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
  18 #include <trace/events/bcachefs.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, "found duplicate but non identical journal entries (seq %llu)",
 145                          le64_to_cpu(j->seq));
 146                 i = dup;
 147                 goto found;
 148         }
 149 replace:
 150         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 151         if (!i)
 152                 return -BCH_ERR_ENOMEM_journal_entry_add;
 153
 154         i->nr_ptrs      = 0;
 155         i->csum_good    = entry_ptr.csum_good;
 156         i->ignore       = false;
 157         unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
 158         i->ptrs[i->nr_ptrs++] = entry_ptr;
 159
 160         if (dup) {
 161                 if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) {
 162                         bch_err(c, "found too many copies of journal entry %llu",
 163                                 le64_to_cpu(i->j.seq));
 164                         dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1;
 165                 }
 166
 167                 /* The first ptr should represent the jset we kept: */
 168                 memcpy(i->ptrs + i->nr_ptrs,
 169                        dup->ptrs,
 170                        sizeof(dup->ptrs[0]) * dup->nr_ptrs);
 171                 i->nr_ptrs += dup->nr_ptrs;
 172                 __journal_replay_free(c, dup);
 173         }
 174
 175         *_i = i;
 176         return 0;
 177 found:
 178         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 179                 if (ptr->dev == ca->dev_idx) {
 180                         bch_err(c, "duplicate journal entry %llu on same device",
 181                                 le64_to_cpu(i->j.seq));
 182                         goto out;
 183                 }
 184         }
 185
 186         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 187                 bch_err(c, "found too many copies of journal entry %llu",
 188                         le64_to_cpu(i->j.seq));
 189                 goto out;
 190         }
 191
 192         i->ptrs[i->nr_ptrs++] = entry_ptr;
 193 out:
 194 fsck_err:
 195         return ret;
 196 }
 197
 198 /* this fills in a range with empty jset_entries: */
 199 static void journal_entry_null_range(void *start, void *end)
 200 {
 201         struct jset_entry *entry;
 202
 203         for (entry = start; entry != end; entry = vstruct_next(entry))
 204                 memset(entry, 0, sizeof(*entry));
 205 }
 206
 207 #define JOURNAL_ENTRY_REREAD    5
 208 #define JOURNAL_ENTRY_NONE      6
 209 #define JOURNAL_ENTRY_BAD       7
 210
 211 static void journal_entry_err_msg(struct printbuf *out,
 212                                   struct jset *jset,
 213                                   struct jset_entry *entry)
 214 {
 215         prt_str(out, "invalid journal entry ");
 216         if (entry)
 217                 prt_printf(out, "%s ", bch2_jset_entry_types[entry->type]);
 218
 219         if (!jset)
 220                 prt_printf(out, "in superblock");
 221         else if (!entry)
 222                 prt_printf(out, "at seq %llu", le64_to_cpu(jset->seq));
 223         else
 224                 prt_printf(out, "at offset %zi/%u seq %llu",
 225                            (u64 *) entry - jset->_data,
 226                            le32_to_cpu(jset->u64s),
 227                            le64_to_cpu(jset->seq));
 228         prt_str(out, ": ");
 229 }
 230
 231 #define journal_entry_err(c, jset, entry, msg, ...)                     \
 232 ({                                                                      \
 233         struct printbuf buf = PRINTBUF;                                 \
 234                                                                         \
 235         journal_entry_err_msg(&buf, jset, entry);                       \
 236         prt_printf(&buf, msg, ##__VA_ARGS__);                           \
 237                                                                         \
 238         switch (write) {                                                \
 239         case READ:                                                      \
 240                 mustfix_fsck_err(c, "%s", buf.buf);                     \
 241                 break;                                                  \
 242         case WRITE:                                                     \
 243                 bch_err(c, "corrupt metadata before write: %s\n", buf.buf);\
 244                 if (bch2_fs_inconsistent(c)) {                          \
 245                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 246                         goto fsck_err;                                  \
 247                 }                                                       \
 248                 break;                                                  \
 249         }                                                               \
 250                                                                         \
 251         printbuf_exit(&buf);                                            \
 252         true;                                                           \
 253 })
 254
 255 #define journal_entry_err_on(cond, c, jset, entry, msg, ...)            \
 256         ((cond) ? journal_entry_err(c, jset, entry, msg, ##__VA_ARGS__) : false)
 257
 258 #define FSCK_DELETED_KEY        5
 259
 260 static int journal_validate_key(struct bch_fs *c,
 261                                 struct jset *jset,
 262                                 struct jset_entry *entry,
 263                                 unsigned level, enum btree_id btree_id,
 264                                 struct bkey_i *k,
 265                                 unsigned version, int big_endian, int write)
 266 {
 267         void *next = vstruct_next(entry);
 268         struct printbuf buf = PRINTBUF;
 269         int ret = 0;
 270
 271         if (journal_entry_err_on(!k->k.u64s, c, jset, entry, "k->u64s 0")) {
 272                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 273                 journal_entry_null_range(vstruct_next(entry), next);
 274                 return FSCK_DELETED_KEY;
 275         }
 276
 277         if (journal_entry_err_on((void *) bkey_next(k) >
 278                                  (void *) vstruct_next(entry),
 279                                  c, jset, entry,
 280                                  "extends past end of journal entry")) {
 281                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 282                 journal_entry_null_range(vstruct_next(entry), next);
 283                 return FSCK_DELETED_KEY;
 284         }
 285
 286         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
 287                                  c, jset, entry,
 288                                  "bad format %u", k->k.format)) {
 289                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 290                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 291                 journal_entry_null_range(vstruct_next(entry), next);
 292                 return FSCK_DELETED_KEY;
 293         }
 294
 295         if (!write)
 296                 bch2_bkey_compat(level, btree_id, version, big_endian,
 297                                  write, NULL, bkey_to_packed(k));
 298
 299         if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 300                               __btree_node_type(level, btree_id), write, &buf)) {
 301                 printbuf_reset(&buf);
 302                 prt_printf(&buf, "invalid journal entry %s at offset %zi/%u seq %llu:",
 303                            bch2_jset_entry_types[entry->type],
 304                            (u64 *) entry - jset->_data,
 305                            le32_to_cpu(jset->u64s),
 306                            le64_to_cpu(jset->seq));
 307                 prt_newline(&buf);
 308                 printbuf_indent_add(&buf, 2);
 309
 310                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
 311                 prt_newline(&buf);
 312                 bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 313                                   __btree_node_type(level, btree_id), write, &buf);
 314
 315                 mustfix_fsck_err(c, "%s", buf.buf);
 316
 317                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 318                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 319                 journal_entry_null_range(vstruct_next(entry), next);
 320
 321                 printbuf_exit(&buf);
 322                 return FSCK_DELETED_KEY;
 323         }
 324
 325         if (write)
 326                 bch2_bkey_compat(level, btree_id, version, big_endian,
 327                                  write, NULL, bkey_to_packed(k));
 328 fsck_err:
 329         printbuf_exit(&buf);
 330         return ret;
 331 }
 332
 333 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 334                                              struct jset *jset,
 335                                              struct jset_entry *entry,
 336                                              unsigned version, int big_endian, int write)
 337 {
 338         struct bkey_i *k = entry->start;
 339
 340         while (k != vstruct_last(entry)) {
 341                 int ret = journal_validate_key(c, jset, entry,
 342                                                entry->level,
 343                                                entry->btree_id,
 344                                                k, version, big_endian, write|BKEY_INVALID_FROM_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((version != BCH_JSET_VERSION_OLD &&
 750                                   version < bcachefs_metadata_version_min) ||
 751                                  version >= bcachefs_metadata_version_max,
 752                                  c, jset, NULL,
 753                         "%s sector %llu seq %llu: unknown journal entry version %u",
 754                         ca ? ca->name : c->name,
 755                         sector, le64_to_cpu(jset->seq),
 756                         version)) {
 757                 /* don't try to continue: */
 758                 return -EINVAL;
 759         }
 760
 761         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
 762                                  c, jset, NULL,
 763                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 764                         ca ? ca->name : c->name,
 765                         sector, le64_to_cpu(jset->seq),
 766                         JSET_CSUM_TYPE(jset)))
 767                 ret = JOURNAL_ENTRY_BAD;
 768
 769         /* last_seq is ignored when JSET_NO_FLUSH is true */
 770         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 771                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
 772                                  c, jset, NULL,
 773                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 774                                  le64_to_cpu(jset->last_seq),
 775                                  le64_to_cpu(jset->seq))) {
 776                 jset->last_seq = jset->seq;
 777                 return JOURNAL_ENTRY_BAD;
 778         }
 779
 780         ret = jset_validate_entries(c, jset, write);
 781 fsck_err:
 782         return ret;
 783 }
 784
 785 static int jset_validate_early(struct bch_fs *c,
 786                          struct bch_dev *ca,
 787                          struct jset *jset, u64 sector,
 788                          unsigned bucket_sectors_left,
 789                          unsigned sectors_read)
 790 {
 791         size_t bytes = vstruct_bytes(jset);
 792         unsigned version;
 793         int write = READ;
 794         int ret = 0;
 795
 796         if (le64_to_cpu(jset->magic) != jset_magic(c))
 797                 return JOURNAL_ENTRY_NONE;
 798
 799         version = le32_to_cpu(jset->version);
 800         if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
 801                                   version < bcachefs_metadata_version_min) ||
 802                                  version >= bcachefs_metadata_version_max,
 803                                  c, jset, NULL,
 804                         "%s sector %llu seq %llu: unknown journal entry version %u",
 805                         ca ? ca->name : c->name,
 806                         sector, le64_to_cpu(jset->seq),
 807                         version)) {
 808                 /* don't try to continue: */
 809                 return -EINVAL;
 810         }
 811
 812         if (bytes > (sectors_read << 9) &&
 813             sectors_read < bucket_sectors_left)
 814                 return JOURNAL_ENTRY_REREAD;
 815
 816         if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
 817                                  c, jset, NULL,
 818                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 819                         ca ? ca->name : c->name,
 820                         sector, le64_to_cpu(jset->seq), bytes))
 821                 le32_add_cpu(&jset->u64s,
 822                              -((bytes - (bucket_sectors_left << 9)) / 8));
 823 fsck_err:
 824         return ret;
 825 }
 826
 827 struct journal_read_buf {
 828         void            *data;
 829         size_t          size;
 830 };
 831
 832 static int journal_read_buf_realloc(struct journal_read_buf *b,
 833                                     size_t new_size)
 834 {
 835         void *n;
 836
 837         /* the bios are sized for this many pages, max: */
 838         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 839                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 840
 841         new_size = roundup_pow_of_two(new_size);
 842         n = kvpmalloc(new_size, GFP_KERNEL);
 843         if (!n)
 844                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 845
 846         kvpfree(b->data, b->size);
 847         b->data = n;
 848         b->size = new_size;
 849         return 0;
 850 }
 851
 852 static int journal_read_bucket(struct bch_dev *ca,
 853                                struct journal_read_buf *buf,
 854                                struct journal_list *jlist,
 855                                unsigned bucket)
 856 {
 857         struct bch_fs *c = ca->fs;
 858         struct journal_device *ja = &ca->journal;
 859         struct jset *j = NULL;
 860         unsigned sectors, sectors_read = 0;
 861         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 862             end = offset + ca->mi.bucket_size;
 863         bool saw_bad = false, csum_good;
 864         int ret = 0;
 865
 866         pr_debug("reading %u", bucket);
 867
 868         while (offset < end) {
 869                 if (!sectors_read) {
 870                         struct bio *bio;
 871                         unsigned nr_bvecs;
 872 reread:
 873                         sectors_read = min_t(unsigned,
 874                                 end - offset, buf->size >> 9);
 875                         nr_bvecs = buf_pages(buf->data, sectors_read << 9);
 876
 877                         bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
 878                         bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
 879
 880                         bio->bi_iter.bi_sector = offset;
 881                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 882
 883                         ret = submit_bio_wait(bio);
 884                         kfree(bio);
 885
 886                         if (bch2_dev_io_err_on(ret, ca,
 887                                                "journal read error: sector %llu",
 888                                                offset) ||
 889                             bch2_meta_read_fault("journal")) {
 890                                 /*
 891                                  * We don't error out of the recovery process
 892                                  * here, since the relevant journal entry may be
 893                                  * found on a different device, and missing or
 894                                  * no journal entries will be handled later
 895                                  */
 896                                 return 0;
 897                         }
 898
 899                         j = buf->data;
 900                 }
 901
 902                 ret = jset_validate_early(c, ca, j, offset,
 903                                     end - offset, sectors_read);
 904                 switch (ret) {
 905                 case 0:
 906                         sectors = vstruct_sectors(j, c->block_bits);
 907                         break;
 908                 case JOURNAL_ENTRY_REREAD:
 909                         if (vstruct_bytes(j) > buf->size) {
 910                                 ret = journal_read_buf_realloc(buf,
 911                                                         vstruct_bytes(j));
 912                                 if (ret)
 913                                         return ret;
 914                         }
 915                         goto reread;
 916                 case JOURNAL_ENTRY_NONE:
 917                         if (!saw_bad)
 918                                 return 0;
 919                         /*
 920                          * On checksum error we don't really trust the size
 921                          * field of the journal entry we read, so try reading
 922                          * again at next block boundary:
 923                          */
 924                         sectors = block_sectors(c);
 925                         goto next_block;
 926                 default:
 927                         return ret;
 928                 }
 929
 930                 /*
 931                  * This happens sometimes if we don't have discards on -
 932                  * when we've partially overwritten a bucket with new
 933                  * journal entries. We don't need the rest of the
 934                  * bucket:
 935                  */
 936                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 937                         return 0;
 938
 939                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 940
 941                 csum_good = jset_csum_good(c, j);
 942                 if (!csum_good)
 943                         saw_bad = true;
 944
 945                 ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
 946                              j->encrypted_start,
 947                              vstruct_end(j) - (void *) j->encrypted_start);
 948                 bch2_fs_fatal_err_on(ret, c,
 949                                 "error decrypting journal entry: %i", ret);
 950
 951                 mutex_lock(&jlist->lock);
 952                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
 953                                         .csum_good      = csum_good,
 954                                         .dev            = ca->dev_idx,
 955                                         .bucket         = bucket,
 956                                         .bucket_offset  = offset -
 957                                                 bucket_to_sector(ca, ja->buckets[bucket]),
 958                                         .sector         = offset,
 959                                         }, jlist, j);
 960                 mutex_unlock(&jlist->lock);
 961
 962                 switch (ret) {
 963                 case JOURNAL_ENTRY_ADD_OK:
 964                         break;
 965                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
 966                         break;
 967                 default:
 968                         return ret;
 969                 }
 970 next_block:
 971                 pr_debug("next");
 972                 offset          += sectors;
 973                 sectors_read    -= sectors;
 974                 j = ((void *) j) + (sectors << 9);
 975         }
 976
 977         return 0;
 978 }
 979
 980 static void bch2_journal_read_device(struct closure *cl)
 981 {
 982         struct journal_device *ja =
 983                 container_of(cl, struct journal_device, read);
 984         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
 985         struct bch_fs *c = ca->fs;
 986         struct journal_list *jlist =
 987                 container_of(cl->parent, struct journal_list, cl);
 988         struct journal_replay *r, **_r;
 989         struct genradix_iter iter;
 990         struct journal_read_buf buf = { NULL, 0 };
 991         unsigned i;
 992         int ret = 0;
 993
 994         if (!ja->nr)
 995                 goto out;
 996
 997         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
 998         if (ret)
 999                 goto err;
1000
1001         pr_debug("%u journal buckets", ja->nr);
1002
1003         for (i = 0; i < ja->nr; i++) {
1004                 ret = journal_read_bucket(ca, &buf, jlist, i);
1005                 if (ret)
1006                         goto err;
1007         }
1008
1009         ja->sectors_free = ca->mi.bucket_size;
1010
1011         mutex_lock(&jlist->lock);
1012         genradix_for_each_reverse(&c->journal_entries, iter, _r) {
1013                 r = *_r;
1014
1015                 if (!r)
1016                         continue;
1017
1018                 for (i = 0; i < r->nr_ptrs; i++) {
1019                         if (r->ptrs[i].dev == ca->dev_idx) {
1020                                 unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
1021                                         vstruct_sectors(&r->j, c->block_bits);
1022
1023                                 ja->cur_idx = r->ptrs[i].bucket;
1024                                 ja->sectors_free = ca->mi.bucket_size - wrote;
1025                                 goto found;
1026                         }
1027                 }
1028         }
1029 found:
1030         mutex_unlock(&jlist->lock);
1031
1032         if (ja->bucket_seq[ja->cur_idx] &&
1033             ja->sectors_free == ca->mi.bucket_size) {
1034                 bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
1035                 bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
1036                 for (i = 0; i < 3; i++) {
1037                         unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
1038                         bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
1039                 }
1040                 ja->sectors_free = 0;
1041         }
1042
1043         /*
1044          * Set dirty_idx to indicate the entire journal is full and needs to be
1045          * reclaimed - journal reclaim will immediately reclaim whatever isn't
1046          * pinned when it first runs:
1047          */
1048         ja->discard_idx = ja->dirty_idx_ondisk =
1049                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
1050 out:
1051         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
1052         kvpfree(buf.data, buf.size);
1053         percpu_ref_put(&ca->io_ref);
1054         closure_return(cl);
1055         return;
1056 err:
1057         mutex_lock(&jlist->lock);
1058         jlist->ret = ret;
1059         mutex_unlock(&jlist->lock);
1060         goto out;
1061 }
1062
1063 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1064                                struct journal_replay *j)
1065 {
1066         unsigned i;
1067
1068         for (i = 0; i < j->nr_ptrs; i++) {
1069                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
1070                 u64 offset;
1071
1072                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
1073
1074                 if (i)
1075                         prt_printf(out, " ");
1076                 prt_printf(out, "%u:%u:%u (sector %llu)",
1077                        j->ptrs[i].dev,
1078                        j->ptrs[i].bucket,
1079                        j->ptrs[i].bucket_offset,
1080                        j->ptrs[i].sector);
1081         }
1082 }
1083
1084 int bch2_journal_read(struct bch_fs *c,
1085                       u64 *last_seq,
1086                       u64 *blacklist_seq,
1087                       u64 *start_seq)
1088 {
1089         struct journal_list jlist;
1090         struct journal_replay *i, **_i, *prev = NULL;
1091         struct genradix_iter radix_iter;
1092         struct bch_dev *ca;
1093         unsigned iter;
1094         struct printbuf buf = PRINTBUF;
1095         bool degraded = false, last_write_torn = false;
1096         u64 seq;
1097         int ret = 0;
1098
1099         closure_init_stack(&jlist.cl);
1100         mutex_init(&jlist.lock);
1101         jlist.last_seq = 0;
1102         jlist.ret = 0;
1103
1104         for_each_member_device(ca, c, iter) {
1105                 if (!c->opts.fsck &&
1106                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1107                         continue;
1108
1109                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1110                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1111                     percpu_ref_tryget(&ca->io_ref))
1112                         closure_call(&ca->journal.read,
1113                                      bch2_journal_read_device,
1114                                      system_unbound_wq,
1115                                      &jlist.cl);
1116                 else
1117                         degraded = true;
1118         }
1119
1120         closure_sync(&jlist.cl);
1121
1122         if (jlist.ret)
1123                 return jlist.ret;
1124
1125         *last_seq       = 0;
1126         *start_seq      = 0;
1127         *blacklist_seq  = 0;
1128
1129         /*
1130          * Find most recent flush entry, and ignore newer non flush entries -
1131          * those entries will be blacklisted:
1132          */
1133         genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
1134                 int write = READ;
1135
1136                 i = *_i;
1137
1138                 if (!i || i->ignore)
1139                         continue;
1140
1141                 if (!*start_seq)
1142                         *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
1143
1144                 if (JSET_NO_FLUSH(&i->j)) {
1145                         i->ignore = true;
1146                         continue;
1147                 }
1148
1149                 if (!last_write_torn && !i->csum_good) {
1150                         last_write_torn = true;
1151                         i->ignore = true;
1152                         continue;
1153                 }
1154
1155                 if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
1156                                          c, &i->j, NULL,
1157                                          "invalid journal entry: last_seq > seq (%llu > %llu)",
1158                                          le64_to_cpu(i->j.last_seq),
1159                                          le64_to_cpu(i->j.seq)))
1160                         i->j.last_seq = i->j.seq;
1161
1162                 *last_seq       = le64_to_cpu(i->j.last_seq);
1163                 *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1164                 break;
1165         }
1166
1167         if (!*start_seq) {
1168                 bch_info(c, "journal read done, but no entries found");
1169                 return 0;
1170         }
1171
1172         if (!*last_seq) {
1173                 fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
1174                 return 0;
1175         }
1176
1177         bch_info(c, "journal read done, replaying entries %llu-%llu",
1178                  *last_seq, *blacklist_seq - 1);
1179
1180         if (*start_seq != *blacklist_seq)
1181                 bch_info(c, "dropped unflushed entries %llu-%llu",
1182                          *blacklist_seq, *start_seq - 1);
1183
1184         /* Drop blacklisted entries and entries older than last_seq: */
1185         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1186                 i = *_i;
1187
1188                 if (!i || i->ignore)
1189                         continue;
1190
1191                 seq = le64_to_cpu(i->j.seq);
1192                 if (seq < *last_seq) {
1193                         journal_replay_free(c, i);
1194                         continue;
1195                 }
1196
1197                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1198                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1199                                     "found blacklisted journal entry %llu", seq);
1200                         i->ignore = true;
1201                 }
1202         }
1203
1204         /* Check for missing entries: */
1205         seq = *last_seq;
1206         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1207                 i = *_i;
1208
1209                 if (!i || i->ignore)
1210                         continue;
1211
1212                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1213
1214                 while (seq < le64_to_cpu(i->j.seq)) {
1215                         u64 missing_start, missing_end;
1216                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1217
1218                         while (seq < le64_to_cpu(i->j.seq) &&
1219                                bch2_journal_seq_is_blacklisted(c, seq, false))
1220                                 seq++;
1221
1222                         if (seq == le64_to_cpu(i->j.seq))
1223                                 break;
1224
1225                         missing_start = seq;
1226
1227                         while (seq < le64_to_cpu(i->j.seq) &&
1228                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1229                                 seq++;
1230
1231                         if (prev) {
1232                                 bch2_journal_ptrs_to_text(&buf1, c, prev);
1233                                 prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits));
1234                         } else
1235                                 prt_printf(&buf1, "(none)");
1236                         bch2_journal_ptrs_to_text(&buf2, c, i);
1237
1238                         missing_end = seq - 1;
1239                         fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1240                                  "  prev at %s\n"
1241                                  "  next at %s",
1242                                  missing_start, missing_end,
1243                                  *last_seq, *blacklist_seq - 1,
1244                                  buf1.buf, buf2.buf);
1245
1246                         printbuf_exit(&buf1);
1247                         printbuf_exit(&buf2);
1248                 }
1249
1250                 prev = i;
1251                 seq++;
1252         }
1253
1254         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1255                 struct bch_replicas_padded replicas = {
1256                         .e.data_type = BCH_DATA_journal,
1257                         .e.nr_required = 1,
1258                 };
1259                 unsigned ptr;
1260
1261                 i = *_i;
1262                 if (!i || i->ignore)
1263                         continue;
1264
1265                 for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
1266                         struct bch_dev *ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
1267
1268                         if (!i->ptrs[ptr].csum_good)
1269                                 bch_err_dev_offset(ca, i->ptrs[ptr].sector,
1270                                                    "invalid journal checksum, seq %llu%s",
1271                                                    le64_to_cpu(i->j.seq),
1272                                                    i->csum_good ? " (had good copy on another device)" : "");
1273                 }
1274
1275                 ret = jset_validate(c,
1276                                     bch_dev_bkey_exists(c, i->ptrs[0].dev),
1277                                     &i->j,
1278                                     i->ptrs[0].sector,
1279                                     READ);
1280                 if (ret)
1281                         goto err;
1282
1283                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1284                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1285
1286                 bch2_replicas_entry_sort(&replicas.e);
1287
1288                 /*
1289                  * If we're mounting in degraded mode - if we didn't read all
1290                  * the devices - this is wrong:
1291                  */
1292
1293                 printbuf_reset(&buf);
1294                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1295
1296                 if (!degraded &&
1297                     fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
1298                                 "superblock not marked as containing replicas %s",
1299                                 buf.buf)) {
1300                         ret = bch2_mark_replicas(c, &replicas.e);
1301                         if (ret)
1302                                 goto err;
1303                 }
1304         }
1305 err:
1306 fsck_err:
1307         printbuf_exit(&buf);
1308         return ret;
1309 }
1310
1311 /* journal write: */
1312
1313 static void __journal_write_alloc(struct journal *j,
1314                                   struct journal_buf *w,
1315                                   struct dev_alloc_list *devs_sorted,
1316                                   unsigned sectors,
1317                                   unsigned *replicas,
1318                                   unsigned replicas_want)
1319 {
1320         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1321         struct journal_device *ja;
1322         struct bch_dev *ca;
1323         unsigned i;
1324
1325         if (*replicas >= replicas_want)
1326                 return;
1327
1328         for (i = 0; i < devs_sorted->nr; i++) {
1329                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1330                 if (!ca)
1331                         continue;
1332
1333                 ja = &ca->journal;
1334
1335                 /*
1336                  * Check that we can use this device, and aren't already using
1337                  * it:
1338                  */
1339                 if (!ca->mi.durability ||
1340                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1341                     !ja->nr ||
1342                     bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) ||
1343                     sectors > ja->sectors_free)
1344                         continue;
1345
1346                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1347
1348                 bch2_bkey_append_ptr(&w->key,
1349                         (struct bch_extent_ptr) {
1350                                   .offset = bucket_to_sector(ca,
1351                                         ja->buckets[ja->cur_idx]) +
1352                                         ca->mi.bucket_size -
1353                                         ja->sectors_free,
1354                                   .dev = ca->dev_idx,
1355                 });
1356
1357                 ja->sectors_free -= sectors;
1358                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1359
1360                 *replicas += ca->mi.durability;
1361
1362                 if (*replicas >= replicas_want)
1363                         break;
1364         }
1365 }
1366
1367 /**
1368  * journal_next_bucket - move on to the next journal bucket if possible
1369  */
1370 static int journal_write_alloc(struct journal *j, struct journal_buf *w,
1371                                unsigned sectors)
1372 {
1373         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1374         struct bch_devs_mask devs;
1375         struct journal_device *ja;
1376         struct bch_dev *ca;
1377         struct dev_alloc_list devs_sorted;
1378         unsigned target = c->opts.metadata_target ?:
1379                 c->opts.foreground_target;
1380         unsigned i, replicas = 0, replicas_want =
1381                 READ_ONCE(c->opts.metadata_replicas);
1382
1383         rcu_read_lock();
1384 retry:
1385         devs = target_rw_devs(c, BCH_DATA_journal, target);
1386
1387         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1388
1389         __journal_write_alloc(j, w, &devs_sorted,
1390                               sectors, &replicas, replicas_want);
1391
1392         if (replicas >= replicas_want)
1393                 goto done;
1394
1395         for (i = 0; i < devs_sorted.nr; i++) {
1396                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1397                 if (!ca)
1398                         continue;
1399
1400                 ja = &ca->journal;
1401
1402                 if (sectors > ja->sectors_free &&
1403                     sectors <= ca->mi.bucket_size &&
1404                     bch2_journal_dev_buckets_available(j, ja,
1405                                         journal_space_discarded)) {
1406                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1407                         ja->sectors_free = ca->mi.bucket_size;
1408
1409                         /*
1410                          * ja->bucket_seq[ja->cur_idx] must always have
1411                          * something sensible:
1412                          */
1413                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1414                 }
1415         }
1416
1417         __journal_write_alloc(j, w, &devs_sorted,
1418                               sectors, &replicas, replicas_want);
1419
1420         if (replicas < replicas_want && target) {
1421                 /* Retry from all devices: */
1422                 target = 0;
1423                 goto retry;
1424         }
1425 done:
1426         rcu_read_unlock();
1427
1428         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1429
1430         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1431 }
1432
1433 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1434 {
1435         /* we aren't holding j->lock: */
1436         unsigned new_size = READ_ONCE(j->buf_size_want);
1437         void *new_buf;
1438
1439         if (buf->buf_size >= new_size)
1440                 return;
1441
1442         new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN);
1443         if (!new_buf)
1444                 return;
1445
1446         memcpy(new_buf, buf->data, buf->buf_size);
1447
1448         spin_lock(&j->lock);
1449         swap(buf->data,         new_buf);
1450         swap(buf->buf_size,     new_size);
1451         spin_unlock(&j->lock);
1452
1453         kvpfree(new_buf, new_size);
1454 }
1455
1456 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1457 {
1458         return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
1459 }
1460
1461 static void journal_write_done(struct closure *cl)
1462 {
1463         struct journal *j = container_of(cl, struct journal, io);
1464         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1465         struct journal_buf *w = journal_last_unwritten_buf(j);
1466         struct bch_replicas_padded replicas;
1467         union journal_res_state old, new;
1468         u64 v, seq;
1469         int err = 0;
1470
1471         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1472                                ? j->flush_write_time
1473                                : j->noflush_write_time, j->write_start_time);
1474
1475         if (!w->devs_written.nr) {
1476                 bch_err(c, "unable to write journal to sufficient devices");
1477                 err = -EIO;
1478         } else {
1479                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1480                                          w->devs_written);
1481                 if (bch2_mark_replicas(c, &replicas.e))
1482                         err = -EIO;
1483         }
1484
1485         if (err)
1486                 bch2_fatal_error(c);
1487
1488         spin_lock(&j->lock);
1489         seq = le64_to_cpu(w->data->seq);
1490
1491         if (seq >= j->pin.front)
1492                 journal_seq_pin(j, seq)->devs = w->devs_written;
1493
1494         if (!err) {
1495                 if (!JSET_NO_FLUSH(w->data)) {
1496                         j->flushed_seq_ondisk = seq;
1497                         j->last_seq_ondisk = w->last_seq;
1498
1499                         bch2_do_discards(c);
1500                         closure_wake_up(&c->freelist_wait);
1501
1502                         bch2_reset_alloc_cursors(c);
1503                 }
1504         } else if (!j->err_seq || seq < j->err_seq)
1505                 j->err_seq      = seq;
1506
1507         j->seq_ondisk           = seq;
1508
1509         /*
1510          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1511          * more buckets:
1512          *
1513          * Must come before signaling write completion, for
1514          * bch2_fs_journal_stop():
1515          */
1516         if (j->watermark)
1517                 journal_reclaim_kick(&c->journal);
1518
1519         /* also must come before signalling write completion: */
1520         closure_debug_destroy(cl);
1521
1522         v = atomic64_read(&j->reservations.counter);
1523         do {
1524                 old.v = new.v = v;
1525                 BUG_ON(journal_state_count(new, new.unwritten_idx));
1526
1527                 new.unwritten_idx++;
1528         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1529                                        old.v, new.v)) != old.v);
1530
1531         bch2_journal_space_available(j);
1532
1533         closure_wake_up(&w->wait);
1534         journal_wake(j);
1535
1536         if (!journal_state_count(new, new.unwritten_idx) &&
1537             journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
1538                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1539         } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
1540                    new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
1541                 struct journal_buf *buf = journal_cur_buf(j);
1542                 long delta = buf->expires - jiffies;
1543
1544                 /*
1545                  * We don't close a journal entry to write it while there's
1546                  * previous entries still in flight - the current journal entry
1547                  * might want to be written now:
1548                  */
1549
1550                 mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
1551         }
1552
1553         spin_unlock(&j->lock);
1554 }
1555
1556 static void journal_write_endio(struct bio *bio)
1557 {
1558         struct bch_dev *ca = bio->bi_private;
1559         struct journal *j = &ca->fs->journal;
1560         struct journal_buf *w = journal_last_unwritten_buf(j);
1561         unsigned long flags;
1562
1563         if (bch2_dev_io_err_on(bio->bi_status, ca, "error writing journal entry %llu: %s",
1564                                le64_to_cpu(w->data->seq),
1565                                bch2_blk_status_to_str(bio->bi_status)) ||
1566             bch2_meta_write_fault("journal")) {
1567                 spin_lock_irqsave(&j->err_lock, flags);
1568                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1569                 spin_unlock_irqrestore(&j->err_lock, flags);
1570         }
1571
1572         closure_put(&j->io);
1573         percpu_ref_put(&ca->io_ref);
1574 }
1575
1576 static void do_journal_write(struct closure *cl)
1577 {
1578         struct journal *j = container_of(cl, struct journal, io);
1579         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1580         struct bch_dev *ca;
1581         struct journal_buf *w = journal_last_unwritten_buf(j);
1582         struct bch_extent_ptr *ptr;
1583         struct bio *bio;
1584         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1585
1586         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1587                 ca = bch_dev_bkey_exists(c, ptr->dev);
1588                 if (!percpu_ref_tryget(&ca->io_ref)) {
1589                         /* XXX: fix this */
1590                         bch_err(c, "missing device for journal write\n");
1591                         continue;
1592                 }
1593
1594                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1595                              sectors);
1596
1597                 bio = ca->journal.bio;
1598                 bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
1599                 bio->bi_iter.bi_sector  = ptr->offset;
1600                 bio->bi_end_io          = journal_write_endio;
1601                 bio->bi_private         = ca;
1602
1603                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1604                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1605
1606                 if (!JSET_NO_FLUSH(w->data))
1607                         bio->bi_opf    |= REQ_FUA;
1608                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1609                         bio->bi_opf    |= REQ_PREFLUSH;
1610
1611                 bch2_bio_map(bio, w->data, sectors << 9);
1612
1613                 trace_and_count(c, journal_write, bio);
1614                 closure_bio_submit(bio, cl);
1615
1616                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1617                         le64_to_cpu(w->data->seq);
1618         }
1619
1620         continue_at(cl, journal_write_done, c->io_complete_wq);
1621         return;
1622 }
1623
1624 static void bch2_journal_entries_postprocess(struct bch_fs *c, struct jset *jset)
1625 {
1626         struct jset_entry *i, *next, *prev = NULL;
1627
1628         /*
1629          * Simple compaction, dropping empty jset_entries (from journal
1630          * reservations that weren't fully used) and merging jset_entries that
1631          * can be.
1632          *
1633          * If we wanted to be really fancy here, we could sort all the keys in
1634          * the jset and drop keys that were overwritten - probably not worth it:
1635          */
1636         vstruct_for_each_safe(jset, i, next) {
1637                 unsigned u64s = le16_to_cpu(i->u64s);
1638
1639                 /* Empty entry: */
1640                 if (!u64s)
1641                         continue;
1642
1643                 if (i->type == BCH_JSET_ENTRY_btree_root)
1644                         bch2_journal_entry_to_btree_root(c, i);
1645
1646                 /* Can we merge with previous entry? */
1647                 if (prev &&
1648                     i->btree_id == prev->btree_id &&
1649                     i->level    == prev->level &&
1650                     i->type     == prev->type &&
1651                     i->type     == BCH_JSET_ENTRY_btree_keys &&
1652                     le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
1653                         memmove_u64s_down(vstruct_next(prev),
1654                                           i->_data,
1655                                           u64s);
1656                         le16_add_cpu(&prev->u64s, u64s);
1657                         continue;
1658                 }
1659
1660                 /* Couldn't merge, move i into new position (after prev): */
1661                 prev = prev ? vstruct_next(prev) : jset->start;
1662                 if (i != prev)
1663                         memmove_u64s_down(prev, i, jset_u64s(u64s));
1664         }
1665
1666         prev = prev ? vstruct_next(prev) : jset->start;
1667         jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
1668 }
1669
1670 void bch2_journal_write(struct closure *cl)
1671 {
1672         struct journal *j = container_of(cl, struct journal, io);
1673         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1674         struct bch_dev *ca;
1675         struct journal_buf *w = journal_last_unwritten_buf(j);
1676         struct jset_entry *start, *end;
1677         struct jset *jset;
1678         struct bio *bio;
1679         struct printbuf journal_debug_buf = PRINTBUF;
1680         bool validate_before_checksum = false;
1681         unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
1682         int ret;
1683
1684         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1685
1686         journal_buf_realloc(j, w);
1687         jset = w->data;
1688
1689         j->write_start_time = local_clock();
1690
1691         spin_lock(&j->lock);
1692
1693         /*
1694          * If the journal is in an error state - we did an emergency shutdown -
1695          * we prefer to continue doing journal writes. We just mark them as
1696          * noflush so they'll never be used, but they'll still be visible by the
1697          * list_journal tool - this helps in debugging.
1698          *
1699          * There's a caveat: the first journal write after marking the
1700          * superblock dirty must always be a flush write, because on startup
1701          * from a clean shutdown we didn't necessarily read the journal and the
1702          * new journal write might overwrite whatever was in the journal
1703          * previously - we can't leave the journal without any flush writes in
1704          * it.
1705          *
1706          * So if we're in an error state, and we're still starting up, we don't
1707          * write anything at all.
1708          */
1709         if (!test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags) &&
1710             (bch2_journal_error(j) ||
1711              w->noflush ||
1712              (!w->must_flush &&
1713               (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1714               test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)))) {
1715                 w->noflush = true;
1716                 SET_JSET_NO_FLUSH(jset, true);
1717                 jset->last_seq  = 0;
1718                 w->last_seq     = 0;
1719
1720                 j->nr_noflush_writes++;
1721         } else if (!bch2_journal_error(j)) {
1722                 j->last_flush_write = jiffies;
1723                 j->nr_flush_writes++;
1724                 clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
1725         } else {
1726                 spin_unlock(&j->lock);
1727                 goto err;
1728         }
1729         spin_unlock(&j->lock);
1730
1731         /*
1732          * New btree roots are set by journalling them; when the journal entry
1733          * gets written we have to propagate them to c->btree_roots
1734          *
1735          * But, every journal entry we write has to contain all the btree roots
1736          * (at least for now); so after we copy btree roots to c->btree_roots we
1737          * have to get any missing btree roots and add them to this journal
1738          * entry:
1739          */
1740
1741         bch2_journal_entries_postprocess(c, jset);
1742
1743         start = end = vstruct_last(jset);
1744
1745         end     = bch2_btree_roots_to_journal_entries(c, jset->start, end);
1746
1747         bch2_journal_super_entries_add_common(c, &end,
1748                                 le64_to_cpu(jset->seq));
1749         u64s    = (u64 *) end - (u64 *) start;
1750         BUG_ON(u64s > j->entry_u64s_reserved);
1751
1752         le32_add_cpu(&jset->u64s, u64s);
1753         BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors);
1754
1755         jset->magic             = cpu_to_le64(jset_magic(c));
1756         jset->version           = c->sb.version < bcachefs_metadata_version_bkey_renumber
1757                 ? cpu_to_le32(BCH_JSET_VERSION_OLD)
1758                 : cpu_to_le32(c->sb.version);
1759
1760         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1761         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1762
1763         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1764                 j->last_empty_seq = le64_to_cpu(jset->seq);
1765
1766         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1767                 validate_before_checksum = true;
1768
1769         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1770                 validate_before_checksum = true;
1771
1772         if (validate_before_checksum &&
1773             jset_validate(c, NULL, jset, 0, WRITE))
1774                 goto err;
1775
1776         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1777                     jset->encrypted_start,
1778                     vstruct_end(jset) - (void *) jset->encrypted_start);
1779         if (bch2_fs_fatal_err_on(ret, c,
1780                         "error decrypting journal entry: %i", ret))
1781                 goto err;
1782
1783         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1784                                   journal_nonce(jset), jset);
1785
1786         if (!validate_before_checksum &&
1787             jset_validate(c, NULL, jset, 0, WRITE))
1788                 goto err;
1789
1790         sectors = vstruct_sectors(jset, c->block_bits);
1791         BUG_ON(sectors > w->sectors);
1792
1793         bytes = vstruct_bytes(jset);
1794         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1795
1796 retry_alloc:
1797         spin_lock(&j->lock);
1798         ret = journal_write_alloc(j, w, sectors);
1799
1800         if (ret && j->can_discard) {
1801                 spin_unlock(&j->lock);
1802                 bch2_journal_do_discards(j);
1803                 goto retry_alloc;
1804         }
1805
1806         if (ret)
1807                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1808
1809         /*
1810          * write is allocated, no longer need to account for it in
1811          * bch2_journal_space_available():
1812          */
1813         w->sectors = 0;
1814
1815         /*
1816          * journal entry has been compacted and allocated, recalculate space
1817          * available:
1818          */
1819         bch2_journal_space_available(j);
1820         spin_unlock(&j->lock);
1821
1822         if (ret) {
1823                 bch_err(c, "Unable to allocate journal write:\n%s",
1824                         journal_debug_buf.buf);
1825                 printbuf_exit(&journal_debug_buf);
1826                 bch2_fatal_error(c);
1827                 continue_at(cl, journal_write_done, c->io_complete_wq);
1828                 return;
1829         }
1830
1831         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1832
1833         if (c->opts.nochanges)
1834                 goto no_io;
1835
1836         for_each_rw_member(ca, c, i)
1837                 nr_rw_members++;
1838
1839         if (nr_rw_members > 1)
1840                 w->separate_flush = true;
1841
1842         if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
1843                 for_each_rw_member(ca, c, i) {
1844                         percpu_ref_get(&ca->io_ref);
1845
1846                         bio = ca->journal.bio;
1847                         bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
1848                         bio->bi_end_io          = journal_write_endio;
1849                         bio->bi_private         = ca;
1850                         closure_bio_submit(bio, cl);
1851                 }
1852         }
1853
1854         continue_at(cl, do_journal_write, c->io_complete_wq);
1855         return;
1856 no_io:
1857         continue_at(cl, journal_write_done, c->io_complete_wq);
1858         return;
1859 err:
1860         bch2_fatal_error(c);
1861         continue_at(cl, journal_write_done, c->io_complete_wq);
1862 }