned when it first runs:
1119          */
1120         ja->discard_idx = ja->dirty_idx_ondisk =
1121                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
1122 out:
1123         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
1124         kvpfree(buf.data, buf.size);
1125         percpu_ref_put(&ca->io_ref);
1126         closure_return(cl);
1127         return;
1128 err:
1129         mutex_lock(&jlist->lock);
1130         jlist->ret = ret;
1131         mutex_unlock(&jlist->lock);
1132         goto out;
1133 }
1134
1135 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1136                                struct journal_replay *j)
1137 {
1138         unsigned i;
1139
1140         for (i = 0; i < j->nr_ptrs; i++) {
1141                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
1142                 u64 offset;
1143
1144                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
1145
1146                 if (i)
1147                         prt_printf(out, " ");
1148                 prt_printf(out, "%u:%u:%u (sector %llu)",
1149                        j->ptrs[i].dev,
1150                        j->ptrs[i].bucket,
1151                        j->ptrs[i].bucket_offset,
1152                        j->ptrs[i].sector);
1153         }
1154 }
1155
1156 int bch2_journal_read(struct bch_fs *c,
1157                       u64 *last_seq,
1158                       u64 *blacklist_seq,
1159                       u64 *start_seq)
1160 {
1161         struct journal_list jlist;
1162         struct journal_replay *i, **_i, *prev = NULL;
1163         struct genradix_iter radix_iter;
1164         struct bch_dev *ca;
1165         unsigned iter;
1166         struct printbuf buf = PRINTBUF;
1167         bool degraded = false, last_write_torn = false;
1168         u64 seq;
1169         int ret = 0;
1170
1171         closure_init_stack(&jlist.cl);
1172         mutex_init(&jlist.lock);
1173         jlist.last_seq = 0;
1174         jlist.ret = 0;
1175
1176         for_each_member_device(ca, c, iter) {
1177                 if (!c->opts.fsck &&
1178                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1179                         continue;
1180
1181                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1182                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1183                     percpu_ref_tryget(&ca->io_ref))
1184                         closure_call(&ca->journal.read,
1185                                      bch2_journal_read_device,
1186                                      system_unbound_wq,
1187                                      &jlist.cl);
1188                 else
1189                         degraded = true;
1190         }
1191
1192         closure_sync(&jlist.cl);
1193
1194         if (jlist.ret)
1195                 return jlist.ret;
1196
1197         *last_seq       = 0;
1198         *start_seq      = 0;
1199         *blacklist_seq  = 0;
1200
1201         /*
1202          * Find most recent flush entry, and ignore newer non flush entries -
1203          * those entries will be blacklisted:
1204          */
1205         genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
1206                 enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
1207
1208                 i = *_i;
1209
1210                 if (!i || i->ignore)
1211                         continue;
1212
1213                 if (!*start_seq)
1214                         *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
1215
1216                 if (JSET_NO_FLUSH(&i->j)) {
1217                         i->ignore = true;
1218                         continue;
1219                 }
1220
1221                 if (!last_write_torn && !i->csum_good) {
1222                         last_write_torn = true;
1223                         i->ignore = true;
1224                         continue;
1225                 }
1226
1227                 if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
1228                                          c, le32_to_cpu(i->j.version), &i->j, NULL,
1229                                          jset_last_seq_newer_than_seq,
1230                                          "invalid journal entry: last_seq > seq (%llu > %llu)",
1231                                          le64_to_cpu(i->j.last_seq),
1232                                          le64_to_cpu(i->j.seq)))
1233                         i->j.last_seq = i->j.seq;
1234
1235                 *last_seq       = le64_to_cpu(i->j.last_seq);
1236                 *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1237                 break;
1238         }
1239
1240         if (!*start_seq) {
1241                 bch_info(c, "journal read done, but no entries found");
1242                 return 0;
1243         }
1244
1245         if (!*last_seq) {
1246                 fsck_err(c, dirty_but_no_journal_entries_post_drop_nonflushes,
1247                          "journal read done, but no entries found after dropping non-flushes");
1248                 return 0;
1249         }
1250
1251         bch_info(c, "journal read done, replaying entries %llu-%llu",
1252                  *last_seq, *blacklist_seq - 1);
1253
1254         if (*start_seq != *blacklist_seq)
1255                 bch_info(c, "dropped unflushed entries %llu-%llu",
1256                          *blacklist_seq, *start_seq - 1);
1257
1258         /* Drop blacklisted entries and entries older than last_seq: */
1259         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1260                 i = *_i;
1261
1262                 if (!i || i->ignore)
1263                         continue;
1264
1265                 seq = le64_to_cpu(i->j.seq);
1266                 if (seq < *last_seq) {
1267                         journal_replay_free(c, i);
1268                         continue;
1269                 }
1270
1271                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1272                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1273                                     jset_seq_blacklisted,
1274                                     "found blacklisted journal entry %llu", seq);
1275                         i->ignore = true;
1276                 }
1277         }
1278
1279         /* Check for missing entries: */
1280         seq = *last_seq;
1281         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1282                 i = *_i;
1283
1284                 if (!i || i->ignore)
1285                         continue;
1286
1287                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1288
1289                 while (seq < le64_to_cpu(i->j.seq)) {
1290                         u64 missing_start, missing_end;
1291                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1292
1293                         while (seq < le64_to_cpu(i->j.seq) &&
1294                                bch2_journal_seq_is_blacklisted(c, seq, false))
1295                                 seq++;
1296
1297                         if (seq == le64_to_cpu(i->j.seq))
1298                                 break;
1299
1300                         missing_start = seq;
1301
1302                         while (seq < le64_to_cpu(i->j.seq) &&
1303                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1304                                 seq++;
1305
1306                         if (prev) {
1307                                 bch2_journal_ptrs_to_text(&buf1, c, prev);
1308                                 prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits));
1309                         } else
1310                                 prt_printf(&buf1, "(none)");
1311                         bch2_journal_ptrs_to_text(&buf2, c, i);
1312
1313                         missing_end = seq - 1;
1314                         fsck_err(c, journal_entries_missing,
1315                                  "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1316                                  "  prev at %s\n"
1317                                  "  next at %s",
1318                                  missing_start, missing_end,
1319                                  *last_seq, *blacklist_seq - 1,
1320                                  buf1.buf, buf2.buf);
1321
1322                         printbuf_exit(&buf1);
1323                         printbuf_exit(&buf2);
1324                 }
1325
1326                 prev = i;
1327                 seq++;
1328         }
1329
1330         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1331                 struct bch_replicas_padded replicas = {
1332                         .e.data_type = BCH_DATA_journal,
1333                         .e.nr_required = 1,
1334                 };
1335                 unsigned ptr;
1336
1337                 i = *_i;
1338                 if (!i || i->ignore)
1339                         continue;
1340
1341                 for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
1342                         ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
1343
1344                         if (!i->ptrs[ptr].csum_good)
1345                                 bch_err_dev_offset(ca, i->ptrs[ptr].sector,
1346                                                    "invalid journal checksum, seq %llu%s",
1347                                                    le64_to_cpu(i->j.seq),
1348                                                    i->csum_good ? " (had good copy on another device)" : "");
1349                 }
1350
1351                 ret = jset_validate(c,
1352                                     bch_dev_bkey_exists(c, i->ptrs[0].dev),
1353                                     &i->j,
1354                                     i->ptrs[0].sector,
1355                                     READ);
1356                 if (ret)
1357                         goto err;
1358
1359                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1360                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1361
1362                 bch2_replicas_entry_sort(&replicas.e);
1363
1364                 printbuf_reset(&buf);
1365                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1366
1367                 if (!degraded &&
1368                     !bch2_replicas_marked(c, &replicas.e) &&
1369                     (le64_to_cpu(i->j.seq) == *last_seq ||
1370                      fsck_err(c, journal_entry_replicas_not_marked,
1371                               "superblock not marked as containing replicas for journal entry %llu\n  %s",
1372                               le64_to_cpu(i->j.seq), buf.buf))) {
1373                         ret = bch2_mark_replicas(c, &replicas.e);
1374                         if (ret)
1375                                 goto err;
1376                 }
1377         }
1378 err:
1379 fsck_err:
1380         printbuf_exit(&buf);
1381         return ret;
1382 }
1383
1384 /* journal write: */
1385
1386 static void __journal_write_alloc(struct journal *j,
1387                                   struct journal_buf *w,
1388                                   struct dev_alloc_list *devs_sorted,
1389                                   unsigned sectors,
1390                                   unsigned *replicas,
1391                                   unsigned replicas_want)
1392 {
1393         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1394         struct journal_device *ja;
1395         struct bch_dev *ca;
1396         unsigned i;
1397
1398         if (*replicas >= replicas_want)
1399                 return;
1400
1401         for (i = 0; i < devs_sorted->nr; i++) {
1402                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1403                 if (!ca)
1404                         continue;
1405
1406                 ja = &ca->journal;
1407
1408                 /*
1409                  * Check that we can use this device, and aren't already using
1410                  * it:
1411                  */
1412                 if (!ca->mi.durability ||
1413                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1414                     !ja->nr ||
1415                     bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) ||
1416                     sectors > ja->sectors_free)
1417                         continue;
1418
1419                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1420
1421                 bch2_bkey_append_ptr(&w->key,
1422                         (struct bch_extent_ptr) {
1423                                   .offset = bucket_to_sector(ca,
1424                                         ja->buckets[ja->cur_idx]) +
1425                                         ca->mi.bucket_size -
1426                                         ja->sectors_free,
1427                                   .dev = ca->dev_idx,
1428                 });
1429
1430                 ja->sectors_free -= sectors;
1431                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1432
1433                 *replicas += ca->mi.durability;
1434
1435                 if (*replicas >= replicas_want)
1436                         break;
1437         }
1438 }
1439
1440 /**
1441  * journal_write_alloc - decide where to write next journal entry
1442  *
1443  * @j:          journal object
1444  * @w:          journal buf (entry to be written)
1445  *
1446  * Returns: 0 on success, or -EROFS on failure
1447  */
1448 static int journal_write_alloc(struct journal *j, struct journal_buf *w)
1449 {
1450         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1451         struct bch_devs_mask devs;
1452         struct journal_device *ja;
1453         struct bch_dev *ca;
1454         struct dev_alloc_list devs_sorted;
1455         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1456         unsigned target = c->opts.metadata_target ?:
1457                 c->opts.foreground_target;
1458         unsigned i, replicas = 0, replicas_want =
1459                 READ_ONCE(c->opts.metadata_replicas);
1460
1461         rcu_read_lock();
1462 retry:
1463         devs = target_rw_devs(c, BCH_DATA_journal, target);
1464
1465         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1466
1467         __journal_write_alloc(j, w, &devs_sorted,
1468                               sectors, &replicas, replicas_want);
1469
1470         if (replicas >= replicas_want)
1471                 goto done;
1472
1473         for (i = 0; i < devs_sorted.nr; i++) {
1474                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1475                 if (!ca)
1476                         continue;
1477
1478                 ja = &ca->journal;
1479
1480                 if (sectors > ja->sectors_free &&
1481                     sectors <= ca->mi.bucket_size &&
1482                     bch2_journal_dev_buckets_available(j, ja,
1483                                         journal_space_discarded)) {
1484                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1485                         ja->sectors_free = ca->mi.bucket_size;
1486
1487                         /*
1488                          * ja->bucket_seq[ja->cur_idx] must always have
1489                          * something sensible:
1490                          */
1491                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1492                 }
1493         }
1494
1495         __journal_write_alloc(j, w, &devs_sorted,
1496                               sectors, &replicas, replicas_want);
1497
1498         if (replicas < replicas_want && target) {
1499                 /* Retry from all devices: */
1500                 target = 0;
1501                 goto retry;
1502         }
1503 done:
1504         rcu_read_unlock();
1505
1506         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1507
1508         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1509 }
1510
1511 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1512 {
1513         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1514
1515         /* we aren't holding j->lock: */
1516         unsigned new_size = READ_ONCE(j->buf_size_want);
1517         void *new_buf;
1518
1519         if (buf->buf_size >= new_size)
1520                 return;
1521
1522         size_t btree_write_buffer_size = new_size / 64;
1523
1524         if (bch2_btree_write_buffer_resize(c, btree_write_buffer_size))
1525                 return;
1526
1527         new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
1528         if (!new_buf)
1529                 return;
1530
1531         memcpy(new_buf, buf->data, buf->buf_size);
1532
1533         spin_lock(&j->lock);
1534         swap(buf->data,         new_buf);
1535         swap(buf->buf_size,     new_size);
1536         spin_unlock(&j->lock);
1537
1538         kvpfree(new_buf, new_size);
1539 }
1540
1541 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1542 {
1543         return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
1544 }
1545
1546 static CLOSURE_CALLBACK(journal_write_done)
1547 {
1548         closure_type(j, struct journal, io);
1549         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1550         struct journal_buf *w = journal_last_unwritten_buf(j);
1551         struct bch_replicas_padded replicas;
1552         union journal_res_state old, new;
1553         u64 v, seq;
1554         int err = 0;
1555
1556         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1557                                ? j->flush_write_time
1558                                : j->noflush_write_time, j->write_start_time);
1559
1560         if (!w->devs_written.nr) {
1561                 bch_err(c, "unable to write journal to sufficient devices");
1562                 err = -EIO;
1563         } else {
1564                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1565                                          w->devs_written);
1566                 if (bch2_mark_replicas(c, &replicas.e))
1567                         err = -EIO;
1568         }
1569
1570         if (err)
1571                 bch2_fatal_error(c);
1572
1573         spin_lock(&j->lock);
1574         seq = le64_to_cpu(w->data->seq);
1575
1576         if (seq >= j->pin.front)
1577                 journal_seq_pin(j, seq)->devs = w->devs_written;
1578
1579         if (!err) {
1580                 if (!JSET_NO_FLUSH(w->data)) {
1581                         j->flushed_seq_ondisk = seq;
1582                         j->last_seq_ondisk = w->last_seq;
1583
1584                         bch2_do_discards(c);
1585                         closure_wake_up(&c->freelist_wait);
1586
1587                         bch2_reset_alloc_cursors(c);
1588                 }
1589         } else if (!j->err_seq || seq < j->err_seq)
1590                 j->err_seq      = seq;
1591
1592         j->seq_ondisk           = seq;
1593
1594         /*
1595          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1596          * more buckets:
1597          *
1598          * Must come before signaling write completion, for
1599          * bch2_fs_journal_stop():
1600          */
1601         if (j->watermark != BCH_WATERMARK_stripe)
1602                 journal_reclaim_kick(&c->journal);
1603
1604         /* also must come before signalling write completion: */
1605         closure_debug_destroy(cl);
1606
1607         v = atomic64_read(&j->reservations.counter);
1608         do {
1609                 old.v = new.v = v;
1610                 BUG_ON(journal_state_count(new, new.unwritten_idx));
1611
1612                 new.unwritten_idx++;
1613         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1614                                        old.v, new.v)) != old.v);
1615
1616         bch2_journal_reclaim_fast(j);
1617         bch2_journal_space_available(j);
1618
1619         track_event_change(&c->times[BCH_TIME_blocked_journal_max_in_flight],
1620                            &j->max_in_flight_start, false);
1621
1622         closure_wake_up(&w->wait);
1623         journal_wake(j);
1624
1625         if (!journal_state_count(new, new.unwritten_idx) &&
1626             journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
1627                 spin_unlock(&j->lock);
1628                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1629         } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
1630                    new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
1631                 struct journal_buf *buf = journal_cur_buf(j);
1632                 long delta = buf->expires - jiffies;
1633
1634                 /*
1635                  * We don't close a journal entry to write it while there's
1636                  * previous entries still in flight - the current journal entry
1637                  * might want to be written now:
1638                  */
1639
1640                 spin_unlock(&j->lock);
1641                 mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
1642         } else {
1643                 spin_unlock(&j->lock);
1644         }
1645 }
1646
1647 static void journal_write_endio(struct bio *bio)
1648 {
1649         struct bch_dev *ca = bio->bi_private;
1650         struct journal *j = &ca->fs->journal;
1651         struct journal_buf *w = journal_last_unwritten_buf(j);
1652         unsigned long flags;
1653
1654         if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1655                                "error writing journal entry %llu: %s",
1656                                le64_to_cpu(w->data->seq),
1657                                bch2_blk_status_to_str(bio->bi_status)) ||
1658             bch2_meta_write_fault("journal")) {
1659                 spin_lock_irqsave(&j->err_lock, flags);
1660                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1661                 spin_unlock_irqrestore(&j->err_lock, flags);
1662         }
1663
1664         closure_put(&j->io);
1665         percpu_ref_put(&ca->io_ref);
1666 }
1667
1668 static CLOSURE_CALLBACK(do_journal_write)
1669 {
1670         closure_type(j, struct journal, io);
1671         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1672         struct bch_dev *ca;
1673         struct journal_buf *w = journal_last_unwritten_buf(j);
1674         struct bch_extent_ptr *ptr;
1675         struct bio *bio;
1676         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1677
1678         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1679                 ca = bch_dev_bkey_exists(c, ptr->dev);
1680                 if (!percpu_ref_tryget(&ca->io_ref)) {
1681                         /* XXX: fix this */
1682                         bch_err(c, "missing device for journal write\n");
1683                         continue;
1684                 }
1685
1686                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1687                              sectors);
1688
1689                 bio = ca->journal.bio;
1690                 bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
1691                 bio->bi_iter.bi_sector  = ptr->offset;
1692                 bio->bi_end_io          = journal_write_endio;
1693                 bio->bi_private         = ca;
1694
1695                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1696                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1697
1698                 if (!JSET_NO_FLUSH(w->data))
1699                         bio->bi_opf    |= REQ_FUA;
1700                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1701                         bio->bi_opf    |= REQ_PREFLUSH;
1702
1703                 bch2_bio_map(bio, w->data, sectors << 9);
1704
1705                 trace_and_count(c, journal_write, bio);
1706                 closure_bio_submit(bio, cl);
1707
1708                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1709                         le64_to_cpu(w->data->seq);
1710         }
1711
1712         continue_at(cl, journal_write_done, c->io_complete_wq);
1713 }
1714
1715 static int bch2_journal_write_prep(struct journal *j, struct journal_buf *w)
1716 {
1717         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1718         struct jset_entry *start, *end, *i, *next, *prev = NULL;
1719         struct jset *jset = w->data;
1720         struct journal_keys_to_wb wb = { NULL };
1721         unsigned sectors, bytes, u64s;
1722         unsigned long btree_roots_have = 0;
1723         bool validate_before_checksum = false;
1724         u64 seq = le64_to_cpu(jset->seq);
1725         int ret;
1726
1727         /*
1728          * Simple compaction, dropping empty jset_entries (from journal
1729          * reservations that weren't fully used) and merging jset_entries that
1730          * can be.
1731          *
1732          * If we wanted to be really fancy here, we could sort all the keys in
1733          * the jset and drop keys that were overwritten - probably not worth it:
1734          */
1735         vstruct_for_each_safe(jset, i, next) {
1736                 unsigned u64s = le16_to_cpu(i->u64s);
1737
1738                 /* Empty entry: */
1739                 if (!u64s)
1740                         continue;
1741
1742                 /*
1743                  * New btree roots are set by journalling them; when the journal
1744                  * entry gets written we have to propagate them to
1745                  * c->btree_roots
1746                  *
1747                  * But, every journal entry we write has to contain all the
1748                  * btree roots (at least for now); so after we copy btree roots
1749                  * to c->btree_roots we have to get any missing btree roots and
1750                  * add them to this journal entry:
1751                  */
1752                 switch (i->type) {
1753                 case BCH_JSET_ENTRY_btree_root:
1754                         bch2_journal_entry_to_btree_root(c, i);
1755                         __set_bit(i->btree_id, &btree_roots_have);
1756                         break;
1757                 case BCH_JSET_ENTRY_write_buffer_keys:
1758                         EBUG_ON(!w->need_flush_to_write_buffer);
1759
1760                         if (!wb.wb)
1761                                 bch2_journal_keys_to_write_buffer_start(c, &wb, seq);
1762
1763                         struct bkey_i *k;
1764                         jset_entry_for_each_key(i, k) {
1765                                 ret = bch2_journal_key_to_wb(c, &wb, i->btree_id, k);
1766                                 if (ret) {
1767                                         bch2_fs_fatal_error(c, "-ENOMEM flushing journal keys to btree write buffer");
1768                                         bch2_journal_keys_to_write_buffer_end(c, &wb);
1769                                         return ret;
1770                                 }
1771                         }
1772                         i->type = BCH_JSET_ENTRY_btree_keys;
1773                         break;
1774                 }
1775
1776                 /* Can we merge with previous entry? */
1777                 if (prev &&
1778                     i->btree_id == prev->btree_id &&
1779                     i->level    == prev->level &&
1780                     i->type     == prev->type &&
1781                     i->type     == BCH_JSET_ENTRY_btree_keys &&
1782                     le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
1783                         memmove_u64s_down(vstruct_next(prev),
1784                                           i->_data,
1785                                           u64s);
1786                         le16_add_cpu(&prev->u64s, u64s);
1787                         continue;
1788                 }
1789
1790                 /* Couldn't merge, move i into new position (after prev): */
1791                 prev = prev ? vstruct_next(prev) : jset->start;
1792                 if (i != prev)
1793                         memmove_u64s_down(prev, i, jset_u64s(u64s));
1794         }
1795
1796         if (wb.wb)
1797                 bch2_journal_keys_to_write_buffer_end(c, &wb);
1798         w->need_flush_to_write_buffer = false;
1799
1800         prev = prev ? vstruct_next(prev) : jset->start;
1801         jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
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 bch_dev *ca;
1909         struct journal_buf *w = journal_last_unwritten_buf(j);
1910         struct bch_replicas_padded replicas;
1911         struct bio *bio;
1912         struct printbuf journal_debug_buf = PRINTBUF;
1913         unsigned i, nr_rw_members = 0;
1914         int ret;
1915
1916         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1917
1918         j->write_start_time = local_clock();
1919
1920         spin_lock(&j->lock);
1921         ret = bch2_journal_write_pick_flush(j, w);
1922         spin_unlock(&j->lock);
1923         if (ret)
1924                 goto err;
1925
1926         mutex_lock(&j->buf_lock);
1927         journal_buf_realloc(j, w);
1928
1929         ret = bch2_journal_write_prep(j, w);
1930         mutex_unlock(&j->buf_lock);
1931         if (ret)
1932                 goto err;
1933
1934         j->entry_bytes_written += vstruct_bytes(w->data);
1935
1936         while (1) {
1937                 spin_lock(&j->lock);
1938                 ret = journal_write_alloc(j, w);
1939                 if (!ret || !j->can_discard)
1940                         break;
1941
1942                 spin_unlock(&j->lock);
1943                 bch2_journal_do_discards(j);
1944         }
1945
1946         if (ret) {
1947                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1948                 spin_unlock(&j->lock);
1949                 bch_err(c, "Unable to allocate journal write:\n%s",
1950                         journal_debug_buf.buf);
1951                 printbuf_exit(&journal_debug_buf);
1952                 goto err;
1953         }
1954
1955         /*
1956          * write is allocated, no longer need to account for it in
1957          * bch2_journal_space_available():
1958          */
1959         w->sectors = 0;
1960
1961         /*
1962          * journal entry has been compacted and allocated, recalculate space
1963          * available:
1964          */
1965         bch2_journal_space_available(j);
1966         spin_unlock(&j->lock);
1967
1968         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1969
1970         if (c->opts.nochanges)
1971                 goto no_io;
1972
1973         for_each_rw_member(ca, c, i)
1974                 nr_rw_members++;
1975
1976         if (nr_rw_members > 1)
1977                 w->separate_flush = true;
1978
1979         /*
1980          * Mark journal replicas before we submit the write to guarantee
1981          * recovery will find the journal entries after a crash.
1982          */
1983         bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1984                                  w->devs_written);
1985         ret = bch2_mark_replicas(c, &replicas.e);
1986         if (ret)
1987                 goto err;
1988
1989         if (!JSET_NO_FLUSH(w->data) && w->separate_flush) {
1990                 for_each_rw_member(ca, c, i) {
1991                         percpu_ref_get(&ca->io_ref);
1992
1993                         bio = ca->journal.bio;
1994                         bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
1995                         bio->bi_end_io          = journal_write_endio;
1996                         bio->bi_private         = ca;
1997                         closure_bio_submit(bio, cl);
1998                 }
1999         }
2000
2001         continue_at(cl, do_journal_write, c->io_complete_wq);
2002         return;
2003 no_io:
2004         continue_at(cl, journal_write_done, c->io_complete_wq);
2005         return;
2006 err:
2007         bch2_fatal_error(c);
2008         continue_at(cl, journal_write_done, c->io_complete_wq);
2009 }