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 "io.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 "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, "found duplicate but non identical journal entries (seq %llu)",
 144                          le64_to_cpu(j->seq));
 145                 i = dup;
 146                 goto found;
 147         }
 148 replace:
 149         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 150         if (!i)
 151                 return -BCH_ERR_ENOMEM_journal_entry_add;
 152
 153         i->nr_ptrs      = 0;
 154         i->csum_good    = entry_ptr.csum_good;
 155         i->ignore       = false;
 156         unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
 157         i->ptrs[i->nr_ptrs++] = entry_ptr;
 158
 159         if (dup) {
 160                 if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) {
 161                         bch_err(c, "found too many copies of journal entry %llu",
 162                                 le64_to_cpu(i->j.seq));
 163                         dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1;
 164                 }
 165
 166                 /* The first ptr should represent the jset we kept: */
 167                 memcpy(i->ptrs + i->nr_ptrs,
 168                        dup->ptrs,
 169                        sizeof(dup->ptrs[0]) * dup->nr_ptrs);
 170                 i->nr_ptrs += dup->nr_ptrs;
 171                 __journal_replay_free(c, dup);
 172         }
 173
 174         *_i = i;
 175         return 0;
 176 found:
 177         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 178                 if (ptr->dev == ca->dev_idx) {
 179                         bch_err(c, "duplicate journal entry %llu on same device",
 180                                 le64_to_cpu(i->j.seq));
 181                         goto out;
 182                 }
 183         }
 184
 185         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 186                 bch_err(c, "found too many copies of journal entry %llu",
 187                         le64_to_cpu(i->j.seq));
 188                 goto out;
 189         }
 190
 191         i->ptrs[i->nr_ptrs++] = entry_ptr;
 192 out:
 193 fsck_err:
 194         return ret;
 195 }
 196
 197 /* this fills in a range with empty jset_entries: */
 198 static void journal_entry_null_range(void *start, void *end)
 199 {
 200         struct jset_entry *entry;
 201
 202         for (entry = start; entry != end; entry = vstruct_next(entry))
 203                 memset(entry, 0, sizeof(*entry));
 204 }
 205
 206 #define JOURNAL_ENTRY_REREAD    5
 207 #define JOURNAL_ENTRY_NONE      6
 208 #define JOURNAL_ENTRY_BAD       7
 209
 210 static void journal_entry_err_msg(struct printbuf *out,
 211                                   struct jset *jset,
 212                                   struct jset_entry *entry)
 213 {
 214         prt_str(out, "invalid journal entry ");
 215         if (entry)
 216                 prt_printf(out, "%s ", bch2_jset_entry_types[entry->type]);
 217
 218         if (!jset)
 219                 prt_printf(out, "in superblock");
 220         else if (!entry)
 221                 prt_printf(out, "at seq %llu", le64_to_cpu(jset->seq));
 222         else
 223                 prt_printf(out, "at offset %zi/%u seq %llu",
 224                            (u64 *) entry - jset->_data,
 225                            le32_to_cpu(jset->u64s),
 226                            le64_to_cpu(jset->seq));
 227         prt_str(out, ": ");
 228 }
 229
 230 #define journal_entry_err(c, jset, entry, msg, ...)                     \
 231 ({                                                                      \
 232         struct printbuf buf = PRINTBUF;                                 \
 233                                                                         \
 234         journal_entry_err_msg(&buf, jset, entry);                       \
 235         prt_printf(&buf, msg, ##__VA_ARGS__);                           \
 236                                                                         \
 237         switch (write) {                                                \
 238         case READ:                                                      \
 239                 mustfix_fsck_err(c, "%s", buf.buf);                     \
 240                 break;                                                  \
 241         case WRITE:                                                     \
 242                 bch_err(c, "corrupt metadata before write: %s\n", buf.buf);\
 243                 if (bch2_fs_inconsistent(c)) {                          \
 244                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 245                         goto fsck_err;                                  \
 246                 }                                                       \
 247                 break;                                                  \
 248         }                                                               \
 249                                                                         \
 250         printbuf_exit(&buf);                                            \
 251         true;                                                           \
 252 })
 253
 254 #define journal_entry_err_on(cond, c, jset, entry, msg, ...)            \
 255         ((cond) ? journal_entry_err(c, jset, entry, msg, ##__VA_ARGS__) : false)
 256
 257 #define FSCK_DELETED_KEY        5
 258
 259 static int journal_validate_key(struct bch_fs *c,
 260                                 struct jset *jset,
 261                                 struct jset_entry *entry,
 262                                 unsigned level, enum btree_id btree_id,
 263                                 struct bkey_i *k,
 264                                 unsigned version, int big_endian, int write)
 265 {
 266         void *next = vstruct_next(entry);
 267         struct printbuf buf = PRINTBUF;
 268         int ret = 0;
 269
 270         if (journal_entry_err_on(!k->k.u64s, c, jset, entry, "k->u64s 0")) {
 271                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 272                 journal_entry_null_range(vstruct_next(entry), next);
 273                 return FSCK_DELETED_KEY;
 274         }
 275
 276         if (journal_entry_err_on((void *) bkey_next(k) >
 277                                  (void *) vstruct_next(entry),
 278                                  c, jset, entry,
 279                                  "extends past end of journal entry")) {
 280                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 281                 journal_entry_null_range(vstruct_next(entry), next);
 282                 return FSCK_DELETED_KEY;
 283         }
 284
 285         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
 286                                  c, jset, entry,
 287                                  "bad format %u", k->k.format)) {
 288                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 289                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 290                 journal_entry_null_range(vstruct_next(entry), next);
 291                 return FSCK_DELETED_KEY;
 292         }
 293
 294         if (!write)
 295                 bch2_bkey_compat(level, btree_id, version, big_endian,
 296                                  write, NULL, bkey_to_packed(k));
 297
 298         if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 299                               __btree_node_type(level, btree_id), write, &buf)) {
 300                 printbuf_reset(&buf);
 301                 prt_printf(&buf, "invalid journal entry %s at offset %zi/%u seq %llu:",
 302                            bch2_jset_entry_types[entry->type],
 303                            (u64 *) entry - jset->_data,
 304                            le32_to_cpu(jset->u64s),
 305                            le64_to_cpu(jset->seq));
 306                 prt_newline(&buf);
 307                 printbuf_indent_add(&buf, 2);
 308
 309                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
 310                 prt_newline(&buf);
 311                 bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 312                                   __btree_node_type(level, btree_id), write, &buf);
 313
 314                 mustfix_fsck_err(c, "%s", buf.buf);
 315
 316                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 317                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 318                 journal_entry_null_range(vstruct_next(entry), next);
 319
 320                 printbuf_exit(&buf);
 321                 return FSCK_DELETED_KEY;
 322         }
 323
 324         if (write)
 325                 bch2_bkey_compat(level, btree_id, version, big_endian,
 326                                  write, NULL, bkey_to_packed(k));
 327 fsck_err:
 328         printbuf_exit(&buf);
 329         return ret;
 330 }
 331
 332 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 333                                              struct jset *jset,
 334                                              struct jset_entry *entry,
 335                                              unsigned version, int big_endian, int write)
 336 {
 337         struct bkey_i *k = entry->start;
 338
 339         while (k != vstruct_last(entry)) {
 340                 int ret = journal_validate_key(c, jset, entry,
 341                                                entry->level,
 342                                                entry->btree_id,
 343                                                k, version, big_endian,
 344                                                write|BKEY_INVALID_JOURNAL);
 345                 if (ret == FSCK_DELETED_KEY)
 346                         continue;
 347
 348                 k = bkey_next(k);
 349         }
 350
 351         return 0;
 352 }
 353
 354 static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
 355                                              struct jset_entry *entry)
 356 {
 357         struct bkey_i *k;
 358         bool first = true;
 359
 360         jset_entry_for_each_key(entry, k) {
 361                 if (!first) {
 362                         prt_newline(out);
 363                         prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 364                 }
 365                 prt_printf(out, "btree=%s l=%u ", bch2_btree_ids[entry->btree_id], entry->level);
 366                 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 367                 first = false;
 368         }
 369 }
 370
 371 static int journal_entry_btree_root_validate(struct bch_fs *c,
 372                                              struct jset *jset,
 373                                              struct jset_entry *entry,
 374                                              unsigned version, int big_endian, int write)
 375 {
 376         struct bkey_i *k = entry->start;
 377         int ret = 0;
 378
 379         if (journal_entry_err_on(!entry->u64s ||
 380                                  le16_to_cpu(entry->u64s) != k->k.u64s,
 381                                  c, jset, entry,
 382                                  "invalid btree root journal entry: wrong number of keys")) {
 383                 void *next = vstruct_next(entry);
 384                 /*
 385                  * we don't want to null out this jset_entry,
 386                  * just the contents, so that later we can tell
 387                  * we were _supposed_ to have a btree root
 388                  */
 389                 entry->u64s = 0;
 390                 journal_entry_null_range(vstruct_next(entry), next);
 391                 return 0;
 392         }
 393
 394         return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
 395                                     version, big_endian, write);
 396 fsck_err:
 397         return ret;
 398 }
 399
 400 static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
 401                                              struct jset_entry *entry)
 402 {
 403         journal_entry_btree_keys_to_text(out, c, entry);
 404 }
 405
 406 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 407                                             struct jset *jset,
 408                                             struct jset_entry *entry,
 409                                             unsigned version, int big_endian, int write)
 410 {
 411         /* obsolete, don't care: */
 412         return 0;
 413 }
 414
 415 static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 416                                             struct jset_entry *entry)
 417 {
 418 }
 419
 420 static int journal_entry_blacklist_validate(struct bch_fs *c,
 421                                             struct jset *jset,
 422                                             struct jset_entry *entry,
 423                                             unsigned version, int big_endian, int write)
 424 {
 425         int ret = 0;
 426
 427         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1,
 428                                  c, jset, entry,
 429                 "invalid journal seq blacklist entry: bad size")) {
 430                 journal_entry_null_range(entry, vstruct_next(entry));
 431         }
 432 fsck_err:
 433         return ret;
 434 }
 435
 436 static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
 437                                             struct jset_entry *entry)
 438 {
 439         struct jset_entry_blacklist *bl =
 440                 container_of(entry, struct jset_entry_blacklist, entry);
 441
 442         prt_printf(out, "seq=%llu", le64_to_cpu(bl->seq));
 443 }
 444
 445 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 446                                                struct jset *jset,
 447                                                struct jset_entry *entry,
 448                                                unsigned version, int big_endian, int write)
 449 {
 450         struct jset_entry_blacklist_v2 *bl_entry;
 451         int ret = 0;
 452
 453         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2,
 454                                  c, jset, entry,
 455                 "invalid journal seq blacklist entry: bad size")) {
 456                 journal_entry_null_range(entry, vstruct_next(entry));
 457                 goto out;
 458         }
 459
 460         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 461
 462         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 463                                  le64_to_cpu(bl_entry->end),
 464                                  c, jset, entry,
 465                 "invalid journal seq blacklist entry: start > end")) {
 466                 journal_entry_null_range(entry, vstruct_next(entry));
 467         }
 468 out:
 469 fsck_err:
 470         return ret;
 471 }
 472
 473 static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
 474                                                struct jset_entry *entry)
 475 {
 476         struct jset_entry_blacklist_v2 *bl =
 477                 container_of(entry, struct jset_entry_blacklist_v2, entry);
 478
 479         prt_printf(out, "start=%llu end=%llu",
 480                le64_to_cpu(bl->start),
 481                le64_to_cpu(bl->end));
 482 }
 483
 484 static int journal_entry_usage_validate(struct bch_fs *c,
 485                                         struct jset *jset,
 486                                         struct jset_entry *entry,
 487                                         unsigned version, int big_endian, int write)
 488 {
 489         struct jset_entry_usage *u =
 490                 container_of(entry, struct jset_entry_usage, entry);
 491         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 492         int ret = 0;
 493
 494         if (journal_entry_err_on(bytes < sizeof(*u),
 495                                  c, jset, entry,
 496                                  "invalid journal entry usage: bad size")) {
 497                 journal_entry_null_range(entry, vstruct_next(entry));
 498                 return ret;
 499         }
 500
 501 fsck_err:
 502         return ret;
 503 }
 504
 505 static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 506                                         struct jset_entry *entry)
 507 {
 508         struct jset_entry_usage *u =
 509                 container_of(entry, struct jset_entry_usage, entry);
 510
 511         prt_printf(out, "type=%s v=%llu",
 512                bch2_fs_usage_types[u->entry.btree_id],
 513                le64_to_cpu(u->v));
 514 }
 515
 516 static int journal_entry_data_usage_validate(struct bch_fs *c,
 517                                         struct jset *jset,
 518                                         struct jset_entry *entry,
 519                                         unsigned version, int big_endian, int write)
 520 {
 521         struct jset_entry_data_usage *u =
 522                 container_of(entry, struct jset_entry_data_usage, entry);
 523         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 524         int ret = 0;
 525
 526         if (journal_entry_err_on(bytes < sizeof(*u) ||
 527                                  bytes < sizeof(*u) + u->r.nr_devs,
 528                                  c, jset, entry,
 529                                  "invalid journal entry usage: bad size")) {
 530                 journal_entry_null_range(entry, vstruct_next(entry));
 531                 return ret;
 532         }
 533
 534 fsck_err:
 535         return ret;
 536 }
 537
 538 static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
 539                                              struct jset_entry *entry)
 540 {
 541         struct jset_entry_data_usage *u =
 542                 container_of(entry, struct jset_entry_data_usage, entry);
 543
 544         bch2_replicas_entry_to_text(out, &u->r);
 545         prt_printf(out, "=%llu", le64_to_cpu(u->v));
 546 }
 547
 548 static int journal_entry_clock_validate(struct bch_fs *c,
 549                                         struct jset *jset,
 550                                         struct jset_entry *entry,
 551                                         unsigned version, int big_endian, int write)
 552 {
 553         struct jset_entry_clock *clock =
 554                 container_of(entry, struct jset_entry_clock, entry);
 555         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 556         int ret = 0;
 557
 558         if (journal_entry_err_on(bytes != sizeof(*clock),
 559                                  c, jset, entry, "bad size")) {
 560                 journal_entry_null_range(entry, vstruct_next(entry));
 561                 return ret;
 562         }
 563
 564         if (journal_entry_err_on(clock->rw > 1,
 565                                  c, jset, entry, "bad rw")) {
 566                 journal_entry_null_range(entry, vstruct_next(entry));
 567                 return ret;
 568         }
 569
 570 fsck_err:
 571         return ret;
 572 }
 573
 574 static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 575                                         struct jset_entry *entry)
 576 {
 577         struct jset_entry_clock *clock =
 578                 container_of(entry, struct jset_entry_clock, entry);
 579
 580         prt_printf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
 581 }
 582
 583 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 584                                             struct jset *jset,
 585                                             struct jset_entry *entry,
 586                                             unsigned version, int big_endian, int write)
 587 {
 588         struct jset_entry_dev_usage *u =
 589                 container_of(entry, struct jset_entry_dev_usage, entry);
 590         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 591         unsigned expected = sizeof(*u);
 592         unsigned dev;
 593         int ret = 0;
 594
 595         if (journal_entry_err_on(bytes < expected,
 596                                  c, jset, entry, "bad size (%u < %u)",
 597                                  bytes, expected)) {
 598                 journal_entry_null_range(entry, vstruct_next(entry));
 599                 return ret;
 600         }
 601
 602         dev = le32_to_cpu(u->dev);
 603
 604         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 605                                  c, jset, entry, "bad dev")) {
 606                 journal_entry_null_range(entry, vstruct_next(entry));
 607                 return ret;
 608         }
 609
 610         if (journal_entry_err_on(u->pad,
 611                                  c, jset, entry, "bad pad")) {
 612                 journal_entry_null_range(entry, vstruct_next(entry));
 613                 return ret;
 614         }
 615
 616 fsck_err:
 617         return ret;
 618 }
 619
 620 static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
 621                                             struct jset_entry *entry)
 622 {
 623         struct jset_entry_dev_usage *u =
 624                 container_of(entry, struct jset_entry_dev_usage, entry);
 625         unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 626
 627         prt_printf(out, "dev=%u", le32_to_cpu(u->dev));
 628
 629         for (i = 0; i < nr_types; i++) {
 630                 if (i < BCH_DATA_NR)
 631                         prt_printf(out, " %s", bch2_data_types[i]);
 632                 else
 633                         prt_printf(out, " (unknown data type %u)", i);
 634                 prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 635                        le64_to_cpu(u->d[i].buckets),
 636                        le64_to_cpu(u->d[i].sectors),
 637                        le64_to_cpu(u->d[i].fragmented));
 638         }
 639
 640         prt_printf(out, " buckets_ec: %llu", le64_to_cpu(u->buckets_ec));
 641 }
 642
 643 static int journal_entry_log_validate(struct bch_fs *c,
 644                                       struct jset *jset,
 645                                       struct jset_entry *entry,
 646                                       unsigned version, int big_endian, int write)
 647 {
 648         return 0;
 649 }
 650
 651 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 652                                       struct jset_entry *entry)
 653 {
 654         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 655         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 656
 657         prt_printf(out, "%.*s", bytes, l->d);
 658 }
 659
 660 static int journal_entry_overwrite_validate(struct bch_fs *c,
 661                                       struct jset *jset,
 662                                       struct jset_entry *entry,
 663                                       unsigned version, int big_endian, int write)
 664 {
 665         return journal_entry_btree_keys_validate(c, jset, entry,
 666                                 version, big_endian, READ);
 667 }
 668
 669 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
 670                                             struct jset_entry *entry)
 671 {
 672         journal_entry_btree_keys_to_text(out, c, entry);
 673 }
 674
 675 struct jset_entry_ops {
 676         int (*validate)(struct bch_fs *, struct jset *,
 677                         struct jset_entry *, unsigned, int, int);
 678         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 679 };
 680
 681 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 682 #define x(f, nr)                                                \
 683         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 684                 .validate       = journal_entry_##f##_validate, \
 685                 .to_text        = journal_entry_##f##_to_text,  \
 686         },
 687         BCH_JSET_ENTRY_TYPES()
 688 #undef x
 689 };
 690
 691 int bch2_journal_entry_validate(struct bch_fs *c,
 692                                 struct jset *jset,
 693                                 struct jset_entry *entry,
 694                                 unsigned version, int big_endian, int write)
 695 {
 696         return entry->type < BCH_JSET_ENTRY_NR
 697                 ? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
 698                                 version, big_endian, write)
 699                 : 0;
 700 }
 701
 702 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 703                                 struct jset_entry *entry)
 704 {
 705         if (entry->type < BCH_JSET_ENTRY_NR) {
 706                 prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 707                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 708         } else {
 709                 prt_printf(out, "(unknown type %u)", entry->type);
 710         }
 711 }
 712
 713 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 714                                  int write)
 715 {
 716         struct jset_entry *entry;
 717         int ret = 0;
 718
 719         vstruct_for_each(jset, entry) {
 720                 if (journal_entry_err_on(vstruct_next(entry) >
 721                                          vstruct_last(jset), c, jset, entry,
 722                                 "journal entry extends past end of jset")) {
 723                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 724                         break;
 725                 }
 726
 727                 ret = bch2_journal_entry_validate(c, jset, entry,
 728                                         le32_to_cpu(jset->version),
 729                                         JSET_BIG_ENDIAN(jset), write);
 730                 if (ret)
 731                         break;
 732         }
 733 fsck_err:
 734         return ret;
 735 }
 736
 737 static int jset_validate(struct bch_fs *c,
 738                          struct bch_dev *ca,
 739                          struct jset *jset, u64 sector,
 740                          int write)
 741 {
 742         unsigned version;
 743         int ret = 0;
 744
 745         if (le64_to_cpu(jset->magic) != jset_magic(c))
 746                 return JOURNAL_ENTRY_NONE;
 747
 748         version = le32_to_cpu(jset->version);
 749         if (journal_entry_err_on(!bch2_version_compatible(version), c, jset, NULL,
 750                         "%s sector %llu seq %llu: incompatible journal entry version %u",
 751                         ca ? ca->name : c->name,
 752                         sector, le64_to_cpu(jset->seq), version)) {
 753                 /* don't try to continue: */
 754                 return -EINVAL;
 755         }
 756
 757         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
 758                                  c, jset, NULL,
 759                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 760                         ca ? ca->name : c->name,
 761                         sector, le64_to_cpu(jset->seq),
 762                         JSET_CSUM_TYPE(jset)))
 763                 ret = JOURNAL_ENTRY_BAD;
 764
 765         /* last_seq is ignored when JSET_NO_FLUSH is true */
 766         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 767                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
 768                                  c, jset, NULL,
 769                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 770                                  le64_to_cpu(jset->last_seq),
 771                                  le64_to_cpu(jset->seq))) {
 772                 jset->last_seq = jset->seq;
 773                 return JOURNAL_ENTRY_BAD;
 774         }
 775
 776         ret = jset_validate_entries(c, jset, write);
 777 fsck_err:
 778         return ret;
 779 }
 780
 781 static int jset_validate_early(struct bch_fs *c,
 782                          struct bch_dev *ca,
 783                          struct jset *jset, u64 sector,
 784                          unsigned bucket_sectors_left,
 785                          unsigned sectors_read)
 786 {
 787         size_t bytes = vstruct_bytes(jset);
 788         unsigned version;
 789         int write = READ;
 790         int ret = 0;
 791
 792         if (le64_to_cpu(jset->magic) != jset_magic(c))
 793                 return JOURNAL_ENTRY_NONE;
 794
 795         version = le32_to_cpu(jset->version);
 796         if (journal_entry_err_on(!bch2_version_compatible(version), c, jset, NULL,
 797                         "%s sector %llu seq %llu: unknown journal entry version %u",
 798                         ca ? ca->name : c->name,
 799                         sector, le64_to_cpu(jset->seq), version)) {
 800                 /* don't try to continue: */
 801                 return -EINVAL;
 802         }
 803
 804         if (bytes > (sectors_read << 9) &&
 805             sectors_read < bucket_sectors_left)
 806                 return JOURNAL_ENTRY_REREAD;
 807
 808         if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
 809                                  c, jset, NULL,
 810                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 811                         ca ? ca->name : c->name,
 812                         sector, le64_to_cpu(jset->seq), bytes))
 813                 le32_add_cpu(&jset->u64s,
 814                              -((bytes - (bucket_sectors_left << 9)) / 8));
 815 fsck_err:
 816         return ret;
 817 }
 818
 819 struct journal_read_buf {
 820         void            *data;
 821         size_t          size;
 822 };
 823
 824 static int journal_read_buf_realloc(struct journal_read_buf *b,
 825                                     size_t new_size)
 826 {
 827         void *n;
 828
 829         /* the bios are sized for this many pages, max: */
 830         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 831                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 832
 833         new_size = roundup_pow_of_two(new_size);
 834         n = kvpmalloc(new_size, GFP_KERNEL);
 835         if (!n)
 836                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 837
 838         kvpfree(b->data, b->size);
 839         b->data = n;
 840         b->size = new_size;
 841         return 0;
 842 }
 843
 844 static int journal_read_bucket(struct bch_dev *ca,
 845                                struct journal_read_buf *buf,
 846                                struct journal_list *jlist,
 847                                unsigned bucket)
 848 {
 849         struct bch_fs *c = ca->fs;
 850         struct journal_device *ja = &ca->journal;
 851         struct jset *j = NULL;
 852         unsigned sectors, sectors_read = 0;
 853         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 854             end = offset + ca->mi.bucket_size;
 855         bool saw_bad = false, csum_good;
 856         int ret = 0;
 857
 858         pr_debug("reading %u", bucket);
 859
 860         while (offset < end) {
 861                 if (!sectors_read) {
 862                         struct bio *bio;
 863                         unsigned nr_bvecs;
 864 reread:
 865                         sectors_read = min_t(unsigned,
 866                                 end - offset, buf->size >> 9);
 867                         nr_bvecs = buf_pages(buf->data, sectors_read << 9);
 868
 869                         bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
 870                         bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
 871
 872                         bio->bi_iter.bi_sector = offset;
 873                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 874
 875                         ret = submit_bio_wait(bio);
 876                         kfree(bio);
 877
 878                         if (bch2_dev_io_err_on(ret, ca,
 879                                                "journal read error: sector %llu",
 880                                                offset) ||
 881                             bch2_meta_read_fault("journal")) {
 882                                 /*
 883                                  * We don't error out of the recovery process
 884                                  * here, since the relevant journal entry may be
 885                                  * found on a different device, and missing or
 886                                  * no journal entries will be handled later
 887                                  */
 888                                 return 0;
 889                         }
 890
 891                         j = buf->data;
 892                 }
 893
 894                 ret = jset_validate_early(c, ca, j, offset,
 895                                     end - offset, sectors_read);
 896                 switch (ret) {
 897                 case 0:
 898                         sectors = vstruct_sectors(j, c->block_bits);
 899                         break;
 900                 case JOURNAL_ENTRY_REREAD:
 901                         if (vstruct_bytes(j) > buf->size) {
 902                                 ret = journal_read_buf_realloc(buf,
 903                                                         vstruct_bytes(j));
 904                                 if (ret)
 905                                         return ret;
 906                         }
 907                         goto reread;
 908                 case JOURNAL_ENTRY_NONE:
 909                         if (!saw_bad)
 910                                 return 0;
 911                         /*
 912                          * On checksum error we don't really trust the size
 913                          * field of the journal entry we read, so try reading
 914                          * again at next block boundary:
 915                          */
 916                         sectors = block_sectors(c);
 917                         goto next_block;
 918                 default:
 919                         return ret;
 920                 }
 921
 922                 /*
 923                  * This happens sometimes if we don't have discards on -
 924                  * when we've partially overwritten a bucket with new
 925                  * journal entries. We don't need the rest of the
 926                  * bucket:
 927                  */
 928                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 929                         return 0;
 930
 931                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 932
 933                 csum_good = jset_csum_good(c, j);
 934                 if (!csum_good)
 935                         saw_bad = true;
 936
 937                 ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
 938                              j->encrypted_start,
 939                              vstruct_end(j) - (void *) j->encrypted_start);
 940                 bch2_fs_fatal_err_on(ret, c,
 941                                 "error decrypting journal entry: %i", ret);
 942
 943                 mutex_lock(&jlist->lock);
 944                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
 945                                         .csum_good      = csum_good,
 946                                         .dev            = ca->dev_idx,
 947                                         .bucket         = bucket,
 948                                         .bucket_offset  = offset -
 949                                                 bucket_to_sector(ca, ja->buckets[bucket]),
 950                                         .sector         = offset,
 951                                         }, jlist, j);
 952                 mutex_unlock(&jlist->lock);
 953
 954                 switch (ret) {
 955                 case JOURNAL_ENTRY_ADD_OK:
 956                         break;
 957                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
 958                         break;
 959                 default:
 960                         return ret;
 961                 }
 962 next_block:
 963                 pr_debug("next");
 964                 offset          += sectors;
 965                 sectors_read    -= sectors;
 966                 j = ((void *) j) + (sectors << 9);
 967         }
 968
 969         return 0;
 970 }
 971
 972 static void bch2_journal_read_device(struct closure *cl)
 973 {
 974         struct journal_device *ja =
 975                 container_of(cl, struct journal_device, read);
 976         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
 977         struct bch_fs *c = ca->fs;
 978         struct journal_list *jlist =
 979                 container_of(cl->parent, struct journal_list, cl);
 980         struct journal_replay *r, **_r;
 981         struct genradix_iter iter;
 982         struct journal_read_buf buf = { NULL, 0 };
 983         unsigned i;
 984         int ret = 0;
 985
 986         if (!ja->nr)
 987                 goto out;
 988
 989         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
 990         if (ret)
 991                 goto err;
 992
 993         pr_debug("%u journal buckets", ja->nr);
 994
 995         for (i = 0; i < ja->nr; i++) {
 996                 ret = journal_read_bucket(ca, &buf, jlist, i);
 997                 if (ret)
 998                         goto err;
 999         }
1000
1001         ja->sectors_free = ca->mi.bucket_size;
1002
1003         mutex_lock(&jlist->lock);
1004         genradix_for_each_reverse(&c->journal_entries, iter, _r) {
1005                 r = *_r;
1006
1007                 if (!r)
1008                         continue;
1009
1010                 for (i = 0; i < r->nr_ptrs; i++) {
1011                         if (r->ptrs[i].dev == ca->dev_idx) {
1012                                 unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
1013                                         vstruct_sectors(&r->j, c->block_bits);
1014
1015                                 ja->cur_idx = r->ptrs[i].bucket;
1016                                 ja->sectors_free = ca->mi.bucket_size - wrote;
1017                                 goto found;
1018                         }
1019                 }
1020         }
1021 found:
1022         mutex_unlock(&jlist->lock);
1023
1024         if (ja->bucket_seq[ja->cur_idx] &&
1025             ja->sectors_free == ca->mi.bucket_size) {
1026                 bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
1027                 bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
1028                 for (i = 0; i < 3; i++) {
1029                         unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
1030                         bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
1031                 }
1032                 ja->sectors_free = 0;
1033         }
1034
1035         /*
1036          * Set dirty_idx to indicate the entire journal is full and needs to be
1037          * reclaimed - journal reclaim will immediately reclaim whatever isn't
1038          * pinned when it first runs:
1039          */
1040         ja->discard_idx = ja->dirty_idx_ondisk =
1041                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
1042 out:
1043         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
1044         kvpfree(buf.data, buf.size);
1045         percpu_ref_put(&ca->io_ref);
1046         closure_return(cl);
1047         return;
1048 err:
1049         mutex_lock(&jlist->lock);
1050         jlist->ret = ret;
1051         mutex_unlock(&jlist->lock);
1052         goto out;
1053 }
1054
1055 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1056                                struct journal_replay *j)
1057 {
1058         unsigned i;
1059
1060         for (i = 0; i < j->nr_ptrs; i++) {
1061                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
1062                 u64 offset;
1063
1064                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
1065
1066                 if (i)
1067                         prt_printf(out, " ");
1068                 prt_printf(out, "%u:%u:%u (sector %llu)",
1069                        j->ptrs[i].dev,
1070                        j->ptrs[i].bucket,
1071                        j->ptrs[i].bucket_offset,
1072                        j->ptrs[i].sector);
1073         }
1074 }
1075
1076 int bch2_journal_read(struct bch_fs *c,
1077                       u64 *last_seq,
1078                       u64 *blacklist_seq,
1079                       u64 *start_seq)
1080 {
1081         struct journal_list jlist;
1082         struct journal_replay *i, **_i, *prev = NULL;
1083         struct genradix_iter radix_iter;
1084         struct bch_dev *ca;
1085         unsigned iter;
1086         struct printbuf buf = PRINTBUF;
1087         bool degraded = false, last_write_torn = false;
1088         u64 seq;
1089         int ret = 0;
1090
1091         closure_init_stack(&jlist.cl);
1092         mutex_init(&jlist.lock);
1093         jlist.last_seq = 0;
1094         jlist.ret = 0;
1095
1096         for_each_member_device(ca, c, iter) {
1097                 if (!c->opts.fsck &&
1098                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1099                         continue;
1100
1101                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1102                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1103                     percpu_ref_tryget(&ca->io_ref))
1104                         closure_call(&ca->journal.read,
1105                                      bch2_journal_read_device,
1106                                      system_unbound_wq,
1107                                      &jlist.cl);
1108                 else
1109                         degraded = true;
1110         }
1111
1112         closure_sync(&jlist.cl);
1113
1114         if (jlist.ret)
1115                 return jlist.ret;
1116
1117         *last_seq       = 0;
1118         *start_seq      = 0;
1119         *blacklist_seq  = 0;
1120
1121         /*
1122          * Find most recent flush entry, and ignore newer non flush entries -
1123          * those entries will be blacklisted:
1124          */
1125         genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
1126                 int write = READ;
1127
1128                 i = *_i;
1129
1130                 if (!i || i->ignore)
1131                         continue;
1132
1133                 if (!*start_seq)
1134                         *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
1135
1136                 if (JSET_NO_FLUSH(&i->j)) {
1137                         i->ignore = true;
1138                         continue;
1139                 }
1140
1141                 if (!last_write_torn && !i->csum_good) {
1142                         last_write_torn = true;
1143                         i->ignore = true;
1144                         continue;
1145                 }
1146
1147                 if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
1148                                          c, &i->j, NULL,
1149                                          "invalid journal entry: last_seq > seq (%llu > %llu)",
1150                                          le64_to_cpu(i->j.last_seq),
1151                                          le64_to_cpu(i->j.seq)))
1152                         i->j.last_seq = i->j.seq;
1153
1154                 *last_seq       = le64_to_cpu(i->j.last_seq);
1155                 *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1156                 break;
1157         }
1158
1159         if (!*start_seq) {
1160                 bch_info(c, "journal read done, but no entries found");
1161                 return 0;
1162         }
1163
1164         if (!*last_seq) {
1165                 fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
1166                 return 0;
1167         }
1168
1169         bch_info(c, "journal read done, replaying entries %llu-%llu",
1170                  *last_seq, *blacklist_seq - 1);
1171
1172         if (*start_seq != *blacklist_seq)
1173                 bch_info(c, "dropped unflushed entries %llu-%llu",
1174                          *blacklist_seq, *start_seq - 1);
1175
1176         /* Drop blacklisted entries and entries older than last_seq: */
1177         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1178                 i = *_i;
1179
1180                 if (!i || i->ignore)
1181                         continue;
1182
1183                 seq = le64_to_cpu(i->j.seq);
1184                 if (seq < *last_seq) {
1185                         journal_replay_free(c, i);
1186                         continue;
1187                 }
1188
1189                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1190                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1191                                     "found blacklisted journal entry %llu", seq);
1192                         i->ignore = true;
1193                 }
1194         }
1195
1196         /* Check for missing entries: */
1197         seq = *last_seq;
1198         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1199                 i = *_i;
1200
1201                 if (!i || i->ignore)
1202                         continue;
1203
1204                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1205
1206                 while (seq < le64_to_cpu(i->j.seq)) {
1207                         u64 missing_start, missing_end;
1208                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1209
1210                         while (seq < le64_to_cpu(i->j.seq) &&
1211                                bch2_journal_seq_is_blacklisted(c, seq, false))
1212                                 seq++;
1213
1214                         if (seq == le64_to_cpu(i->j.seq))
1215                                 break;
1216
1217                         missing_start = seq;
1218
1219                         while (seq < le64_to_cpu(i->j.seq) &&
1220                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1221                                 seq++;
1222
1223                         if (prev) {
1224                                 bch2_journal_ptrs_to_text(&buf1, c, prev);
1225                                 prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits));
1226                         } else
1227                                 prt_printf(&buf1, "(none)");
1228                         bch2_journal_ptrs_to_text(&buf2, c, i);
1229
1230                         missing_end = seq - 1;
1231                         fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1232                                  "  prev at %s\n"
1233                                  "  next at %s",
1234                                  missing_start, missing_end,
1235                                  *last_seq, *blacklist_seq - 1,
1236                                  buf1.buf, buf2.buf);
1237
1238                         printbuf_exit(&buf1);
1239                         printbuf_exit(&buf2);
1240                 }
1241
1242                 prev = i;
1243                 seq++;
1244         }
1245
1246         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1247                 struct bch_replicas_padded replicas = {
1248                         .e.data_type = BCH_DATA_journal,
1249                         .e.nr_required = 1,
1250                 };
1251                 unsigned ptr;
1252
1253                 i = *_i;
1254                 if (!i || i->ignore)
1255                         continue;
1256
1257                 for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
1258                         struct bch_dev *ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
1259
1260                         if (!i->ptrs[ptr].csum_good)
1261                                 bch_err_dev_offset(ca, i->ptrs[ptr].sector,
1262                                                    "invalid journal checksum, seq %llu%s",
1263                                                    le64_to_cpu(i->j.seq),
1264                                                    i->csum_good ? " (had good copy on another device)" : "");
1265                 }
1266
1267                 ret = jset_validate(c,
1268                                     bch_dev_bkey_exists(c, i->ptrs[0].dev),
1269                                     &i->j,
1270                                     i->ptrs[0].sector,
1271                                     READ);
1272                 if (ret)
1273                         goto err;
1274
1275                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1276                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1277
1278                 bch2_replicas_entry_sort(&replicas.e);
1279
1280                 /*
1281                  * If we're mounting in degraded mode - if we didn't read all
1282                  * the devices - this is wrong:
1283                  */
1284
1285                 printbuf_reset(&buf);
1286                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1287
1288                 if (!degraded &&
1289                     fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
1290                                 "superblock not marked as containing replicas %s",
1291                                 buf.buf)) {
1292                         ret = bch2_mark_replicas(c, &replicas.e);
1293                         if (ret)
1294                                 goto err;
1295                 }
1296         }
1297 err:
1298 fsck_err:
1299         printbuf_exit(&buf);
1300         return ret;
1301 }
1302
1303 /* journal write: */
1304
1305 static void __journal_write_alloc(struct journal *j,
1306                                   struct journal_buf *w,
1307                                   struct dev_alloc_list *devs_sorted,
1308                                   unsigned sectors,
1309                                   unsigned *replicas,
1310                                   unsigned replicas_want)
1311 {
1312         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1313         struct journal_device *ja;
1314         struct bch_dev *ca;
1315         unsigned i;
1316
1317         if (*replicas >= replicas_want)
1318                 return;
1319
1320         for (i = 0; i < devs_sorted->nr; i++) {
1321                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1322                 if (!ca)
1323                         continue;
1324
1325                 ja = &ca->journal;
1326
1327                 /*
1328                  * Check that we can use this device, and aren't already using
1329                  * it:
1330                  */
1331                 if (!ca->mi.durability ||
1332                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1333                     !ja->nr ||
1334                     bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) ||
1335                     sectors > ja->sectors_free)
1336                         continue;
1337
1338                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1339
1340                 bch2_bkey_append_ptr(&w->key,
1341                         (struct bch_extent_ptr) {
1342                                   .offset = bucket_to_sector(ca,
1343                                         ja->buckets[ja->cur_idx]) +
1344                                         ca->mi.bucket_size -
1345                                         ja->sectors_free,
1346                                   .dev = ca->dev_idx,
1347                 });
1348
1349                 ja->sectors_free -= sectors;
1350                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1351
1352                 *replicas += ca->mi.durability;
1353
1354                 if (*replicas >= replicas_want)
1355                         break;
1356         }
1357 }
1358
1359 /**
1360  * journal_next_bucket - move on to the next journal bucket if possible
1361  */
1362 static int journal_write_alloc(struct journal *j, struct journal_buf *w,
1363                                unsigned sectors)
1364 {
1365         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1366         struct bch_devs_mask devs;
1367         struct journal_device *ja;
1368         struct bch_dev *ca;
1369         struct dev_alloc_list devs_sorted;
1370         unsigned target = c->opts.metadata_target ?:
1371                 c->opts.foreground_target;
1372         unsigned i, replicas = 0, replicas_want =
1373                 READ_ONCE(c->opts.metadata_replicas);
1374
1375         rcu_read_lock();
1376 retry:
1377         devs = target_rw_devs(c, BCH_DATA_journal, target);
1378
1379         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1380
1381         __journal_write_alloc(j, w, &devs_sorted,
1382                               sectors, &replicas, replicas_want);
1383
1384         if (replicas >= replicas_want)
1385                 goto done;
1386
1387         for (i = 0; i < devs_sorted.nr; i++) {
1388                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1389                 if (!ca)
1390                         continue;
1391
1392                 ja = &ca->journal;
1393
1394                 if (sectors > ja->sectors_free &&
1395                     sectors <= ca->mi.bucket_size &&
1396                     bch2_journal_dev_buckets_available(j, ja,
1397                                         journal_space_discarded)) {
1398                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1399                         ja->sectors_free = ca->mi.bucket_size;
1400
1401                         /*
1402                          * ja->bucket_seq[ja->cur_idx] must always have
1403                          * something sensible:
1404                          */
1405                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1406                 }
1407         }
1408
1409         __journal_write_alloc(j, w, &devs_sorted,
1410                               sectors, &replicas, replicas_want);
1411
1412         if (replicas < replicas_want && target) {
1413                 /* Retry from all devices: */
1414                 target = 0;
1415                 goto retry;
1416         }
1417 done:
1418         rcu_read_unlock();
1419
1420         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1421
1422         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1423 }
1424
1425 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1426 {
1427         /* we aren't holding j->lock: */
1428         unsigned new_size = READ_ONCE(j->buf_size_want);
1429         void *new_buf;
1430
1431         if (buf->buf_size >= new_size)
1432                 return;
1433
1434         new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
1435         if (!new_buf)
1436                 return;
1437
1438         memcpy(new_buf, buf->data, buf->buf_size);
1439
1440         spin_lock(&j->lock);
1441         swap(buf->data,         new_buf);
1442         swap(buf->buf_size,     new_size);
1443         spin_unlock(&j->lock);
1444
1445         kvpfree(new_buf, new_size);
1446 }
1447
1448 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1449 {
1450         return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
1451 }
1452
1453 static void journal_write_done(struct closure *cl)
1454 {
1455         struct journal *j = container_of(cl, struct journal, io);
1456         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1457         struct journal_buf *w = journal_last_unwritten_buf(j);
1458         union journal_res_state old, new;
1459         u64 v, seq;
1460         int err = 0;
1461
1462         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1463                                ? j->flush_write_time
1464                                : j->noflush_write_time, j->write_start_time);
1465
1466         if (!w->devs_written.nr) {
1467                 bch_err(c, "unable to write journal to sufficient devices");
1468                 err = -EIO;
1469         }
1470         if (err)
1471                 bch2_fatal_error(c);
1472
1473         spin_lock(&j->lock);
1474         seq = le64_to_cpu(w->data->seq);
1475
1476         if (seq >= j->pin.front)
1477                 journal_seq_pin(j, seq)->devs = w->devs_written;
1478
1479         if (!err) {
1480                 if (!JSET_NO_FLUSH(w->data)) {
1481                         j->flushed_seq_ondisk = seq;
1482                         j->last_seq_ondisk = w->last_seq;
1483
1484                         bch2_do_discards(c);
1485                         closure_wake_up(&c->freelist_wait);
1486
1487                         bch2_reset_alloc_cursors(c);
1488                 }
1489         } else if (!j->err_seq || seq < j->err_seq)
1490                 j->err_seq      = seq;
1491
1492         j->seq_ondisk           = seq;
1493
1494         /*
1495          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1496          * more buckets:
1497          *
1498          * Must come before signaling write completion, for
1499          * bch2_fs_journal_stop():
1500          */
1501         if (j->watermark != BCH_WATERMARK_stripe)
1502                 journal_reclaim_kick(&c->journal);
1503
1504         /* also must come before signalling write completion: */
1505         closure_debug_destroy(cl);
1506
1507         v = atomic64_read(&j->reservations.counter);
1508         do {
1509                 old.v = new.v = v;
1510                 BUG_ON(journal_state_count(new, new.unwritten_idx));
1511
1512                 new.unwritten_idx++;
1513         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1514                                        old.v, new.v)) != old.v);
1515
1516         bch2_journal_space_available(j);
1517
1518         closure_wake_up(&w->wait);
1519         journal_wake(j);
1520
1521         if (!journal_state_count(new, new.unwritten_idx) &&
1522             journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
1523                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1524         } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
1525                    new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
1526                 struct journal_buf *buf = journal_cur_buf(j);
1527                 long delta = buf->expires - jiffies;
1528
1529                 /*
1530                  * We don't close a journal entry to write it while there's
1531                  * previous entries still in flight - the current journal entry
1532                  * might want to be written now:
1533                  */
1534
1535                 mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
1536         }
1537
1538         spin_unlock(&j->lock);
1539 }
1540
1541 static void journal_write_endio(struct bio *bio)
1542 {
1543         struct bch_dev *ca = bio->bi_private;
1544         struct journal *j = &ca->fs->journal;
1545         struct journal_buf *w = journal_last_unwritten_buf(j);
1546         unsigned long flags;
1547
1548         if (bch2_dev_io_err_on(bio->bi_status, ca, "error writing journal entry %llu: %s",
1549                                le64_to_cpu(w->data->seq),
1550                                bch2_blk_status_to_str(bio->bi_status)) ||
1551             bch2_meta_write_fault("journal")) {
1552                 spin_lock_irqsave(&j->err_lock, flags);
1553                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1554                 spin_unlock_irqrestore(&j->err_lock, flags);
1555         }
1556
1557         closure_put(&j->io);
1558         percpu_ref_put(&ca->io_ref);
1559 }
1560
1561 static void do_journal_write(struct closure *cl)
1562 {
1563         struct journal *j = container_of(cl, struct journal, io);
1564         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1565         struct bch_dev *ca;
1566         struct journal_buf *w = journal_last_unwritten_buf(j);
1567         struct bch_extent_ptr *ptr;
1568         struct bio *bio;
1569         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1570
1571         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1572                 ca = bch_dev_bkey_exists(c, ptr->dev);
1573                 if (!percpu_ref_tryget(&ca->io_ref)) {
1574                         /* XXX: fix this */
1575                         bch_err(c, "missing device for journal write\n");
1576                         continue;
1577                 }
1578
1579                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1580                              sectors);
1581
1582                 bio = ca->journal.bio;
1583                 bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
1584                 bio->bi_iter.bi_sector  = ptr->offset;
1585                 bio->bi_end_io          = journal_write_endio;
1586                 bio->bi_private         = ca;
1587
1588                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1589                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1590
1591                 if (!JSET_NO_FLUSH(w->data))
1592                         bio->bi_opf    |= REQ_FUA;
1593                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1594                         bio->bi_opf    |= REQ_PREFLUSH;
1595
1596                 bch2_bio_map(bio, w->data, sectors << 9);
1597
1598                 trace_and_count(c, journal_write, bio);
1599                 closure_bio_submit(bio, cl);
1600
1601                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1602                         le64_to_cpu(w->data->seq);
1603         }
1604
1605         continue_at(cl, journal_write_done, c->io_complete_wq);
1606         return;
1607 }
1608
1609 static void bch2_journal_entries_postprocess(struct bch_fs *c, struct jset *jset)
1610 {
1611         struct jset_entry *i, *next, *prev = NULL;
1612
1613         /*
1614          * Simple compaction, dropping empty jset_entries (from journal
1615          * reservations that weren't fully used) and merging jset_entries that
1616          * can be.
1617          *
1618          * If we wanted to be really fancy here, we could sort all the keys in
1619          * the jset and drop keys that were overwritten - probably not worth it:
1620          */
1621         vstruct_for_each_safe(jset, i, next) {
1622                 unsigned u64s = le16_to_cpu(i->u64s);
1623
1624                 /* Empty entry: */
1625                 if (!u64s)
1626                         continue;
1627
1628                 if (i->type == BCH_JSET_ENTRY_btree_root)
1629                         bch2_journal_entry_to_btree_root(c, i);
1630
1631                 /* Can we merge with previous entry? */
1632                 if (prev &&
1633                     i->btree_id == prev->btree_id &&
1634                     i->level    == prev->level &&
1635                     i->type     == prev->type &&
1636                     i->type     == BCH_JSET_ENTRY_btree_keys &&
1637                     le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
1638                         memmove_u64s_down(vstruct_next(prev),
1639                                           i->_data,
1640                                           u64s);
1641                         le16_add_cpu(&prev->u64s, u64s);
1642                         continue;
1643                 }
1644
1645                 /* Couldn't merge, move i into new position (after prev): */
1646                 prev = prev ? vstruct_next(prev) : jset->start;
1647                 if (i != prev)
1648                         memmove_u64s_down(prev, i, jset_u64s(u64s));
1649         }
1650
1651         prev = prev ? vstruct_next(prev) : jset->start;
1652         jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
1653 }
1654
1655 void bch2_journal_write(struct closure *cl)
1656 {
1657         struct journal *j = container_of(cl, struct journal, io);
1658         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1659         struct bch_dev *ca;
1660         struct journal_buf *w = journal_last_unwritten_buf(j);
1661         struct bch_replicas_padded replicas;
1662         struct jset_entry *start, *end;
1663         struct jset *jset;
1664         struct bio *bio;
1665         struct printbuf journal_debug_buf = PRINTBUF;
1666         bool validate_before_checksum = false;
1667         unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
1668         int ret;
1669
1670         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1671
1672         journal_buf_realloc(j, w);
1673         jset = w->data;
1674
1675         j->write_start_time = local_clock();
1676
1677         spin_lock(&j->lock);
1678
1679         /*
1680          * If the journal is in an error state - we did an emergency shutdown -
1681          * we prefer to continue doing journal writes. We just mark them as
1682          * noflush so they'll never be used, but they'll still be visible by the
1683          * list_journal tool - this helps in debugging.
1684          *
1685          * There's a caveat: the first journal write after marking the
1686          * superblock dirty must always be a flush write, because on startup
1687          * from a clean shutdown we didn't necessarily read the journal and the
1688          * new journal write might overwrite whatever was in the journal
1689          * previously - we can't leave the journal without any flush writes in
1690          * it.
1691          *
1692          * So if we're in an error state, and we're still starting up, we don't
1693          * write anything at all.
1694          */
1695         if (!test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags) &&
1696             (bch2_journal_error(j) ||
1697              w->noflush ||
1698              (!w->must_flush &&
1699               (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1700               test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)))) {
1701                 w->noflush = true;
1702                 SET_JSET_NO_FLUSH(jset, true);
1703                 jset->last_seq  = 0;
1704                 w->last_seq     = 0;
1705
1706                 j->nr_noflush_writes++;
1707         } else if (!bch2_journal_error(j)) {
1708                 j->last_flush_write = jiffies;
1709                 j->nr_flush_writes++;
1710                 clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
1711         } else {
1712                 spin_unlock(&j->lock);
1713                 goto err;
1714         }
1715         spin_unlock(&j->lock);
1716
1717         /*
1718          * New btree roots are set by journalling them; when the journal entry
1719          * gets written we have to propagate them to c->btree_roots
1720          *
1721          * But, every journal entry we write has to contain all the btree roots
1722          * (at least for now); so after we copy btree roots to c->btree_roots we
1723          * have to get any missing btree roots and add them to this journal
1724          * entry:
1725          */
1726
1727         bch2_journal_entries_postprocess(c, jset);
1728
1729         start = end = vstruct_last(jset);
1730
1731         end     = bch2_btree_roots_to_journal_entries(c, jset->start, end);
1732
1733         bch2_journal_super_entries_add_common(c, &end,
1734                                 le64_to_cpu(jset->seq));
1735         u64s    = (u64 *) end - (u64 *) start;
1736         BUG_ON(u64s > j->entry_u64s_reserved);
1737
1738         le32_add_cpu(&jset->u64s, u64s);
1739
1740         sectors = vstruct_sectors(jset, c->block_bits);
1741         bytes   = vstruct_bytes(jset);
1742
1743         if (sectors > w->sectors) {
1744                 bch2_fs_fatal_error(c, "aieeee! journal write overran available space, %zu > %u (extra %u reserved %u/%u)",
1745                                     vstruct_bytes(jset), w->sectors << 9,
1746                                     u64s, w->u64s_reserved, j->entry_u64s_reserved);
1747                 goto err;
1748         }
1749
1750         jset->magic             = cpu_to_le64(jset_magic(c));
1751         jset->version           = cpu_to_le32(c->sb.version);
1752
1753         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1754         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1755
1756         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1757                 j->last_empty_seq = le64_to_cpu(jset->seq);
1758
1759         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1760                 validate_before_checksum = true;
1761
1762         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1763                 validate_before_checksum = true;
1764
1765         if (validate_before_checksum &&
1766             jset_validate(c, NULL, jset, 0, WRITE))
1767                 goto err;
1768
1769         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1770                     jset->encrypted_start,
1771                     vstruct_end(jset) - (void *) jset->encrypted_start);
1772         if (bch2_fs_fatal_err_on(ret, c,
1773                         "error decrypting journal entry: %i", ret))
1774                 goto err;
1775
1776         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1777                                   journal_nonce(jset), jset);
1778
1779         if (!validate_before_checksum &&
1780             jset_validate(c, NULL, jset, 0, WRITE))
1781                 goto err;
1782
1783         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1784
1785 retry_alloc:
1786         spin_lock(&j->lock);
1787         ret = journal_write_alloc(j, w, sectors);
1788
1789         if (ret && j->can_discard) {
1790                 spin_unlock(&j->lock);
1791                 bch2_journal_do_discards(j);
1792                 goto retry_alloc;
1793         }
1794
1795         if (ret)
1796                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1797
1798         /*
1799          * write is allocated, no longer need to account for it in
1800          * bch2_journal_space_available():
1801          */
1802         w->sectors = 0;
1803
1804         /*
1805          * journal entry has been compacted and allocated, recalculate space
1806          * available:
1807          */
1808         bch2_journal_space_available(j);
1809         spin_unlock(&j->lock);
1810
1811         if (ret) {
1812                 bch_err(c, "Unable to allocate journal write:\n%s",
1813                         journal_debug_buf.buf);
1814                 printbuf_exit(&journal_debug_buf);
1815                 goto err;
1816         }
1817
1818         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1819
1820         if (c->opts.nochanges)
1821                 goto no_io;
1822
1823         for_each_rw_member(ca, c, i)
1824                 nr_rw_members++;
1825
1826         if (nr_rw_members > 1)
1827                 w->separate_flush = true;
1828
1829         /*
1830          * Mark journal replicas before we submit the write to guarantee
1831          * recovery will find the journal entries after a crash.
1832          */
1833         bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1834                                  w->devs_written);
1835         ret = bch2_mark_replicas(c, &replicas.e);
1836         if (ret)
1837                 goto err;
1838
1839         if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
1840                 for_each_rw_member(ca, c, i) {
1841                         percpu_ref_get(&ca->io_ref);
1842
1843                         bio = ca->journal.bio;
1844                         bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
1845                         bio->bi_end_io          = journal_write_endio;
1846                         bio->bi_private         = ca;
1847                         closure_bio_submit(bio, cl);
1848                 }
1849         }
1850
1851         continue_at(cl, do_journal_write, c->io_complete_wq);
1852         return;
1853 no_io:
1854         continue_at(cl, journal_write_done, c->io_complete_wq);
1855         return;
1856 err:
1857         bch2_fatal_error(c);
1858         continue_at(cl, journal_write_done, c->io_complete_wq);
1859 }