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 bch_extent_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 bch_extent_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                 char buf[160];
 256
 257                 bch2_bkey_val_to_text(&PBUF(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);
 263
 264                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 265                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 266                 journal_entry_null_range(vstruct_next(entry), next);
 267                 return FSCK_DELETED_KEY;
 268         }
 269
 270         if (write)
 271                 bch2_bkey_compat(level, btree_id, version, big_endian,
 272                                  write, NULL, bkey_to_packed(k));
 273 fsck_err:
 274         return ret;
 275 }
 276
 277 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 278                                              const char *where,
 279                                              struct jset_entry *entry,
 280                                              unsigned version, int big_endian, int write)
 281 {
 282         struct bkey_i *k = entry->start;
 283
 284         while (k != vstruct_last(entry)) {
 285                 int ret = journal_validate_key(c, where, entry,
 286                                                entry->level,
 287                                                entry->btree_id,
 288                                                k, "key", version, big_endian, write);
 289                 if (ret == FSCK_DELETED_KEY)
 290                         continue;
 291
 292                 k = bkey_next(k);
 293         }
 294
 295         return 0;
 296 }
 297
 298 static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
 299                                              struct jset_entry *entry)
 300 {
 301         struct bkey_i *k;
 302
 303         pr_buf(out, "btree=%s l=%u ", bch2_btree_ids[entry->btree_id], entry->level);
 304
 305         vstruct_for_each(entry, k)
 306                 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 307 }
 308
 309 static int journal_entry_btree_root_validate(struct bch_fs *c,
 310                                              const char *where,
 311                                              struct jset_entry *entry,
 312                                              unsigned version, int big_endian, int write)
 313 {
 314         struct bkey_i *k = entry->start;
 315         int ret = 0;
 316
 317         if (journal_entry_err_on(!entry->u64s ||
 318                                  le16_to_cpu(entry->u64s) != k->k.u64s, c,
 319                                  "invalid btree root journal entry: wrong number of keys")) {
 320                 void *next = vstruct_next(entry);
 321                 /*
 322                  * we don't want to null out this jset_entry,
 323                  * just the contents, so that later we can tell
 324                  * we were _supposed_ to have a btree root
 325                  */
 326                 entry->u64s = 0;
 327                 journal_entry_null_range(vstruct_next(entry), next);
 328                 return 0;
 329         }
 330
 331         return journal_validate_key(c, where, entry, 1, entry->btree_id, k,
 332                                     "btree root", version, big_endian, write);
 333 fsck_err:
 334         return ret;
 335 }
 336
 337 static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
 338                                              struct jset_entry *entry)
 339 {
 340         journal_entry_btree_keys_to_text(out, c, entry);
 341 }
 342
 343 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 344                                             const char *where,
 345                                             struct jset_entry *entry,
 346                                             unsigned version, int big_endian, int write)
 347 {
 348         /* obsolete, don't care: */
 349         return 0;
 350 }
 351
 352 static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 353                                             struct jset_entry *entry)
 354 {
 355 }
 356
 357 static int journal_entry_blacklist_validate(struct bch_fs *c,
 358                                             const char *where,
 359                                             struct jset_entry *entry,
 360                                             unsigned version, int big_endian, int write)
 361 {
 362         int ret = 0;
 363
 364         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1, c,
 365                 "invalid journal seq blacklist entry: bad size")) {
 366                 journal_entry_null_range(entry, vstruct_next(entry));
 367         }
 368 fsck_err:
 369         return ret;
 370 }
 371
 372 static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
 373                                             struct jset_entry *entry)
 374 {
 375         struct jset_entry_blacklist *bl =
 376                 container_of(entry, struct jset_entry_blacklist, entry);
 377
 378         pr_buf(out, "seq=%llu", le64_to_cpu(bl->seq));
 379 }
 380
 381 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 382                                                const char *where,
 383                                                struct jset_entry *entry,
 384                                                unsigned version, int big_endian, int write)
 385 {
 386         struct jset_entry_blacklist_v2 *bl_entry;
 387         int ret = 0;
 388
 389         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2, c,
 390                 "invalid journal seq blacklist entry: bad size")) {
 391                 journal_entry_null_range(entry, vstruct_next(entry));
 392                 goto out;
 393         }
 394
 395         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 396
 397         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 398                                  le64_to_cpu(bl_entry->end), c,
 399                 "invalid journal seq blacklist entry: start > end")) {
 400                 journal_entry_null_range(entry, vstruct_next(entry));
 401         }
 402 out:
 403 fsck_err:
 404         return ret;
 405 }
 406
 407 static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
 408                                                struct jset_entry *entry)
 409 {
 410         struct jset_entry_blacklist_v2 *bl =
 411                 container_of(entry, struct jset_entry_blacklist_v2, entry);
 412
 413         pr_buf(out, "start=%llu end=%llu",
 414                le64_to_cpu(bl->start),
 415                le64_to_cpu(bl->end));
 416 }
 417
 418 static int journal_entry_usage_validate(struct bch_fs *c,
 419                                         const char *where,
 420                                         struct jset_entry *entry,
 421                                         unsigned version, int big_endian, int write)
 422 {
 423         struct jset_entry_usage *u =
 424                 container_of(entry, struct jset_entry_usage, entry);
 425         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 426         int ret = 0;
 427
 428         if (journal_entry_err_on(bytes < sizeof(*u),
 429                                  c,
 430                                  "invalid journal entry usage: bad size")) {
 431                 journal_entry_null_range(entry, vstruct_next(entry));
 432                 return ret;
 433         }
 434
 435 fsck_err:
 436         return ret;
 437 }
 438
 439 static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 440                                         struct jset_entry *entry)
 441 {
 442         struct jset_entry_usage *u =
 443                 container_of(entry, struct jset_entry_usage, entry);
 444
 445         pr_buf(out, "type=%s v=%llu",
 446                bch2_fs_usage_types[u->entry.btree_id],
 447                le64_to_cpu(u->v));
 448 }
 449
 450 static int journal_entry_data_usage_validate(struct bch_fs *c,
 451                                         const char *where,
 452                                         struct jset_entry *entry,
 453                                         unsigned version, int big_endian, int write)
 454 {
 455         struct jset_entry_data_usage *u =
 456                 container_of(entry, struct jset_entry_data_usage, entry);
 457         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 458         int ret = 0;
 459
 460         if (journal_entry_err_on(bytes < sizeof(*u) ||
 461                                  bytes < sizeof(*u) + u->r.nr_devs,
 462                                  c,
 463                                  "invalid journal entry usage: bad size")) {
 464                 journal_entry_null_range(entry, vstruct_next(entry));
 465                 return ret;
 466         }
 467
 468 fsck_err:
 469         return ret;
 470 }
 471
 472 static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
 473                                              struct jset_entry *entry)
 474 {
 475         struct jset_entry_data_usage *u =
 476                 container_of(entry, struct jset_entry_data_usage, entry);
 477
 478         bch2_replicas_entry_to_text(out, &u->r);
 479         pr_buf(out, "=%llu", le64_to_cpu(u->v));
 480 }
 481
 482 static int journal_entry_clock_validate(struct bch_fs *c,
 483                                         const char *where,
 484                                         struct jset_entry *entry,
 485                                         unsigned version, int big_endian, int write)
 486 {
 487         struct jset_entry_clock *clock =
 488                 container_of(entry, struct jset_entry_clock, entry);
 489         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 490         int ret = 0;
 491
 492         if (journal_entry_err_on(bytes != sizeof(*clock),
 493                                  c, "invalid journal entry clock: bad size")) {
 494                 journal_entry_null_range(entry, vstruct_next(entry));
 495                 return ret;
 496         }
 497
 498         if (journal_entry_err_on(clock->rw > 1,
 499                                  c, "invalid journal entry clock: bad rw")) {
 500                 journal_entry_null_range(entry, vstruct_next(entry));
 501                 return ret;
 502         }
 503
 504 fsck_err:
 505         return ret;
 506 }
 507
 508 static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 509                                         struct jset_entry *entry)
 510 {
 511         struct jset_entry_clock *clock =
 512                 container_of(entry, struct jset_entry_clock, entry);
 513
 514         pr_buf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
 515 }
 516
 517 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 518                                             const char *where,
 519                                             struct jset_entry *entry,
 520                                             unsigned version, int big_endian, int write)
 521 {
 522         struct jset_entry_dev_usage *u =
 523                 container_of(entry, struct jset_entry_dev_usage, entry);
 524         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 525         unsigned expected = sizeof(*u);
 526         unsigned dev;
 527         int ret = 0;
 528
 529         if (journal_entry_err_on(bytes < expected,
 530                                  c, "invalid journal entry dev usage: bad size (%u < %u)",
 531                                  bytes, expected)) {
 532                 journal_entry_null_range(entry, vstruct_next(entry));
 533                 return ret;
 534         }
 535
 536         dev = le32_to_cpu(u->dev);
 537
 538         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 539                                  c, "invalid journal entry dev usage: bad dev")) {
 540                 journal_entry_null_range(entry, vstruct_next(entry));
 541                 return ret;
 542         }
 543
 544         if (journal_entry_err_on(u->pad,
 545                                  c, "invalid journal entry dev usage: bad pad")) {
 546                 journal_entry_null_range(entry, vstruct_next(entry));
 547                 return ret;
 548         }
 549
 550 fsck_err:
 551         return ret;
 552 }
 553
 554 static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
 555                                             struct jset_entry *entry)
 556 {
 557         struct jset_entry_dev_usage *u =
 558                 container_of(entry, struct jset_entry_dev_usage, entry);
 559         unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 560
 561         pr_buf(out, "dev=%u", le32_to_cpu(u->dev));
 562
 563         for (i = 0; i < nr_types; i++) {
 564                 if (i < BCH_DATA_NR)
 565                         pr_buf(out, " %s", bch2_data_types[i]);
 566                 else
 567                         pr_buf(out, " (unknown data type %u)", i);
 568                 pr_buf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 569                        le64_to_cpu(u->d[i].buckets),
 570                        le64_to_cpu(u->d[i].sectors),
 571                        le64_to_cpu(u->d[i].fragmented));
 572         }
 573
 574         pr_buf(out, " buckets_ec: %llu buckets_unavailable: %llu",
 575                le64_to_cpu(u->buckets_ec),
 576                le64_to_cpu(u->buckets_unavailable));
 577 }
 578
 579 static int journal_entry_log_validate(struct bch_fs *c,
 580                                       const char *where,
 581                                       struct jset_entry *entry,
 582                                       unsigned version, int big_endian, int write)
 583 {
 584         return 0;
 585 }
 586
 587 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 588                                       struct jset_entry *entry)
 589 {
 590         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 591         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 592
 593         bch_scnmemcpy(out, l->d, strnlen(l->d, bytes));
 594 }
 595
 596 struct jset_entry_ops {
 597         int (*validate)(struct bch_fs *, const char *,
 598                         struct jset_entry *, unsigned, int, int);
 599         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 600 };
 601
 602 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 603 #define x(f, nr)                                                \
 604         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 605                 .validate       = journal_entry_##f##_validate, \
 606                 .to_text        = journal_entry_##f##_to_text,  \
 607         },
 608         BCH_JSET_ENTRY_TYPES()
 609 #undef x
 610 };
 611
 612 int bch2_journal_entry_validate(struct bch_fs *c, const char *where,
 613                                 struct jset_entry *entry,
 614                                 unsigned version, int big_endian, int write)
 615 {
 616         return entry->type < BCH_JSET_ENTRY_NR
 617                 ? bch2_jset_entry_ops[entry->type].validate(c, where, entry,
 618                                 version, big_endian, write)
 619                 : 0;
 620 }
 621
 622 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 623                                 struct jset_entry *entry)
 624 {
 625         if (entry->type < BCH_JSET_ENTRY_NR) {
 626                 pr_buf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 627                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 628         } else {
 629                 pr_buf(out, "(unknown type %u)", entry->type);
 630         }
 631 }
 632
 633 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 634                                  int write)
 635 {
 636         char buf[100];
 637         struct jset_entry *entry;
 638         int ret = 0;
 639
 640         vstruct_for_each(jset, entry) {
 641                 scnprintf(buf, sizeof(buf), "jset %llu entry offset %zi/%u",
 642                           le64_to_cpu(jset->seq),
 643                           (u64 *) entry - jset->_data,
 644                           le32_to_cpu(jset->u64s));
 645
 646                 if (journal_entry_err_on(vstruct_next(entry) >
 647                                          vstruct_last(jset), c,
 648                                 "journal entry extends past end of jset")) {
 649                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 650                         break;
 651                 }
 652
 653                 ret = bch2_journal_entry_validate(c, buf, entry,
 654                                         le32_to_cpu(jset->version),
 655                                         JSET_BIG_ENDIAN(jset), write);
 656                 if (ret)
 657                         break;
 658         }
 659 fsck_err:
 660         return ret;
 661 }
 662
 663 static int jset_validate(struct bch_fs *c,
 664                          struct bch_dev *ca,
 665                          struct jset *jset, u64 sector,
 666                          unsigned bucket_sectors_left,
 667                          unsigned sectors_read,
 668                          int write)
 669 {
 670         size_t bytes = vstruct_bytes(jset);
 671         struct bch_csum csum;
 672         unsigned version;
 673         int ret = 0;
 674
 675         if (le64_to_cpu(jset->magic) != jset_magic(c))
 676                 return JOURNAL_ENTRY_NONE;
 677
 678         version = le32_to_cpu(jset->version);
 679         if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
 680                                   version < bcachefs_metadata_version_min) ||
 681                                  version >= bcachefs_metadata_version_max, c,
 682                         "%s sector %llu seq %llu: unknown journal entry version %u",
 683                         ca ? ca->name : c->name,
 684                         sector, le64_to_cpu(jset->seq),
 685                         version)) {
 686                 /* don't try to continue: */
 687                 return EINVAL;
 688         }
 689
 690         if (bytes > (sectors_read << 9) &&
 691             sectors_read < bucket_sectors_left)
 692                 return JOURNAL_ENTRY_REREAD;
 693
 694         if (journal_entry_err_on(bytes > bucket_sectors_left << 9, c,
 695                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 696                         ca ? ca->name : c->name,
 697                         sector, le64_to_cpu(jset->seq), bytes)) {
 698                 ret = JOURNAL_ENTRY_BAD;
 699                 le32_add_cpu(&jset->u64s,
 700                              -((bytes - (bucket_sectors_left << 9)) / 8));
 701         }
 702
 703         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)), c,
 704                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 705                         ca ? ca->name : c->name,
 706                         sector, le64_to_cpu(jset->seq),
 707                         JSET_CSUM_TYPE(jset))) {
 708                 ret = JOURNAL_ENTRY_BAD;
 709                 goto csum_done;
 710         }
 711
 712         if (write)
 713                 goto csum_done;
 714
 715         csum = csum_vstruct(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), jset);
 716         if (journal_entry_err_on(bch2_crc_cmp(csum, jset->csum), c,
 717                                  "%s sector %llu seq %llu: journal checksum bad",
 718                                  ca ? ca->name : c->name,
 719                                  sector, le64_to_cpu(jset->seq)))
 720                 ret = JOURNAL_ENTRY_BAD;
 721
 722         bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
 723                      jset->encrypted_start,
 724                      vstruct_end(jset) - (void *) jset->encrypted_start);
 725 csum_done:
 726         /* last_seq is ignored when JSET_NO_FLUSH is true */
 727         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 728                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq), c,
 729                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 730                                  le64_to_cpu(jset->last_seq),
 731                                  le64_to_cpu(jset->seq))) {
 732                 jset->last_seq = jset->seq;
 733                 return JOURNAL_ENTRY_BAD;
 734         }
 735 fsck_err:
 736         return ret;
 737 }
 738
 739 static int jset_validate_for_write(struct bch_fs *c, struct jset *jset)
 740 {
 741         unsigned sectors = vstruct_sectors(jset, c->block_bits);
 742
 743         return jset_validate(c, NULL, jset, 0, sectors, sectors, WRITE) ?:
 744                 jset_validate_entries(c, jset, WRITE);
 745 }
 746
 747 struct journal_read_buf {
 748         void            *data;
 749         size_t          size;
 750 };
 751
 752 static int journal_read_buf_realloc(struct journal_read_buf *b,
 753                                     size_t new_size)
 754 {
 755         void *n;
 756
 757         /* the bios are sized for this many pages, max: */
 758         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 759                 return -ENOMEM;
 760
 761         new_size = roundup_pow_of_two(new_size);
 762         n = kvpmalloc(new_size, GFP_KERNEL);
 763         if (!n)
 764                 return -ENOMEM;
 765
 766         kvpfree(b->data, b->size);
 767         b->data = n;
 768         b->size = new_size;
 769         return 0;
 770 }
 771
 772 static int journal_read_bucket(struct bch_dev *ca,
 773                                struct journal_read_buf *buf,
 774                                struct journal_list *jlist,
 775                                unsigned bucket)
 776 {
 777         struct bch_fs *c = ca->fs;
 778         struct journal_device *ja = &ca->journal;
 779         struct jset *j = NULL;
 780         unsigned sectors, sectors_read = 0;
 781         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 782             end = offset + ca->mi.bucket_size;
 783         bool saw_bad = false;
 784         int ret = 0;
 785
 786         pr_debug("reading %u", bucket);
 787
 788         while (offset < end) {
 789                 if (!sectors_read) {
 790                         struct bio *bio;
 791 reread:
 792                         sectors_read = min_t(unsigned,
 793                                 end - offset, buf->size >> 9);
 794
 795                         bio = bio_kmalloc(GFP_KERNEL,
 796                                           buf_pages(buf->data,
 797                                                     sectors_read << 9));
 798                         bio_set_dev(bio, ca->disk_sb.bdev);
 799                         bio->bi_iter.bi_sector  = offset;
 800                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 801                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 802
 803                         ret = submit_bio_wait(bio);
 804                         bio_put(bio);
 805
 806                         if (bch2_dev_io_err_on(ret, ca,
 807                                                "journal read error: sector %llu",
 808                                                offset) ||
 809                             bch2_meta_read_fault("journal")) {
 810                                 /*
 811                                  * We don't error out of the recovery process
 812                                  * here, since the relevant journal entry may be
 813                                  * found on a different device, and missing or
 814                                  * no journal entries will be handled later
 815                                  */
 816                                 return 0;
 817                         }
 818
 819                         j = buf->data;
 820                 }
 821
 822                 ret = jset_validate(c, ca, j, offset,
 823                                     end - offset, sectors_read,
 824                                     READ);
 825                 switch (ret) {
 826                 case BCH_FSCK_OK:
 827                         sectors = vstruct_sectors(j, c->block_bits);
 828                         break;
 829                 case JOURNAL_ENTRY_REREAD:
 830                         if (vstruct_bytes(j) > buf->size) {
 831                                 ret = journal_read_buf_realloc(buf,
 832                                                         vstruct_bytes(j));
 833                                 if (ret)
 834                                         return ret;
 835                         }
 836                         goto reread;
 837                 case JOURNAL_ENTRY_NONE:
 838                         if (!saw_bad)
 839                                 return 0;
 840                         sectors = block_sectors(c);
 841                         goto next_block;
 842                 case JOURNAL_ENTRY_BAD:
 843                         saw_bad = true;
 844                         /*
 845                          * On checksum error we don't really trust the size
 846                          * field of the journal entry we read, so try reading
 847                          * again at next block boundary:
 848                          */
 849                         sectors = block_sectors(c);
 850                         break;
 851                 default:
 852                         return ret;
 853                 }
 854
 855                 /*
 856                  * This happens sometimes if we don't have discards on -
 857                  * when we've partially overwritten a bucket with new
 858                  * journal entries. We don't need the rest of the
 859                  * bucket:
 860                  */
 861                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 862                         return 0;
 863
 864                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 865
 866                 mutex_lock(&jlist->lock);
 867                 ret = journal_entry_add(c, ca, (struct bch_extent_ptr) {
 868                                         .dev = ca->dev_idx,
 869                                         .offset = offset,
 870                                         }, jlist, j, ret != 0);
 871                 mutex_unlock(&jlist->lock);
 872
 873                 switch (ret) {
 874                 case JOURNAL_ENTRY_ADD_OK:
 875                         break;
 876                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
 877                         break;
 878                 default:
 879                         return ret;
 880                 }
 881 next_block:
 882                 pr_debug("next");
 883                 offset          += sectors;
 884                 sectors_read    -= sectors;
 885                 j = ((void *) j) + (sectors << 9);
 886         }
 887
 888         return 0;
 889 }
 890
 891 static void bch2_journal_read_device(struct closure *cl)
 892 {
 893         struct journal_device *ja =
 894                 container_of(cl, struct journal_device, read);
 895         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
 896         struct bch_fs *c = ca->fs;
 897         struct journal_list *jlist =
 898                 container_of(cl->parent, struct journal_list, cl);
 899         struct journal_read_buf buf = { NULL, 0 };
 900         u64 min_seq = U64_MAX;
 901         unsigned i;
 902         int ret = 0;
 903
 904         if (!ja->nr)
 905                 goto out;
 906
 907         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
 908         if (ret)
 909                 goto err;
 910
 911         pr_debug("%u journal buckets", ja->nr);
 912
 913         for (i = 0; i < ja->nr; i++) {
 914                 ret = journal_read_bucket(ca, &buf, jlist, i);
 915                 if (ret)
 916                         goto err;
 917         }
 918
 919         /* Find the journal bucket with the highest sequence number: */
 920         for (i = 0; i < ja->nr; i++) {
 921                 if (ja->bucket_seq[i] > ja->bucket_seq[ja->cur_idx])
 922                         ja->cur_idx = i;
 923
 924                 min_seq = min(ja->bucket_seq[i], min_seq);
 925         }
 926
 927         /*
 928          * If there's duplicate journal entries in multiple buckets (which
 929          * definitely isn't supposed to happen, but...) - make sure to start
 930          * cur_idx at the last of those buckets, so we don't deadlock trying to
 931          * allocate
 932          */
 933         while (ja->bucket_seq[ja->cur_idx] > min_seq &&
 934                ja->bucket_seq[ja->cur_idx] >
 935                ja->bucket_seq[(ja->cur_idx + 1) % ja->nr])
 936                 ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 937
 938         ja->sectors_free = 0;
 939
 940         /*
 941          * Set dirty_idx to indicate the entire journal is full and needs to be
 942          * reclaimed - journal reclaim will immediately reclaim whatever isn't
 943          * pinned when it first runs:
 944          */
 945         ja->discard_idx = ja->dirty_idx_ondisk =
 946                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
 947 out:
 948         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
 949         kvpfree(buf.data, buf.size);
 950         percpu_ref_put(&ca->io_ref);
 951         closure_return(cl);
 952         return;
 953 err:
 954         mutex_lock(&jlist->lock);
 955         jlist->ret = ret;
 956         mutex_unlock(&jlist->lock);
 957         goto out;
 958 }
 959
 960 static void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 961                                       struct journal_replay *j)
 962 {
 963         unsigned i;
 964
 965         for (i = 0; i < j->nr_ptrs; i++) {
 966                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
 967                 u64 offset;
 968
 969                 div64_u64_rem(j->ptrs[i].offset, ca->mi.bucket_size, &offset);
 970
 971                 if (i)
 972                         pr_buf(out, " ");
 973                 pr_buf(out, "%u:%llu (offset %llu)",
 974                        j->ptrs[i].dev,
 975                        (u64) j->ptrs[i].offset, offset);
 976         }
 977 }
 978
 979 int bch2_journal_read(struct bch_fs *c, struct list_head *list,
 980                       u64 *blacklist_seq, u64 *start_seq)
 981 {
 982         struct journal_list jlist;
 983         struct journal_replay *i, *t;
 984         struct bch_dev *ca;
 985         unsigned iter;
 986         size_t keys = 0, entries = 0;
 987         bool degraded = false;
 988         u64 seq, last_seq = 0;
 989         int ret = 0;
 990
 991         closure_init_stack(&jlist.cl);
 992         mutex_init(&jlist.lock);
 993         jlist.head = list;
 994         jlist.ret = 0;
 995
 996         for_each_member_device(ca, c, iter) {
 997                 if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
 998                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
 999                         continue;
1000
1001                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1002                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1003                     percpu_ref_tryget(&ca->io_ref))
1004                         closure_call(&ca->journal.read,
1005                                      bch2_journal_read_device,
1006                                      system_unbound_wq,
1007                                      &jlist.cl);
1008                 else
1009                         degraded = true;
1010         }
1011
1012         closure_sync(&jlist.cl);
1013
1014         if (jlist.ret)
1015                 return jlist.ret;
1016
1017         if (list_empty(list)) {
1018                 bch_info(c, "journal read done, but no entries found");
1019                 return 0;
1020         }
1021
1022         i = list_last_entry(list, struct journal_replay, list);
1023         *start_seq = le64_to_cpu(i->j.seq) + 1;
1024
1025         /*
1026          * Find most recent flush entry, and ignore newer non flush entries -
1027          * those entries will be blacklisted:
1028          */
1029         list_for_each_entry_safe_reverse(i, t, list, list) {
1030                 if (i->ignore)
1031                         continue;
1032
1033                 if (!JSET_NO_FLUSH(&i->j)) {
1034                         last_seq        = le64_to_cpu(i->j.last_seq);
1035                         *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1036                         break;
1037                 }
1038
1039                 journal_replay_free(c, i);
1040         }
1041
1042         if (!last_seq) {
1043                 fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
1044                 return -1;
1045         }
1046
1047         /* Drop blacklisted entries and entries older than last_seq: */
1048         list_for_each_entry_safe(i, t, list, list) {
1049                 if (i->ignore)
1050                         continue;
1051
1052                 seq = le64_to_cpu(i->j.seq);
1053                 if (seq < last_seq) {
1054                         journal_replay_free(c, i);
1055                         continue;
1056                 }
1057
1058                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1059                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1060                                     "found blacklisted journal entry %llu", seq);
1061
1062                         journal_replay_free(c, i);
1063                 }
1064         }
1065
1066         /* Check for missing entries: */
1067         seq = last_seq;
1068         list_for_each_entry(i, list, list) {
1069                 if (i->ignore)
1070                         continue;
1071
1072                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1073
1074                 while (seq < le64_to_cpu(i->j.seq)) {
1075                         u64 missing_start, missing_end;
1076                         char buf1[200], buf2[200];
1077
1078                         while (seq < le64_to_cpu(i->j.seq) &&
1079                                bch2_journal_seq_is_blacklisted(c, seq, false))
1080                                 seq++;
1081
1082                         if (seq == le64_to_cpu(i->j.seq))
1083                                 break;
1084
1085                         missing_start = seq;
1086
1087                         while (seq < le64_to_cpu(i->j.seq) &&
1088                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1089                                 seq++;
1090
1091                         if (i->list.prev != list) {
1092                                 struct printbuf out = PBUF(buf1);
1093                                 struct journal_replay *p = list_prev_entry(i, list);
1094
1095                                 bch2_journal_ptrs_to_text(&out, c, p);
1096                                 pr_buf(&out, " size %llu", vstruct_sectors(&p->j, c->block_bits));
1097                         } else
1098                                 sprintf(buf1, "(none)");
1099                         bch2_journal_ptrs_to_text(&PBUF(buf2), c, i);
1100
1101                         missing_end = seq - 1;
1102                         fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1103                                  "  prev at %s\n"
1104                                  "  next at %s",
1105                                  missing_start, missing_end,
1106                                  last_seq, *blacklist_seq - 1,
1107                                  buf1, buf2);
1108                 }
1109
1110                 seq++;
1111         }
1112
1113         list_for_each_entry(i, list, list) {
1114                 struct jset_entry *entry;
1115                 struct bkey_i *k, *_n;
1116                 struct bch_replicas_padded replicas = {
1117                         .e.data_type = BCH_DATA_journal,
1118                         .e.nr_required = 1,
1119                 };
1120                 unsigned ptr;
1121                 char buf[80];
1122
1123                 if (i->ignore)
1124                         continue;
1125
1126                 ret = jset_validate_entries(c, &i->j, READ);
1127                 if (ret)
1128                         goto fsck_err;
1129
1130                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1131                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1132
1133                 bch2_replicas_entry_sort(&replicas.e);
1134
1135                 /*
1136                  * If we're mounting in degraded mode - if we didn't read all
1137                  * the devices - this is wrong:
1138                  */
1139
1140                 if (!degraded &&
1141                     (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
1142                      fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
1143                                  "superblock not marked as containing replicas %s",
1144                                  (bch2_replicas_entry_to_text(&PBUF(buf),
1145                                                               &replicas.e), buf)))) {
1146                         ret = bch2_mark_replicas(c, &replicas.e);
1147                         if (ret)
1148                                 return ret;
1149                 }
1150
1151                 for_each_jset_key(k, _n, entry, &i->j)
1152                         keys++;
1153                 entries++;
1154         }
1155
1156         bch_info(c, "journal read done, %zu keys in %zu entries, seq %llu",
1157                  keys, entries, *start_seq);
1158
1159         if (*start_seq != *blacklist_seq)
1160                 bch_info(c, "dropped unflushed entries %llu-%llu",
1161                          *blacklist_seq, *start_seq - 1);
1162 fsck_err:
1163         return ret;
1164 }
1165
1166 /* journal write: */
1167
1168 static void __journal_write_alloc(struct journal *j,
1169                                   struct journal_buf *w,
1170                                   struct dev_alloc_list *devs_sorted,
1171                                   unsigned sectors,
1172                                   unsigned *replicas,
1173                                   unsigned replicas_want)
1174 {
1175         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1176         struct journal_device *ja;
1177         struct bch_dev *ca;
1178         unsigned i;
1179
1180         if (*replicas >= replicas_want)
1181                 return;
1182
1183         for (i = 0; i < devs_sorted->nr; i++) {
1184                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1185                 if (!ca)
1186                         continue;
1187
1188                 ja = &ca->journal;
1189
1190                 /*
1191                  * Check that we can use this device, and aren't already using
1192                  * it:
1193                  */
1194                 if (!ca->mi.durability ||
1195                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1196                     !ja->nr ||
1197                     bch2_bkey_has_device(bkey_i_to_s_c(&w->key),
1198                                          ca->dev_idx) ||
1199                     sectors > ja->sectors_free)
1200                         continue;
1201
1202                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1203
1204                 bch2_bkey_append_ptr(&w->key,
1205                         (struct bch_extent_ptr) {
1206                                   .offset = bucket_to_sector(ca,
1207                                         ja->buckets[ja->cur_idx]) +
1208                                         ca->mi.bucket_size -
1209                                         ja->sectors_free,
1210                                   .dev = ca->dev_idx,
1211                 });
1212
1213                 ja->sectors_free -= sectors;
1214                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1215
1216                 *replicas += ca->mi.durability;
1217
1218                 if (*replicas >= replicas_want)
1219                         break;
1220         }
1221 }
1222
1223 /**
1224  * journal_next_bucket - move on to the next journal bucket if possible
1225  */
1226 static int journal_write_alloc(struct journal *j, struct journal_buf *w,
1227                                unsigned sectors)
1228 {
1229         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1230         struct bch_devs_mask devs;
1231         struct journal_device *ja;
1232         struct bch_dev *ca;
1233         struct dev_alloc_list devs_sorted;
1234         unsigned target = c->opts.metadata_target ?:
1235                 c->opts.foreground_target;
1236         unsigned i, replicas = 0, replicas_want =
1237                 READ_ONCE(c->opts.metadata_replicas);
1238
1239         rcu_read_lock();
1240 retry:
1241         devs = target_rw_devs(c, BCH_DATA_journal, target);
1242
1243         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1244
1245         __journal_write_alloc(j, w, &devs_sorted,
1246                               sectors, &replicas, replicas_want);
1247
1248         if (replicas >= replicas_want)
1249                 goto done;
1250
1251         for (i = 0; i < devs_sorted.nr; i++) {
1252                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1253                 if (!ca)
1254                         continue;
1255
1256                 ja = &ca->journal;
1257
1258                 if (sectors > ja->sectors_free &&
1259                     sectors <= ca->mi.bucket_size &&
1260                     bch2_journal_dev_buckets_available(j, ja,
1261                                         journal_space_discarded)) {
1262                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1263                         ja->sectors_free = ca->mi.bucket_size;
1264
1265                         /*
1266                          * ja->bucket_seq[ja->cur_idx] must always have
1267                          * something sensible:
1268                          */
1269                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1270                 }
1271         }
1272
1273         __journal_write_alloc(j, w, &devs_sorted,
1274                               sectors, &replicas, replicas_want);
1275
1276         if (replicas < replicas_want && target) {
1277                 /* Retry from all devices: */
1278                 target = 0;
1279                 goto retry;
1280         }
1281 done:
1282         rcu_read_unlock();
1283
1284         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1285
1286         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1287 }
1288
1289 static void journal_write_compact(struct jset *jset)
1290 {
1291         struct jset_entry *i, *next, *prev = NULL;
1292
1293         /*
1294          * Simple compaction, dropping empty jset_entries (from journal
1295          * reservations that weren't fully used) and merging jset_entries that
1296          * can be.
1297          *
1298          * If we wanted to be really fancy here, we could sort all the keys in
1299          * the jset and drop keys that were overwritten - probably not worth it:
1300          */
1301         vstruct_for_each_safe(jset, i, next) {
1302                 unsigned u64s = le16_to_cpu(i->u64s);
1303
1304                 /* Empty entry: */
1305                 if (!u64s)
1306                         continue;
1307
1308                 /* Can we merge with previous entry? */
1309                 if (prev &&
1310                     i->btree_id == prev->btree_id &&
1311                     i->level    == prev->level &&
1312                     i->type     == prev->type &&
1313                     i->type     == BCH_JSET_ENTRY_btree_keys &&
1314                     le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
1315                         memmove_u64s_down(vstruct_next(prev),
1316                                           i->_data,
1317                                           u64s);
1318                         le16_add_cpu(&prev->u64s, u64s);
1319                         continue;
1320                 }
1321
1322                 /* Couldn't merge, move i into new position (after prev): */
1323                 prev = prev ? vstruct_next(prev) : jset->start;
1324                 if (i != prev)
1325                         memmove_u64s_down(prev, i, jset_u64s(u64s));
1326         }
1327
1328         prev = prev ? vstruct_next(prev) : jset->start;
1329         jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
1330 }
1331
1332 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1333 {
1334         /* we aren't holding j->lock: */
1335         unsigned new_size = READ_ONCE(j->buf_size_want);
1336         void *new_buf;
1337
1338         if (buf->buf_size >= new_size)
1339                 return;
1340
1341         new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN);
1342         if (!new_buf)
1343                 return;
1344
1345         memcpy(new_buf, buf->data, buf->buf_size);
1346
1347         spin_lock(&j->lock);
1348         swap(buf->data,         new_buf);
1349         swap(buf->buf_size,     new_size);
1350         spin_unlock(&j->lock);
1351
1352         kvpfree(new_buf, new_size);
1353 }
1354
1355 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1356 {
1357         return j->buf + j->reservations.unwritten_idx;
1358 }
1359
1360 static void journal_write_done(struct closure *cl)
1361 {
1362         struct journal *j = container_of(cl, struct journal, io);
1363         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1364         struct journal_buf *w = journal_last_unwritten_buf(j);
1365         struct bch_replicas_padded replicas;
1366         union journal_res_state old, new;
1367         u64 v, seq;
1368         int err = 0;
1369
1370         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1371                                ? j->flush_write_time
1372                                : j->noflush_write_time, j->write_start_time);
1373
1374         if (!w->devs_written.nr) {
1375                 bch_err(c, "unable to write journal to sufficient devices");
1376                 err = -EIO;
1377         } else {
1378                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1379                                          w->devs_written);
1380                 if (bch2_mark_replicas(c, &replicas.e))
1381                         err = -EIO;
1382         }
1383
1384         if (err)
1385                 bch2_fatal_error(c);
1386
1387         spin_lock(&j->lock);
1388         seq = le64_to_cpu(w->data->seq);
1389
1390         if (seq >= j->pin.front)
1391                 journal_seq_pin(j, seq)->devs = w->devs_written;
1392
1393         if (!err) {
1394                 j->seq_ondisk           = seq;
1395
1396                 if (!JSET_NO_FLUSH(w->data)) {
1397                         j->flushed_seq_ondisk = seq;
1398                         j->last_seq_ondisk = w->last_seq;
1399                 }
1400         } else if (!j->err_seq || seq < j->err_seq)
1401                 j->err_seq      = seq;
1402
1403         /*
1404          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1405          * more buckets:
1406          *
1407          * Must come before signaling write completion, for
1408          * bch2_fs_journal_stop():
1409          */
1410         journal_reclaim_kick(&c->journal);
1411
1412         /* also must come before signalling write completion: */
1413         closure_debug_destroy(cl);
1414
1415         v = atomic64_read(&j->reservations.counter);
1416         do {
1417                 old.v = new.v = v;
1418                 BUG_ON(new.idx == new.unwritten_idx);
1419
1420                 new.unwritten_idx++;
1421         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1422                                        old.v, new.v)) != old.v);
1423
1424         bch2_journal_space_available(j);
1425
1426         closure_wake_up(&w->wait);
1427         journal_wake(j);
1428
1429         if (test_bit(JOURNAL_NEED_WRITE, &j->flags))
1430                 mod_delayed_work(c->io_complete_wq, &j->write_work, 0);
1431         spin_unlock(&j->lock);
1432
1433         if (new.unwritten_idx != new.idx &&
1434             !journal_state_count(new, new.unwritten_idx))
1435                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1436 }
1437
1438 static void journal_write_endio(struct bio *bio)
1439 {
1440         struct bch_dev *ca = bio->bi_private;
1441         struct journal *j = &ca->fs->journal;
1442         struct journal_buf *w = journal_last_unwritten_buf(j);
1443         unsigned long flags;
1444
1445         if (bch2_dev_io_err_on(bio->bi_status, ca, "error writing journal entry %llu: %s",
1446                                le64_to_cpu(w->data->seq),
1447                                bch2_blk_status_to_str(bio->bi_status)) ||
1448             bch2_meta_write_fault("journal")) {
1449                 spin_lock_irqsave(&j->err_lock, flags);
1450                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1451                 spin_unlock_irqrestore(&j->err_lock, flags);
1452         }
1453
1454         closure_put(&j->io);
1455         percpu_ref_put(&ca->io_ref);
1456 }
1457
1458 static void do_journal_write(struct closure *cl)
1459 {
1460         struct journal *j = container_of(cl, struct journal, io);
1461         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1462         struct bch_dev *ca;
1463         struct journal_buf *w = journal_last_unwritten_buf(j);
1464         struct bch_extent_ptr *ptr;
1465         struct bio *bio;
1466         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1467
1468         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1469                 ca = bch_dev_bkey_exists(c, ptr->dev);
1470                 if (!percpu_ref_tryget(&ca->io_ref)) {
1471                         /* XXX: fix this */
1472                         bch_err(c, "missing device for journal write\n");
1473                         continue;
1474                 }
1475
1476                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1477                              sectors);
1478
1479                 bio = ca->journal.bio;
1480                 bio_reset(bio);
1481                 bio_set_dev(bio, ca->disk_sb.bdev);
1482                 bio->bi_iter.bi_sector  = ptr->offset;
1483                 bio->bi_end_io          = journal_write_endio;
1484                 bio->bi_private         = ca;
1485                 bio->bi_opf             = REQ_OP_WRITE|REQ_SYNC|REQ_META;
1486
1487                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1488                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1489
1490                 if (!JSET_NO_FLUSH(w->data))
1491                         bio->bi_opf    |= REQ_FUA;
1492                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1493                         bio->bi_opf    |= REQ_PREFLUSH;
1494
1495                 bch2_bio_map(bio, w->data, sectors << 9);
1496
1497                 trace_journal_write(bio);
1498                 closure_bio_submit(bio, cl);
1499
1500                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1501                         le64_to_cpu(w->data->seq);
1502         }
1503
1504         continue_at(cl, journal_write_done, c->io_complete_wq);
1505         return;
1506 }
1507
1508 void bch2_journal_write(struct closure *cl)
1509 {
1510         struct journal *j = container_of(cl, struct journal, io);
1511         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1512         struct bch_dev *ca;
1513         struct journal_buf *w = journal_last_unwritten_buf(j);
1514         struct jset_entry *start, *end;
1515         struct jset *jset;
1516         struct bio *bio;
1517         char *journal_debug_buf = NULL;
1518         bool validate_before_checksum = false;
1519         unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
1520         int ret;
1521
1522         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1523
1524         journal_buf_realloc(j, w);
1525         jset = w->data;
1526
1527         j->write_start_time = local_clock();
1528
1529         spin_lock(&j->lock);
1530         if (c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush) &&
1531             (w->noflush ||
1532              (!w->must_flush &&
1533               (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1534               test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)))) {
1535                 w->noflush = true;
1536                 SET_JSET_NO_FLUSH(jset, true);
1537                 jset->last_seq  = 0;
1538                 w->last_seq     = 0;
1539
1540                 j->nr_noflush_writes++;
1541         } else {
1542                 j->last_flush_write = jiffies;
1543                 j->nr_flush_writes++;
1544         }
1545         spin_unlock(&j->lock);
1546
1547         /*
1548          * New btree roots are set by journalling them; when the journal entry
1549          * gets written we have to propagate them to c->btree_roots
1550          *
1551          * But, every journal entry we write has to contain all the btree roots
1552          * (at least for now); so after we copy btree roots to c->btree_roots we
1553          * have to get any missing btree roots and add them to this journal
1554          * entry:
1555          */
1556
1557         bch2_journal_entries_to_btree_roots(c, jset);
1558
1559         start = end = vstruct_last(jset);
1560
1561         end     = bch2_btree_roots_to_journal_entries(c, jset->start, end);
1562
1563         bch2_journal_super_entries_add_common(c, &end,
1564                                 le64_to_cpu(jset->seq));
1565         u64s    = (u64 *) end - (u64 *) start;
1566         BUG_ON(u64s > j->entry_u64s_reserved);
1567
1568         le32_add_cpu(&jset->u64s, u64s);
1569         BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors);
1570
1571         journal_write_compact(jset);
1572
1573         jset->magic             = cpu_to_le64(jset_magic(c));
1574         jset->version           = c->sb.version < bcachefs_metadata_version_new_versioning
1575                 ? cpu_to_le32(BCH_JSET_VERSION_OLD)
1576                 : cpu_to_le32(c->sb.version);
1577
1578         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1579         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1580
1581         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1582                 j->last_empty_seq = le64_to_cpu(jset->seq);
1583
1584         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1585                 validate_before_checksum = true;
1586
1587         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1588                 validate_before_checksum = true;
1589
1590         if (validate_before_checksum &&
1591             jset_validate_for_write(c, jset))
1592                 goto err;
1593
1594         bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1595                     jset->encrypted_start,
1596                     vstruct_end(jset) - (void *) jset->encrypted_start);
1597
1598         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1599                                   journal_nonce(jset), jset);
1600
1601         if (!validate_before_checksum &&
1602             jset_validate_for_write(c, jset))
1603                 goto err;
1604
1605         sectors = vstruct_sectors(jset, c->block_bits);
1606         BUG_ON(sectors > w->sectors);
1607
1608         bytes = vstruct_bytes(jset);
1609         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1610
1611 retry_alloc:
1612         spin_lock(&j->lock);
1613         ret = journal_write_alloc(j, w, sectors);
1614
1615         if (ret && j->can_discard) {
1616                 spin_unlock(&j->lock);
1617                 bch2_journal_do_discards(j);
1618                 goto retry_alloc;
1619         }
1620
1621         if (ret) {
1622                 journal_debug_buf = kmalloc(4096, GFP_ATOMIC);
1623                 if (journal_debug_buf)
1624                         __bch2_journal_debug_to_text(&_PBUF(journal_debug_buf, 4096), j);
1625         }
1626
1627         /*
1628          * write is allocated, no longer need to account for it in
1629          * bch2_journal_space_available():
1630          */
1631         w->sectors = 0;
1632
1633         /*
1634          * journal entry has been compacted and allocated, recalculate space
1635          * available:
1636          */
1637         bch2_journal_space_available(j);
1638         spin_unlock(&j->lock);
1639
1640         if (ret) {
1641                 bch_err(c, "Unable to allocate journal write:\n%s",
1642                         journal_debug_buf);
1643                 kfree(journal_debug_buf);
1644                 bch2_fatal_error(c);
1645                 continue_at(cl, journal_write_done, c->io_complete_wq);
1646                 return;
1647         }
1648
1649         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1650
1651         if (test_bit(JOURNAL_NOCHANGES, &j->flags))
1652                 goto no_io;
1653
1654         for_each_rw_member(ca, c, i)
1655                 nr_rw_members++;
1656
1657         if (nr_rw_members > 1)
1658                 w->separate_flush = true;
1659
1660         if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
1661                 for_each_rw_member(ca, c, i) {
1662                         percpu_ref_get(&ca->io_ref);
1663
1664                         bio = ca->journal.bio;
1665                         bio_reset(bio);
1666                         bio_set_dev(bio, ca->disk_sb.bdev);
1667                         bio->bi_opf             = REQ_OP_FLUSH;
1668                         bio->bi_end_io          = journal_write_endio;
1669                         bio->bi_private         = ca;
1670                         closure_bio_submit(bio, cl);
1671                 }
1672         }
1673
1674         continue_at(cl, do_journal_write, c->io_complete_wq);
1675         return;
1676 no_io:
1677         continue_at(cl, journal_write_done, c->io_complete_wq);
1678         return;
1679 err:
1680         bch2_fatal_error(c);
1681         continue_at(cl, journal_write_done, c->io_complete_wq);
1682 }