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_foreground.h"
   4 #include "btree_io.h"
   5 #include "btree_update_interior.h"
   6 #include "buckets.h"
   7 #include "checksum.h"
   8 #include "disk_groups.h"
   9 #include "error.h"
  10 #include "io.h"
  11 #include "journal.h"
  12 #include "journal_io.h"
  13 #include "journal_reclaim.h"
  14 #include "journal_seq_blacklist.h"
  15 #include "replicas.h"
  16
  17 #include <trace/events/bcachefs.h>
  18
  19 static void __journal_replay_free(struct journal_replay *i)
  20 {
  21         list_del(&i->list);
  22         kvpfree(i, offsetof(struct journal_replay, j) +
  23                 vstruct_bytes(&i->j));
  24
  25 }
  26
  27 static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
  28 {
  29         i->ignore = true;
  30
  31         if (!c->opts.read_entire_journal)
  32                 __journal_replay_free(i);
  33 }
  34
  35 struct journal_list {
  36         struct closure          cl;
  37         struct mutex            lock;
  38         struct list_head        *head;
  39         int                     ret;
  40 };
  41
  42 #define JOURNAL_ENTRY_ADD_OK            0
  43 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  44
  45 /*
  46  * Given a journal entry we just read, add it to the list of journal entries to
  47  * be replayed:
  48  */
  49 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  50                              struct journal_ptr entry_ptr,
  51                              struct journal_list *jlist, struct jset *j,
  52                              bool bad)
  53 {
  54         struct journal_replay *i, *pos, *dup = NULL;
  55         struct journal_ptr *ptr;
  56         struct list_head *where;
  57         size_t bytes = vstruct_bytes(j);
  58         u64 last_seq = 0;
  59         int ret = JOURNAL_ENTRY_ADD_OK;
  60
  61         list_for_each_entry_reverse(i, jlist->head, list) {
  62                 if (!JSET_NO_FLUSH(&i->j)) {
  63                         last_seq = le64_to_cpu(i->j.last_seq);
  64                         break;
  65                 }
  66         }
  67
  68         /* Is this entry older than the range we need? */
  69         if (!c->opts.read_entire_journal &&
  70             le64_to_cpu(j->seq) < last_seq) {
  71                 ret = JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  72                 goto out;
  73         }
  74
  75         /* Drop entries we don't need anymore */
  76         if (!JSET_NO_FLUSH(j)) {
  77                 list_for_each_entry_safe(i, pos, jlist->head, list) {
  78                         if (le64_to_cpu(i->j.seq) >= le64_to_cpu(j->last_seq))
  79                                 break;
  80                         journal_replay_free(c, i);
  81                 }
  82         }
  83
  84         list_for_each_entry_reverse(i, jlist->head, list) {
  85                 if (le64_to_cpu(j->seq) > le64_to_cpu(i->j.seq)) {
  86                         where = &i->list;
  87                         goto add;
  88                 }
  89         }
  90
  91         where = jlist->head;
  92 add:
  93         dup = where->next != jlist->head
  94                 ? container_of(where->next, struct journal_replay, list)
  95                 : NULL;
  96
  97         if (dup && le64_to_cpu(j->seq) != le64_to_cpu(dup->j.seq))
  98                 dup = NULL;
  99
 100         /*
 101          * Duplicate journal entries? If so we want the one that didn't have a
 102          * checksum error:
 103          */
 104         if (dup) {
 105                 if (dup->bad) {
 106                         /* we'll replace @dup: */
 107                 } else if (bad) {
 108                         i = dup;
 109                         goto found;
 110                 } else {
 111                         fsck_err_on(bytes != vstruct_bytes(&dup->j) ||
 112                                     memcmp(j, &dup->j, bytes), c,
 113                                     "found duplicate but non identical journal entries (seq %llu)",
 114                                     le64_to_cpu(j->seq));
 115                         i = dup;
 116                         goto found;
 117                 }
 118         }
 119
 120         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 121         if (!i) {
 122                 ret = -ENOMEM;
 123                 goto out;
 124         }
 125
 126         i->nr_ptrs       = 0;
 127         i->bad          = bad;
 128         i->ignore       = false;
 129         memcpy(&i->j, j, bytes);
 130
 131         if (dup) {
 132                 i->nr_ptrs = dup->nr_ptrs;
 133                 memcpy(i->ptrs, dup->ptrs, sizeof(dup->ptrs));
 134                 __journal_replay_free(dup);
 135         }
 136
 137         list_add(&i->list, where);
 138 found:
 139         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 140                 if (ptr->dev == ca->dev_idx) {
 141                         bch_err(c, "duplicate journal entry %llu on same device",
 142                                 le64_to_cpu(i->j.seq));
 143                         goto out;
 144                 }
 145         }
 146
 147         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 148                 bch_err(c, "found too many copies of journal entry %llu",
 149                         le64_to_cpu(i->j.seq));
 150                 goto out;
 151         }
 152
 153         i->ptrs[i->nr_ptrs++] = entry_ptr;
 154 out:
 155 fsck_err:
 156         return ret;
 157 }
 158
 159 static struct nonce journal_nonce(const struct jset *jset)
 160 {
 161         return (struct nonce) {{
 162                 [0] = 0,
 163                 [1] = ((__le32 *) &jset->seq)[0],
 164                 [2] = ((__le32 *) &jset->seq)[1],
 165                 [3] = BCH_NONCE_JOURNAL,
 166         }};
 167 }
 168
 169 /* this fills in a range with empty jset_entries: */
 170 static void journal_entry_null_range(void *start, void *end)
 171 {
 172         struct jset_entry *entry;
 173
 174         for (entry = start; entry != end; entry = vstruct_next(entry))
 175                 memset(entry, 0, sizeof(*entry));
 176 }
 177
 178 #define JOURNAL_ENTRY_REREAD    5
 179 #define JOURNAL_ENTRY_NONE      6
 180 #define JOURNAL_ENTRY_BAD       7
 181
 182 #define journal_entry_err(c, msg, ...)                                  \
 183 ({                                                                      \
 184         switch (write) {                                                \
 185         case READ:                                                      \
 186                 mustfix_fsck_err(c, msg, ##__VA_ARGS__);                \
 187                 break;                                                  \
 188         case WRITE:                                                     \
 189                 bch_err(c, "corrupt metadata before write:\n"           \
 190                         msg, ##__VA_ARGS__);                            \
 191                 if (bch2_fs_inconsistent(c)) {                          \
 192                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 193                         goto fsck_err;                                  \
 194                 }                                                       \
 195                 break;                                                  \
 196         }                                                               \
 197         true;                                                           \
 198 })
 199
 200 #define journal_entry_err_on(cond, c, msg, ...)                         \
 201         ((cond) ? journal_entry_err(c, msg, ##__VA_ARGS__) : false)
 202
 203 #define FSCK_DELETED_KEY        5
 204
 205 static int journal_validate_key(struct bch_fs *c, const char *where,
 206                                 struct jset_entry *entry,
 207                                 unsigned level, enum btree_id btree_id,
 208                                 struct bkey_i *k, const char *type,
 209                                 unsigned version, int big_endian, int write)
 210 {
 211         void *next = vstruct_next(entry);
 212         const char *invalid;
 213         int ret = 0;
 214
 215         if (journal_entry_err_on(!k->k.u64s, c,
 216                         "invalid %s in %s entry offset %zi/%u: k->u64s 0",
 217                         type, where,
 218                         (u64 *) k - entry->_data,
 219                         le16_to_cpu(entry->u64s))) {
 220                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 221                 journal_entry_null_range(vstruct_next(entry), next);
 222                 return FSCK_DELETED_KEY;
 223         }
 224
 225         if (journal_entry_err_on((void *) bkey_next(k) >
 226                                 (void *) vstruct_next(entry), c,
 227                         "invalid %s in %s entry offset %zi/%u: extends past end of journal entry",
 228                         type, where,
 229                         (u64 *) k - entry->_data,
 230                         le16_to_cpu(entry->u64s))) {
 231                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 232                 journal_entry_null_range(vstruct_next(entry), next);
 233                 return FSCK_DELETED_KEY;
 234         }
 235
 236         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT, c,
 237                         "invalid %s in %s entry offset %zi/%u: bad format %u",
 238                         type, where,
 239                         (u64 *) k - entry->_data,
 240                         le16_to_cpu(entry->u64s),
 241                         k->k.format)) {
 242                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 243                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 244                 journal_entry_null_range(vstruct_next(entry), next);
 245                 return FSCK_DELETED_KEY;
 246         }
 247
 248         if (!write)
 249                 bch2_bkey_compat(level, btree_id, version, big_endian,
 250                                  write, NULL, bkey_to_packed(k));
 251
 252         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 253                                     __btree_node_type(level, btree_id));
 254         if (invalid) {
 255                 struct printbuf buf = PRINTBUF;
 256
 257                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
 258                 mustfix_fsck_err(c, "invalid %s in %s entry offset %zi/%u: %s\n%s",
 259                                  type, where,
 260                                  (u64 *) k - entry->_data,
 261                                  le16_to_cpu(entry->u64s),
 262                                  invalid, buf.buf);
 263                 printbuf_exit(&buf);
 264
 265                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 266                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 267                 journal_entry_null_range(vstruct_next(entry), next);
 268                 return FSCK_DELETED_KEY;
 269         }
 270
 271         if (write)
 272                 bch2_bkey_compat(level, btree_id, version, big_endian,
 273                                  write, NULL, bkey_to_packed(k));
 274 fsck_err:
 275         return ret;
 276 }
 277
 278 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 279                                              const char *where,
 280                                              struct jset_entry *entry,
 281                                              unsigned version, int big_endian, int write)
 282 {
 283         struct bkey_i *k = entry->start;
 284
 285         while (k != vstruct_last(entry)) {
 286                 int ret = journal_validate_key(c, where, entry,
 287                                                entry->level,
 288                                                entry->btree_id,
 289                                                k, "key", version, big_endian, write);
 290                 if (ret == FSCK_DELETED_KEY)
 291                         continue;
 292
 293                 k = bkey_next(k);
 294         }
 295
 296         return 0;
 297 }
 298
 299 static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
 300                                              struct jset_entry *entry)
 301 {
 302         struct bkey_i *k;
 303         bool first = true;
 304
 305         vstruct_for_each(entry, k) {
 306                 if (!first) {
 307                         pr_newline(out);
 308                         pr_buf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 309                 }
 310                 pr_buf(out, "btree=%s l=%u ", bch2_btree_ids[entry->btree_id], entry->level);
 311                 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 312                 first = false;
 313         }
 314 }
 315
 316 static int journal_entry_btree_root_validate(struct bch_fs *c,
 317                                              const char *where,
 318                                              struct jset_entry *entry,
 319                                              unsigned version, int big_endian, int write)
 320 {
 321         struct bkey_i *k = entry->start;
 322         int ret = 0;
 323
 324         if (journal_entry_err_on(!entry->u64s ||
 325                                  le16_to_cpu(entry->u64s) != k->k.u64s, c,
 326                                  "invalid btree root journal entry: wrong number of keys")) {
 327                 void *next = vstruct_next(entry);
 328                 /*
 329                  * we don't want to null out this jset_entry,
 330                  * just the contents, so that later we can tell
 331                  * we were _supposed_ to have a btree root
 332                  */
 333                 entry->u64s = 0;
 334                 journal_entry_null_range(vstruct_next(entry), next);
 335                 return 0;
 336         }
 337
 338         return journal_validate_key(c, where, entry, 1, entry->btree_id, k,
 339                                     "btree root", version, big_endian, write);
 340 fsck_err:
 341         return ret;
 342 }
 343
 344 static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
 345                                              struct jset_entry *entry)
 346 {
 347         journal_entry_btree_keys_to_text(out, c, entry);
 348 }
 349
 350 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 351                                             const char *where,
 352                                             struct jset_entry *entry,
 353                                             unsigned version, int big_endian, int write)
 354 {
 355         /* obsolete, don't care: */
 356         return 0;
 357 }
 358
 359 static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 360                                             struct jset_entry *entry)
 361 {
 362 }
 363
 364 static int journal_entry_blacklist_validate(struct bch_fs *c,
 365                                             const char *where,
 366                                             struct jset_entry *entry,
 367                                             unsigned version, int big_endian, int write)
 368 {
 369         int ret = 0;
 370
 371         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1, c,
 372                 "invalid journal seq blacklist entry: bad size")) {
 373                 journal_entry_null_range(entry, vstruct_next(entry));
 374         }
 375 fsck_err:
 376         return ret;
 377 }
 378
 379 static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
 380                                             struct jset_entry *entry)
 381 {
 382         struct jset_entry_blacklist *bl =
 383                 container_of(entry, struct jset_entry_blacklist, entry);
 384
 385         pr_buf(out, "seq=%llu", le64_to_cpu(bl->seq));
 386 }
 387
 388 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 389                                                const char *where,
 390                                                struct jset_entry *entry,
 391                                                unsigned version, int big_endian, int write)
 392 {
 393         struct jset_entry_blacklist_v2 *bl_entry;
 394         int ret = 0;
 395
 396         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2, c,
 397                 "invalid journal seq blacklist entry: bad size")) {
 398                 journal_entry_null_range(entry, vstruct_next(entry));
 399                 goto out;
 400         }
 401
 402         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 403
 404         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 405                                  le64_to_cpu(bl_entry->end), c,
 406                 "invalid journal seq blacklist entry: start > end")) {
 407                 journal_entry_null_range(entry, vstruct_next(entry));
 408         }
 409 out:
 410 fsck_err:
 411         return ret;
 412 }
 413
 414 static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
 415                                                struct jset_entry *entry)
 416 {
 417         struct jset_entry_blacklist_v2 *bl =
 418                 container_of(entry, struct jset_entry_blacklist_v2, entry);
 419
 420         pr_buf(out, "start=%llu end=%llu",
 421                le64_to_cpu(bl->start),
 422                le64_to_cpu(bl->end));
 423 }
 424
 425 static int journal_entry_usage_validate(struct bch_fs *c,
 426                                         const char *where,
 427                                         struct jset_entry *entry,
 428                                         unsigned version, int big_endian, int write)
 429 {
 430         struct jset_entry_usage *u =
 431                 container_of(entry, struct jset_entry_usage, entry);
 432         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 433         int ret = 0;
 434
 435         if (journal_entry_err_on(bytes < sizeof(*u),
 436                                  c,
 437                                  "invalid journal entry usage: bad size")) {
 438                 journal_entry_null_range(entry, vstruct_next(entry));
 439                 return ret;
 440         }
 441
 442 fsck_err:
 443         return ret;
 444 }
 445
 446 static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 447                                         struct jset_entry *entry)
 448 {
 449         struct jset_entry_usage *u =
 450                 container_of(entry, struct jset_entry_usage, entry);
 451
 452         pr_buf(out, "type=%s v=%llu",
 453                bch2_fs_usage_types[u->entry.btree_id],
 454                le64_to_cpu(u->v));
 455 }
 456
 457 static int journal_entry_data_usage_validate(struct bch_fs *c,
 458                                         const char *where,
 459                                         struct jset_entry *entry,
 460                                         unsigned version, int big_endian, int write)
 461 {
 462         struct jset_entry_data_usage *u =
 463                 container_of(entry, struct jset_entry_data_usage, entry);
 464         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 465         int ret = 0;
 466
 467         if (journal_entry_err_on(bytes < sizeof(*u) ||
 468                                  bytes < sizeof(*u) + u->r.nr_devs,
 469                                  c,
 470                                  "invalid journal entry usage: bad size")) {
 471                 journal_entry_null_range(entry, vstruct_next(entry));
 472                 return ret;
 473         }
 474
 475 fsck_err:
 476         return ret;
 477 }
 478
 479 static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
 480                                              struct jset_entry *entry)
 481 {
 482         struct jset_entry_data_usage *u =
 483                 container_of(entry, struct jset_entry_data_usage, entry);
 484
 485         bch2_replicas_entry_to_text(out, &u->r);
 486         pr_buf(out, "=%llu", le64_to_cpu(u->v));
 487 }
 488
 489 static int journal_entry_clock_validate(struct bch_fs *c,
 490                                         const char *where,
 491                                         struct jset_entry *entry,
 492                                         unsigned version, int big_endian, int write)
 493 {
 494         struct jset_entry_clock *clock =
 495                 container_of(entry, struct jset_entry_clock, entry);
 496         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 497         int ret = 0;
 498
 499         if (journal_entry_err_on(bytes != sizeof(*clock),
 500                                  c, "invalid journal entry clock: bad size")) {
 501                 journal_entry_null_range(entry, vstruct_next(entry));
 502                 return ret;
 503         }
 504
 505         if (journal_entry_err_on(clock->rw > 1,
 506                                  c, "invalid journal entry clock: bad rw")) {
 507                 journal_entry_null_range(entry, vstruct_next(entry));
 508                 return ret;
 509         }
 510
 511 fsck_err:
 512         return ret;
 513 }
 514
 515 static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 516                                         struct jset_entry *entry)
 517 {
 518         struct jset_entry_clock *clock =
 519                 container_of(entry, struct jset_entry_clock, entry);
 520
 521         pr_buf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
 522 }
 523
 524 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 525                                             const char *where,
 526                                             struct jset_entry *entry,
 527                                             unsigned version, int big_endian, int write)
 528 {
 529         struct jset_entry_dev_usage *u =
 530                 container_of(entry, struct jset_entry_dev_usage, entry);
 531         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 532         unsigned expected = sizeof(*u);
 533         unsigned dev;
 534         int ret = 0;
 535
 536         if (journal_entry_err_on(bytes < expected,
 537                                  c, "invalid journal entry dev usage: bad size (%u < %u)",
 538                                  bytes, expected)) {
 539                 journal_entry_null_range(entry, vstruct_next(entry));
 540                 return ret;
 541         }
 542
 543         dev = le32_to_cpu(u->dev);
 544
 545         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 546                                  c, "invalid journal entry dev usage: bad dev")) {
 547                 journal_entry_null_range(entry, vstruct_next(entry));
 548                 return ret;
 549         }
 550
 551         if (journal_entry_err_on(u->pad,
 552                                  c, "invalid journal entry dev usage: bad pad")) {
 553                 journal_entry_null_range(entry, vstruct_next(entry));
 554                 return ret;
 555         }
 556
 557 fsck_err:
 558         return ret;
 559 }
 560
 561 static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
 562                                             struct jset_entry *entry)
 563 {
 564         struct jset_entry_dev_usage *u =
 565                 container_of(entry, struct jset_entry_dev_usage, entry);
 566         unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 567
 568         pr_buf(out, "dev=%u", le32_to_cpu(u->dev));
 569
 570         for (i = 0; i < nr_types; i++) {
 571                 if (i < BCH_DATA_NR)
 572                         pr_buf(out, " %s", bch2_data_types[i]);
 573                 else
 574                         pr_buf(out, " (unknown data type %u)", i);
 575                 pr_buf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 576                        le64_to_cpu(u->d[i].buckets),
 577                        le64_to_cpu(u->d[i].sectors),
 578                        le64_to_cpu(u->d[i].fragmented));
 579         }
 580
 581         pr_buf(out, " buckets_ec: %llu buckets_unavailable: %llu",
 582                le64_to_cpu(u->buckets_ec),
 583                le64_to_cpu(u->buckets_unavailable));
 584 }
 585
 586 static int journal_entry_log_validate(struct bch_fs *c,
 587                                       const char *where,
 588                                       struct jset_entry *entry,
 589                                       unsigned version, int big_endian, int write)
 590 {
 591         return 0;
 592 }
 593
 594 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 595                                       struct jset_entry *entry)
 596 {
 597         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 598         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 599
 600         pr_buf(out, "%.*s", bytes, l->d);
 601 }
 602
 603 struct jset_entry_ops {
 604         int (*validate)(struct bch_fs *, const char *,
 605                         struct jset_entry *, unsigned, int, int);
 606         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 607 };
 608
 609 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 610 #define x(f, nr)                                                \
 611         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 612                 .validate       = journal_entry_##f##_validate, \
 613                 .to_text        = journal_entry_##f##_to_text,  \
 614         },
 615         BCH_JSET_ENTRY_TYPES()
 616 #undef x
 617 };
 618
 619 int bch2_journal_entry_validate(struct bch_fs *c, const char *where,
 620                                 struct jset_entry *entry,
 621                                 unsigned version, int big_endian, int write)
 622 {
 623         return entry->type < BCH_JSET_ENTRY_NR
 624                 ? bch2_jset_entry_ops[entry->type].validate(c, where, entry,
 625                                 version, big_endian, write)
 626                 : 0;
 627 }
 628
 629 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 630                                 struct jset_entry *entry)
 631 {
 632         if (entry->type < BCH_JSET_ENTRY_NR) {
 633                 pr_buf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 634                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 635         } else {
 636                 pr_buf(out, "(unknown type %u)", entry->type);
 637         }
 638 }
 639
 640 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 641                                  int write)
 642 {
 643         char buf[100];
 644         struct jset_entry *entry;
 645         int ret = 0;
 646
 647         vstruct_for_each(jset, entry) {
 648                 scnprintf(buf, sizeof(buf), "jset %llu entry offset %zi/%u",
 649                           le64_to_cpu(jset->seq),
 650                           (u64 *) entry - jset->_data,
 651                           le32_to_cpu(jset->u64s));
 652
 653                 if (journal_entry_err_on(vstruct_next(entry) >
 654                                          vstruct_last(jset), c,
 655                                 "journal entry extends past end of jset")) {
 656                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 657                         break;
 658                 }
 659
 660                 ret = bch2_journal_entry_validate(c, buf, entry,
 661                                         le32_to_cpu(jset->version),
 662                                         JSET_BIG_ENDIAN(jset), write);
 663                 if (ret)
 664                         break;
 665         }
 666 fsck_err:
 667         return ret;
 668 }
 669
 670 static int jset_validate(struct bch_fs *c,
 671                          struct bch_dev *ca,
 672                          struct jset *jset, u64 sector,
 673                          unsigned bucket_sectors_left,
 674                          unsigned sectors_read,
 675                          int write)
 676 {
 677         size_t bytes = vstruct_bytes(jset);
 678         struct bch_csum csum;
 679         unsigned version;
 680         int ret = 0;
 681
 682         if (le64_to_cpu(jset->magic) != jset_magic(c))
 683                 return JOURNAL_ENTRY_NONE;
 684
 685         version = le32_to_cpu(jset->version);
 686         if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
 687                                   version < bcachefs_metadata_version_min) ||
 688                                  version >= bcachefs_metadata_version_max, c,
 689                         "%s sector %llu seq %llu: unknown journal entry version %u",
 690                         ca ? ca->name : c->name,
 691                         sector, le64_to_cpu(jset->seq),
 692                         version)) {
 693                 /* don't try to continue: */
 694                 return EINVAL;
 695         }
 696
 697         if (bytes > (sectors_read << 9) &&
 698             sectors_read < bucket_sectors_left)
 699                 return JOURNAL_ENTRY_REREAD;
 700
 701         if (journal_entry_err_on(bytes > bucket_sectors_left << 9, c,
 702                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 703                         ca ? ca->name : c->name,
 704                         sector, le64_to_cpu(jset->seq), bytes)) {
 705                 ret = JOURNAL_ENTRY_BAD;
 706                 le32_add_cpu(&jset->u64s,
 707                              -((bytes - (bucket_sectors_left << 9)) / 8));
 708         }
 709
 710         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)), c,
 711                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 712                         ca ? ca->name : c->name,
 713                         sector, le64_to_cpu(jset->seq),
 714                         JSET_CSUM_TYPE(jset))) {
 715                 ret = JOURNAL_ENTRY_BAD;
 716                 goto csum_done;
 717         }
 718
 719         if (write)
 720                 goto csum_done;
 721
 722         csum = csum_vstruct(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), jset);
 723         if (journal_entry_err_on(bch2_crc_cmp(csum, jset->csum), c,
 724                                  "%s sector %llu seq %llu: journal checksum bad",
 725                                  ca ? ca->name : c->name,
 726                                  sector, le64_to_cpu(jset->seq)))
 727                 ret = JOURNAL_ENTRY_BAD;
 728
 729         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
 730                      jset->encrypted_start,
 731                      vstruct_end(jset) - (void *) jset->encrypted_start);
 732         bch2_fs_fatal_err_on(ret, c,
 733                         "error decrypting journal entry: %i", ret);
 734 csum_done:
 735         /* last_seq is ignored when JSET_NO_FLUSH is true */
 736         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 737                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq), c,
 738                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 739                                  le64_to_cpu(jset->last_seq),
 740                                  le64_to_cpu(jset->seq))) {
 741                 jset->last_seq = jset->seq;
 742                 return JOURNAL_ENTRY_BAD;
 743         }
 744 fsck_err:
 745         return ret;
 746 }
 747
 748 static int jset_validate_for_write(struct bch_fs *c, struct jset *jset)
 749 {
 750         unsigned sectors = vstruct_sectors(jset, c->block_bits);
 751
 752         return jset_validate(c, NULL, jset, 0, sectors, sectors, WRITE) ?:
 753                 jset_validate_entries(c, jset, WRITE);
 754 }
 755
 756 struct journal_read_buf {
 757         void            *data;
 758         size_t          size;
 759 };
 760
 761 static int journal_read_buf_realloc(struct journal_read_buf *b,
 762                                     size_t new_size)
 763 {
 764         void *n;
 765
 766         /* the bios are sized for this many pages, max: */
 767         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 768                 return -ENOMEM;
 769
 770         new_size = roundup_pow_of_two(new_size);
 771         n = kvpmalloc(new_size, GFP_KERNEL);
 772         if (!n)
 773                 return -ENOMEM;
 774
 775         kvpfree(b->data, b->size);
 776         b->data = n;
 777         b->size = new_size;
 778         return 0;
 779 }
 780
 781 static int journal_read_bucket(struct bch_dev *ca,
 782                                struct journal_read_buf *buf,
 783                                struct journal_list *jlist,
 784                                unsigned bucket)
 785 {
 786         struct bch_fs *c = ca->fs;
 787         struct journal_device *ja = &ca->journal;
 788         struct jset *j = NULL;
 789         unsigned sectors, sectors_read = 0;
 790         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 791             end = offset + ca->mi.bucket_size;
 792         bool saw_bad = false;
 793         int ret = 0;
 794
 795         pr_debug("reading %u", bucket);
 796
 797         while (offset < end) {
 798                 if (!sectors_read) {
 799                         struct bio *bio;
 800 reread:
 801                         sectors_read = min_t(unsigned,
 802                                 end - offset, buf->size >> 9);
 803
 804                         bio = bio_kmalloc(GFP_KERNEL,
 805                                           buf_pages(buf->data,
 806                                                     sectors_read << 9));
 807                         bio_set_dev(bio, ca->disk_sb.bdev);
 808                         bio->bi_iter.bi_sector  = offset;
 809                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 810                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 811
 812                         ret = submit_bio_wait(bio);
 813                         bio_put(bio);
 814
 815                         if (bch2_dev_io_err_on(ret, ca,
 816                                                "journal read error: sector %llu",
 817                                                offset) ||
 818                             bch2_meta_read_fault("journal")) {
 819                                 /*
 820                                  * We don't error out of the recovery process
 821                                  * here, since the relevant journal entry may be
 822                                  * found on a different device, and missing or
 823                                  * no journal entries will be handled later
 824                                  */
 825                                 return 0;
 826                         }
 827
 828                         j = buf->data;
 829                 }
 830
 831                 ret = jset_validate(c, ca, j, offset,
 832                                     end - offset, sectors_read,
 833                                     READ);
 834                 switch (ret) {
 835                 case BCH_FSCK_OK:
 836                         sectors = vstruct_sectors(j, c->block_bits);
 837                         break;
 838                 case JOURNAL_ENTRY_REREAD:
 839                         if (vstruct_bytes(j) > buf->size) {
 840                                 ret = journal_read_buf_realloc(buf,
 841                                                         vstruct_bytes(j));
 842                                 if (ret)
 843                                         return ret;
 844                         }
 845                         goto reread;
 846                 case JOURNAL_ENTRY_NONE:
 847                         if (!saw_bad)
 848                                 return 0;
 849                         sectors = block_sectors(c);
 850                         goto next_block;
 851                 case JOURNAL_ENTRY_BAD:
 852                         saw_bad = true;
 853                         /*
 854                          * On checksum error we don't really trust the size
 855                          * field of the journal entry we read, so try reading
 856                          * again at next block boundary:
 857                          */
 858                         sectors = block_sectors(c);
 859                         break;
 860                 default:
 861                         return ret;
 862                 }
 863
 864                 /*
 865                  * This happens sometimes if we don't have discards on -
 866                  * when we've partially overwritten a bucket with new
 867                  * journal entries. We don't need the rest of the
 868                  * bucket:
 869                  */
 870                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 871                         return 0;
 872
 873                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 874
 875                 mutex_lock(&jlist->lock);
 876                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
 877                                         .dev            = ca->dev_idx,
 878                                         .bucket         = bucket,
 879                                         .bucket_offset  = offset -
 880                                                 bucket_to_sector(ca, ja->buckets[bucket]),
 881                                         .sector         = offset,
 882                                         }, jlist, j, ret != 0);
 883                 mutex_unlock(&jlist->lock);
 884
 885                 switch (ret) {
 886                 case JOURNAL_ENTRY_ADD_OK:
 887                         break;
 888                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
 889                         break;
 890                 default:
 891                         return ret;
 892                 }
 893 next_block:
 894                 pr_debug("next");
 895                 offset          += sectors;
 896                 sectors_read    -= sectors;
 897                 j = ((void *) j) + (sectors << 9);
 898         }
 899
 900         return 0;
 901 }
 902
 903 static void bch2_journal_read_device(struct closure *cl)
 904 {
 905         struct journal_device *ja =
 906                 container_of(cl, struct journal_device, read);
 907         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
 908         struct bch_fs *c = ca->fs;
 909         struct journal_list *jlist =
 910                 container_of(cl->parent, struct journal_list, cl);
 911         struct journal_read_buf buf = { NULL, 0 };
 912         u64 min_seq = U64_MAX;
 913         unsigned i;
 914         int ret = 0;
 915
 916         if (!ja->nr)
 917                 goto out;
 918
 919         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
 920         if (ret)
 921                 goto err;
 922
 923         pr_debug("%u journal buckets", ja->nr);
 924
 925         for (i = 0; i < ja->nr; i++) {
 926                 ret = journal_read_bucket(ca, &buf, jlist, i);
 927                 if (ret)
 928                         goto err;
 929         }
 930
 931         /* Find the journal bucket with the highest sequence number: */
 932         for (i = 0; i < ja->nr; i++) {
 933                 if (ja->bucket_seq[i] > ja->bucket_seq[ja->cur_idx])
 934                         ja->cur_idx = i;
 935
 936                 min_seq = min(ja->bucket_seq[i], min_seq);
 937         }
 938
 939         /*
 940          * If there's duplicate journal entries in multiple buckets (which
 941          * definitely isn't supposed to happen, but...) - make sure to start
 942          * cur_idx at the last of those buckets, so we don't deadlock trying to
 943          * allocate
 944          */
 945         while (ja->bucket_seq[ja->cur_idx] > min_seq &&
 946                ja->bucket_seq[ja->cur_idx] >
 947                ja->bucket_seq[(ja->cur_idx + 1) % ja->nr])
 948                 ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 949
 950         ja->sectors_free = 0;
 951
 952         /*
 953          * Set dirty_idx to indicate the entire journal is full and needs to be
 954          * reclaimed - journal reclaim will immediately reclaim whatever isn't
 955          * pinned when it first runs:
 956          */
 957         ja->discard_idx = ja->dirty_idx_ondisk =
 958                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
 959 out:
 960         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
 961         kvpfree(buf.data, buf.size);
 962         percpu_ref_put(&ca->io_ref);
 963         closure_return(cl);
 964         return;
 965 err:
 966         mutex_lock(&jlist->lock);
 967         jlist->ret = ret;
 968         mutex_unlock(&jlist->lock);
 969         goto out;
 970 }
 971
 972 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 973                                struct journal_replay *j)
 974 {
 975         unsigned i;
 976
 977         for (i = 0; i < j->nr_ptrs; i++) {
 978                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
 979                 u64 offset;
 980
 981                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
 982
 983                 if (i)
 984                         pr_buf(out, " ");
 985                 pr_buf(out, "%u:%u:%u (sector %llu)",
 986                        j->ptrs[i].dev,
 987                        j->ptrs[i].bucket,
 988                        j->ptrs[i].bucket_offset,
 989                        j->ptrs[i].sector);
 990         }
 991 }
 992
 993 int bch2_journal_read(struct bch_fs *c, struct list_head *list,
 994                       u64 *blacklist_seq, u64 *start_seq)
 995 {
 996         struct journal_list jlist;
 997         struct journal_replay *i, *t;
 998         struct bch_dev *ca;
 999         unsigned iter;
1000         struct printbuf buf = PRINTBUF;
1001         size_t keys = 0, entries = 0;
1002         bool degraded = false;
1003         u64 seq, last_seq = 0;
1004         int ret = 0;
1005
1006         closure_init_stack(&jlist.cl);
1007         mutex_init(&jlist.lock);
1008         jlist.head = list;
1009         jlist.ret = 0;
1010
1011         for_each_member_device(ca, c, iter) {
1012                 if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
1013                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1014                         continue;
1015
1016                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1017                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1018                     percpu_ref_tryget(&ca->io_ref))
1019                         closure_call(&ca->journal.read,
1020                                      bch2_journal_read_device,
1021                                      system_unbound_wq,
1022                                      &jlist.cl);
1023                 else
1024                         degraded = true;
1025         }
1026
1027         closure_sync(&jlist.cl);
1028
1029         if (jlist.ret)
1030                 return jlist.ret;
1031
1032         if (list_empty(list)) {
1033                 bch_info(c, "journal read done, but no entries found");
1034                 return 0;
1035         }
1036
1037         i = list_last_entry(list, struct journal_replay, list);
1038         *start_seq = le64_to_cpu(i->j.seq) + 1;
1039
1040         /*
1041          * Find most recent flush entry, and ignore newer non flush entries -
1042          * those entries will be blacklisted:
1043          */
1044         list_for_each_entry_safe_reverse(i, t, list, list) {
1045                 if (i->ignore)
1046                         continue;
1047
1048                 if (!JSET_NO_FLUSH(&i->j)) {
1049                         last_seq        = le64_to_cpu(i->j.last_seq);
1050                         *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1051                         break;
1052                 }
1053
1054                 journal_replay_free(c, i);
1055         }
1056
1057         if (!last_seq) {
1058                 fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
1059                 ret = -1;
1060                 goto err;
1061         }
1062
1063         /* Drop blacklisted entries and entries older than last_seq: */
1064         list_for_each_entry_safe(i, t, list, list) {
1065                 if (i->ignore)
1066                         continue;
1067
1068                 seq = le64_to_cpu(i->j.seq);
1069                 if (seq < last_seq) {
1070                         journal_replay_free(c, i);
1071                         continue;
1072                 }
1073
1074                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1075                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1076                                     "found blacklisted journal entry %llu", seq);
1077
1078                         journal_replay_free(c, i);
1079                 }
1080         }
1081
1082         /* Check for missing entries: */
1083         seq = last_seq;
1084         list_for_each_entry(i, list, list) {
1085                 if (i->ignore)
1086                         continue;
1087
1088                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1089
1090                 while (seq < le64_to_cpu(i->j.seq)) {
1091                         u64 missing_start, missing_end;
1092                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1093
1094                         while (seq < le64_to_cpu(i->j.seq) &&
1095                                bch2_journal_seq_is_blacklisted(c, seq, false))
1096                                 seq++;
1097
1098                         if (seq == le64_to_cpu(i->j.seq))
1099                                 break;
1100
1101                         missing_start = seq;
1102
1103                         while (seq < le64_to_cpu(i->j.seq) &&
1104                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1105                                 seq++;
1106
1107                         if (i->list.prev != list) {
1108                                 struct journal_replay *p = list_prev_entry(i, list);
1109
1110                                 bch2_journal_ptrs_to_text(&buf1, c, p);
1111                                 pr_buf(&buf1, " size %zu", vstruct_sectors(&p->j, c->block_bits));
1112                         } else
1113                                 pr_buf(&buf1, "(none)");
1114                         bch2_journal_ptrs_to_text(&buf2, c, i);
1115
1116                         missing_end = seq - 1;
1117                         fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1118                                  "  prev at %s\n"
1119                                  "  next at %s",
1120                                  missing_start, missing_end,
1121                                  last_seq, *blacklist_seq - 1,
1122                                  buf1.buf, buf2.buf);
1123
1124                         printbuf_exit(&buf1);
1125                         printbuf_exit(&buf2);
1126                 }
1127
1128                 seq++;
1129         }
1130
1131         list_for_each_entry(i, list, list) {
1132                 struct jset_entry *entry;
1133                 struct bkey_i *k, *_n;
1134                 struct bch_replicas_padded replicas = {
1135                         .e.data_type = BCH_DATA_journal,
1136                         .e.nr_required = 1,
1137                 };
1138                 unsigned ptr;
1139
1140                 if (i->ignore)
1141                         continue;
1142
1143                 ret = jset_validate_entries(c, &i->j, READ);
1144                 if (ret)
1145                         goto err;
1146
1147                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1148                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1149
1150                 bch2_replicas_entry_sort(&replicas.e);
1151
1152                 /*
1153                  * If we're mounting in degraded mode - if we didn't read all
1154                  * the devices - this is wrong:
1155                  */
1156
1157                 printbuf_reset(&buf);
1158                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1159
1160                 if (!degraded &&
1161                     (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
1162                      fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
1163                                  "superblock not marked as containing replicas %s",
1164                                  buf.buf))) {
1165                         ret = bch2_mark_replicas(c, &replicas.e);
1166                         if (ret)
1167                                 goto err;
1168                 }
1169
1170                 for_each_jset_key(k, _n, entry, &i->j)
1171                         keys++;
1172                 entries++;
1173         }
1174
1175         bch_info(c, "journal read done, %zu keys in %zu entries, seq %llu",
1176                  keys, entries, *start_seq);
1177
1178         if (*start_seq != *blacklist_seq)
1179                 bch_info(c, "dropped unflushed entries %llu-%llu",
1180                          *blacklist_seq, *start_seq - 1);
1181 err:
1182 fsck_err:
1183         printbuf_exit(&buf);
1184         return ret;
1185 }
1186
1187 /* journal write: */
1188
1189 static void __journal_write_alloc(struct journal *j,
1190                                   struct journal_buf *w,
1191                                   struct dev_alloc_list *devs_sorted,
1192                                   unsigned sectors,
1193                                   unsigned *replicas,
1194                                   unsigned replicas_want)
1195 {
1196         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1197         struct journal_device *ja;
1198         struct bch_dev *ca;
1199         unsigned i;
1200
1201         if (*replicas >= replicas_want)
1202                 return;
1203
1204         for (i = 0; i < devs_sorted->nr; i++) {
1205                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1206                 if (!ca)
1207                         continue;
1208
1209                 ja = &ca->journal;
1210
1211                 /*
1212                  * Check that we can use this device, and aren't already using
1213                  * it:
1214                  */
1215                 if (!ca->mi.durability ||
1216                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1217                     !ja->nr ||
1218                     bch2_bkey_has_device(bkey_i_to_s_c(&w->key),
1219                                          ca->dev_idx) ||
1220                     sectors > ja->sectors_free)
1221                         continue;
1222
1223                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1224
1225                 bch2_bkey_append_ptr(&w->key,
1226                         (struct bch_extent_ptr) {
1227                                   .offset = bucket_to_sector(ca,
1228                                         ja->buckets[ja->cur_idx]) +
1229                                         ca->mi.bucket_size -
1230                                         ja->sectors_free,
1231                                   .dev = ca->dev_idx,
1232                 });
1233
1234                 ja->sectors_free -= sectors;
1235                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1236
1237                 *replicas += ca->mi.durability;
1238
1239                 if (*replicas >= replicas_want)
1240                         break;
1241         }
1242 }
1243
1244 /**
1245  * journal_next_bucket - move on to the next journal bucket if possible
1246  */
1247 static int journal_write_alloc(struct journal *j, struct journal_buf *w,
1248                                unsigned sectors)
1249 {
1250         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1251         struct bch_devs_mask devs;
1252         struct journal_device *ja;
1253         struct bch_dev *ca;
1254         struct dev_alloc_list devs_sorted;
1255         unsigned target = c->opts.metadata_target ?:
1256                 c->opts.foreground_target;
1257         unsigned i, replicas = 0, replicas_want =
1258                 READ_ONCE(c->opts.metadata_replicas);
1259
1260         rcu_read_lock();
1261 retry:
1262         devs = target_rw_devs(c, BCH_DATA_journal, target);
1263
1264         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1265
1266         __journal_write_alloc(j, w, &devs_sorted,
1267                               sectors, &replicas, replicas_want);
1268
1269         if (replicas >= replicas_want)
1270                 goto done;
1271
1272         for (i = 0; i < devs_sorted.nr; i++) {
1273                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1274                 if (!ca)
1275                         continue;
1276
1277                 ja = &ca->journal;
1278
1279                 if (sectors > ja->sectors_free &&
1280                     sectors <= ca->mi.bucket_size &&
1281                     bch2_journal_dev_buckets_available(j, ja,
1282                                         journal_space_discarded)) {
1283                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1284                         ja->sectors_free = ca->mi.bucket_size;
1285
1286                         /*
1287                          * ja->bucket_seq[ja->cur_idx] must always have
1288                          * something sensible:
1289                          */
1290                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1291                 }
1292         }
1293
1294         __journal_write_alloc(j, w, &devs_sorted,
1295                               sectors, &replicas, replicas_want);
1296
1297         if (replicas < replicas_want && target) {
1298                 /* Retry from all devices: */
1299                 target = 0;
1300                 goto retry;
1301         }
1302 done:
1303         rcu_read_unlock();
1304
1305         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1306
1307         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1308 }
1309
1310 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1311 {
1312         /* we aren't holding j->lock: */
1313         unsigned new_size = READ_ONCE(j->buf_size_want);
1314         void *new_buf;
1315
1316         if (buf->buf_size >= new_size)
1317                 return;
1318
1319         new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN);
1320         if (!new_buf)
1321                 return;
1322
1323         memcpy(new_buf, buf->data, buf->buf_size);
1324
1325         spin_lock(&j->lock);
1326         swap(buf->data,         new_buf);
1327         swap(buf->buf_size,     new_size);
1328         spin_unlock(&j->lock);
1329
1330         kvpfree(new_buf, new_size);
1331 }
1332
1333 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1334 {
1335         return j->buf + j->reservations.unwritten_idx;
1336 }
1337
1338 static void journal_write_done(struct closure *cl)
1339 {
1340         struct journal *j = container_of(cl, struct journal, io);
1341         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1342         struct journal_buf *w = journal_last_unwritten_buf(j);
1343         struct bch_replicas_padded replicas;
1344         union journal_res_state old, new;
1345         u64 v, seq;
1346         int err = 0;
1347
1348         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1349                                ? j->flush_write_time
1350                                : j->noflush_write_time, j->write_start_time);
1351
1352         if (!w->devs_written.nr) {
1353                 bch_err(c, "unable to write journal to sufficient devices");
1354                 err = -EIO;
1355         } else {
1356                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1357                                          w->devs_written);
1358                 if (bch2_mark_replicas(c, &replicas.e))
1359                         err = -EIO;
1360         }
1361
1362         if (err)
1363                 bch2_fatal_error(c);
1364
1365         spin_lock(&j->lock);
1366         seq = le64_to_cpu(w->data->seq);
1367
1368         if (seq >= j->pin.front)
1369                 journal_seq_pin(j, seq)->devs = w->devs_written;
1370
1371         if (!err) {
1372                 j->seq_ondisk           = seq;
1373
1374                 if (!JSET_NO_FLUSH(w->data)) {
1375                         j->flushed_seq_ondisk = seq;
1376                         j->last_seq_ondisk = w->last_seq;
1377                 }
1378         } else if (!j->err_seq || seq < j->err_seq)
1379                 j->err_seq      = seq;
1380
1381         /*
1382          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1383          * more buckets:
1384          *
1385          * Must come before signaling write completion, for
1386          * bch2_fs_journal_stop():
1387          */
1388         journal_reclaim_kick(&c->journal);
1389
1390         /* also must come before signalling write completion: */
1391         closure_debug_destroy(cl);
1392
1393         v = atomic64_read(&j->reservations.counter);
1394         do {
1395                 old.v = new.v = v;
1396                 BUG_ON(new.idx == new.unwritten_idx);
1397
1398                 new.unwritten_idx++;
1399         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1400                                        old.v, new.v)) != old.v);
1401
1402         bch2_journal_space_available(j);
1403
1404         closure_wake_up(&w->wait);
1405         journal_wake(j);
1406
1407         if (test_bit(JOURNAL_NEED_WRITE, &j->flags))
1408                 mod_delayed_work(c->io_complete_wq, &j->write_work, 0);
1409         spin_unlock(&j->lock);
1410
1411         if (new.unwritten_idx != new.idx &&
1412             !journal_state_count(new, new.unwritten_idx))
1413                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1414 }
1415
1416 static void journal_write_endio(struct bio *bio)
1417 {
1418         struct bch_dev *ca = bio->bi_private;
1419         struct journal *j = &ca->fs->journal;
1420         struct journal_buf *w = journal_last_unwritten_buf(j);
1421         unsigned long flags;
1422
1423         if (bch2_dev_io_err_on(bio->bi_status, ca, "error writing journal entry %llu: %s",
1424                                le64_to_cpu(w->data->seq),
1425                                bch2_blk_status_to_str(bio->bi_status)) ||
1426             bch2_meta_write_fault("journal")) {
1427                 spin_lock_irqsave(&j->err_lock, flags);
1428                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1429                 spin_unlock_irqrestore(&j->err_lock, flags);
1430         }
1431
1432         closure_put(&j->io);
1433         percpu_ref_put(&ca->io_ref);
1434 }
1435
1436 static void do_journal_write(struct closure *cl)
1437 {
1438         struct journal *j = container_of(cl, struct journal, io);
1439         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1440         struct bch_dev *ca;
1441         struct journal_buf *w = journal_last_unwritten_buf(j);
1442         struct bch_extent_ptr *ptr;
1443         struct bio *bio;
1444         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1445
1446         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1447                 ca = bch_dev_bkey_exists(c, ptr->dev);
1448                 if (!percpu_ref_tryget(&ca->io_ref)) {
1449                         /* XXX: fix this */
1450                         bch_err(c, "missing device for journal write\n");
1451                         continue;
1452                 }
1453
1454                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1455                              sectors);
1456
1457                 bio = ca->journal.bio;
1458                 bio_reset(bio);
1459                 bio_set_dev(bio, ca->disk_sb.bdev);
1460                 bio->bi_iter.bi_sector  = ptr->offset;
1461                 bio->bi_end_io          = journal_write_endio;
1462                 bio->bi_private         = ca;
1463                 bio->bi_opf             = REQ_OP_WRITE|REQ_SYNC|REQ_META;
1464
1465                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1466                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1467
1468                 if (!JSET_NO_FLUSH(w->data))
1469                         bio->bi_opf    |= REQ_FUA;
1470                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1471                         bio->bi_opf    |= REQ_PREFLUSH;
1472
1473                 bch2_bio_map(bio, w->data, sectors << 9);
1474
1475                 trace_journal_write(bio);
1476                 closure_bio_submit(bio, cl);
1477
1478                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1479                         le64_to_cpu(w->data->seq);
1480         }
1481
1482         continue_at(cl, journal_write_done, c->io_complete_wq);
1483         return;
1484 }
1485
1486 void bch2_journal_write(struct closure *cl)
1487 {
1488         struct journal *j = container_of(cl, struct journal, io);
1489         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1490         struct bch_dev *ca;
1491         struct journal_buf *w = journal_last_unwritten_buf(j);
1492         struct jset_entry *start, *end;
1493         struct jset *jset;
1494         struct bio *bio;
1495         struct printbuf journal_debug_buf = PRINTBUF;
1496         bool validate_before_checksum = false;
1497         unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
1498         int ret;
1499
1500         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1501
1502         journal_buf_realloc(j, w);
1503         jset = w->data;
1504
1505         j->write_start_time = local_clock();
1506
1507         spin_lock(&j->lock);
1508         if (c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush) &&
1509             (w->noflush ||
1510              (!w->must_flush &&
1511               (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1512               test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)))) {
1513                 w->noflush = true;
1514                 SET_JSET_NO_FLUSH(jset, true);
1515                 jset->last_seq  = 0;
1516                 w->last_seq     = 0;
1517
1518                 j->nr_noflush_writes++;
1519         } else {
1520                 j->last_flush_write = jiffies;
1521                 j->nr_flush_writes++;
1522         }
1523         spin_unlock(&j->lock);
1524
1525         /*
1526          * New btree roots are set by journalling them; when the journal entry
1527          * gets written we have to propagate them to c->btree_roots
1528          *
1529          * But, every journal entry we write has to contain all the btree roots
1530          * (at least for now); so after we copy btree roots to c->btree_roots we
1531          * have to get any missing btree roots and add them to this journal
1532          * entry:
1533          */
1534
1535         bch2_journal_entries_to_btree_roots(c, jset);
1536
1537         start = end = vstruct_last(jset);
1538
1539         end     = bch2_btree_roots_to_journal_entries(c, jset->start, end);
1540
1541         bch2_journal_super_entries_add_common(c, &end,
1542                                 le64_to_cpu(jset->seq));
1543         u64s    = (u64 *) end - (u64 *) start;
1544         BUG_ON(u64s > j->entry_u64s_reserved);
1545
1546         le32_add_cpu(&jset->u64s, u64s);
1547         BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors);
1548
1549         jset->magic             = cpu_to_le64(jset_magic(c));
1550         jset->version           = c->sb.version < bcachefs_metadata_version_new_versioning
1551                 ? cpu_to_le32(BCH_JSET_VERSION_OLD)
1552                 : cpu_to_le32(c->sb.version);
1553
1554         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1555         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1556
1557         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1558                 j->last_empty_seq = le64_to_cpu(jset->seq);
1559
1560         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1561                 validate_before_checksum = true;
1562
1563         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1564                 validate_before_checksum = true;
1565
1566         if (validate_before_checksum &&
1567             jset_validate_for_write(c, jset))
1568                 goto err;
1569
1570         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1571                     jset->encrypted_start,
1572                     vstruct_end(jset) - (void *) jset->encrypted_start);
1573         if (bch2_fs_fatal_err_on(ret, c,
1574                         "error decrypting journal entry: %i", ret))
1575                 goto err;
1576
1577         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1578                                   journal_nonce(jset), jset);
1579
1580         if (!validate_before_checksum &&
1581             jset_validate_for_write(c, jset))
1582                 goto err;
1583
1584         sectors = vstruct_sectors(jset, c->block_bits);
1585         BUG_ON(sectors > w->sectors);
1586
1587         bytes = vstruct_bytes(jset);
1588         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1589
1590 retry_alloc:
1591         spin_lock(&j->lock);
1592         ret = journal_write_alloc(j, w, sectors);
1593
1594         if (ret && j->can_discard) {
1595                 spin_unlock(&j->lock);
1596                 bch2_journal_do_discards(j);
1597                 goto retry_alloc;
1598         }
1599
1600         if (ret)
1601                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1602
1603         /*
1604          * write is allocated, no longer need to account for it in
1605          * bch2_journal_space_available():
1606          */
1607         w->sectors = 0;
1608
1609         /*
1610          * journal entry has been compacted and allocated, recalculate space
1611          * available:
1612          */
1613         bch2_journal_space_available(j);
1614         spin_unlock(&j->lock);
1615
1616         if (ret) {
1617                 bch_err(c, "Unable to allocate journal write:\n%s",
1618                         journal_debug_buf.buf);
1619                 printbuf_exit(&journal_debug_buf);
1620                 bch2_fatal_error(c);
1621                 continue_at(cl, journal_write_done, c->io_complete_wq);
1622                 return;
1623         }
1624
1625         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1626
1627         if (c->opts.nochanges)
1628                 goto no_io;
1629
1630         for_each_rw_member(ca, c, i)
1631                 nr_rw_members++;
1632
1633         if (nr_rw_members > 1)
1634                 w->separate_flush = true;
1635
1636         if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
1637                 for_each_rw_member(ca, c, i) {
1638                         percpu_ref_get(&ca->io_ref);
1639
1640                         bio = ca->journal.bio;
1641                         bio_reset(bio);
1642                         bio_set_dev(bio, ca->disk_sb.bdev);
1643                         bio->bi_opf             = REQ_OP_FLUSH;
1644                         bio->bi_end_io          = journal_write_endio;
1645                         bio->bi_private         = ca;
1646                         closure_bio_submit(bio, cl);
1647                 }
1648         }
1649
1650         continue_at(cl, do_journal_write, c->io_complete_wq);
1651         return;
1652 no_io:
1653         continue_at(cl, journal_write_done, c->io_complete_wq);
1654         return;
1655 err:
1656         bch2_fatal_error(c);
1657         continue_at(cl, journal_write_done, c->io_complete_wq);
1658 }