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 "buckets.h"
   5 #include "checksum.h"
   6 #include "error.h"
   7 #include "journal.h"
   8 #include "journal_io.h"
   9 #include "journal_reclaim.h"
  10 #include "replicas.h"
  11
  12 #include <trace/events/bcachefs.h>
  13
  14 struct journal_list {
  15         struct closure          cl;
  16         struct mutex            lock;
  17         struct list_head        *head;
  18         int                     ret;
  19 };
  20
  21 #define JOURNAL_ENTRY_ADD_OK            0
  22 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  23
  24 /*
  25  * Given a journal entry we just read, add it to the list of journal entries to
  26  * be replayed:
  27  */
  28 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  29                              struct journal_list *jlist, struct jset *j)
  30 {
  31         struct journal_replay *i, *pos;
  32         struct list_head *where;
  33         size_t bytes = vstruct_bytes(j);
  34         __le64 last_seq;
  35         int ret;
  36
  37         last_seq = !list_empty(jlist->head)
  38                 ? list_last_entry(jlist->head, struct journal_replay,
  39                                   list)->j.last_seq
  40                 : 0;
  41
  42         /* Is this entry older than the range we need? */
  43         if (le64_to_cpu(j->seq) < le64_to_cpu(last_seq)) {
  44                 ret = JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  45                 goto out;
  46         }
  47
  48         /* Drop entries we don't need anymore */
  49         list_for_each_entry_safe(i, pos, jlist->head, list) {
  50                 if (le64_to_cpu(i->j.seq) >= le64_to_cpu(j->last_seq))
  51                         break;
  52                 list_del(&i->list);
  53                 kvpfree(i, offsetof(struct journal_replay, j) +
  54                         vstruct_bytes(&i->j));
  55         }
  56
  57         list_for_each_entry_reverse(i, jlist->head, list) {
  58                 /* Duplicate? */
  59                 if (le64_to_cpu(j->seq) == le64_to_cpu(i->j.seq)) {
  60                         fsck_err_on(bytes != vstruct_bytes(&i->j) ||
  61                                     memcmp(j, &i->j, bytes), c,
  62                                     "found duplicate but non identical journal entries (seq %llu)",
  63                                     le64_to_cpu(j->seq));
  64                         goto found;
  65                 }
  66
  67                 if (le64_to_cpu(j->seq) > le64_to_cpu(i->j.seq)) {
  68                         where = &i->list;
  69                         goto add;
  70                 }
  71         }
  72
  73         where = jlist->head;
  74 add:
  75         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
  76         if (!i) {
  77                 ret = -ENOMEM;
  78                 goto out;
  79         }
  80
  81         list_add(&i->list, where);
  82         i->devs.nr = 0;
  83         memcpy(&i->j, j, bytes);
  84 found:
  85         if (!bch2_dev_list_has_dev(i->devs, ca->dev_idx))
  86                 bch2_dev_list_add_dev(&i->devs, ca->dev_idx);
  87         else
  88                 fsck_err_on(1, c, "duplicate journal entries on same device");
  89         ret = JOURNAL_ENTRY_ADD_OK;
  90 out:
  91 fsck_err:
  92         return ret;
  93 }
  94
  95 static struct nonce journal_nonce(const struct jset *jset)
  96 {
  97         return (struct nonce) {{
  98                 [0] = 0,
  99                 [1] = ((__le32 *) &jset->seq)[0],
 100                 [2] = ((__le32 *) &jset->seq)[1],
 101                 [3] = BCH_NONCE_JOURNAL,
 102         }};
 103 }
 104
 105 /* this fills in a range with empty jset_entries: */
 106 static void journal_entry_null_range(void *start, void *end)
 107 {
 108         struct jset_entry *entry;
 109
 110         for (entry = start; entry != end; entry = vstruct_next(entry))
 111                 memset(entry, 0, sizeof(*entry));
 112 }
 113
 114 #define JOURNAL_ENTRY_REREAD    5
 115 #define JOURNAL_ENTRY_NONE      6
 116 #define JOURNAL_ENTRY_BAD       7
 117
 118 #define journal_entry_err(c, msg, ...)                                  \
 119 ({                                                                      \
 120         switch (write) {                                                \
 121         case READ:                                                      \
 122                 mustfix_fsck_err(c, msg, ##__VA_ARGS__);                \
 123                 break;                                                  \
 124         case WRITE:                                                     \
 125                 bch_err(c, "corrupt metadata before write:\n"           \
 126                         msg, ##__VA_ARGS__);                            \
 127                 if (bch2_fs_inconsistent(c)) {                          \
 128                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 129                         goto fsck_err;                                  \
 130                 }                                                       \
 131                 break;                                                  \
 132         }                                                               \
 133         true;                                                           \
 134 })
 135
 136 #define journal_entry_err_on(cond, c, msg, ...)                         \
 137         ((cond) ? journal_entry_err(c, msg, ##__VA_ARGS__) : false)
 138
 139 static int journal_validate_key(struct bch_fs *c, struct jset *jset,
 140                                 struct jset_entry *entry,
 141                                 struct bkey_i *k, enum btree_node_type key_type,
 142                                 const char *type, int write)
 143 {
 144         void *next = vstruct_next(entry);
 145         const char *invalid;
 146         unsigned version = le32_to_cpu(jset->version);
 147         int ret = 0;
 148
 149         if (journal_entry_err_on(!k->k.u64s, c,
 150                         "invalid %s in journal: k->u64s 0", type)) {
 151                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 152                 journal_entry_null_range(vstruct_next(entry), next);
 153                 return 0;
 154         }
 155
 156         if (journal_entry_err_on((void *) bkey_next(k) >
 157                                 (void *) vstruct_next(entry), c,
 158                         "invalid %s in journal: extends past end of journal entry",
 159                         type)) {
 160                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 161                 journal_entry_null_range(vstruct_next(entry), next);
 162                 return 0;
 163         }
 164
 165         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT, c,
 166                         "invalid %s in journal: bad format %u",
 167                         type, k->k.format)) {
 168                 le16_add_cpu(&entry->u64s, -k->k.u64s);
 169                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 170                 journal_entry_null_range(vstruct_next(entry), next);
 171                 return 0;
 172         }
 173
 174         if (JSET_BIG_ENDIAN(jset) != CPU_BIG_ENDIAN) {
 175                 bch2_bkey_swab_key(NULL, bkey_to_packed(k));
 176                 bch2_bkey_swab_val(bkey_i_to_s(k));
 177         }
 178
 179         if (!write &&
 180             version < bcachefs_metadata_version_bkey_renumber)
 181                 bch2_bkey_renumber(key_type, bkey_to_packed(k), write);
 182
 183         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(k), key_type);
 184         if (invalid) {
 185                 char buf[160];
 186
 187                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(k));
 188                 mustfix_fsck_err(c, "invalid %s in journal: %s\n%s",
 189                                  type, invalid, buf);
 190
 191                 le16_add_cpu(&entry->u64s, -k->k.u64s);
 192                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 193                 journal_entry_null_range(vstruct_next(entry), next);
 194                 return 0;
 195         }
 196
 197         if (write &&
 198             version < bcachefs_metadata_version_bkey_renumber)
 199                 bch2_bkey_renumber(key_type, bkey_to_packed(k), write);
 200 fsck_err:
 201         return ret;
 202 }
 203
 204 static int journal_entry_validate_btree_keys(struct bch_fs *c,
 205                                              struct jset *jset,
 206                                              struct jset_entry *entry,
 207                                              int write)
 208 {
 209         struct bkey_i *k;
 210
 211         vstruct_for_each(entry, k) {
 212                 int ret = journal_validate_key(c, jset, entry, k,
 213                                 __btree_node_type(entry->level,
 214                                                   entry->btree_id),
 215                                 "key", write);
 216                 if (ret)
 217                         return ret;
 218         }
 219
 220         return 0;
 221 }
 222
 223 static int journal_entry_validate_btree_root(struct bch_fs *c,
 224                                              struct jset *jset,
 225                                              struct jset_entry *entry,
 226                                              int write)
 227 {
 228         struct bkey_i *k = entry->start;
 229         int ret = 0;
 230
 231         if (journal_entry_err_on(!entry->u64s ||
 232                                  le16_to_cpu(entry->u64s) != k->k.u64s, c,
 233                                  "invalid btree root journal entry: wrong number of keys")) {
 234                 void *next = vstruct_next(entry);
 235                 /*
 236                  * we don't want to null out this jset_entry,
 237                  * just the contents, so that later we can tell
 238                  * we were _supposed_ to have a btree root
 239                  */
 240                 entry->u64s = 0;
 241                 journal_entry_null_range(vstruct_next(entry), next);
 242                 return 0;
 243         }
 244
 245         return journal_validate_key(c, jset, entry, k, BKEY_TYPE_BTREE,
 246                                     "btree root", write);
 247 fsck_err:
 248         return ret;
 249 }
 250
 251 static int journal_entry_validate_prio_ptrs(struct bch_fs *c,
 252                                             struct jset *jset,
 253                                             struct jset_entry *entry,
 254                                             int write)
 255 {
 256         /* obsolete, don't care: */
 257         return 0;
 258 }
 259
 260 static int journal_entry_validate_blacklist(struct bch_fs *c,
 261                                             struct jset *jset,
 262                                             struct jset_entry *entry,
 263                                             int write)
 264 {
 265         int ret = 0;
 266
 267         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1, c,
 268                 "invalid journal seq blacklist entry: bad size")) {
 269                 journal_entry_null_range(entry, vstruct_next(entry));
 270         }
 271 fsck_err:
 272         return ret;
 273 }
 274
 275 static int journal_entry_validate_blacklist_v2(struct bch_fs *c,
 276                                                struct jset *jset,
 277                                                struct jset_entry *entry,
 278                                                int write)
 279 {
 280         struct jset_entry_blacklist_v2 *bl_entry;
 281         int ret = 0;
 282
 283         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2, c,
 284                 "invalid journal seq blacklist entry: bad size")) {
 285                 journal_entry_null_range(entry, vstruct_next(entry));
 286                 goto out;
 287         }
 288
 289         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 290
 291         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 292                                  le64_to_cpu(bl_entry->end), c,
 293                 "invalid journal seq blacklist entry: start > end")) {
 294                 journal_entry_null_range(entry, vstruct_next(entry));
 295         }
 296 out:
 297 fsck_err:
 298         return ret;
 299 }
 300
 301 static int journal_entry_validate_usage(struct bch_fs *c,
 302                                         struct jset *jset,
 303                                         struct jset_entry *entry,
 304                                         int write)
 305 {
 306         struct jset_entry_usage *u =
 307                 container_of(entry, struct jset_entry_usage, entry);
 308         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 309         int ret = 0;
 310
 311         if (journal_entry_err_on(bytes < sizeof(*u),
 312                                  c,
 313                                  "invalid journal entry usage: bad size")) {
 314                 journal_entry_null_range(entry, vstruct_next(entry));
 315                 return ret;
 316         }
 317
 318 fsck_err:
 319         return ret;
 320 }
 321
 322 static int journal_entry_validate_data_usage(struct bch_fs *c,
 323                                         struct jset *jset,
 324                                         struct jset_entry *entry,
 325                                         int write)
 326 {
 327         struct jset_entry_data_usage *u =
 328                 container_of(entry, struct jset_entry_data_usage, entry);
 329         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 330         int ret = 0;
 331
 332         if (journal_entry_err_on(bytes < sizeof(*u) ||
 333                                  bytes < sizeof(*u) + u->r.nr_devs,
 334                                  c,
 335                                  "invalid journal entry usage: bad size")) {
 336                 journal_entry_null_range(entry, vstruct_next(entry));
 337                 return ret;
 338         }
 339
 340 fsck_err:
 341         return ret;
 342 }
 343
 344 struct jset_entry_ops {
 345         int (*validate)(struct bch_fs *, struct jset *,
 346                         struct jset_entry *, int);
 347 };
 348
 349 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 350 #define x(f, nr)                                                \
 351         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 352                 .validate       = journal_entry_validate_##f,   \
 353         },
 354         BCH_JSET_ENTRY_TYPES()
 355 #undef x
 356 };
 357
 358 static int journal_entry_validate(struct bch_fs *c, struct jset *jset,
 359                                   struct jset_entry *entry, int write)
 360 {
 361         return entry->type < BCH_JSET_ENTRY_NR
 362                 ? bch2_jset_entry_ops[entry->type].validate(c, jset,
 363                                                             entry, write)
 364                 : 0;
 365 }
 366
 367 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 368                                  int write)
 369 {
 370         struct jset_entry *entry;
 371         int ret = 0;
 372
 373         vstruct_for_each(jset, entry) {
 374                 if (journal_entry_err_on(vstruct_next(entry) >
 375                                          vstruct_last(jset), c,
 376                                 "journal entry extends past end of jset")) {
 377                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 378                         break;
 379                 }
 380
 381                 ret = journal_entry_validate(c, jset, entry, write);
 382                 if (ret)
 383                         break;
 384         }
 385 fsck_err:
 386         return ret;
 387 }
 388
 389 static int jset_validate(struct bch_fs *c,
 390                          struct jset *jset, u64 sector,
 391                          unsigned bucket_sectors_left,
 392                          unsigned sectors_read,
 393                          int write)
 394 {
 395         size_t bytes = vstruct_bytes(jset);
 396         struct bch_csum csum;
 397         unsigned version;
 398         int ret = 0;
 399
 400         if (le64_to_cpu(jset->magic) != jset_magic(c))
 401                 return JOURNAL_ENTRY_NONE;
 402
 403         version = le32_to_cpu(jset->version);
 404         if ((version != BCH_JSET_VERSION_OLD &&
 405              version < bcachefs_metadata_version_min) ||
 406             version >= bcachefs_metadata_version_max) {
 407                 bch_err(c, "unknown journal entry version %u", jset->version);
 408                 return BCH_FSCK_UNKNOWN_VERSION;
 409         }
 410
 411         if (journal_entry_err_on(bytes > bucket_sectors_left << 9, c,
 412                                  "journal entry too big (%zu bytes), sector %lluu",
 413                                  bytes, sector)) {
 414                 /* XXX: note we might have missing journal entries */
 415                 return JOURNAL_ENTRY_BAD;
 416         }
 417
 418         if (bytes > sectors_read << 9)
 419                 return JOURNAL_ENTRY_REREAD;
 420
 421         if (fsck_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)), c,
 422                         "journal entry with unknown csum type %llu sector %lluu",
 423                         JSET_CSUM_TYPE(jset), sector))
 424                 return JOURNAL_ENTRY_BAD;
 425
 426         csum = csum_vstruct(c, JSET_CSUM_TYPE(jset), journal_nonce(jset), jset);
 427         if (journal_entry_err_on(bch2_crc_cmp(csum, jset->csum), c,
 428                                  "journal checksum bad, sector %llu", sector)) {
 429                 /* XXX: retry IO, when we start retrying checksum errors */
 430                 /* XXX: note we might have missing journal entries */
 431                 return JOURNAL_ENTRY_BAD;
 432         }
 433
 434         bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
 435                      jset->encrypted_start,
 436                      vstruct_end(jset) - (void *) jset->encrypted_start);
 437
 438         if (journal_entry_err_on(le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq), c,
 439                                  "invalid journal entry: last_seq > seq"))
 440                 jset->last_seq = jset->seq;
 441
 442         return 0;
 443 fsck_err:
 444         return ret;
 445 }
 446
 447 struct journal_read_buf {
 448         void            *data;
 449         size_t          size;
 450 };
 451
 452 static int journal_read_buf_realloc(struct journal_read_buf *b,
 453                                     size_t new_size)
 454 {
 455         void *n;
 456
 457         /* the bios are sized for this many pages, max: */
 458         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 459                 return -ENOMEM;
 460
 461         new_size = roundup_pow_of_two(new_size);
 462         n = kvpmalloc(new_size, GFP_KERNEL);
 463         if (!n)
 464                 return -ENOMEM;
 465
 466         kvpfree(b->data, b->size);
 467         b->data = n;
 468         b->size = new_size;
 469         return 0;
 470 }
 471
 472 static int journal_read_bucket(struct bch_dev *ca,
 473                                struct journal_read_buf *buf,
 474                                struct journal_list *jlist,
 475                                unsigned bucket)
 476 {
 477         struct bch_fs *c = ca->fs;
 478         struct journal_device *ja = &ca->journal;
 479         struct jset *j = NULL;
 480         unsigned sectors, sectors_read = 0;
 481         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 482             end = offset + ca->mi.bucket_size;
 483         bool saw_bad = false;
 484         int ret = 0;
 485
 486         pr_debug("reading %u", bucket);
 487
 488         while (offset < end) {
 489                 if (!sectors_read) {
 490                         struct bio *bio;
 491 reread:
 492                         sectors_read = min_t(unsigned,
 493                                 end - offset, buf->size >> 9);
 494
 495                         bio = bio_kmalloc(GFP_KERNEL,
 496                                           buf_pages(buf->data,
 497                                                     sectors_read << 9));
 498                         bio_set_dev(bio, ca->disk_sb.bdev);
 499                         bio->bi_iter.bi_sector  = offset;
 500                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
 501                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 502
 503                         ret = submit_bio_wait(bio);
 504                         bio_put(bio);
 505
 506                         if (bch2_dev_io_err_on(ret, ca,
 507                                                "journal read from sector %llu",
 508                                                offset) ||
 509                             bch2_meta_read_fault("journal"))
 510                                 return -EIO;
 511
 512                         j = buf->data;
 513                 }
 514
 515                 ret = jset_validate(c, j, offset,
 516                                     end - offset, sectors_read,
 517                                     READ);
 518                 switch (ret) {
 519                 case BCH_FSCK_OK:
 520                         break;
 521                 case JOURNAL_ENTRY_REREAD:
 522                         if (vstruct_bytes(j) > buf->size) {
 523                                 ret = journal_read_buf_realloc(buf,
 524                                                         vstruct_bytes(j));
 525                                 if (ret)
 526                                         return ret;
 527                         }
 528                         goto reread;
 529                 case JOURNAL_ENTRY_NONE:
 530                         if (!saw_bad)
 531                                 return 0;
 532                         sectors = c->opts.block_size;
 533                         goto next_block;
 534                 case JOURNAL_ENTRY_BAD:
 535                         saw_bad = true;
 536                         sectors = c->opts.block_size;
 537                         goto next_block;
 538                 default:
 539                         return ret;
 540                 }
 541
 542                 /*
 543                  * This happens sometimes if we don't have discards on -
 544                  * when we've partially overwritten a bucket with new
 545                  * journal entries. We don't need the rest of the
 546                  * bucket:
 547                  */
 548                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 549                         return 0;
 550
 551                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 552
 553                 mutex_lock(&jlist->lock);
 554                 ret = journal_entry_add(c, ca, jlist, j);
 555                 mutex_unlock(&jlist->lock);
 556
 557                 switch (ret) {
 558                 case JOURNAL_ENTRY_ADD_OK:
 559                         break;
 560                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
 561                         break;
 562                 default:
 563                         return ret;
 564                 }
 565
 566                 sectors = vstruct_sectors(j, c->block_bits);
 567 next_block:
 568                 pr_debug("next");
 569                 offset          += sectors;
 570                 sectors_read    -= sectors;
 571                 j = ((void *) j) + (sectors << 9);
 572         }
 573
 574         return 0;
 575 }
 576
 577 static void bch2_journal_read_device(struct closure *cl)
 578 {
 579         struct journal_device *ja =
 580                 container_of(cl, struct journal_device, read);
 581         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
 582         struct journal_list *jlist =
 583                 container_of(cl->parent, struct journal_list, cl);
 584         struct journal_read_buf buf = { NULL, 0 };
 585         u64 min_seq = U64_MAX;
 586         unsigned i;
 587         int ret;
 588
 589         if (!ja->nr)
 590                 goto out;
 591
 592         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
 593         if (ret)
 594                 goto err;
 595
 596         pr_debug("%u journal buckets", ja->nr);
 597
 598         for (i = 0; i < ja->nr; i++) {
 599                 ret = journal_read_bucket(ca, &buf, jlist, i);
 600                 if (ret)
 601                         goto err;
 602         }
 603
 604         /* Find the journal bucket with the highest sequence number: */
 605         for (i = 0; i < ja->nr; i++) {
 606                 if (ja->bucket_seq[i] > ja->bucket_seq[ja->cur_idx])
 607                         ja->cur_idx = i;
 608
 609                 min_seq = min(ja->bucket_seq[i], min_seq);
 610         }
 611
 612         /*
 613          * If there's duplicate journal entries in multiple buckets (which
 614          * definitely isn't supposed to happen, but...) - make sure to start
 615          * cur_idx at the last of those buckets, so we don't deadlock trying to
 616          * allocate
 617          */
 618         while (ja->bucket_seq[ja->cur_idx] > min_seq &&
 619                ja->bucket_seq[ja->cur_idx] >
 620                ja->bucket_seq[(ja->cur_idx + 1) % ja->nr])
 621                 ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 622
 623         ja->sectors_free = 0;
 624
 625         /*
 626          * Set dirty_idx to indicate the entire journal is full and needs to be
 627          * reclaimed - journal reclaim will immediately reclaim whatever isn't
 628          * pinned when it first runs:
 629          */
 630         ja->discard_idx = ja->dirty_idx_ondisk =
 631                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
 632 out:
 633         kvpfree(buf.data, buf.size);
 634         percpu_ref_put(&ca->io_ref);
 635         closure_return(cl);
 636         return;
 637 err:
 638         mutex_lock(&jlist->lock);
 639         jlist->ret = ret;
 640         mutex_unlock(&jlist->lock);
 641         goto out;
 642 }
 643
 644 int bch2_journal_read(struct bch_fs *c, struct list_head *list)
 645 {
 646         struct journal_list jlist;
 647         struct journal_replay *i;
 648         struct bch_dev *ca;
 649         unsigned iter;
 650         size_t keys = 0, entries = 0;
 651         bool degraded = false;
 652         int ret = 0;
 653
 654         closure_init_stack(&jlist.cl);
 655         mutex_init(&jlist.lock);
 656         jlist.head = list;
 657         jlist.ret = 0;
 658
 659         for_each_member_device(ca, c, iter) {
 660                 if (!test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) &&
 661                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_JOURNAL)))
 662                         continue;
 663
 664                 if ((ca->mi.state == BCH_MEMBER_STATE_RW ||
 665                      ca->mi.state == BCH_MEMBER_STATE_RO) &&
 666                     percpu_ref_tryget(&ca->io_ref))
 667                         closure_call(&ca->journal.read,
 668                                      bch2_journal_read_device,
 669                                      system_unbound_wq,
 670                                      &jlist.cl);
 671                 else
 672                         degraded = true;
 673         }
 674
 675         closure_sync(&jlist.cl);
 676
 677         if (jlist.ret)
 678                 return jlist.ret;
 679
 680         list_for_each_entry(i, list, list) {
 681                 struct jset_entry *entry;
 682                 struct bkey_i *k, *_n;
 683                 struct bch_replicas_padded replicas;
 684                 char buf[80];
 685
 686                 ret = jset_validate_entries(c, &i->j, READ);
 687                 if (ret)
 688                         goto fsck_err;
 689
 690                 /*
 691                  * If we're mounting in degraded mode - if we didn't read all
 692                  * the devices - this is wrong:
 693                  */
 694
 695                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_JOURNAL, i->devs);
 696
 697                 if (!degraded &&
 698                     (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 699                      fsck_err_on(!bch2_replicas_marked(c, &replicas.e, false), c,
 700                                  "superblock not marked as containing replicas %s",
 701                                  (bch2_replicas_entry_to_text(&PBUF(buf),
 702                                                               &replicas.e), buf)))) {
 703                         ret = bch2_mark_replicas(c, &replicas.e);
 704                         if (ret)
 705                                 return ret;
 706                 }
 707
 708                 for_each_jset_key(k, _n, entry, &i->j)
 709                         keys++;
 710                 entries++;
 711         }
 712
 713         if (!list_empty(list)) {
 714                 i = list_last_entry(list, struct journal_replay, list);
 715
 716                 bch_info(c, "journal read done, %zu keys in %zu entries, seq %llu",
 717                          keys, entries, le64_to_cpu(i->j.seq));
 718         }
 719 fsck_err:
 720         return ret;
 721 }
 722
 723 /* journal write: */
 724
 725 static void __journal_write_alloc(struct journal *j,
 726                                   struct journal_buf *w,
 727                                   struct dev_alloc_list *devs_sorted,
 728                                   unsigned sectors,
 729                                   unsigned *replicas,
 730                                   unsigned replicas_want)
 731 {
 732         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 733         struct journal_device *ja;
 734         struct bch_dev *ca;
 735         unsigned i;
 736
 737         if (*replicas >= replicas_want)
 738                 return;
 739
 740         for (i = 0; i < devs_sorted->nr; i++) {
 741                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
 742                 if (!ca)
 743                         continue;
 744
 745                 ja = &ca->journal;
 746
 747                 /*
 748                  * Check that we can use this device, and aren't already using
 749                  * it:
 750                  */
 751                 if (!ca->mi.durability ||
 752                     ca->mi.state != BCH_MEMBER_STATE_RW ||
 753                     !ja->nr ||
 754                     bch2_bkey_has_device(bkey_i_to_s_c(&w->key),
 755                                          ca->dev_idx) ||
 756                     sectors > ja->sectors_free)
 757                         continue;
 758
 759                 bch2_dev_stripe_increment(c, ca, &j->wp.stripe);
 760
 761                 bch2_bkey_append_ptr(&w->key,
 762                         (struct bch_extent_ptr) {
 763                                   .offset = bucket_to_sector(ca,
 764                                         ja->buckets[ja->cur_idx]) +
 765                                         ca->mi.bucket_size -
 766                                         ja->sectors_free,
 767                                   .dev = ca->dev_idx,
 768                 });
 769
 770                 ja->sectors_free -= sectors;
 771                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
 772
 773                 *replicas += ca->mi.durability;
 774
 775                 if (*replicas >= replicas_want)
 776                         break;
 777         }
 778 }
 779
 780 /**
 781  * journal_next_bucket - move on to the next journal bucket if possible
 782  */
 783 static int journal_write_alloc(struct journal *j, struct journal_buf *w,
 784                                unsigned sectors)
 785 {
 786         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 787         struct journal_device *ja;
 788         struct bch_dev *ca;
 789         struct dev_alloc_list devs_sorted;
 790         unsigned i, replicas = 0, replicas_want =
 791                 READ_ONCE(c->opts.metadata_replicas);
 792
 793         rcu_read_lock();
 794
 795         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe,
 796                                           &c->rw_devs[BCH_DATA_JOURNAL]);
 797
 798         __journal_write_alloc(j, w, &devs_sorted,
 799                               sectors, &replicas, replicas_want);
 800
 801         if (replicas >= replicas_want)
 802                 goto done;
 803
 804         for (i = 0; i < devs_sorted.nr; i++) {
 805                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 806                 if (!ca)
 807                         continue;
 808
 809                 ja = &ca->journal;
 810
 811                 if (sectors > ja->sectors_free &&
 812                     sectors <= ca->mi.bucket_size &&
 813                     bch2_journal_dev_buckets_available(j, ja,
 814                                         journal_space_discarded)) {
 815                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
 816                         ja->sectors_free = ca->mi.bucket_size;
 817
 818                         /*
 819                          * ja->bucket_seq[ja->cur_idx] must always have
 820                          * something sensible:
 821                          */
 822                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
 823                 }
 824         }
 825
 826         __journal_write_alloc(j, w, &devs_sorted,
 827                               sectors, &replicas, replicas_want);
 828 done:
 829         rcu_read_unlock();
 830
 831         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
 832 }
 833
 834 static void journal_write_compact(struct jset *jset)
 835 {
 836         struct jset_entry *i, *next, *prev = NULL;
 837
 838         /*
 839          * Simple compaction, dropping empty jset_entries (from journal
 840          * reservations that weren't fully used) and merging jset_entries that
 841          * can be.
 842          *
 843          * If we wanted to be really fancy here, we could sort all the keys in
 844          * the jset and drop keys that were overwritten - probably not worth it:
 845          */
 846         vstruct_for_each_safe(jset, i, next) {
 847                 unsigned u64s = le16_to_cpu(i->u64s);
 848
 849                 /* Empty entry: */
 850                 if (!u64s)
 851                         continue;
 852
 853                 /* Can we merge with previous entry? */
 854                 if (prev &&
 855                     i->btree_id == prev->btree_id &&
 856                     i->level    == prev->level &&
 857                     i->type     == prev->type &&
 858                     i->type     == BCH_JSET_ENTRY_btree_keys &&
 859                     le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
 860                         memmove_u64s_down(vstruct_next(prev),
 861                                           i->_data,
 862                                           u64s);
 863                         le16_add_cpu(&prev->u64s, u64s);
 864                         continue;
 865                 }
 866
 867                 /* Couldn't merge, move i into new position (after prev): */
 868                 prev = prev ? vstruct_next(prev) : jset->start;
 869                 if (i != prev)
 870                         memmove_u64s_down(prev, i, jset_u64s(u64s));
 871         }
 872
 873         prev = prev ? vstruct_next(prev) : jset->start;
 874         jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
 875 }
 876
 877 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
 878 {
 879         /* we aren't holding j->lock: */
 880         unsigned new_size = READ_ONCE(j->buf_size_want);
 881         void *new_buf;
 882
 883         if (buf->buf_size >= new_size)
 884                 return;
 885
 886         new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN);
 887         if (!new_buf)
 888                 return;
 889
 890         memcpy(new_buf, buf->data, buf->buf_size);
 891         kvpfree(buf->data, buf->buf_size);
 892         buf->data       = new_buf;
 893         buf->buf_size   = new_size;
 894 }
 895
 896 static void journal_write_done(struct closure *cl)
 897 {
 898         struct journal *j = container_of(cl, struct journal, io);
 899         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 900         struct journal_buf *w = journal_prev_buf(j);
 901         struct bch_devs_list devs =
 902                 bch2_bkey_devs(bkey_i_to_s_c(&w->key));
 903         struct bch_replicas_padded replicas;
 904         u64 seq = le64_to_cpu(w->data->seq);
 905         u64 last_seq = le64_to_cpu(w->data->last_seq);
 906
 907         bch2_time_stats_update(j->write_time, j->write_start_time);
 908
 909         if (!devs.nr) {
 910                 bch_err(c, "unable to write journal to sufficient devices");
 911                 goto err;
 912         }
 913
 914         bch2_devlist_to_replicas(&replicas.e, BCH_DATA_JOURNAL, devs);
 915
 916         if (bch2_mark_replicas(c, &replicas.e))
 917                 goto err;
 918
 919         spin_lock(&j->lock);
 920         if (seq >= j->pin.front)
 921                 journal_seq_pin(j, seq)->devs = devs;
 922
 923         j->seq_ondisk           = seq;
 924         j->last_seq_ondisk      = last_seq;
 925         bch2_journal_space_available(j);
 926
 927         /*
 928          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
 929          * more buckets:
 930          *
 931          * Must come before signaling write completion, for
 932          * bch2_fs_journal_stop():
 933          */
 934         mod_delayed_work(c->journal_reclaim_wq, &j->reclaim_work, 0);
 935 out:
 936         /* also must come before signalling write completion: */
 937         closure_debug_destroy(cl);
 938
 939         BUG_ON(!j->reservations.prev_buf_unwritten);
 940         atomic64_sub(((union journal_res_state) { .prev_buf_unwritten = 1 }).v,
 941                      &j->reservations.counter);
 942
 943         closure_wake_up(&w->wait);
 944         journal_wake(j);
 945
 946         if (test_bit(JOURNAL_NEED_WRITE, &j->flags))
 947                 mod_delayed_work(system_freezable_wq, &j->write_work, 0);
 948         spin_unlock(&j->lock);
 949         return;
 950 err:
 951         bch2_fatal_error(c);
 952         spin_lock(&j->lock);
 953         goto out;
 954 }
 955
 956 static void journal_write_endio(struct bio *bio)
 957 {
 958         struct bch_dev *ca = bio->bi_private;
 959         struct journal *j = &ca->fs->journal;
 960
 961         if (bch2_dev_io_err_on(bio->bi_status, ca, "journal write") ||
 962             bch2_meta_write_fault("journal")) {
 963                 struct journal_buf *w = journal_prev_buf(j);
 964                 unsigned long flags;
 965
 966                 spin_lock_irqsave(&j->err_lock, flags);
 967                 bch2_bkey_drop_device(bkey_i_to_s(&w->key), ca->dev_idx);
 968                 spin_unlock_irqrestore(&j->err_lock, flags);
 969         }
 970
 971         closure_put(&j->io);
 972         percpu_ref_put(&ca->io_ref);
 973 }
 974
 975 void bch2_journal_write(struct closure *cl)
 976 {
 977         struct journal *j = container_of(cl, struct journal, io);
 978         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 979         struct bch_dev *ca;
 980         struct journal_buf *w = journal_prev_buf(j);
 981         struct jset_entry *start, *end;
 982         struct jset *jset;
 983         struct bio *bio;
 984         struct bch_extent_ptr *ptr;
 985         bool validate_before_checksum = false;
 986         unsigned i, sectors, bytes, u64s;
 987         int ret;
 988
 989         bch2_journal_pin_put(j, le64_to_cpu(w->data->seq));
 990
 991         journal_buf_realloc(j, w);
 992         jset = w->data;
 993
 994         j->write_start_time = local_clock();
 995
 996         start   = vstruct_last(jset);
 997         end     = bch2_journal_super_entries_add_common(c, start,
 998                                                 le64_to_cpu(jset->seq));
 999         u64s    = (u64 *) end - (u64 *) start;
1000         BUG_ON(u64s > j->entry_u64s_reserved);
1001
1002         le32_add_cpu(&jset->u64s, u64s);
1003         BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors);
1004
1005         journal_write_compact(jset);
1006
1007         jset->read_clock        = cpu_to_le16(c->bucket_clock[READ].hand);
1008         jset->write_clock       = cpu_to_le16(c->bucket_clock[WRITE].hand);
1009         jset->magic             = cpu_to_le64(jset_magic(c));
1010
1011         jset->version           = c->sb.version < bcachefs_metadata_version_new_versioning
1012                 ? cpu_to_le32(BCH_JSET_VERSION_OLD)
1013                 : cpu_to_le32(c->sb.version);
1014
1015         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1016         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1017
1018         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1019                 validate_before_checksum = true;
1020
1021         if (le32_to_cpu(jset->version) <
1022             bcachefs_metadata_version_bkey_renumber)
1023                 validate_before_checksum = true;
1024
1025         if (validate_before_checksum &&
1026             jset_validate_entries(c, jset, WRITE))
1027                 goto err;
1028
1029         bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1030                     jset->encrypted_start,
1031                     vstruct_end(jset) - (void *) jset->encrypted_start);
1032
1033         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1034                                   journal_nonce(jset), jset);
1035
1036         if (!validate_before_checksum &&
1037             jset_validate_entries(c, jset, WRITE))
1038                 goto err;
1039
1040         sectors = vstruct_sectors(jset, c->block_bits);
1041         BUG_ON(sectors > w->sectors);
1042
1043         bytes = vstruct_bytes(jset);
1044         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1045
1046         spin_lock(&j->lock);
1047         ret = journal_write_alloc(j, w, sectors);
1048
1049         /*
1050          * write is allocated, no longer need to account for it in
1051          * bch2_journal_space_available():
1052          */
1053         w->sectors = 0;
1054
1055         /*
1056          * journal entry has been compacted and allocated, recalculate space
1057          * available:
1058          */
1059         bch2_journal_space_available(j);
1060         spin_unlock(&j->lock);
1061
1062         if (ret) {
1063                 bch_err(c, "Unable to allocate journal write");
1064                 bch2_fatal_error(c);
1065                 continue_at(cl, journal_write_done, system_highpri_wq);
1066                 return;
1067         }
1068
1069         /*
1070          * XXX: we really should just disable the entire journal in nochanges
1071          * mode
1072          */
1073         if (c->opts.nochanges)
1074                 goto no_io;
1075
1076         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1077                 ca = bch_dev_bkey_exists(c, ptr->dev);
1078                 if (!percpu_ref_tryget(&ca->io_ref)) {
1079                         /* XXX: fix this */
1080                         bch_err(c, "missing device for journal write\n");
1081                         continue;
1082                 }
1083
1084                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_JOURNAL],
1085                              sectors);
1086
1087                 bio = ca->journal.bio;
1088                 bio_reset(bio);
1089                 bio_set_dev(bio, ca->disk_sb.bdev);
1090                 bio->bi_iter.bi_sector  = ptr->offset;
1091                 bio->bi_end_io          = journal_write_endio;
1092                 bio->bi_private         = ca;
1093                 bio_set_op_attrs(bio, REQ_OP_WRITE,
1094                                  REQ_SYNC|REQ_META|REQ_PREFLUSH|REQ_FUA);
1095                 bch2_bio_map(bio, jset, sectors << 9);
1096
1097                 trace_journal_write(bio);
1098                 closure_bio_submit(bio, cl);
1099
1100                 ca->journal.bucket_seq[ca->journal.cur_idx] = le64_to_cpu(jset->seq);
1101         }
1102
1103         for_each_rw_member(ca, c, i)
1104                 if (journal_flushes_device(ca) &&
1105                     !bch2_bkey_has_device(bkey_i_to_s_c(&w->key), i)) {
1106                         percpu_ref_get(&ca->io_ref);
1107
1108                         bio = ca->journal.bio;
1109                         bio_reset(bio);
1110                         bio_set_dev(bio, ca->disk_sb.bdev);
1111                         bio->bi_opf             = REQ_OP_FLUSH;
1112                         bio->bi_end_io          = journal_write_endio;
1113                         bio->bi_private         = ca;
1114                         closure_bio_submit(bio, cl);
1115                 }
1116
1117 no_io:
1118         bch2_bucket_seq_cleanup(c);
1119
1120         continue_at(cl, journal_write_done, system_highpri_wq);
1121         return;
1122 err:
1123         bch2_inconsistent_error(c);
1124         continue_at(cl, journal_write_done, system_highpri_wq);
1125 }