git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/journal_io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_io.h"
   6 #include "btree_update_interior.h"
   7 #include "buckets.h"
   8 #include "checksum.h"
   9 #include "disk_groups.h"
  10 #include "error.h"
  11 #include "io.h"
  12 #include "journal.h"
  13 #include "journal_io.h"
  14 #include "journal_reclaim.h"
  15 #include "journal_seq_blacklist.h"
  16 #include "replicas.h"
  17
  18 #include <trace/events/bcachefs.h>
  19
  20 static struct nonce journal_nonce(const struct jset *jset)
  21 {
  22         return (struct nonce) {{
  23                 [0] = 0,
  24                 [1] = ((__le32 *) &jset->seq)[0],
  25                 [2] = ((__le32 *) &jset->seq)[1],
  26                 [3] = BCH_NONCE_JOURNAL,
  27         }};
  28 }
  29
  30 static bool jset_csum_good(struct bch_fs *c, struct jset *j)
  31 {
  32         return bch2_checksum_type_valid(c, JSET_CSUM_TYPE(j)) &&
  33                 !bch2_crc_cmp(j->csum,
  34                               csum_vstruct(c, JSET_CSUM_TYPE(j), journal_nonce(j), j));
  35 }
  36
  37 static inline u32 journal_entry_radix_idx(struct bch_fs *c, u64 seq)
  38 {
  39         return (seq - c->journal_entries_base_seq) & (~0U >> 1);
  40 }
  41
  42 static void __journal_replay_free(struct bch_fs *c,
  43                                   struct journal_replay *i)
  44 {
  45         struct journal_replay **p =
  46                 genradix_ptr(&c->journal_entries,
  47                              journal_entry_radix_idx(c, le64_to_cpu(i->j.seq)));
  48
  49         BUG_ON(*p != i);
  50         *p = NULL;
  51         kvpfree(i, offsetof(struct journal_replay, j) +
  52                 vstruct_bytes(&i->j));
  53 }
  54
  55 static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
  56 {
  57         i->ignore = true;
  58
  59         if (!c->opts.read_entire_journal)
  60                 __journal_replay_free(c, i);
  61 }
  62
  63 struct journal_list {
  64         struct closure          cl;
  65         u64                     last_seq;
  66         struct mutex            lock;
  67         int                     ret;
  68 };
  69
  70 #define JOURNAL_ENTRY_ADD_OK            0
  71 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  72
  73 /*
  74  * Given a journal entry we just read, add it to the list of journal entries to
  75  * be replayed:
  76  */
  77 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  78                              struct journal_ptr entry_ptr,
  79                              struct journal_list *jlist, struct jset *j)
  80 {
  81         struct genradix_iter iter;
  82         struct journal_replay **_i, *i, *dup;
  83         struct journal_ptr *ptr;
  84         size_t bytes = vstruct_bytes(j);
  85         u64 last_seq = !JSET_NO_FLUSH(j) ? le64_to_cpu(j->last_seq) : 0;
  86         int ret = JOURNAL_ENTRY_ADD_OK;
  87
  88         /* Is this entry older than the range we need? */
  89         if (!c->opts.read_entire_journal &&
  90             le64_to_cpu(j->seq) < jlist->last_seq)
  91                 return JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  92
  93         /*
  94          * genradixes are indexed by a ulong, not a u64, so we can't index them
  95          * by sequence number directly: Assume instead that they will all fall
  96          * within the range of +-2billion of the filrst one we find.
  97          */
  98         if (!c->journal_entries_base_seq)
  99                 c->journal_entries_base_seq = max_t(s64, 1, le64_to_cpu(j->seq) - S32_MAX);
 100
 101         /* Drop entries we don't need anymore */
 102         if (last_seq > jlist->last_seq && !c->opts.read_entire_journal) {
 103                 genradix_for_each_from(&c->journal_entries, iter, _i,
 104                                        journal_entry_radix_idx(c, jlist->last_seq)) {
 105                         i = *_i;
 106
 107                         if (!i || i->ignore)
 108                                 continue;
 109
 110                         if (le64_to_cpu(i->j.seq) >= last_seq)
 111                                 break;
 112                         journal_replay_free(c, i);
 113                 }
 114         }
 115
 116         jlist->last_seq = max(jlist->last_seq, last_seq);
 117
 118         _i = genradix_ptr_alloc(&c->journal_entries,
 119                                 journal_entry_radix_idx(c, le64_to_cpu(j->seq)),
 120                                 GFP_KERNEL);
 121         if (!_i)
 122                 return -ENOMEM;
 123
 124         /*
 125          * Duplicate journal entries? If so we want the one that didn't have a
 126          * checksum error:
 127          */
 128         dup = *_i;
 129         if (dup) {
 130                 if (bytes == vstruct_bytes(&dup->j) &&
 131                     !memcmp(j, &dup->j, bytes)) {
 132                         i = dup;
 133                         goto found;
 134                 }
 135
 136                 if (!entry_ptr.csum_good) {
 137                         i = dup;
 138                         goto found;
 139                 }
 140
 141                 if (!dup->csum_good)
 142                         goto replace;
 143
 144                 fsck_err(c, "found duplicate but non identical journal entries (seq %llu)",
 145                          le64_to_cpu(j->seq));
 146                 i = dup;
 147                 goto found;
 148         }
 149 replace:
 150         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 151         if (!i)
 152                 return -ENOMEM;
 153
 154         i->nr_ptrs      = 0;
 155         i->csum_good    = entry_ptr.csum_good;
 156         i->ignore       = false;
 157         memcpy(&i->j, j, bytes);
 158         i->ptrs[i->nr_ptrs++] = entry_ptr;
 159
 160         if (dup) {
 161                 if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) {
 162                         bch_err(c, "found too many copies of journal entry %llu",
 163                                 le64_to_cpu(i->j.seq));
 164                         dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1;
 165                 }
 166
 167                 /* The first ptr should represent the jset we kept: */
 168                 memcpy(i->ptrs + i->nr_ptrs,
 169                        dup->ptrs,
 170                        sizeof(dup->ptrs[0]) * dup->nr_ptrs);
 171                 i->nr_ptrs += dup->nr_ptrs;
 172                 __journal_replay_free(c, dup);
 173         }
 174
 175         *_i = i;
 176         return 0;
 177 found:
 178         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 179                 if (ptr->dev == ca->dev_idx) {
 180                         bch_err(c, "duplicate journal entry %llu on same device",
 181                                 le64_to_cpu(i->j.seq));
 182                         goto out;
 183                 }
 184         }
 185
 186         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 187                 bch_err(c, "found too many copies of journal entry %llu",
 188                         le64_to_cpu(i->j.seq));
 189                 goto out;
 190         }
 191
 192         i->ptrs[i->nr_ptrs++] = entry_ptr;
 193 out:
 194 fsck_err:
 195         return ret;
 196 }
 197
 198 /* this fills in a range with empty jset_entries: */
 199 static void journal_entry_null_range(void *start, void *end)
 200 {
 201         struct jset_entry *entry;
 202
 203         for (entry = start; entry != end; entry = vstruct_next(entry))
 204                 memset(entry, 0, sizeof(*entry));
 205 }
 206
 207 #define JOURNAL_ENTRY_REREAD    5
 208 #define JOURNAL_ENTRY_NONE      6
 209 #define JOURNAL_ENTRY_BAD       7
 210
 211 static void journal_entry_err_msg(struct printbuf *out,
 212                                   struct jset *jset,
 213                                   struct jset_entry *entry)
 214 {
 215         prt_str(out, "invalid journal entry ");
 216         if (entry)
 217                 prt_printf(out, "%s ", bch2_jset_entry_types[entry->type]);
 218
 219         if (!jset)
 220                 prt_printf(out, "in superblock");
 221         else if (!entry)
 222                 prt_printf(out, "at seq %llu", le64_to_cpu(jset->seq));
 223         else
 224                 prt_printf(out, "at offset %zi/%u seq %llu",
 225                            (u64 *) entry - jset->_data,
 226                            le32_to_cpu(jset->u64s),
 227                            le64_to_cpu(jset->seq));
 228         prt_str(out, ": ");
 229 }
 230
 231 #define journal_entry_err(c, jset, entry, msg, ...)                     \
 232 ({                                                                      \
 233         struct printbuf buf = PRINTBUF;                                 \
 234                                                                         \
 235         journal_entry_err_msg(&buf, jset, entry);                       \
 236         prt_printf(&buf, msg, ##__VA_ARGS__);                           \
 237                                                                         \
 238         switch (write) {                                                \
 239         case READ:                                                      \
 240                 mustfix_fsck_err(c, "%s", buf.buf);                     \
 241                 break;                                                  \
 242         case WRITE:                                                     \
 243                 bch_err(c, "corrupt metadata before write: %s\n", buf.buf);\
 244                 if (bch2_fs_inconsistent(c)) {                          \
 245                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 246                         goto fsck_err;                                  \
 247                 }                                                       \
 248                 break;                                                  \
 249         }                                                               \
 250                                                                         \
 251         printbuf_exit(&buf);                                            \
 252         true;                                                           \
 253 })
 254
 255 #define journal_entry_err_on(cond, c, jset, entry, msg, ...)            \
 256         ((cond) ? journal_entry_err(c, jset, entry, msg, ##__VA_ARGS__) : false)
 257
 258 #define FSCK_DELETED_KEY        5
 259
 260 static int journal_validate_key(struct bch_fs *c,
 261                                 struct jset *jset,
 262                                 struct jset_entry *entry,
 263                                 unsigned level, enum btree_id btree_id,
 264                                 struct bkey_i *k,
 265                                 unsigned version, int big_endian, int write)
 266 {
 267         void *next = vstruct_next(entry);
 268         struct printbuf buf = PRINTBUF;
 269         int ret = 0;
 270
 271         if (journal_entry_err_on(!k->k.u64s, c, jset, entry, "k->u64s 0")) {
 272                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 273                 journal_entry_null_range(vstruct_next(entry), next);
 274                 return FSCK_DELETED_KEY;
 275         }
 276
 277         if (journal_entry_err_on((void *) bkey_next(k) >
 278                                  (void *) vstruct_next(entry),
 279                                  c, jset, entry,
 280                                  "extends past end of journal entry")) {
 281                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 282                 journal_entry_null_range(vstruct_next(entry), next);
 283                 return FSCK_DELETED_KEY;
 284         }
 285
 286         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
 287                                  c, jset, entry,
 288                                  "bad format %u", k->k.format)) {
 289                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 290                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 291                 journal_entry_null_range(vstruct_next(entry), next);
 292                 return FSCK_DELETED_KEY;
 293         }
 294
 295         if (!write)
 296                 bch2_bkey_compat(level, btree_id, version, big_endian,
 297                                  write, NULL, bkey_to_packed(k));
 298
 299         if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 300                               __btree_node_type(level, btree_id), write, &buf)) {
 301                 printbuf_reset(&buf);
 302                 prt_printf(&buf, "invalid journal entry %s at offset %zi/%u seq %llu:",
 303                            bch2_jset_entry_types[entry->type],
 304                            (u64 *) entry - jset->_data,
 305                            le32_to_cpu(jset->u64s),
 306                            le64_to_cpu(jset->seq));
 307                 prt_newline(&buf);
 308                 printbuf_indent_add(&buf, 2);
 309
 310                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
 311                 prt_newline(&buf);
 312                 bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 313                                   __btree_node_type(level, btree_id), write, &buf);
 314
 315                 mustfix_fsck_err(c, "%s", buf.buf);
 316
 317                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 318                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 319                 journal_entry_null_range(vstruct_next(entry), next);
 320
 321                 printbuf_exit(&buf);
 322                 return FSCK_DELETED_KEY;
 323         }
 324
 325         if (write)
 326                 bch2_bkey_compat(level, btree_id, version, big_endian,
 327                                  write, NULL, bkey_to_packed(k));
 328 fsck_err:
 329         printbuf_exit(&buf);
 330         return ret;
 331 }
 332
 333 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 334                                              struct jset *jset,
 335                                              struct jset_entry *entry,
 336                                              unsigned version, int big_endian, int write)
 337 {
 338         struct bkey_i *k = entry->start;
 339
 340         while (k != vstruct_last(entry)) {
 341                 int ret = journal_validate_key(c, jset, entry,
 342                                                entry->level,
 343                                                entry->btree_id,
 344                                                k, version, big_endian, write);
 345                 if (ret == FSCK_DELETED_KEY)
 346                         continue;
 347
 348                 k = bkey_next(k);
 349         }
 350
 351         return 0;
 352 }
 353
 354 static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
 355                                              struct jset_entry *entry)
 356 {
 357         struct bkey_i *k;
 358         bool first = true;
 359
 360         vstruct_for_each(entry, k) {
 361                 if (!first) {
 362                         prt_newline(out);
 363                         prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 364                 }
 365                 prt_printf(out, "btree=%s l=%u ", bch2_btree_ids[entry->btree_id], entry->level);
 366                 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 367                 first = false;
 368         }
 369 }
 370
 371 static int journal_entry_btree_root_validate(struct bch_fs *c,
 372                                              struct jset *jset,
 373                                              struct jset_entry *entry,
 374                                              unsigned version, int big_endian, int write)
 375 {
 376         struct bkey_i *k = entry->start;
 377         int ret = 0;
 378
 379         if (journal_entry_err_on(!entry->u64s ||
 380                                  le16_to_cpu(entry->u64s) != k->k.u64s,
 381                                  c, jset, entry,
 382                                  "invalid btree root journal entry: wrong number of keys")) {
 383                 void *next = vstruct_next(entry);
 384                 /*
 385                  * we don't want to null out this jset_entry,
 386                  * just the contents, so that later we can tell
 387                  * we were _supposed_ to have a btree root
 388                  */
 389                 entry->u64s = 0;
 390                 journal_entry_null_range(vstruct_next(entry), next);
 391                 return 0;
 392         }
 393
 394         return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
 395                                     version, big_endian, write);
 396 fsck_err:
 397         return ret;
 398 }
 399
 400 static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
 401                                              struct jset_entry *entry)
 402 {
 403         journal_entry_btree_keys_to_text(out, c, entry);
 404 }
 405
 406 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 407                                             struct jset *jset,
 408                                             struct jset_entry *entry,
 409                                             unsigned version, int big_endian, int write)
 410 {
 411         /* obsolete, don't care: */
 412         return 0;
 413 }
 414
 415 static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 416                                             struct jset_entry *entry)
 417 {
 418 }
 419
 420 static int journal_entry_blacklist_validate(struct bch_fs *c,
 421                                             struct jset *jset,
 422                                             struct jset_entry *entry,
 423                                             unsigned version, int big_endian, int write)
 424 {
 425         int ret = 0;
 426
 427         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1,
 428                                  c, jset, entry,
 429                 "invalid journal seq blacklist entry: bad size")) {
 430                 journal_entry_null_range(entry, vstruct_next(entry));
 431         }
 432 fsck_err:
 433         return ret;
 434 }
 435
 436 static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
 437                                             struct jset_entry *entry)
 438 {
 439         struct jset_entry_blacklist *bl =
 440                 container_of(entry, struct jset_entry_blacklist, entry);
 441
 442         prt_printf(out, "seq=%llu", le64_to_cpu(bl->seq));
 443 }
 444
 445 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 446                                                struct jset *jset,
 447                                                struct jset_entry *entry,
 448                                                unsigned version, int big_endian, int write)
 449 {
 450         struct jset_entry_blacklist_v2 *bl_entry;
 451         int ret = 0;
 452
 453         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2,
 454                                  c, jset, entry,
 455                 "invalid journal seq blacklist entry: bad size")) {
 456                 journal_entry_null_range(entry, vstruct_next(entry));
 457                 goto out;
 458         }
 459
 460         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 461
 462         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 463                                  le64_to_cpu(bl_entry->end),
 464                                  c, jset, entry,
 465                 "invalid journal seq blacklist entry: start > end")) {
 466                 journal_entry_null_range(entry, vstruct_next(entry));
 467         }
 468 out:
 469 fsck_err:
 470         return ret;
 471 }
 472
 473 static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
 474                                                struct jset_entry *entry)
 475 {
 476         struct jset_entry_blacklist_v2 *bl =
 477                 container_of(entry, struct jset_entry_blacklist_v2, entry);
 478
 479         prt_printf(out, "start=%llu end=%llu",
 480                le64_to_cpu(bl->start),
 481                le64_to_cpu(bl->end));
 482 }
 483
 484 static int journal_entry_usage_validate(struct bch_fs *c,
 485                                         struct jset *jset,
 486                                         struct jset_entry *entry,
 487                                         unsigned version, int big_endian, int write)
 488 {
 489         struct jset_entry_usage *u =
 490                 container_of(entry, struct jset_entry_usage, entry);
 491         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 492         int ret = 0;
 493
 494         if (journal_entry_err_on(bytes < sizeof(*u),
 495                                  c, jset, entry,
 496                                  "invalid journal entry usage: bad size")) {
 497                 journal_entry_null_range(entry, vstruct_next(entry));
 498                 return ret;
 499         }
 500
 501 fsck_err:
 502         return ret;
 503 }
 504
 505 static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 506                                         struct jset_entry *entry)
 507 {
 508         struct jset_entry_usage *u =
 509                 container_of(entry, struct jset_entry_usage, entry);
 510
 511         prt_printf(out, "type=%s v=%llu",
 512                bch2_fs_usage_types[u->entry.btree_id],
 513                le64_to_cpu(u->v));
 514 }
 515
 516 static int journal_entry_data_usage_validate(struct bch_fs *c,
 517                                         struct jset *jset,
 518                                         struct jset_entry *entry,
 519                                         unsigned version, int big_endian, int write)
 520 {
 521         struct jset_entry_data_usage *u =
 522                 container_of(entry, struct jset_entry_data_usage, entry);
 523         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 524         int ret = 0;
 525
 526         if (journal_entry_err_on(bytes < sizeof(*u) ||
 527                                  bytes < sizeof(*u) + u->r.nr_devs,
 528                                  c, jset, entry,
 529                                  "invalid journal entry usage: bad size")) {
 530                 journal_entry_null_range(entry, vstruct_next(entry));
 531                 return ret;
 532         }
 533
 534 fsck_err:
 535         return ret;
 536 }
 537
 538 static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
 539                                              struct jset_entry *entry)
 540 {
 541         struct jset_entry_data_usage *u =
 542                 container_of(entry, struct jset_entry_data_usage, entry);
 543
 544         bch2_replicas_entry_to_text(out, &u->r);
 545         prt_printf(out, "=%llu", le64_to_cpu(u->v));
 546 }
 547
 548 static int journal_entry_clock_validate(struct bch_fs *c,
 549                                         struct jset *jset,
 550                                         struct jset_entry *entry,
 551                                         unsigned version, int big_endian, int write)
 552 {
 553         struct jset_entry_clock *clock =
 554                 container_of(entry, struct jset_entry_clock, entry);
 555         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 556         int ret = 0;
 557
 558         if (journal_entry_err_on(bytes != sizeof(*clock),
 559                                  c, jset, entry, "bad size")) {
 560                 journal_entry_null_range(entry, vstruct_next(entry));
 561                 return ret;
 562         }
 563
 564         if (journal_entry_err_on(clock->rw > 1,
 565                                  c, jset, entry, "bad rw")) {
 566                 journal_entry_null_range(entry, vstruct_next(entry));
 567                 return ret;
 568         }
 569
 570 fsck_err:
 571         return ret;
 572 }
 573
 574 static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 575                                         struct jset_entry *entry)
 576 {
 577         struct jset_entry_clock *clock =
 578                 container_of(entry, struct jset_entry_clock, entry);
 579
 580         prt_printf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
 581 }
 582
 583 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 584                                             struct jset *jset,
 585                                             struct jset_entry *entry,
 586                                             unsigned version, int big_endian, int write)
 587 {
 588         struct jset_entry_dev_usage *u =
 589                 container_of(entry, struct jset_entry_dev_usage, entry);
 590         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 591         unsigned expected = sizeof(*u);
 592         unsigned dev;
 593         int ret = 0;
 594
 595         if (journal_entry_err_on(bytes < expected,
 596                                  c, jset, entry, "bad size (%u < %u)",
 597                                  bytes, expected)) {
 598                 journal_entry_null_range(entry, vstruct_next(entry));
 599                 return ret;
 600         }
 601
 602         dev = le32_to_cpu(u->dev);
 603
 604         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 605                                  c, jset, entry, "bad dev")) {
 606                 journal_entry_null_range(entry, vstruct_next(entry));
 607                 return ret;
 608         }
 609
 610         if (journal_entry_err_on(u->pad,
 611                                  c, jset, entry, "bad pad")) {
 612                 journal_entry_null_range(entry, vstruct_next(entry));
 613                 return ret;
 614         }
 615
 616 fsck_err:
 617         return ret;
 618 }
 619
 620 static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
 621                                             struct jset_entry *entry)
 622 {
 623         struct jset_entry_dev_usage *u =
 624                 container_of(entry, struct jset_entry_dev_usage, entry);
 625         unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 626
 627         prt_printf(out, "dev=%u", le32_to_cpu(u->dev));
 628
 629         for (i = 0; i < nr_types; i++) {
 630                 if (i < BCH_DATA_NR)
 631                         prt_printf(out, " %s", bch2_data_types[i]);
 632                 else
 633                         prt_printf(out, " (unknown data type %u)", i);
 634                 prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 635                        le64_to_cpu(u->d[i].buckets),
 636                        le64_to_cpu(u->d[i].sectors),
 637                        le64_to_cpu(u->d[i].fragmented));
 638         }
 639
 640         prt_printf(out, " buckets_ec: %llu", le64_to_cpu(u->buckets_ec));
 641 }
 642
 643 static int journal_entry_log_validate(struct bch_fs *c,
 644                                       struct jset *jset,
 645                                       struct jset_entry *entry,
 646                                       unsigned version, int big_endian, int write)
 647 {
 648         return 0;
 649 }
 650
 651 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 652                                       struct jset_entry *entry)
 653 {
 654         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 655         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 656
 657         prt_printf(out, "%.*s", bytes, l->d);
 658 }
 659
 660 static int journal_entry_overwrite_validate(struct bch_fs *c,
 661                                       struct jset *jset,
 662                                       struct jset_entry *entry,
 663                                       unsigned version, int big_endian, int write)
 664 {
 665         return journal_entry_btree_keys_validate(c, jset, entry, version, big_endian, write);
 666 }
 667
 668 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
 669                                             struct jset_entry *entry)
 670 {
 671         journal_entry_btree_keys_to_text(out, c, entry);
 672 }
 673
 674 struct jset_entry_ops {
 675         int (*validate)(struct bch_fs *, struct jset *,
 676                         struct jset_entry *, unsigned, int, int);
 677         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 678 };
 679
 680 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 681 #define x(f, nr)                                                \
 682         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 683                 .validate       = journal_entry_##f##_validate, \
 684                 .to_text        = journal_entry_##f##_to_text,  \
 685         },
 686         BCH_JSET_ENTRY_TYPES()
 687 #undef x
 688 };
 689
 690 int bch2_journal_entry_validate(struct bch_fs *c,
 691                                 struct jset *jset,
 692                                 struct jset_entry *entry,
 693                                 unsigned version, int big_endian, int write)
 694 {
 695         return entry->type < BCH_JSET_ENTRY_NR
 696                 ? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
 697                                 version, big_endian, write)
 698                 : 0;
 699 }
 700
 701 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 702                                 struct jset_entry *entry)
 703 {
 704         if (entry->type < BCH_JSET_ENTRY_NR) {
 705                 prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 706                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 707         } else {
 708                 prt_printf(out, "(unknown type %u)", entry->type);
 709         }
 710 }
 711
 712 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 713                                  int write)
 714 {
 715         struct jset_entry *entry;
 716         int ret = 0;
 717
 718         vstruct_for_each(jset, entry) {
 719                 if (journal_entry_err_on(vstruct_next(entry) >
 720                                          vstruct_last(jset), c, jset, entry,
 721                                 "journal entry extends past end of jset")) {
 722                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 723                         break;
 724                 }
 725
 726                 ret = bch2_journal_entry_validate(c, jset, entry,
 727                                         le32_to_cpu(jset->version),
 728                                         JSET_BIG_ENDIAN(jset), write);
 729                 if (ret)
 730                         break;
 731         }
 732 fsck_err:
 733         return ret;
 734 }
 735
 736 static int jset_validate(struct bch_fs *c,
 737                          struct bch_dev *ca,
 738                          struct jset *jset, u64 sector,
 739                          int write)
 740 {
 741         unsigned version;
 742         int ret = 0;
 743
 744         if (le64_to_cpu(jset->magic) != jset_magic(c))
 745                 return JOURNAL_ENTRY_NONE;
 746
 747         version = le32_to_cpu(jset->version);
 748         if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
 749                                   version < bcachefs_metadata_version_min) ||
 750                                  version >= bcachefs_metadata_version_max,
 751                                  c, jset, NULL,
 752                         "%s sector %llu seq %llu: unknown journal entry version %u",
 753                         ca ? ca->name : c->name,
 754                         sector, le64_to_cpu(jset->seq),
 755                         version)) {
 756                 /* don't try to continue: */
 757                 return -EINVAL;
 758         }
 759
 760         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
 761                                  c, jset, NULL,
 762                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 763                         ca ? ca->name : c->name,
 764                         sector, le64_to_cpu(jset->seq),
 765                         JSET_CSUM_TYPE(jset)))
 766                 ret = JOURNAL_ENTRY_BAD;
 767
 768         /* last_seq is ignored when JSET_NO_FLUSH is true */
 769         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 770                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
 771                                  c, jset, NULL,
 772                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 773                                  le64_to_cpu(jset->last_seq),
 774                                  le64_to_cpu(jset->seq))) {
 775                 jset->last_seq = jset->seq;
 776                 return JOURNAL_ENTRY_BAD;
 777         }
 778
 779         ret = jset_validate_entries(c, jset, write);
 780 fsck_err:
 781         return ret;
 782 }
 783
 784 static int jset_validate_early(struct bch_fs *c,
 785                          struct bch_dev *ca,
 786                          struct jset *jset, u64 sector,
 787                          unsigned bucket_sectors_left,
 788                          unsigned sectors_read)
 789 {
 790         size_t bytes = vstruct_bytes(jset);
 791         unsigned version;
 792         int write = READ;
 793         int ret = 0;
 794
 795         if (le64_to_cpu(jset->magic) != jset_magic(c))
 796                 return JOURNAL_ENTRY_NONE;
 797
 798         version = le32_to_cpu(jset->version);
 799         if (journal_entry_err_on((version != BCH_JSET_VERSION_OLD &&
 800                                   version < bcachefs_metadata_version_min) ||
 801                                  version >= bcachefs_metadata_version_max,
 802                                  c, jset, NULL,
 803                         "%s sector %llu seq %llu: unknown journal entry version %u",
 804                         ca ? ca->name : c->name,
 805                         sector, le64_to_cpu(jset->seq),
 806                         version)) {
 807                 /* don't try to continue: */
 808                 return -EINVAL;
 809         }
 810
 811         if (bytes > (sectors_read << 9) &&
 812             sectors_read < bucket_sectors_left)
 813                 return JOURNAL_ENTRY_REREAD;
 814
 815         if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
 816                                  c, jset, NULL,
 817                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 818                         ca ? ca->name : c->name,
 819                         sector, le64_to_cpu(jset->seq), bytes))
 820                 le32_add_cpu(&jset->u64s,
 821                              -((bytes - (bucket_sectors_left << 9)) / 8));
 822 fsck_err:
 823         return ret;
 824 }
 825
 826 struct journal_read_buf {
 827         void            *data;
 828         size_t          size;
 829 };
 830
 831 static int journal_read_buf_realloc(struct journal_read_buf *b,
 832                                     size_t new_size)
 833 {
 834         void *n;
 835
 836         /* the bios are sized for this many pages, max: */
 837         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 838                 return -ENOMEM;
 839
 840         new_size = roundup_pow_of_two(new_size);
 841         n = kvpmalloc(new_size, GFP_KERNEL);
 842         if (!n)
 843                 return -ENOMEM;
 844
 845         kvpfree(b->data, b->size);
 846         b->data = n;
 847         b->size = new_size;
 848         return 0;
 849 }
 850
 851 static int journal_read_bucket(struct bch_dev *ca,
 852                                struct journal_read_buf *buf,
 853                                struct journal_list *jlist,
 854                                unsigned bucket)
 855 {
 856         struct bch_fs *c = ca->fs;
 857         struct journal_device *ja = &ca->journal;
 858         struct jset *j = NULL;
 859         unsigned sectors, sectors_read = 0;
 860         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 861             end = offset + ca->mi.bucket_size;
 862         bool saw_bad = false, csum_good;
 863         int ret = 0;
 864
 865         pr_debug("reading %u", bucket);
 866
 867         while (offset < end) {
 868                 if (!sectors_read) {
 869                         struct bio *bio;
 870                         unsigned nr_bvecs;
 871 reread:
 872                         sectors_read = min_t(unsigned,
 873                                 end - offset, buf->size >> 9);
 874                         nr_bvecs = buf_pages(buf->data, sectors_read << 9);
 875
 876                         bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
 877                         bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
 878
 879                         bio->bi_iter.bi_sector = offset;
 880                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 881
 882                         ret = submit_bio_wait(bio);
 883                         kfree(bio);
 884
 885                         if (bch2_dev_io_err_on(ret, ca,
 886                                                "journal read error: sector %llu",
 887                                                offset) ||
 888                             bch2_meta_read_fault("journal")) {
 889                                 /*
 890                                  * We don't error out of the recovery process
 891                                  * here, since the relevant journal entry may be
 892                                  * found on a different device, and missing or
 893                                  * no journal entries will be handled later
 894                                  */
 895                                 return 0;
 896                         }
 897
 898                         j = buf->data;
 899                 }
 900
 901                 ret = jset_validate_early(c, ca, j, offset,
 902                                     end - offset, sectors_read);
 903                 switch (ret) {
 904                 case 0:
 905                         sectors = vstruct_sectors(j, c->block_bits);
 906                         break;
 907                 case JOURNAL_ENTRY_REREAD:
 908                         if (vstruct_bytes(j) > buf->size) {
 909                                 ret = journal_read_buf_realloc(buf,
 910                                                         vstruct_bytes(j));
 911                                 if (ret)
 912                                         return ret;
 913                         }
 914                         goto reread;
 915                 case JOURNAL_ENTRY_NONE:
 916                         if (!saw_bad)
 917                                 return 0;
 918                         /*
 919                          * On checksum error we don't really trust the size
 920                          * field of the journal entry we read, so try reading
 921                          * again at next block boundary:
 922                          */
 923                         sectors = block_sectors(c);
 924                         goto next_block;
 925                 default:
 926                         return ret;
 927                 }
 928
 929                 /*
 930                  * This happens sometimes if we don't have discards on -
 931                  * when we've partially overwritten a bucket with new
 932                  * journal entries. We don't need the rest of the
 933                  * bucket:
 934                  */
 935                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 936                         return 0;
 937
 938                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 939
 940                 csum_good = jset_csum_good(c, j);
 941                 if (!csum_good)
 942                         saw_bad = true;
 943
 944                 ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
 945                              j->encrypted_start,
 946                              vstruct_end(j) - (void *) j->encrypted_start);
 947                 bch2_fs_fatal_err_on(ret, c,
 948                                 "error decrypting journal entry: %i", ret);
 949
 950                 mutex_lock(&jlist->lock);
 951                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
 952                                         .csum_good      = csum_good,
 953                                         .dev            = ca->dev_idx,
 954                                         .bucket         = bucket,
 955                                         .bucket_offset  = offset -
 956                                                 bucket_to_sector(ca, ja->buckets[bucket]),
 957                                         .sector         = offset,
 958                                         }, jlist, j);
 959                 mutex_unlock(&jlist->lock);
 960
 961                 switch (ret) {
 962                 case JOURNAL_ENTRY_ADD_OK:
 963                         break;
 964                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
 965                         break;
 966                 default:
 967                         return ret;
 968                 }
 969 next_block:
 970                 pr_debug("next");
 971                 offset          += sectors;
 972                 sectors_read    -= sectors;
 973                 j = ((void *) j) + (sectors << 9);
 974         }
 975
 976         return 0;
 977 }
 978
 979 static void bch2_journal_read_device(struct closure *cl)
 980 {
 981         struct journal_device *ja =
 982                 container_of(cl, struct journal_device, read);
 983         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
 984         struct bch_fs *c = ca->fs;
 985         struct journal_list *jlist =
 986                 container_of(cl->parent, struct journal_list, cl);
 987         struct journal_replay *r, **_r;
 988         struct genradix_iter iter;
 989         struct journal_read_buf buf = { NULL, 0 };
 990         unsigned i;
 991         int ret = 0;
 992
 993         if (!ja->nr)
 994                 goto out;
 995
 996         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
 997         if (ret)
 998                 goto err;
 999
1000         pr_debug("%u journal buckets", ja->nr);
1001
1002         for (i = 0; i < ja->nr; i++) {
1003                 ret = journal_read_bucket(ca, &buf, jlist, i);
1004                 if (ret)
1005                         goto err;
1006         }
1007
1008         ja->sectors_free = ca->mi.bucket_size;
1009
1010         mutex_lock(&jlist->lock);
1011         genradix_for_each_reverse(&c->journal_entries, iter, _r) {
1012                 r = *_r;
1013
1014                 if (!r)
1015                         continue;
1016
1017                 for (i = 0; i < r->nr_ptrs; i++) {
1018                         if (r->ptrs[i].dev == ca->dev_idx) {
1019                                 unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
1020                                         vstruct_sectors(&r->j, c->block_bits);
1021
1022                                 ja->cur_idx = r->ptrs[i].bucket;
1023                                 ja->sectors_free = ca->mi.bucket_size - wrote;
1024                                 goto found;
1025                         }
1026                 }
1027         }
1028 found:
1029         mutex_unlock(&jlist->lock);
1030
1031         if (ja->bucket_seq[ja->cur_idx] &&
1032             ja->sectors_free == ca->mi.bucket_size) {
1033                 bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
1034                 bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
1035                 for (i = 0; i < 3; i++) {
1036                         unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
1037                         bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
1038                 }
1039                 ja->sectors_free = 0;
1040         }
1041
1042         /*
1043          * Set dirty_idx to indicate the entire journal is full and needs to be
1044          * reclaimed - journal reclaim will immediately reclaim whatever isn't
1045          * pinned when it first runs:
1046          */
1047         ja->discard_idx = ja->dirty_idx_ondisk =
1048                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
1049 out:
1050         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
1051         kvpfree(buf.data, buf.size);
1052         percpu_ref_put(&ca->io_ref);
1053         closure_return(cl);
1054         return;
1055 err:
1056         mutex_lock(&jlist->lock);
1057         jlist->ret = ret;
1058         mutex_unlock(&jlist->lock);
1059         goto out;
1060 }
1061
1062 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1063                                struct journal_replay *j)
1064 {
1065         unsigned i;
1066
1067         for (i = 0; i < j->nr_ptrs; i++) {
1068                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
1069                 u64 offset;
1070
1071                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
1072
1073                 if (i)
1074                         prt_printf(out, " ");
1075                 prt_printf(out, "%u:%u:%u (sector %llu)",
1076                        j->ptrs[i].dev,
1077                        j->ptrs[i].bucket,
1078                        j->ptrs[i].bucket_offset,
1079                        j->ptrs[i].sector);
1080         }
1081 }
1082
1083 int bch2_journal_read(struct bch_fs *c, u64 *blacklist_seq, u64 *start_seq)
1084 {
1085         struct journal_list jlist;
1086         struct journal_replay *i, **_i, *prev = NULL;
1087         struct genradix_iter radix_iter;
1088         struct bch_dev *ca;
1089         unsigned iter;
1090         struct printbuf buf = PRINTBUF;
1091         bool degraded = false, last_write_torn = false;
1092         u64 seq, last_seq = 0;
1093         int ret = 0;
1094
1095         closure_init_stack(&jlist.cl);
1096         mutex_init(&jlist.lock);
1097         jlist.last_seq = 0;
1098         jlist.ret = 0;
1099
1100         for_each_member_device(ca, c, iter) {
1101                 if (!c->opts.fsck &&
1102                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1103                         continue;
1104
1105                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1106                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1107                     percpu_ref_tryget(&ca->io_ref))
1108                         closure_call(&ca->journal.read,
1109                                      bch2_journal_read_device,
1110                                      system_unbound_wq,
1111                                      &jlist.cl);
1112                 else
1113                         degraded = true;
1114         }
1115
1116         closure_sync(&jlist.cl);
1117
1118         if (jlist.ret)
1119                 return jlist.ret;
1120
1121         *start_seq      = 0;
1122         *blacklist_seq  = 0;
1123
1124         /*
1125          * Find most recent flush entry, and ignore newer non flush entries -
1126          * those entries will be blacklisted:
1127          *
1128          *
1129          * XXX check for torn write on last journal entry
1130          */
1131         genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
1132                 int write = READ;
1133
1134                 i = *_i;
1135
1136                 if (!i || i->ignore)
1137                         continue;
1138
1139                 if (!*start_seq)
1140                         *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
1141
1142                 if (JSET_NO_FLUSH(&i->j)) {
1143                         journal_replay_free(c, i);
1144                         continue;
1145                 }
1146
1147                 if (!last_write_torn && !i->csum_good) {
1148                         last_write_torn = true;
1149                         journal_replay_free(c, i);
1150                         continue;
1151                 }
1152
1153                 if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
1154                                          c, &i->j, NULL,
1155                                          "invalid journal entry: last_seq > seq (%llu > %llu)",
1156                                          le64_to_cpu(i->j.last_seq),
1157                                          le64_to_cpu(i->j.seq)))
1158                         i->j.last_seq = i->j.seq;
1159
1160                 last_seq        = le64_to_cpu(i->j.last_seq);
1161                 *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1162                 break;
1163         }
1164
1165         if (!*start_seq) {
1166                 bch_info(c, "journal read done, but no entries found");
1167                 return 0;
1168         }
1169
1170         if (!last_seq) {
1171                 fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
1172                 return 0;
1173         }
1174
1175         bch_info(c, "journal read done, replaying entries %llu-%llu",
1176                  last_seq, *blacklist_seq - 1);
1177
1178         if (*start_seq != *blacklist_seq)
1179                 bch_info(c, "dropped unflushed entries %llu-%llu",
1180                          *blacklist_seq, *start_seq - 1);
1181
1182         /* Drop blacklisted entries and entries older than last_seq: */
1183         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1184                 i = *_i;
1185
1186                 if (!i || i->ignore)
1187                         continue;
1188
1189                 seq = le64_to_cpu(i->j.seq);
1190                 if (seq < last_seq) {
1191                         journal_replay_free(c, i);
1192                         continue;
1193                 }
1194
1195                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1196                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1197                                     "found blacklisted journal entry %llu", seq);
1198
1199                         journal_replay_free(c, i);
1200                 }
1201         }
1202
1203         /* Check for missing entries: */
1204         seq = last_seq;
1205         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1206                 i = *_i;
1207
1208                 if (!i || i->ignore)
1209                         continue;
1210
1211                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1212
1213                 while (seq < le64_to_cpu(i->j.seq)) {
1214                         u64 missing_start, missing_end;
1215                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1216
1217                         while (seq < le64_to_cpu(i->j.seq) &&
1218                                bch2_journal_seq_is_blacklisted(c, seq, false))
1219                                 seq++;
1220
1221                         if (seq == le64_to_cpu(i->j.seq))
1222                                 break;
1223
1224                         missing_start = seq;
1225
1226                         while (seq < le64_to_cpu(i->j.seq) &&
1227                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1228                                 seq++;
1229
1230                         if (prev) {
1231                                 bch2_journal_ptrs_to_text(&buf1, c, prev);
1232                                 prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits));
1233                         } else
1234                                 prt_printf(&buf1, "(none)");
1235                         bch2_journal_ptrs_to_text(&buf2, c, i);
1236
1237                         missing_end = seq - 1;
1238                         fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1239                                  "  prev at %s\n"
1240                                  "  next at %s",
1241                                  missing_start, missing_end,
1242                                  last_seq, *blacklist_seq - 1,
1243                                  buf1.buf, buf2.buf);
1244
1245                         printbuf_exit(&buf1);
1246                         printbuf_exit(&buf2);
1247                 }
1248
1249                 prev = i;
1250                 seq++;
1251         }
1252
1253         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1254                 struct bch_replicas_padded replicas = {
1255                         .e.data_type = BCH_DATA_journal,
1256                         .e.nr_required = 1,
1257                 };
1258                 unsigned ptr;
1259
1260                 i = *_i;
1261                 if (!i || i->ignore)
1262                         continue;
1263
1264                 for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
1265                         struct bch_dev *ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
1266
1267                         if (!i->ptrs[ptr].csum_good)
1268                                 bch_err_dev_offset(ca, i->ptrs[ptr].sector,
1269                                                    "invalid journal checksum, seq %llu%s",
1270                                                    le64_to_cpu(i->j.seq),
1271                                                    i->csum_good ? " (had good copy on another device)" : "");
1272                 }
1273
1274                 ret = jset_validate(c,
1275                                     bch_dev_bkey_exists(c, i->ptrs[0].dev),
1276                                     &i->j,
1277                                     i->ptrs[0].sector,
1278                                     READ);
1279                 if (ret)
1280                         goto err;
1281
1282                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1283                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1284
1285                 bch2_replicas_entry_sort(&replicas.e);
1286
1287                 /*
1288                  * If we're mounting in degraded mode - if we didn't read all
1289                  * the devices - this is wrong:
1290                  */
1291
1292                 printbuf_reset(&buf);
1293                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1294
1295                 if (!degraded &&
1296                     fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
1297                                 "superblock not marked as containing replicas %s",
1298                                 buf.buf)) {
1299                         ret = bch2_mark_replicas(c, &replicas.e);
1300                         if (ret)
1301                                 goto err;
1302                 }
1303         }
1304 err:
1305 fsck_err:
1306         printbuf_exit(&buf);
1307         return ret;
1308 }
1309
1310 /* journal write: */
1311
1312 static void __journal_write_alloc(struct journal *j,
1313                                   struct journal_buf *w,
1314                                   struct dev_alloc_list *devs_sorted,
1315                                   unsigned sectors,
1316                                   unsigned *replicas,
1317                                   unsigned replicas_want)
1318 {
1319         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1320         struct journal_device *ja;
1321         struct bch_dev *ca;
1322         unsigned i;
1323
1324         if (*replicas >= replicas_want)
1325                 return;
1326
1327         for (i = 0; i < devs_sorted->nr; i++) {
1328                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1329                 if (!ca)
1330                         continue;
1331
1332                 ja = &ca->journal;
1333
1334                 /*
1335                  * Check that we can use this device, and aren't already using
1336                  * it:
1337                  */
1338                 if (!ca->mi.durability ||
1339                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1340                     !ja->nr ||
1341                     bch2_bkey_has_device(bkey_i_to_s_c(&w->key),
1342                                          ca->dev_idx) ||
1343                     sectors > ja->sectors_free)
1344                         continue;
1345
1346                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1347
1348                 bch2_bkey_append_ptr(&w->key,
1349                         (struct bch_extent_ptr) {
1350                                   .offset = bucket_to_sector(ca,
1351                                         ja->buckets[ja->cur_idx]) +
1352                                         ca->mi.bucket_size -
1353                                         ja->sectors_free,
1354                                   .dev = ca->dev_idx,
1355                 });
1356
1357                 ja->sectors_free -= sectors;
1358                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1359
1360                 *replicas += ca->mi.durability;
1361
1362                 if (*replicas >= replicas_want)
1363                         break;
1364         }
1365 }
1366
1367 /**
1368  * journal_next_bucket - move on to the next journal bucket if possible
1369  */
1370 static int journal_write_alloc(struct journal *j, struct journal_buf *w,
1371                                unsigned sectors)
1372 {
1373         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1374         struct bch_devs_mask devs;
1375         struct journal_device *ja;
1376         struct bch_dev *ca;
1377         struct dev_alloc_list devs_sorted;
1378         unsigned target = c->opts.metadata_target ?:
1379                 c->opts.foreground_target;
1380         unsigned i, replicas = 0, replicas_want =
1381                 READ_ONCE(c->opts.metadata_replicas);
1382
1383         rcu_read_lock();
1384 retry:
1385         devs = target_rw_devs(c, BCH_DATA_journal, target);
1386
1387         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1388
1389         __journal_write_alloc(j, w, &devs_sorted,
1390                               sectors, &replicas, replicas_want);
1391
1392         if (replicas >= replicas_want)
1393                 goto done;
1394
1395         for (i = 0; i < devs_sorted.nr; i++) {
1396                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1397                 if (!ca)
1398                         continue;
1399
1400                 ja = &ca->journal;
1401
1402                 if (sectors > ja->sectors_free &&
1403                     sectors <= ca->mi.bucket_size &&
1404                     bch2_journal_dev_buckets_available(j, ja,
1405                                         journal_space_discarded)) {
1406                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1407                         ja->sectors_free = ca->mi.bucket_size;
1408
1409                         /*
1410                          * ja->bucket_seq[ja->cur_idx] must always have
1411                          * something sensible:
1412                          */
1413                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1414                 }
1415         }
1416
1417         __journal_write_alloc(j, w, &devs_sorted,
1418                               sectors, &replicas, replicas_want);
1419
1420         if (replicas < replicas_want && target) {
1421                 /* Retry from all devices: */
1422                 target = 0;
1423                 goto retry;
1424         }
1425 done:
1426         rcu_read_unlock();
1427
1428         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1429
1430         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1431 }
1432
1433 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1434 {
1435         /* we aren't holding j->lock: */
1436         unsigned new_size = READ_ONCE(j->buf_size_want);
1437         void *new_buf;
1438
1439         if (buf->buf_size >= new_size)
1440                 return;
1441
1442         new_buf = kvpmalloc(new_size, GFP_NOIO|__GFP_NOWARN);
1443         if (!new_buf)
1444                 return;
1445
1446         memcpy(new_buf, buf->data, buf->buf_size);
1447
1448         spin_lock(&j->lock);
1449         swap(buf->data,         new_buf);
1450         swap(buf->buf_size,     new_size);
1451         spin_unlock(&j->lock);
1452
1453         kvpfree(new_buf, new_size);
1454 }
1455
1456 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1457 {
1458         return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
1459 }
1460
1461 static void journal_write_done(struct closure *cl)
1462 {
1463         struct journal *j = container_of(cl, struct journal, io);
1464         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1465         struct journal_buf *w = journal_last_unwritten_buf(j);
1466         struct bch_replicas_padded replicas;
1467         union journal_res_state old, new;
1468         u64 v, seq;
1469         int err = 0;
1470
1471         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1472                                ? j->flush_write_time
1473                                : j->noflush_write_time, j->write_start_time);
1474
1475         if (!w->devs_written.nr) {
1476                 bch_err(c, "unable to write journal to sufficient devices");
1477                 err = -EIO;
1478         } else {
1479                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1480                                          w->devs_written);
1481                 if (bch2_mark_replicas(c, &replicas.e))
1482                         err = -EIO;
1483         }
1484
1485         if (err)
1486                 bch2_fatal_error(c);
1487
1488         spin_lock(&j->lock);
1489         seq = le64_to_cpu(w->data->seq);
1490
1491         if (seq >= j->pin.front)
1492                 journal_seq_pin(j, seq)->devs = w->devs_written;
1493
1494         if (!err) {
1495                 if (!JSET_NO_FLUSH(w->data)) {
1496                         j->flushed_seq_ondisk = seq;
1497                         j->last_seq_ondisk = w->last_seq;
1498
1499                         bch2_do_discards(c);
1500                         closure_wake_up(&c->freelist_wait);
1501                 }
1502         } else if (!j->err_seq || seq < j->err_seq)
1503                 j->err_seq      = seq;
1504
1505         j->seq_ondisk           = seq;
1506
1507         /*
1508          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1509          * more buckets:
1510          *
1511          * Must come before signaling write completion, for
1512          * bch2_fs_journal_stop():
1513          */
1514         if (j->watermark)
1515                 journal_reclaim_kick(&c->journal);
1516
1517         /* also must come before signalling write completion: */
1518         closure_debug_destroy(cl);
1519
1520         v = atomic64_read(&j->reservations.counter);
1521         do {
1522                 old.v = new.v = v;
1523                 BUG_ON(journal_state_count(new, new.unwritten_idx));
1524
1525                 new.unwritten_idx++;
1526         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1527                                        old.v, new.v)) != old.v);
1528
1529         bch2_journal_space_available(j);
1530
1531         closure_wake_up(&w->wait);
1532         journal_wake(j);
1533
1534         if (!journal_state_count(new, new.unwritten_idx) &&
1535             journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
1536                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1537         } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
1538                    new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
1539                 struct journal_buf *buf = journal_cur_buf(j);
1540                 long delta = buf->expires - jiffies;
1541
1542                 /*
1543                  * We don't close a journal entry to write it while there's
1544                  * previous entries still in flight - the current journal entry
1545                  * might want to be written now:
1546                  */
1547
1548                 mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
1549         }
1550
1551         spin_unlock(&j->lock);
1552 }
1553
1554 static void journal_write_endio(struct bio *bio)
1555 {
1556         struct bch_dev *ca = bio->bi_private;
1557         struct journal *j = &ca->fs->journal;
1558         struct journal_buf *w = journal_last_unwritten_buf(j);
1559         unsigned long flags;
1560
1561         if (bch2_dev_io_err_on(bio->bi_status, ca, "error writing journal entry %llu: %s",
1562                                le64_to_cpu(w->data->seq),
1563                                bch2_blk_status_to_str(bio->bi_status)) ||
1564             bch2_meta_write_fault("journal")) {
1565                 spin_lock_irqsave(&j->err_lock, flags);
1566                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1567                 spin_unlock_irqrestore(&j->err_lock, flags);
1568         }
1569
1570         closure_put(&j->io);
1571         percpu_ref_put(&ca->io_ref);
1572 }
1573
1574 static void do_journal_write(struct closure *cl)
1575 {
1576         struct journal *j = container_of(cl, struct journal, io);
1577         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1578         struct bch_dev *ca;
1579         struct journal_buf *w = journal_last_unwritten_buf(j);
1580         struct bch_extent_ptr *ptr;
1581         struct bio *bio;
1582         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1583
1584         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1585                 ca = bch_dev_bkey_exists(c, ptr->dev);
1586                 if (!percpu_ref_tryget(&ca->io_ref)) {
1587                         /* XXX: fix this */
1588                         bch_err(c, "missing device for journal write\n");
1589                         continue;
1590                 }
1591
1592                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1593                              sectors);
1594
1595                 bio = ca->journal.bio;
1596                 bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
1597                 bio->bi_iter.bi_sector  = ptr->offset;
1598                 bio->bi_end_io          = journal_write_endio;
1599                 bio->bi_private         = ca;
1600
1601                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1602                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1603
1604                 if (!JSET_NO_FLUSH(w->data))
1605                         bio->bi_opf    |= REQ_FUA;
1606                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1607                         bio->bi_opf    |= REQ_PREFLUSH;
1608
1609                 bch2_bio_map(bio, w->data, sectors << 9);
1610
1611                 trace_and_count(c, journal_write, bio);
1612                 closure_bio_submit(bio, cl);
1613
1614                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1615                         le64_to_cpu(w->data->seq);
1616         }
1617
1618         continue_at(cl, journal_write_done, c->io_complete_wq);
1619         return;
1620 }
1621
1622 void bch2_journal_write(struct closure *cl)
1623 {
1624         struct journal *j = container_of(cl, struct journal, io);
1625         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1626         struct bch_dev *ca;
1627         struct journal_buf *w = journal_last_unwritten_buf(j);
1628         struct jset_entry *start, *end;
1629         struct jset *jset;
1630         struct bio *bio;
1631         struct printbuf journal_debug_buf = PRINTBUF;
1632         bool validate_before_checksum = false;
1633         unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
1634         int ret;
1635
1636         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1637
1638         journal_buf_realloc(j, w);
1639         jset = w->data;
1640
1641         j->write_start_time = local_clock();
1642
1643         spin_lock(&j->lock);
1644
1645         /*
1646          * If the journal is in an error state - we did an emergency shutdown -
1647          * we prefer to continue doing journal writes. We just mark them as
1648          * noflush so they'll never be used, but they'll still be visible by the
1649          * list_journal tool - this helps in debugging.
1650          *
1651          * There's a caveat: the first journal write after marking the
1652          * superblock dirty must always be a flush write, because on startup
1653          * from a clean shutdown we didn't necessarily read the journal and the
1654          * new journal write might overwrite whatever was in the journal
1655          * previously - we can't leave the journal without any flush writes in
1656          * it.
1657          *
1658          * So if we're in an error state, and we're still starting up, we don't
1659          * write anything at all.
1660          */
1661         if (!test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags) &&
1662             (bch2_journal_error(j) ||
1663              w->noflush ||
1664              (!w->must_flush &&
1665               (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1666               test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)))) {
1667                 w->noflush = true;
1668                 SET_JSET_NO_FLUSH(jset, true);
1669                 jset->last_seq  = 0;
1670                 w->last_seq     = 0;
1671
1672                 j->nr_noflush_writes++;
1673         } else if (!bch2_journal_error(j)) {
1674                 j->last_flush_write = jiffies;
1675                 j->nr_flush_writes++;
1676                 clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
1677         } else {
1678                 spin_unlock(&j->lock);
1679                 goto err;
1680         }
1681         spin_unlock(&j->lock);
1682
1683         /*
1684          * New btree roots are set by journalling them; when the journal entry
1685          * gets written we have to propagate them to c->btree_roots
1686          *
1687          * But, every journal entry we write has to contain all the btree roots
1688          * (at least for now); so after we copy btree roots to c->btree_roots we
1689          * have to get any missing btree roots and add them to this journal
1690          * entry:
1691          */
1692
1693         bch2_journal_entries_to_btree_roots(c, jset);
1694
1695         start = end = vstruct_last(jset);
1696
1697         end     = bch2_btree_roots_to_journal_entries(c, jset->start, end);
1698
1699         bch2_journal_super_entries_add_common(c, &end,
1700                                 le64_to_cpu(jset->seq));
1701         u64s    = (u64 *) end - (u64 *) start;
1702         BUG_ON(u64s > j->entry_u64s_reserved);
1703
1704         le32_add_cpu(&jset->u64s, u64s);
1705         BUG_ON(vstruct_sectors(jset, c->block_bits) > w->sectors);
1706
1707         jset->magic             = cpu_to_le64(jset_magic(c));
1708         jset->version           = c->sb.version < bcachefs_metadata_version_bkey_renumber
1709                 ? cpu_to_le32(BCH_JSET_VERSION_OLD)
1710                 : cpu_to_le32(c->sb.version);
1711
1712         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1713         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1714
1715         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1716                 j->last_empty_seq = le64_to_cpu(jset->seq);
1717
1718         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1719                 validate_before_checksum = true;
1720
1721         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1722                 validate_before_checksum = true;
1723
1724         if (validate_before_checksum &&
1725             jset_validate(c, NULL, jset, 0, WRITE))
1726                 goto err;
1727
1728         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1729                     jset->encrypted_start,
1730                     vstruct_end(jset) - (void *) jset->encrypted_start);
1731         if (bch2_fs_fatal_err_on(ret, c,
1732                         "error decrypting journal entry: %i", ret))
1733                 goto err;
1734
1735         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1736                                   journal_nonce(jset), jset);
1737
1738         if (!validate_before_checksum &&
1739             jset_validate(c, NULL, jset, 0, WRITE))
1740                 goto err;
1741
1742         sectors = vstruct_sectors(jset, c->block_bits);
1743         BUG_ON(sectors > w->sectors);
1744
1745         bytes = vstruct_bytes(jset);
1746         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1747
1748 retry_alloc:
1749         spin_lock(&j->lock);
1750         ret = journal_write_alloc(j, w, sectors);
1751
1752         if (ret && j->can_discard) {
1753                 spin_unlock(&j->lock);
1754                 bch2_journal_do_discards(j);
1755                 goto retry_alloc;
1756         }
1757
1758         if (ret)
1759                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1760
1761         /*
1762          * write is allocated, no longer need to account for it in
1763          * bch2_journal_space_available():
1764          */
1765         w->sectors = 0;
1766
1767         /*
1768          * journal entry has been compacted and allocated, recalculate space
1769          * available:
1770          */
1771         bch2_journal_space_available(j);
1772         spin_unlock(&j->lock);
1773
1774         if (ret) {
1775                 bch_err(c, "Unable to allocate journal write:\n%s",
1776                         journal_debug_buf.buf);
1777                 printbuf_exit(&journal_debug_buf);
1778                 bch2_fatal_error(c);
1779                 continue_at(cl, journal_write_done, c->io_complete_wq);
1780                 return;
1781         }
1782
1783         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1784
1785         if (c->opts.nochanges)
1786                 goto no_io;
1787
1788         for_each_rw_member(ca, c, i)
1789                 nr_rw_members++;
1790
1791         if (nr_rw_members > 1)
1792                 w->separate_flush = true;
1793
1794         if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
1795                 for_each_rw_member(ca, c, i) {
1796                         percpu_ref_get(&ca->io_ref);
1797
1798                         bio = ca->journal.bio;
1799                         bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
1800                         bio->bi_end_io          = journal_write_endio;
1801                         bio->bi_private         = ca;
1802                         closure_bio_submit(bio, cl);
1803                 }
1804         }
1805
1806         continue_at(cl, do_journal_write, c->io_complete_wq);
1807         return;
1808 no_io:
1809         continue_at(cl, journal_write_done, c->io_complete_wq);
1810         return;
1811 err:
1812         bch2_fatal_error(c);
1813         continue_at(cl, journal_write_done, c->io_complete_wq);
1814 }