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 #include "sb-clean.h"
  18 #include "trace.h"
  19
  20 static struct nonce journal_nonce(const struct jset *jset)
  21 {
  22         return (struct nonce) {{
  23                 [0] = 0,
  24                 [1] = ((__le32 *) &jset->seq)[0],
  25                 [2] = ((__le32 *) &jset->seq)[1],
  26                 [3] = BCH_NONCE_JOURNAL,
  27         }};
  28 }
  29
  30 static bool jset_csum_good(struct bch_fs *c, struct jset *j)
  31 {
  32         return bch2_checksum_type_valid(c, JSET_CSUM_TYPE(j)) &&
  33                 !bch2_crc_cmp(j->csum,
  34                               csum_vstruct(c, JSET_CSUM_TYPE(j), journal_nonce(j), j));
  35 }
  36
  37 static inline u32 journal_entry_radix_idx(struct bch_fs *c, u64 seq)
  38 {
  39         return (seq - c->journal_entries_base_seq) & (~0U >> 1);
  40 }
  41
  42 static void __journal_replay_free(struct bch_fs *c,
  43                                   struct journal_replay *i)
  44 {
  45         struct journal_replay **p =
  46                 genradix_ptr(&c->journal_entries,
  47                              journal_entry_radix_idx(c, le64_to_cpu(i->j.seq)));
  48
  49         BUG_ON(*p != i);
  50         *p = NULL;
  51         kvpfree(i, offsetof(struct journal_replay, j) +
  52                 vstruct_bytes(&i->j));
  53 }
  54
  55 static void journal_replay_free(struct bch_fs *c, struct journal_replay *i)
  56 {
  57         i->ignore = true;
  58
  59         if (!c->opts.read_entire_journal)
  60                 __journal_replay_free(c, i);
  61 }
  62
  63 struct journal_list {
  64         struct closure          cl;
  65         u64                     last_seq;
  66         struct mutex            lock;
  67         int                     ret;
  68 };
  69
  70 #define JOURNAL_ENTRY_ADD_OK            0
  71 #define JOURNAL_ENTRY_ADD_OUT_OF_RANGE  5
  72
  73 /*
  74  * Given a journal entry we just read, add it to the list of journal entries to
  75  * be replayed:
  76  */
  77 static int journal_entry_add(struct bch_fs *c, struct bch_dev *ca,
  78                              struct journal_ptr entry_ptr,
  79                              struct journal_list *jlist, struct jset *j)
  80 {
  81         struct genradix_iter iter;
  82         struct journal_replay **_i, *i, *dup;
  83         struct journal_ptr *ptr;
  84         size_t bytes = vstruct_bytes(j);
  85         u64 last_seq = !JSET_NO_FLUSH(j) ? le64_to_cpu(j->last_seq) : 0;
  86         int ret = JOURNAL_ENTRY_ADD_OK;
  87
  88         /* Is this entry older than the range we need? */
  89         if (!c->opts.read_entire_journal &&
  90             le64_to_cpu(j->seq) < jlist->last_seq)
  91                 return JOURNAL_ENTRY_ADD_OUT_OF_RANGE;
  92
  93         /*
  94          * genradixes are indexed by a ulong, not a u64, so we can't index them
  95          * by sequence number directly: Assume instead that they will all fall
  96          * within the range of +-2billion of the filrst one we find.
  97          */
  98         if (!c->journal_entries_base_seq)
  99                 c->journal_entries_base_seq = max_t(s64, 1, le64_to_cpu(j->seq) - S32_MAX);
 100
 101         /* Drop entries we don't need anymore */
 102         if (last_seq > jlist->last_seq && !c->opts.read_entire_journal) {
 103                 genradix_for_each_from(&c->journal_entries, iter, _i,
 104                                        journal_entry_radix_idx(c, jlist->last_seq)) {
 105                         i = *_i;
 106
 107                         if (!i || i->ignore)
 108                                 continue;
 109
 110                         if (le64_to_cpu(i->j.seq) >= last_seq)
 111                                 break;
 112                         journal_replay_free(c, i);
 113                 }
 114         }
 115
 116         jlist->last_seq = max(jlist->last_seq, last_seq);
 117
 118         _i = genradix_ptr_alloc(&c->journal_entries,
 119                                 journal_entry_radix_idx(c, le64_to_cpu(j->seq)),
 120                                 GFP_KERNEL);
 121         if (!_i)
 122                 return -BCH_ERR_ENOMEM_journal_entry_add;
 123
 124         /*
 125          * Duplicate journal entries? If so we want the one that didn't have a
 126          * checksum error:
 127          */
 128         dup = *_i;
 129         if (dup) {
 130                 if (bytes == vstruct_bytes(&dup->j) &&
 131                     !memcmp(j, &dup->j, bytes)) {
 132                         i = dup;
 133                         goto found;
 134                 }
 135
 136                 if (!entry_ptr.csum_good) {
 137                         i = dup;
 138                         goto found;
 139                 }
 140
 141                 if (!dup->csum_good)
 142                         goto replace;
 143
 144                 fsck_err(c, "found duplicate but non identical journal entries (seq %llu)",
 145                          le64_to_cpu(j->seq));
 146                 i = dup;
 147                 goto found;
 148         }
 149 replace:
 150         i = kvpmalloc(offsetof(struct journal_replay, j) + bytes, GFP_KERNEL);
 151         if (!i)
 152                 return -BCH_ERR_ENOMEM_journal_entry_add;
 153
 154         i->nr_ptrs      = 0;
 155         i->csum_good    = entry_ptr.csum_good;
 156         i->ignore       = false;
 157         unsafe_memcpy(&i->j, j, bytes, "embedded variable length struct");
 158         i->ptrs[i->nr_ptrs++] = entry_ptr;
 159
 160         if (dup) {
 161                 if (dup->nr_ptrs >= ARRAY_SIZE(dup->ptrs)) {
 162                         bch_err(c, "found too many copies of journal entry %llu",
 163                                 le64_to_cpu(i->j.seq));
 164                         dup->nr_ptrs = ARRAY_SIZE(dup->ptrs) - 1;
 165                 }
 166
 167                 /* The first ptr should represent the jset we kept: */
 168                 memcpy(i->ptrs + i->nr_ptrs,
 169                        dup->ptrs,
 170                        sizeof(dup->ptrs[0]) * dup->nr_ptrs);
 171                 i->nr_ptrs += dup->nr_ptrs;
 172                 __journal_replay_free(c, dup);
 173         }
 174
 175         *_i = i;
 176         return 0;
 177 found:
 178         for (ptr = i->ptrs; ptr < i->ptrs + i->nr_ptrs; ptr++) {
 179                 if (ptr->dev == ca->dev_idx) {
 180                         bch_err(c, "duplicate journal entry %llu on same device",
 181                                 le64_to_cpu(i->j.seq));
 182                         goto out;
 183                 }
 184         }
 185
 186         if (i->nr_ptrs >= ARRAY_SIZE(i->ptrs)) {
 187                 bch_err(c, "found too many copies of journal entry %llu",
 188                         le64_to_cpu(i->j.seq));
 189                 goto out;
 190         }
 191
 192         i->ptrs[i->nr_ptrs++] = entry_ptr;
 193 out:
 194 fsck_err:
 195         return ret;
 196 }
 197
 198 /* this fills in a range with empty jset_entries: */
 199 static void journal_entry_null_range(void *start, void *end)
 200 {
 201         struct jset_entry *entry;
 202
 203         for (entry = start; entry != end; entry = vstruct_next(entry))
 204                 memset(entry, 0, sizeof(*entry));
 205 }
 206
 207 #define JOURNAL_ENTRY_REREAD    5
 208 #define JOURNAL_ENTRY_NONE      6
 209 #define JOURNAL_ENTRY_BAD       7
 210
 211 static void journal_entry_err_msg(struct printbuf *out,
 212                                   u32 version,
 213                                   struct jset *jset,
 214                                   struct jset_entry *entry)
 215 {
 216         prt_str(out, "invalid journal entry, version=");
 217         bch2_version_to_text(out, version);
 218
 219         if (entry) {
 220                 prt_str(out, " type=");
 221                 prt_str(out, bch2_jset_entry_types[entry->type]);
 222         }
 223
 224         if (!jset) {
 225                 prt_printf(out, " in superblock");
 226         } else {
 227
 228                 prt_printf(out, " seq=%llu", le64_to_cpu(jset->seq));
 229
 230                 if (entry)
 231                         prt_printf(out, " offset=%zi/%u",
 232                                    (u64 *) entry - jset->_data,
 233                                    le32_to_cpu(jset->u64s));
 234         }
 235
 236         prt_str(out, ": ");
 237 }
 238
 239 #define journal_entry_err(c, version, jset, entry, msg, ...)            \
 240 ({                                                                      \
 241         struct printbuf buf = PRINTBUF;                                 \
 242                                                                         \
 243         journal_entry_err_msg(&buf, version, jset, entry);              \
 244         prt_printf(&buf, msg, ##__VA_ARGS__);                           \
 245                                                                         \
 246         switch (flags & BKEY_INVALID_WRITE) {                           \
 247         case READ:                                                      \
 248                 mustfix_fsck_err(c, "%s", buf.buf);                     \
 249                 break;                                                  \
 250         case WRITE:                                                     \
 251                 bch_err(c, "corrupt metadata before write: %s\n", buf.buf);\
 252                 if (bch2_fs_inconsistent(c)) {                          \
 253                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 254                         goto fsck_err;                                  \
 255                 }                                                       \
 256                 break;                                                  \
 257         }                                                               \
 258                                                                         \
 259         printbuf_exit(&buf);                                            \
 260         true;                                                           \
 261 })
 262
 263 #define journal_entry_err_on(cond, c, version, jset, entry, msg, ...)   \
 264         ((cond) ? journal_entry_err(c, version, jset, entry, msg, ##__VA_ARGS__) : false)
 265
 266 #define FSCK_DELETED_KEY        5
 267
 268 static int journal_validate_key(struct bch_fs *c,
 269                                 struct jset *jset,
 270                                 struct jset_entry *entry,
 271                                 unsigned level, enum btree_id btree_id,
 272                                 struct bkey_i *k,
 273                                 unsigned version, int big_endian,
 274                                 enum bkey_invalid_flags flags)
 275 {
 276         int write = flags & BKEY_INVALID_WRITE;
 277         void *next = vstruct_next(entry);
 278         struct printbuf buf = PRINTBUF;
 279         int ret = 0;
 280
 281         if (journal_entry_err_on(!k->k.u64s, c, version, jset, entry, "k->u64s 0")) {
 282                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 283                 journal_entry_null_range(vstruct_next(entry), next);
 284                 return FSCK_DELETED_KEY;
 285         }
 286
 287         if (journal_entry_err_on((void *) bkey_next(k) >
 288                                  (void *) vstruct_next(entry),
 289                                  c, version, jset, entry,
 290                                  "extends past end of journal entry")) {
 291                 entry->u64s = cpu_to_le16((u64 *) k - entry->_data);
 292                 journal_entry_null_range(vstruct_next(entry), next);
 293                 return FSCK_DELETED_KEY;
 294         }
 295
 296         if (journal_entry_err_on(k->k.format != KEY_FORMAT_CURRENT,
 297                                  c, version, jset, entry,
 298                                  "bad format %u", k->k.format)) {
 299                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 300                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 301                 journal_entry_null_range(vstruct_next(entry), next);
 302                 return FSCK_DELETED_KEY;
 303         }
 304
 305         if (!write)
 306                 bch2_bkey_compat(level, btree_id, version, big_endian,
 307                                  write, NULL, bkey_to_packed(k));
 308
 309         if (bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 310                               __btree_node_type(level, btree_id), write, &buf)) {
 311                 printbuf_reset(&buf);
 312                 journal_entry_err_msg(&buf, version, jset, entry);
 313                 prt_newline(&buf);
 314                 printbuf_indent_add(&buf, 2);
 315
 316                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(k));
 317                 prt_newline(&buf);
 318                 bch2_bkey_invalid(c, bkey_i_to_s_c(k),
 319                                   __btree_node_type(level, btree_id), write, &buf);
 320
 321                 mustfix_fsck_err(c, "%s", buf.buf);
 322                 BUG();
 323
 324                 le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 325                 memmove(k, bkey_next(k), next - (void *) bkey_next(k));
 326                 journal_entry_null_range(vstruct_next(entry), next);
 327
 328                 printbuf_exit(&buf);
 329                 return FSCK_DELETED_KEY;
 330         }
 331
 332         if (write)
 333                 bch2_bkey_compat(level, btree_id, version, big_endian,
 334                                  write, NULL, bkey_to_packed(k));
 335 fsck_err:
 336         printbuf_exit(&buf);
 337         return ret;
 338 }
 339
 340 static int journal_entry_btree_keys_validate(struct bch_fs *c,
 341                                 struct jset *jset,
 342                                 struct jset_entry *entry,
 343                                 unsigned version, int big_endian,
 344                                 enum bkey_invalid_flags flags)
 345 {
 346         struct bkey_i *k = entry->start;
 347
 348         while (k != vstruct_last(entry)) {
 349                 int ret = journal_validate_key(c, jset, entry,
 350                                                entry->level,
 351                                                entry->btree_id,
 352                                                k, version, big_endian,
 353                                                flags|BKEY_INVALID_JOURNAL);
 354                 if (ret == FSCK_DELETED_KEY)
 355                         continue;
 356
 357                 k = bkey_next(k);
 358         }
 359
 360         return 0;
 361 }
 362
 363 static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs *c,
 364                                              struct jset_entry *entry)
 365 {
 366         struct bkey_i *k;
 367         bool first = true;
 368
 369         jset_entry_for_each_key(entry, k) {
 370                 if (!first) {
 371                         prt_newline(out);
 372                         prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 373                 }
 374                 prt_printf(out, "btree=%s l=%u ", bch2_btree_ids[entry->btree_id], entry->level);
 375                 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 376                 first = false;
 377         }
 378 }
 379
 380 static int journal_entry_btree_root_validate(struct bch_fs *c,
 381                                 struct jset *jset,
 382                                 struct jset_entry *entry,
 383                                 unsigned version, int big_endian,
 384                                 enum bkey_invalid_flags flags)
 385 {
 386         struct bkey_i *k = entry->start;
 387         int ret = 0;
 388
 389         if (journal_entry_err_on(!entry->u64s ||
 390                                  le16_to_cpu(entry->u64s) != k->k.u64s,
 391                                  c, version, jset, entry,
 392                                  "invalid btree root journal entry: wrong number of keys")) {
 393                 void *next = vstruct_next(entry);
 394                 /*
 395                  * we don't want to null out this jset_entry,
 396                  * just the contents, so that later we can tell
 397                  * we were _supposed_ to have a btree root
 398                  */
 399                 entry->u64s = 0;
 400                 journal_entry_null_range(vstruct_next(entry), next);
 401                 return 0;
 402         }
 403
 404         return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
 405                                     version, big_endian, flags);
 406 fsck_err:
 407         return ret;
 408 }
 409
 410 static void journal_entry_btree_root_to_text(struct printbuf *out, struct bch_fs *c,
 411                                              struct jset_entry *entry)
 412 {
 413         journal_entry_btree_keys_to_text(out, c, entry);
 414 }
 415
 416 static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 417                                 struct jset *jset,
 418                                 struct jset_entry *entry,
 419                                 unsigned version, int big_endian,
 420                                 enum bkey_invalid_flags flags)
 421 {
 422         /* obsolete, don't care: */
 423         return 0;
 424 }
 425
 426 static void journal_entry_prio_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 427                                             struct jset_entry *entry)
 428 {
 429 }
 430
 431 static int journal_entry_blacklist_validate(struct bch_fs *c,
 432                                 struct jset *jset,
 433                                 struct jset_entry *entry,
 434                                 unsigned version, int big_endian,
 435                                 enum bkey_invalid_flags flags)
 436 {
 437         int ret = 0;
 438
 439         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 1,
 440                                  c, version, jset, entry,
 441                 "invalid journal seq blacklist entry: bad size")) {
 442                 journal_entry_null_range(entry, vstruct_next(entry));
 443         }
 444 fsck_err:
 445         return ret;
 446 }
 447
 448 static void journal_entry_blacklist_to_text(struct printbuf *out, struct bch_fs *c,
 449                                             struct jset_entry *entry)
 450 {
 451         struct jset_entry_blacklist *bl =
 452                 container_of(entry, struct jset_entry_blacklist, entry);
 453
 454         prt_printf(out, "seq=%llu", le64_to_cpu(bl->seq));
 455 }
 456
 457 static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 458                                 struct jset *jset,
 459                                 struct jset_entry *entry,
 460                                 unsigned version, int big_endian,
 461                                 enum bkey_invalid_flags flags)
 462 {
 463         struct jset_entry_blacklist_v2 *bl_entry;
 464         int ret = 0;
 465
 466         if (journal_entry_err_on(le16_to_cpu(entry->u64s) != 2,
 467                                  c, version, jset, entry,
 468                 "invalid journal seq blacklist entry: bad size")) {
 469                 journal_entry_null_range(entry, vstruct_next(entry));
 470                 goto out;
 471         }
 472
 473         bl_entry = container_of(entry, struct jset_entry_blacklist_v2, entry);
 474
 475         if (journal_entry_err_on(le64_to_cpu(bl_entry->start) >
 476                                  le64_to_cpu(bl_entry->end),
 477                                  c, version, jset, entry,
 478                 "invalid journal seq blacklist entry: start > end")) {
 479                 journal_entry_null_range(entry, vstruct_next(entry));
 480         }
 481 out:
 482 fsck_err:
 483         return ret;
 484 }
 485
 486 static void journal_entry_blacklist_v2_to_text(struct printbuf *out, struct bch_fs *c,
 487                                                struct jset_entry *entry)
 488 {
 489         struct jset_entry_blacklist_v2 *bl =
 490                 container_of(entry, struct jset_entry_blacklist_v2, entry);
 491
 492         prt_printf(out, "start=%llu end=%llu",
 493                le64_to_cpu(bl->start),
 494                le64_to_cpu(bl->end));
 495 }
 496
 497 static int journal_entry_usage_validate(struct bch_fs *c,
 498                                 struct jset *jset,
 499                                 struct jset_entry *entry,
 500                                 unsigned version, int big_endian,
 501                                 enum bkey_invalid_flags flags)
 502 {
 503         struct jset_entry_usage *u =
 504                 container_of(entry, struct jset_entry_usage, entry);
 505         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 506         int ret = 0;
 507
 508         if (journal_entry_err_on(bytes < sizeof(*u),
 509                                  c, version, jset, entry,
 510                                  "invalid journal entry usage: bad size")) {
 511                 journal_entry_null_range(entry, vstruct_next(entry));
 512                 return ret;
 513         }
 514
 515 fsck_err:
 516         return ret;
 517 }
 518
 519 static void journal_entry_usage_to_text(struct printbuf *out, struct bch_fs *c,
 520                                         struct jset_entry *entry)
 521 {
 522         struct jset_entry_usage *u =
 523                 container_of(entry, struct jset_entry_usage, entry);
 524
 525         prt_printf(out, "type=%s v=%llu",
 526                bch2_fs_usage_types[u->entry.btree_id],
 527                le64_to_cpu(u->v));
 528 }
 529
 530 static int journal_entry_data_usage_validate(struct bch_fs *c,
 531                                 struct jset *jset,
 532                                 struct jset_entry *entry,
 533                                 unsigned version, int big_endian,
 534                                 enum bkey_invalid_flags flags)
 535 {
 536         struct jset_entry_data_usage *u =
 537                 container_of(entry, struct jset_entry_data_usage, entry);
 538         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 539         int ret = 0;
 540
 541         if (journal_entry_err_on(bytes < sizeof(*u) ||
 542                                  bytes < sizeof(*u) + u->r.nr_devs,
 543                                  c, version, jset, entry,
 544                                  "invalid journal entry usage: bad size")) {
 545                 journal_entry_null_range(entry, vstruct_next(entry));
 546                 return ret;
 547         }
 548
 549 fsck_err:
 550         return ret;
 551 }
 552
 553 static void journal_entry_data_usage_to_text(struct printbuf *out, struct bch_fs *c,
 554                                              struct jset_entry *entry)
 555 {
 556         struct jset_entry_data_usage *u =
 557                 container_of(entry, struct jset_entry_data_usage, entry);
 558
 559         bch2_replicas_entry_to_text(out, &u->r);
 560         prt_printf(out, "=%llu", le64_to_cpu(u->v));
 561 }
 562
 563 static int journal_entry_clock_validate(struct bch_fs *c,
 564                                 struct jset *jset,
 565                                 struct jset_entry *entry,
 566                                 unsigned version, int big_endian,
 567                                 enum bkey_invalid_flags flags)
 568 {
 569         struct jset_entry_clock *clock =
 570                 container_of(entry, struct jset_entry_clock, entry);
 571         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 572         int ret = 0;
 573
 574         if (journal_entry_err_on(bytes != sizeof(*clock),
 575                                  c, version, jset, entry, "bad size")) {
 576                 journal_entry_null_range(entry, vstruct_next(entry));
 577                 return ret;
 578         }
 579
 580         if (journal_entry_err_on(clock->rw > 1,
 581                                  c, version, jset, entry, "bad rw")) {
 582                 journal_entry_null_range(entry, vstruct_next(entry));
 583                 return ret;
 584         }
 585
 586 fsck_err:
 587         return ret;
 588 }
 589
 590 static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 591                                         struct jset_entry *entry)
 592 {
 593         struct jset_entry_clock *clock =
 594                 container_of(entry, struct jset_entry_clock, entry);
 595
 596         prt_printf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
 597 }
 598
 599 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 600                                 struct jset *jset,
 601                                 struct jset_entry *entry,
 602                                 unsigned version, int big_endian,
 603                                 enum bkey_invalid_flags flags)
 604 {
 605         struct jset_entry_dev_usage *u =
 606                 container_of(entry, struct jset_entry_dev_usage, entry);
 607         unsigned bytes = jset_u64s(le16_to_cpu(entry->u64s)) * sizeof(u64);
 608         unsigned expected = sizeof(*u);
 609         unsigned dev;
 610         int ret = 0;
 611
 612         if (journal_entry_err_on(bytes < expected,
 613                                  c, version, jset, entry, "bad size (%u < %u)",
 614                                  bytes, expected)) {
 615                 journal_entry_null_range(entry, vstruct_next(entry));
 616                 return ret;
 617         }
 618
 619         dev = le32_to_cpu(u->dev);
 620
 621         if (journal_entry_err_on(!bch2_dev_exists2(c, dev),
 622                                  c, version, jset, entry, "bad dev")) {
 623                 journal_entry_null_range(entry, vstruct_next(entry));
 624                 return ret;
 625         }
 626
 627         if (journal_entry_err_on(u->pad,
 628                                  c, version, jset, entry, "bad pad")) {
 629                 journal_entry_null_range(entry, vstruct_next(entry));
 630                 return ret;
 631         }
 632
 633 fsck_err:
 634         return ret;
 635 }
 636
 637 static void journal_entry_dev_usage_to_text(struct printbuf *out, struct bch_fs *c,
 638                                             struct jset_entry *entry)
 639 {
 640         struct jset_entry_dev_usage *u =
 641                 container_of(entry, struct jset_entry_dev_usage, entry);
 642         unsigned i, nr_types = jset_entry_dev_usage_nr_types(u);
 643
 644         prt_printf(out, "dev=%u", le32_to_cpu(u->dev));
 645
 646         for (i = 0; i < nr_types; i++) {
 647                 if (i < BCH_DATA_NR)
 648                         prt_printf(out, " %s", bch2_data_types[i]);
 649                 else
 650                         prt_printf(out, " (unknown data type %u)", i);
 651                 prt_printf(out, ": buckets=%llu sectors=%llu fragmented=%llu",
 652                        le64_to_cpu(u->d[i].buckets),
 653                        le64_to_cpu(u->d[i].sectors),
 654                        le64_to_cpu(u->d[i].fragmented));
 655         }
 656
 657         prt_printf(out, " buckets_ec: %llu", le64_to_cpu(u->buckets_ec));
 658 }
 659
 660 static int journal_entry_log_validate(struct bch_fs *c,
 661                                 struct jset *jset,
 662                                 struct jset_entry *entry,
 663                                 unsigned version, int big_endian,
 664                                 enum bkey_invalid_flags flags)
 665 {
 666         return 0;
 667 }
 668
 669 static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 670                                       struct jset_entry *entry)
 671 {
 672         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 673         unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 674
 675         prt_printf(out, "%.*s", bytes, l->d);
 676 }
 677
 678 static int journal_entry_overwrite_validate(struct bch_fs *c,
 679                                 struct jset *jset,
 680                                 struct jset_entry *entry,
 681                                 unsigned version, int big_endian,
 682                                 enum bkey_invalid_flags flags)
 683 {
 684         return journal_entry_btree_keys_validate(c, jset, entry,
 685                                 version, big_endian, READ);
 686 }
 687
 688 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
 689                                             struct jset_entry *entry)
 690 {
 691         journal_entry_btree_keys_to_text(out, c, entry);
 692 }
 693
 694 struct jset_entry_ops {
 695         int (*validate)(struct bch_fs *, struct jset *,
 696                         struct jset_entry *, unsigned, int,
 697                         enum bkey_invalid_flags);
 698         void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 699 };
 700
 701 static const struct jset_entry_ops bch2_jset_entry_ops[] = {
 702 #define x(f, nr)                                                \
 703         [BCH_JSET_ENTRY_##f]    = (struct jset_entry_ops) {     \
 704                 .validate       = journal_entry_##f##_validate, \
 705                 .to_text        = journal_entry_##f##_to_text,  \
 706         },
 707         BCH_JSET_ENTRY_TYPES()
 708 #undef x
 709 };
 710
 711 int bch2_journal_entry_validate(struct bch_fs *c,
 712                                 struct jset *jset,
 713                                 struct jset_entry *entry,
 714                                 unsigned version, int big_endian,
 715                                 enum bkey_invalid_flags flags)
 716 {
 717         return entry->type < BCH_JSET_ENTRY_NR
 718                 ? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
 719                                 version, big_endian, flags)
 720                 : 0;
 721 }
 722
 723 void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 724                                 struct jset_entry *entry)
 725 {
 726         if (entry->type < BCH_JSET_ENTRY_NR) {
 727                 prt_printf(out, "%s: ", bch2_jset_entry_types[entry->type]);
 728                 bch2_jset_entry_ops[entry->type].to_text(out, c, entry);
 729         } else {
 730                 prt_printf(out, "(unknown type %u)", entry->type);
 731         }
 732 }
 733
 734 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 735                                  enum bkey_invalid_flags flags)
 736 {
 737         struct jset_entry *entry;
 738         unsigned version = le32_to_cpu(jset->version);
 739         int ret = 0;
 740
 741         vstruct_for_each(jset, entry) {
 742                 if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
 743                                          c, version, jset, entry,
 744                                 "journal entry extends past end of jset")) {
 745                         jset->u64s = cpu_to_le32((u64 *) entry - jset->_data);
 746                         break;
 747                 }
 748
 749                 ret = bch2_journal_entry_validate(c, jset, entry,
 750                                         version, JSET_BIG_ENDIAN(jset), flags);
 751                 if (ret)
 752                         break;
 753         }
 754 fsck_err:
 755         return ret;
 756 }
 757
 758 static int jset_validate(struct bch_fs *c,
 759                          struct bch_dev *ca,
 760                          struct jset *jset, u64 sector,
 761                          enum bkey_invalid_flags flags)
 762 {
 763         unsigned version;
 764         int ret = 0;
 765
 766         if (le64_to_cpu(jset->magic) != jset_magic(c))
 767                 return JOURNAL_ENTRY_NONE;
 768
 769         version = le32_to_cpu(jset->version);
 770         if (journal_entry_err_on(!bch2_version_compatible(version),
 771                         c, version, jset, NULL,
 772                         "%s sector %llu seq %llu: incompatible journal entry version %u.%u",
 773                         ca ? ca->name : c->name,
 774                         sector, le64_to_cpu(jset->seq),
 775                         BCH_VERSION_MAJOR(version),
 776                         BCH_VERSION_MINOR(version))) {
 777                 /* don't try to continue: */
 778                 return -EINVAL;
 779         }
 780
 781         if (journal_entry_err_on(!bch2_checksum_type_valid(c, JSET_CSUM_TYPE(jset)),
 782                                  c, version, jset, NULL,
 783                         "%s sector %llu seq %llu: journal entry with unknown csum type %llu",
 784                         ca ? ca->name : c->name,
 785                         sector, le64_to_cpu(jset->seq),
 786                         JSET_CSUM_TYPE(jset)))
 787                 ret = JOURNAL_ENTRY_BAD;
 788
 789         /* last_seq is ignored when JSET_NO_FLUSH is true */
 790         if (journal_entry_err_on(!JSET_NO_FLUSH(jset) &&
 791                                  le64_to_cpu(jset->last_seq) > le64_to_cpu(jset->seq),
 792                                  c, version, jset, NULL,
 793                                  "invalid journal entry: last_seq > seq (%llu > %llu)",
 794                                  le64_to_cpu(jset->last_seq),
 795                                  le64_to_cpu(jset->seq))) {
 796                 jset->last_seq = jset->seq;
 797                 return JOURNAL_ENTRY_BAD;
 798         }
 799
 800         ret = jset_validate_entries(c, jset, flags);
 801 fsck_err:
 802         return ret;
 803 }
 804
 805 static int jset_validate_early(struct bch_fs *c,
 806                          struct bch_dev *ca,
 807                          struct jset *jset, u64 sector,
 808                          unsigned bucket_sectors_left,
 809                          unsigned sectors_read)
 810 {
 811         size_t bytes = vstruct_bytes(jset);
 812         unsigned version;
 813         enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
 814         int ret = 0;
 815
 816         if (le64_to_cpu(jset->magic) != jset_magic(c))
 817                 return JOURNAL_ENTRY_NONE;
 818
 819         version = le32_to_cpu(jset->version);
 820         if (journal_entry_err_on(!bch2_version_compatible(version),
 821                          c, version, jset, NULL,
 822                         "%s sector %llu seq %llu: unknown journal entry version %u.%u",
 823                         ca ? ca->name : c->name,
 824                         sector, le64_to_cpu(jset->seq),
 825                         BCH_VERSION_MAJOR(version),
 826                         BCH_VERSION_MINOR(version))) {
 827                 /* don't try to continue: */
 828                 return -EINVAL;
 829         }
 830
 831         if (bytes > (sectors_read << 9) &&
 832             sectors_read < bucket_sectors_left)
 833                 return JOURNAL_ENTRY_REREAD;
 834
 835         if (journal_entry_err_on(bytes > bucket_sectors_left << 9,
 836                          c, version, jset, NULL,
 837                         "%s sector %llu seq %llu: journal entry too big (%zu bytes)",
 838                         ca ? ca->name : c->name,
 839                         sector, le64_to_cpu(jset->seq), bytes))
 840                 le32_add_cpu(&jset->u64s,
 841                              -((bytes - (bucket_sectors_left << 9)) / 8));
 842 fsck_err:
 843         return ret;
 844 }
 845
 846 struct journal_read_buf {
 847         void            *data;
 848         size_t          size;
 849 };
 850
 851 static int journal_read_buf_realloc(struct journal_read_buf *b,
 852                                     size_t new_size)
 853 {
 854         void *n;
 855
 856         /* the bios are sized for this many pages, max: */
 857         if (new_size > JOURNAL_ENTRY_SIZE_MAX)
 858                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 859
 860         new_size = roundup_pow_of_two(new_size);
 861         n = kvpmalloc(new_size, GFP_KERNEL);
 862         if (!n)
 863                 return -BCH_ERR_ENOMEM_journal_read_buf_realloc;
 864
 865         kvpfree(b->data, b->size);
 866         b->data = n;
 867         b->size = new_size;
 868         return 0;
 869 }
 870
 871 static int journal_read_bucket(struct bch_dev *ca,
 872                                struct journal_read_buf *buf,
 873                                struct journal_list *jlist,
 874                                unsigned bucket)
 875 {
 876         struct bch_fs *c = ca->fs;
 877         struct journal_device *ja = &ca->journal;
 878         struct jset *j = NULL;
 879         unsigned sectors, sectors_read = 0;
 880         u64 offset = bucket_to_sector(ca, ja->buckets[bucket]),
 881             end = offset + ca->mi.bucket_size;
 882         bool saw_bad = false, csum_good;
 883         int ret = 0;
 884
 885         pr_debug("reading %u", bucket);
 886
 887         while (offset < end) {
 888                 if (!sectors_read) {
 889                         struct bio *bio;
 890                         unsigned nr_bvecs;
 891 reread:
 892                         sectors_read = min_t(unsigned,
 893                                 end - offset, buf->size >> 9);
 894                         nr_bvecs = buf_pages(buf->data, sectors_read << 9);
 895
 896                         bio = bio_kmalloc(nr_bvecs, GFP_KERNEL);
 897                         bio_init(bio, ca->disk_sb.bdev, bio->bi_inline_vecs, nr_bvecs, REQ_OP_READ);
 898
 899                         bio->bi_iter.bi_sector = offset;
 900                         bch2_bio_map(bio, buf->data, sectors_read << 9);
 901
 902                         ret = submit_bio_wait(bio);
 903                         kfree(bio);
 904
 905                         if (bch2_dev_io_err_on(ret, ca,
 906                                                "journal read error: sector %llu",
 907                                                offset) ||
 908                             bch2_meta_read_fault("journal")) {
 909                                 /*
 910                                  * We don't error out of the recovery process
 911                                  * here, since the relevant journal entry may be
 912                                  * found on a different device, and missing or
 913                                  * no journal entries will be handled later
 914                                  */
 915                                 return 0;
 916                         }
 917
 918                         j = buf->data;
 919                 }
 920
 921                 ret = jset_validate_early(c, ca, j, offset,
 922                                     end - offset, sectors_read);
 923                 switch (ret) {
 924                 case 0:
 925                         sectors = vstruct_sectors(j, c->block_bits);
 926                         break;
 927                 case JOURNAL_ENTRY_REREAD:
 928                         if (vstruct_bytes(j) > buf->size) {
 929                                 ret = journal_read_buf_realloc(buf,
 930                                                         vstruct_bytes(j));
 931                                 if (ret)
 932                                         return ret;
 933                         }
 934                         goto reread;
 935                 case JOURNAL_ENTRY_NONE:
 936                         if (!saw_bad)
 937                                 return 0;
 938                         /*
 939                          * On checksum error we don't really trust the size
 940                          * field of the journal entry we read, so try reading
 941                          * again at next block boundary:
 942                          */
 943                         sectors = block_sectors(c);
 944                         goto next_block;
 945                 default:
 946                         return ret;
 947                 }
 948
 949                 /*
 950                  * This happens sometimes if we don't have discards on -
 951                  * when we've partially overwritten a bucket with new
 952                  * journal entries. We don't need the rest of the
 953                  * bucket:
 954                  */
 955                 if (le64_to_cpu(j->seq) < ja->bucket_seq[bucket])
 956                         return 0;
 957
 958                 ja->bucket_seq[bucket] = le64_to_cpu(j->seq);
 959
 960                 csum_good = jset_csum_good(c, j);
 961                 if (!csum_good)
 962                         saw_bad = true;
 963
 964                 ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
 965                              j->encrypted_start,
 966                              vstruct_end(j) - (void *) j->encrypted_start);
 967                 bch2_fs_fatal_err_on(ret, c,
 968                                 "error decrypting journal entry: %i", ret);
 969
 970                 mutex_lock(&jlist->lock);
 971                 ret = journal_entry_add(c, ca, (struct journal_ptr) {
 972                                         .csum_good      = csum_good,
 973                                         .dev            = ca->dev_idx,
 974                                         .bucket         = bucket,
 975                                         .bucket_offset  = offset -
 976                                                 bucket_to_sector(ca, ja->buckets[bucket]),
 977                                         .sector         = offset,
 978                                         }, jlist, j);
 979                 mutex_unlock(&jlist->lock);
 980
 981                 switch (ret) {
 982                 case JOURNAL_ENTRY_ADD_OK:
 983                         break;
 984                 case JOURNAL_ENTRY_ADD_OUT_OF_RANGE:
 985                         break;
 986                 default:
 987                         return ret;
 988                 }
 989 next_block:
 990                 pr_debug("next");
 991                 offset          += sectors;
 992                 sectors_read    -= sectors;
 993                 j = ((void *) j) + (sectors << 9);
 994         }
 995
 996         return 0;
 997 }
 998
 999 static void bch2_journal_read_device(struct closure *cl)
1000 {
1001         struct journal_device *ja =
1002                 container_of(cl, struct journal_device, read);
1003         struct bch_dev *ca = container_of(ja, struct bch_dev, journal);
1004         struct bch_fs *c = ca->fs;
1005         struct journal_list *jlist =
1006                 container_of(cl->parent, struct journal_list, cl);
1007         struct journal_replay *r, **_r;
1008         struct genradix_iter iter;
1009         struct journal_read_buf buf = { NULL, 0 };
1010         unsigned i;
1011         int ret = 0;
1012
1013         if (!ja->nr)
1014                 goto out;
1015
1016         ret = journal_read_buf_realloc(&buf, PAGE_SIZE);
1017         if (ret)
1018                 goto err;
1019
1020         pr_debug("%u journal buckets", ja->nr);
1021
1022         for (i = 0; i < ja->nr; i++) {
1023                 ret = journal_read_bucket(ca, &buf, jlist, i);
1024                 if (ret)
1025                         goto err;
1026         }
1027
1028         ja->sectors_free = ca->mi.bucket_size;
1029
1030         mutex_lock(&jlist->lock);
1031         genradix_for_each_reverse(&c->journal_entries, iter, _r) {
1032                 r = *_r;
1033
1034                 if (!r)
1035                         continue;
1036
1037                 for (i = 0; i < r->nr_ptrs; i++) {
1038                         if (r->ptrs[i].dev == ca->dev_idx) {
1039                                 unsigned wrote = bucket_remainder(ca, r->ptrs[i].sector) +
1040                                         vstruct_sectors(&r->j, c->block_bits);
1041
1042                                 ja->cur_idx = r->ptrs[i].bucket;
1043                                 ja->sectors_free = ca->mi.bucket_size - wrote;
1044                                 goto found;
1045                         }
1046                 }
1047         }
1048 found:
1049         mutex_unlock(&jlist->lock);
1050
1051         if (ja->bucket_seq[ja->cur_idx] &&
1052             ja->sectors_free == ca->mi.bucket_size) {
1053                 bch_err(c, "ja->sectors_free == ca->mi.bucket_size");
1054                 bch_err(c, "cur_idx %u/%u", ja->cur_idx, ja->nr);
1055                 for (i = 0; i < 3; i++) {
1056                         unsigned idx = (ja->cur_idx + ja->nr - 1 + i) % ja->nr;
1057                         bch_err(c, "bucket_seq[%u] = %llu", idx, ja->bucket_seq[idx]);
1058                 }
1059                 ja->sectors_free = 0;
1060         }
1061
1062         /*
1063          * Set dirty_idx to indicate the entire journal is full and needs to be
1064          * reclaimed - journal reclaim will immediately reclaim whatever isn't
1065          * pinned when it first runs:
1066          */
1067         ja->discard_idx = ja->dirty_idx_ondisk =
1068                 ja->dirty_idx = (ja->cur_idx + 1) % ja->nr;
1069 out:
1070         bch_verbose(c, "journal read done on device %s, ret %i", ca->name, ret);
1071         kvpfree(buf.data, buf.size);
1072         percpu_ref_put(&ca->io_ref);
1073         closure_return(cl);
1074         return;
1075 err:
1076         mutex_lock(&jlist->lock);
1077         jlist->ret = ret;
1078         mutex_unlock(&jlist->lock);
1079         goto out;
1080 }
1081
1082 void bch2_journal_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
1083                                struct journal_replay *j)
1084 {
1085         unsigned i;
1086
1087         for (i = 0; i < j->nr_ptrs; i++) {
1088                 struct bch_dev *ca = bch_dev_bkey_exists(c, j->ptrs[i].dev);
1089                 u64 offset;
1090
1091                 div64_u64_rem(j->ptrs[i].sector, ca->mi.bucket_size, &offset);
1092
1093                 if (i)
1094                         prt_printf(out, " ");
1095                 prt_printf(out, "%u:%u:%u (sector %llu)",
1096                        j->ptrs[i].dev,
1097                        j->ptrs[i].bucket,
1098                        j->ptrs[i].bucket_offset,
1099                        j->ptrs[i].sector);
1100         }
1101 }
1102
1103 int bch2_journal_read(struct bch_fs *c,
1104                       u64 *last_seq,
1105                       u64 *blacklist_seq,
1106                       u64 *start_seq)
1107 {
1108         struct journal_list jlist;
1109         struct journal_replay *i, **_i, *prev = NULL;
1110         struct genradix_iter radix_iter;
1111         struct bch_dev *ca;
1112         unsigned iter;
1113         struct printbuf buf = PRINTBUF;
1114         bool degraded = false, last_write_torn = false;
1115         u64 seq;
1116         int ret = 0;
1117
1118         closure_init_stack(&jlist.cl);
1119         mutex_init(&jlist.lock);
1120         jlist.last_seq = 0;
1121         jlist.ret = 0;
1122
1123         for_each_member_device(ca, c, iter) {
1124                 if (!c->opts.fsck &&
1125                     !(bch2_dev_has_data(c, ca) & (1 << BCH_DATA_journal)))
1126                         continue;
1127
1128                 if ((ca->mi.state == BCH_MEMBER_STATE_rw ||
1129                      ca->mi.state == BCH_MEMBER_STATE_ro) &&
1130                     percpu_ref_tryget(&ca->io_ref))
1131                         closure_call(&ca->journal.read,
1132                                      bch2_journal_read_device,
1133                                      system_unbound_wq,
1134                                      &jlist.cl);
1135                 else
1136                         degraded = true;
1137         }
1138
1139         closure_sync(&jlist.cl);
1140
1141         if (jlist.ret)
1142                 return jlist.ret;
1143
1144         *last_seq       = 0;
1145         *start_seq      = 0;
1146         *blacklist_seq  = 0;
1147
1148         /*
1149          * Find most recent flush entry, and ignore newer non flush entries -
1150          * those entries will be blacklisted:
1151          */
1152         genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
1153                 enum bkey_invalid_flags flags = BKEY_INVALID_JOURNAL;
1154
1155                 i = *_i;
1156
1157                 if (!i || i->ignore)
1158                         continue;
1159
1160                 if (!*start_seq)
1161                         *blacklist_seq = *start_seq = le64_to_cpu(i->j.seq) + 1;
1162
1163                 if (JSET_NO_FLUSH(&i->j)) {
1164                         i->ignore = true;
1165                         continue;
1166                 }
1167
1168                 if (!last_write_torn && !i->csum_good) {
1169                         last_write_torn = true;
1170                         i->ignore = true;
1171                         continue;
1172                 }
1173
1174                 if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
1175                                          c, le32_to_cpu(i->j.version), &i->j, NULL,
1176                                          "invalid journal entry: last_seq > seq (%llu > %llu)",
1177                                          le64_to_cpu(i->j.last_seq),
1178                                          le64_to_cpu(i->j.seq)))
1179                         i->j.last_seq = i->j.seq;
1180
1181                 *last_seq       = le64_to_cpu(i->j.last_seq);
1182                 *blacklist_seq  = le64_to_cpu(i->j.seq) + 1;
1183                 break;
1184         }
1185
1186         if (!*start_seq) {
1187                 bch_info(c, "journal read done, but no entries found");
1188                 return 0;
1189         }
1190
1191         if (!*last_seq) {
1192                 fsck_err(c, "journal read done, but no entries found after dropping non-flushes");
1193                 return 0;
1194         }
1195
1196         bch_info(c, "journal read done, replaying entries %llu-%llu",
1197                  *last_seq, *blacklist_seq - 1);
1198
1199         if (*start_seq != *blacklist_seq)
1200                 bch_info(c, "dropped unflushed entries %llu-%llu",
1201                          *blacklist_seq, *start_seq - 1);
1202
1203         /* Drop blacklisted entries and entries older than last_seq: */
1204         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1205                 i = *_i;
1206
1207                 if (!i || i->ignore)
1208                         continue;
1209
1210                 seq = le64_to_cpu(i->j.seq);
1211                 if (seq < *last_seq) {
1212                         journal_replay_free(c, i);
1213                         continue;
1214                 }
1215
1216                 if (bch2_journal_seq_is_blacklisted(c, seq, true)) {
1217                         fsck_err_on(!JSET_NO_FLUSH(&i->j), c,
1218                                     "found blacklisted journal entry %llu", seq);
1219                         i->ignore = true;
1220                 }
1221         }
1222
1223         /* Check for missing entries: */
1224         seq = *last_seq;
1225         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1226                 i = *_i;
1227
1228                 if (!i || i->ignore)
1229                         continue;
1230
1231                 BUG_ON(seq > le64_to_cpu(i->j.seq));
1232
1233                 while (seq < le64_to_cpu(i->j.seq)) {
1234                         u64 missing_start, missing_end;
1235                         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1236
1237                         while (seq < le64_to_cpu(i->j.seq) &&
1238                                bch2_journal_seq_is_blacklisted(c, seq, false))
1239                                 seq++;
1240
1241                         if (seq == le64_to_cpu(i->j.seq))
1242                                 break;
1243
1244                         missing_start = seq;
1245
1246                         while (seq < le64_to_cpu(i->j.seq) &&
1247                                !bch2_journal_seq_is_blacklisted(c, seq, false))
1248                                 seq++;
1249
1250                         if (prev) {
1251                                 bch2_journal_ptrs_to_text(&buf1, c, prev);
1252                                 prt_printf(&buf1, " size %zu", vstruct_sectors(&prev->j, c->block_bits));
1253                         } else
1254                                 prt_printf(&buf1, "(none)");
1255                         bch2_journal_ptrs_to_text(&buf2, c, i);
1256
1257                         missing_end = seq - 1;
1258                         fsck_err(c, "journal entries %llu-%llu missing! (replaying %llu-%llu)\n"
1259                                  "  prev at %s\n"
1260                                  "  next at %s",
1261                                  missing_start, missing_end,
1262                                  *last_seq, *blacklist_seq - 1,
1263                                  buf1.buf, buf2.buf);
1264
1265                         printbuf_exit(&buf1);
1266                         printbuf_exit(&buf2);
1267                 }
1268
1269                 prev = i;
1270                 seq++;
1271         }
1272
1273         genradix_for_each(&c->journal_entries, radix_iter, _i) {
1274                 struct bch_replicas_padded replicas = {
1275                         .e.data_type = BCH_DATA_journal,
1276                         .e.nr_required = 1,
1277                 };
1278                 unsigned ptr;
1279
1280                 i = *_i;
1281                 if (!i || i->ignore)
1282                         continue;
1283
1284                 for (ptr = 0; ptr < i->nr_ptrs; ptr++) {
1285                         struct bch_dev *ca = bch_dev_bkey_exists(c, i->ptrs[ptr].dev);
1286
1287                         if (!i->ptrs[ptr].csum_good)
1288                                 bch_err_dev_offset(ca, i->ptrs[ptr].sector,
1289                                                    "invalid journal checksum, seq %llu%s",
1290                                                    le64_to_cpu(i->j.seq),
1291                                                    i->csum_good ? " (had good copy on another device)" : "");
1292                 }
1293
1294                 ret = jset_validate(c,
1295                                     bch_dev_bkey_exists(c, i->ptrs[0].dev),
1296                                     &i->j,
1297                                     i->ptrs[0].sector,
1298                                     READ);
1299                 if (ret)
1300                         goto err;
1301
1302                 for (ptr = 0; ptr < i->nr_ptrs; ptr++)
1303                         replicas.e.devs[replicas.e.nr_devs++] = i->ptrs[ptr].dev;
1304
1305                 bch2_replicas_entry_sort(&replicas.e);
1306
1307                 /*
1308                  * If we're mounting in degraded mode - if we didn't read all
1309                  * the devices - this is wrong:
1310                  */
1311
1312                 printbuf_reset(&buf);
1313                 bch2_replicas_entry_to_text(&buf, &replicas.e);
1314
1315                 if (!degraded &&
1316                     fsck_err_on(!bch2_replicas_marked(c, &replicas.e), c,
1317                                 "superblock not marked as containing replicas %s",
1318                                 buf.buf)) {
1319                         ret = bch2_mark_replicas(c, &replicas.e);
1320                         if (ret)
1321                                 goto err;
1322                 }
1323         }
1324 err:
1325 fsck_err:
1326         printbuf_exit(&buf);
1327         return ret;
1328 }
1329
1330 /* journal write: */
1331
1332 static void __journal_write_alloc(struct journal *j,
1333                                   struct journal_buf *w,
1334                                   struct dev_alloc_list *devs_sorted,
1335                                   unsigned sectors,
1336                                   unsigned *replicas,
1337                                   unsigned replicas_want)
1338 {
1339         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1340         struct journal_device *ja;
1341         struct bch_dev *ca;
1342         unsigned i;
1343
1344         if (*replicas >= replicas_want)
1345                 return;
1346
1347         for (i = 0; i < devs_sorted->nr; i++) {
1348                 ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
1349                 if (!ca)
1350                         continue;
1351
1352                 ja = &ca->journal;
1353
1354                 /*
1355                  * Check that we can use this device, and aren't already using
1356                  * it:
1357                  */
1358                 if (!ca->mi.durability ||
1359                     ca->mi.state != BCH_MEMBER_STATE_rw ||
1360                     !ja->nr ||
1361                     bch2_bkey_has_device_c(bkey_i_to_s_c(&w->key), ca->dev_idx) ||
1362                     sectors > ja->sectors_free)
1363                         continue;
1364
1365                 bch2_dev_stripe_increment(ca, &j->wp.stripe);
1366
1367                 bch2_bkey_append_ptr(&w->key,
1368                         (struct bch_extent_ptr) {
1369                                   .offset = bucket_to_sector(ca,
1370                                         ja->buckets[ja->cur_idx]) +
1371                                         ca->mi.bucket_size -
1372                                         ja->sectors_free,
1373                                   .dev = ca->dev_idx,
1374                 });
1375
1376                 ja->sectors_free -= sectors;
1377                 ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1378
1379                 *replicas += ca->mi.durability;
1380
1381                 if (*replicas >= replicas_want)
1382                         break;
1383         }
1384 }
1385
1386 /**
1387  * journal_next_bucket - move on to the next journal bucket if possible
1388  */
1389 static int journal_write_alloc(struct journal *j, struct journal_buf *w,
1390                                unsigned sectors)
1391 {
1392         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1393         struct bch_devs_mask devs;
1394         struct journal_device *ja;
1395         struct bch_dev *ca;
1396         struct dev_alloc_list devs_sorted;
1397         unsigned target = c->opts.metadata_target ?:
1398                 c->opts.foreground_target;
1399         unsigned i, replicas = 0, replicas_want =
1400                 READ_ONCE(c->opts.metadata_replicas);
1401
1402         rcu_read_lock();
1403 retry:
1404         devs = target_rw_devs(c, BCH_DATA_journal, target);
1405
1406         devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
1407
1408         __journal_write_alloc(j, w, &devs_sorted,
1409                               sectors, &replicas, replicas_want);
1410
1411         if (replicas >= replicas_want)
1412                 goto done;
1413
1414         for (i = 0; i < devs_sorted.nr; i++) {
1415                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1416                 if (!ca)
1417                         continue;
1418
1419                 ja = &ca->journal;
1420
1421                 if (sectors > ja->sectors_free &&
1422                     sectors <= ca->mi.bucket_size &&
1423                     bch2_journal_dev_buckets_available(j, ja,
1424                                         journal_space_discarded)) {
1425                         ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
1426                         ja->sectors_free = ca->mi.bucket_size;
1427
1428                         /*
1429                          * ja->bucket_seq[ja->cur_idx] must always have
1430                          * something sensible:
1431                          */
1432                         ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
1433                 }
1434         }
1435
1436         __journal_write_alloc(j, w, &devs_sorted,
1437                               sectors, &replicas, replicas_want);
1438
1439         if (replicas < replicas_want && target) {
1440                 /* Retry from all devices: */
1441                 target = 0;
1442                 goto retry;
1443         }
1444 done:
1445         rcu_read_unlock();
1446
1447         BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
1448
1449         return replicas >= c->opts.metadata_replicas_required ? 0 : -EROFS;
1450 }
1451
1452 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
1453 {
1454         /* we aren't holding j->lock: */
1455         unsigned new_size = READ_ONCE(j->buf_size_want);
1456         void *new_buf;
1457
1458         if (buf->buf_size >= new_size)
1459                 return;
1460
1461         new_buf = kvpmalloc(new_size, GFP_NOFS|__GFP_NOWARN);
1462         if (!new_buf)
1463                 return;
1464
1465         memcpy(new_buf, buf->data, buf->buf_size);
1466
1467         spin_lock(&j->lock);
1468         swap(buf->data,         new_buf);
1469         swap(buf->buf_size,     new_size);
1470         spin_unlock(&j->lock);
1471
1472         kvpfree(new_buf, new_size);
1473 }
1474
1475 static inline struct journal_buf *journal_last_unwritten_buf(struct journal *j)
1476 {
1477         return j->buf + (journal_last_unwritten_seq(j) & JOURNAL_BUF_MASK);
1478 }
1479
1480 static void journal_write_done(struct closure *cl)
1481 {
1482         struct journal *j = container_of(cl, struct journal, io);
1483         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1484         struct journal_buf *w = journal_last_unwritten_buf(j);
1485         union journal_res_state old, new;
1486         u64 v, seq;
1487         int err = 0;
1488
1489         bch2_time_stats_update(!JSET_NO_FLUSH(w->data)
1490                                ? j->flush_write_time
1491                                : j->noflush_write_time, j->write_start_time);
1492
1493         if (!w->devs_written.nr) {
1494                 bch_err(c, "unable to write journal to sufficient devices");
1495                 err = -EIO;
1496         }
1497         if (err)
1498                 bch2_fatal_error(c);
1499
1500         spin_lock(&j->lock);
1501         seq = le64_to_cpu(w->data->seq);
1502
1503         if (seq >= j->pin.front)
1504                 journal_seq_pin(j, seq)->devs = w->devs_written;
1505
1506         if (!err) {
1507                 if (!JSET_NO_FLUSH(w->data)) {
1508                         j->flushed_seq_ondisk = seq;
1509                         j->last_seq_ondisk = w->last_seq;
1510
1511                         bch2_do_discards(c);
1512                         closure_wake_up(&c->freelist_wait);
1513
1514                         bch2_reset_alloc_cursors(c);
1515                 }
1516         } else if (!j->err_seq || seq < j->err_seq)
1517                 j->err_seq      = seq;
1518
1519         j->seq_ondisk           = seq;
1520
1521         /*
1522          * Updating last_seq_ondisk may let bch2_journal_reclaim_work() discard
1523          * more buckets:
1524          *
1525          * Must come before signaling write completion, for
1526          * bch2_fs_journal_stop():
1527          */
1528         if (j->watermark != BCH_WATERMARK_stripe)
1529                 journal_reclaim_kick(&c->journal);
1530
1531         /* also must come before signalling write completion: */
1532         closure_debug_destroy(cl);
1533
1534         v = atomic64_read(&j->reservations.counter);
1535         do {
1536                 old.v = new.v = v;
1537                 BUG_ON(journal_state_count(new, new.unwritten_idx));
1538
1539                 new.unwritten_idx++;
1540         } while ((v = atomic64_cmpxchg(&j->reservations.counter,
1541                                        old.v, new.v)) != old.v);
1542
1543         bch2_journal_space_available(j);
1544
1545         closure_wake_up(&w->wait);
1546         journal_wake(j);
1547
1548         if (!journal_state_count(new, new.unwritten_idx) &&
1549             journal_last_unwritten_seq(j) <= journal_cur_seq(j)) {
1550                 closure_call(&j->io, bch2_journal_write, c->io_complete_wq, NULL);
1551         } else if (journal_last_unwritten_seq(j) == journal_cur_seq(j) &&
1552                    new.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL) {
1553                 struct journal_buf *buf = journal_cur_buf(j);
1554                 long delta = buf->expires - jiffies;
1555
1556                 /*
1557                  * We don't close a journal entry to write it while there's
1558                  * previous entries still in flight - the current journal entry
1559                  * might want to be written now:
1560                  */
1561
1562                 mod_delayed_work(c->io_complete_wq, &j->write_work, max(0L, delta));
1563         }
1564
1565         spin_unlock(&j->lock);
1566 }
1567
1568 static void journal_write_endio(struct bio *bio)
1569 {
1570         struct bch_dev *ca = bio->bi_private;
1571         struct journal *j = &ca->fs->journal;
1572         struct journal_buf *w = journal_last_unwritten_buf(j);
1573         unsigned long flags;
1574
1575         if (bch2_dev_io_err_on(bio->bi_status, ca, "error writing journal entry %llu: %s",
1576                                le64_to_cpu(w->data->seq),
1577                                bch2_blk_status_to_str(bio->bi_status)) ||
1578             bch2_meta_write_fault("journal")) {
1579                 spin_lock_irqsave(&j->err_lock, flags);
1580                 bch2_dev_list_drop_dev(&w->devs_written, ca->dev_idx);
1581                 spin_unlock_irqrestore(&j->err_lock, flags);
1582         }
1583
1584         closure_put(&j->io);
1585         percpu_ref_put(&ca->io_ref);
1586 }
1587
1588 static void do_journal_write(struct closure *cl)
1589 {
1590         struct journal *j = container_of(cl, struct journal, io);
1591         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1592         struct bch_dev *ca;
1593         struct journal_buf *w = journal_last_unwritten_buf(j);
1594         struct bch_extent_ptr *ptr;
1595         struct bio *bio;
1596         unsigned sectors = vstruct_sectors(w->data, c->block_bits);
1597
1598         extent_for_each_ptr(bkey_i_to_s_extent(&w->key), ptr) {
1599                 ca = bch_dev_bkey_exists(c, ptr->dev);
1600                 if (!percpu_ref_tryget(&ca->io_ref)) {
1601                         /* XXX: fix this */
1602                         bch_err(c, "missing device for journal write\n");
1603                         continue;
1604                 }
1605
1606                 this_cpu_add(ca->io_done->sectors[WRITE][BCH_DATA_journal],
1607                              sectors);
1608
1609                 bio = ca->journal.bio;
1610                 bio_reset(bio, ca->disk_sb.bdev, REQ_OP_WRITE|REQ_SYNC|REQ_META);
1611                 bio->bi_iter.bi_sector  = ptr->offset;
1612                 bio->bi_end_io          = journal_write_endio;
1613                 bio->bi_private         = ca;
1614
1615                 BUG_ON(bio->bi_iter.bi_sector == ca->prev_journal_sector);
1616                 ca->prev_journal_sector = bio->bi_iter.bi_sector;
1617
1618                 if (!JSET_NO_FLUSH(w->data))
1619                         bio->bi_opf    |= REQ_FUA;
1620                 if (!JSET_NO_FLUSH(w->data) && !w->separate_flush)
1621                         bio->bi_opf    |= REQ_PREFLUSH;
1622
1623                 bch2_bio_map(bio, w->data, sectors << 9);
1624
1625                 trace_and_count(c, journal_write, bio);
1626                 closure_bio_submit(bio, cl);
1627
1628                 ca->journal.bucket_seq[ca->journal.cur_idx] =
1629                         le64_to_cpu(w->data->seq);
1630         }
1631
1632         continue_at(cl, journal_write_done, c->io_complete_wq);
1633         return;
1634 }
1635
1636 static void bch2_journal_entries_postprocess(struct bch_fs *c, struct jset *jset)
1637 {
1638         struct jset_entry *i, *next, *prev = NULL;
1639
1640         /*
1641          * Simple compaction, dropping empty jset_entries (from journal
1642          * reservations that weren't fully used) and merging jset_entries that
1643          * can be.
1644          *
1645          * If we wanted to be really fancy here, we could sort all the keys in
1646          * the jset and drop keys that were overwritten - probably not worth it:
1647          */
1648         vstruct_for_each_safe(jset, i, next) {
1649                 unsigned u64s = le16_to_cpu(i->u64s);
1650
1651                 /* Empty entry: */
1652                 if (!u64s)
1653                         continue;
1654
1655                 if (i->type == BCH_JSET_ENTRY_btree_root)
1656                         bch2_journal_entry_to_btree_root(c, i);
1657
1658                 /* Can we merge with previous entry? */
1659                 if (prev &&
1660                     i->btree_id == prev->btree_id &&
1661                     i->level    == prev->level &&
1662                     i->type     == prev->type &&
1663                     i->type     == BCH_JSET_ENTRY_btree_keys &&
1664                     le16_to_cpu(prev->u64s) + u64s <= U16_MAX) {
1665                         memmove_u64s_down(vstruct_next(prev),
1666                                           i->_data,
1667                                           u64s);
1668                         le16_add_cpu(&prev->u64s, u64s);
1669                         continue;
1670                 }
1671
1672                 /* Couldn't merge, move i into new position (after prev): */
1673                 prev = prev ? vstruct_next(prev) : jset->start;
1674                 if (i != prev)
1675                         memmove_u64s_down(prev, i, jset_u64s(u64s));
1676         }
1677
1678         prev = prev ? vstruct_next(prev) : jset->start;
1679         jset->u64s = cpu_to_le32((u64 *) prev - jset->_data);
1680 }
1681
1682 void bch2_journal_write(struct closure *cl)
1683 {
1684         struct journal *j = container_of(cl, struct journal, io);
1685         struct bch_fs *c = container_of(j, struct bch_fs, journal);
1686         struct bch_dev *ca;
1687         struct journal_buf *w = journal_last_unwritten_buf(j);
1688         struct bch_replicas_padded replicas;
1689         struct jset_entry *start, *end;
1690         struct jset *jset;
1691         struct bio *bio;
1692         struct printbuf journal_debug_buf = PRINTBUF;
1693         bool validate_before_checksum = false;
1694         unsigned i, sectors, bytes, u64s, nr_rw_members = 0;
1695         int ret;
1696
1697         BUG_ON(BCH_SB_CLEAN(c->disk_sb.sb));
1698
1699         journal_buf_realloc(j, w);
1700         jset = w->data;
1701
1702         j->write_start_time = local_clock();
1703
1704         spin_lock(&j->lock);
1705
1706         /*
1707          * If the journal is in an error state - we did an emergency shutdown -
1708          * we prefer to continue doing journal writes. We just mark them as
1709          * noflush so they'll never be used, but they'll still be visible by the
1710          * list_journal tool - this helps in debugging.
1711          *
1712          * There's a caveat: the first journal write after marking the
1713          * superblock dirty must always be a flush write, because on startup
1714          * from a clean shutdown we didn't necessarily read the journal and the
1715          * new journal write might overwrite whatever was in the journal
1716          * previously - we can't leave the journal without any flush writes in
1717          * it.
1718          *
1719          * So if we're in an error state, and we're still starting up, we don't
1720          * write anything at all.
1721          */
1722         if (!test_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags) &&
1723             (bch2_journal_error(j) ||
1724              w->noflush ||
1725              (!w->must_flush &&
1726               (jiffies - j->last_flush_write) < msecs_to_jiffies(c->opts.journal_flush_delay) &&
1727               test_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags)))) {
1728                 w->noflush = true;
1729                 SET_JSET_NO_FLUSH(jset, true);
1730                 jset->last_seq  = 0;
1731                 w->last_seq     = 0;
1732
1733                 j->nr_noflush_writes++;
1734         } else if (!bch2_journal_error(j)) {
1735                 j->last_flush_write = jiffies;
1736                 j->nr_flush_writes++;
1737                 clear_bit(JOURNAL_NEED_FLUSH_WRITE, &j->flags);
1738         } else {
1739                 spin_unlock(&j->lock);
1740                 goto err;
1741         }
1742         spin_unlock(&j->lock);
1743
1744         /*
1745          * New btree roots are set by journalling them; when the journal entry
1746          * gets written we have to propagate them to c->btree_roots
1747          *
1748          * But, every journal entry we write has to contain all the btree roots
1749          * (at least for now); so after we copy btree roots to c->btree_roots we
1750          * have to get any missing btree roots and add them to this journal
1751          * entry:
1752          */
1753
1754         bch2_journal_entries_postprocess(c, jset);
1755
1756         start = end = vstruct_last(jset);
1757
1758         end     = bch2_btree_roots_to_journal_entries(c, jset->start, end);
1759
1760         bch2_journal_super_entries_add_common(c, &end,
1761                                 le64_to_cpu(jset->seq));
1762         u64s    = (u64 *) end - (u64 *) start;
1763         BUG_ON(u64s > j->entry_u64s_reserved);
1764
1765         le32_add_cpu(&jset->u64s, u64s);
1766
1767         sectors = vstruct_sectors(jset, c->block_bits);
1768         bytes   = vstruct_bytes(jset);
1769
1770         if (sectors > w->sectors) {
1771                 bch2_fs_fatal_error(c, "aieeee! journal write overran available space, %zu > %u (extra %u reserved %u/%u)",
1772                                     vstruct_bytes(jset), w->sectors << 9,
1773                                     u64s, w->u64s_reserved, j->entry_u64s_reserved);
1774                 goto err;
1775         }
1776
1777         jset->magic             = cpu_to_le64(jset_magic(c));
1778         jset->version           = cpu_to_le32(c->sb.version);
1779
1780         SET_JSET_BIG_ENDIAN(jset, CPU_BIG_ENDIAN);
1781         SET_JSET_CSUM_TYPE(jset, bch2_meta_checksum_type(c));
1782
1783         if (!JSET_NO_FLUSH(jset) && journal_entry_empty(jset))
1784                 j->last_empty_seq = le64_to_cpu(jset->seq);
1785
1786         if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
1787                 validate_before_checksum = true;
1788
1789         if (le32_to_cpu(jset->version) < bcachefs_metadata_version_current)
1790                 validate_before_checksum = true;
1791
1792         if (validate_before_checksum &&
1793             jset_validate(c, NULL, jset, 0, WRITE))
1794                 goto err;
1795
1796         ret = bch2_encrypt(c, JSET_CSUM_TYPE(jset), journal_nonce(jset),
1797                     jset->encrypted_start,
1798                     vstruct_end(jset) - (void *) jset->encrypted_start);
1799         if (bch2_fs_fatal_err_on(ret, c,
1800                         "error decrypting journal entry: %i", ret))
1801                 goto err;
1802
1803         jset->csum = csum_vstruct(c, JSET_CSUM_TYPE(jset),
1804                                   journal_nonce(jset), jset);
1805
1806         if (!validate_before_checksum &&
1807             jset_validate(c, NULL, jset, 0, WRITE))
1808                 goto err;
1809
1810         memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
1811
1812 retry_alloc:
1813         spin_lock(&j->lock);
1814         ret = journal_write_alloc(j, w, sectors);
1815
1816         if (ret && j->can_discard) {
1817                 spin_unlock(&j->lock);
1818                 bch2_journal_do_discards(j);
1819                 goto retry_alloc;
1820         }
1821
1822         if (ret)
1823                 __bch2_journal_debug_to_text(&journal_debug_buf, j);
1824
1825         /*
1826          * write is allocated, no longer need to account for it in
1827          * bch2_journal_space_available():
1828          */
1829         w->sectors = 0;
1830
1831         /*
1832          * journal entry has been compacted and allocated, recalculate space
1833          * available:
1834          */
1835         bch2_journal_space_available(j);
1836         spin_unlock(&j->lock);
1837
1838         if (ret) {
1839                 bch_err(c, "Unable to allocate journal write:\n%s",
1840                         journal_debug_buf.buf);
1841                 printbuf_exit(&journal_debug_buf);
1842                 goto err;
1843         }
1844
1845         w->devs_written = bch2_bkey_devs(bkey_i_to_s_c(&w->key));
1846
1847         if (c->opts.nochanges)
1848                 goto no_io;
1849
1850         for_each_rw_member(ca, c, i)
1851                 nr_rw_members++;
1852
1853         if (nr_rw_members > 1)
1854                 w->separate_flush = true;
1855
1856         /*
1857          * Mark journal replicas before we submit the write to guarantee
1858          * recovery will find the journal entries after a crash.
1859          */
1860         bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
1861                                  w->devs_written);
1862         ret = bch2_mark_replicas(c, &replicas.e);
1863         if (ret)
1864                 goto err;
1865
1866         if (!JSET_NO_FLUSH(jset) && w->separate_flush) {
1867                 for_each_rw_member(ca, c, i) {
1868                         percpu_ref_get(&ca->io_ref);
1869
1870                         bio = ca->journal.bio;
1871                         bio_reset(bio, ca->disk_sb.bdev, REQ_OP_FLUSH);
1872                         bio->bi_end_io          = journal_write_endio;
1873                         bio->bi_private         = ca;
1874                         closure_bio_submit(bio, cl);
1875                 }
1876         }
1877
1878         continue_at(cl, do_journal_write, c->io_complete_wq);
1879         return;
1880 no_io:
1881         continue_at(cl, journal_write_done, c->io_complete_wq);
1882         return;
1883 err:
1884         bch2_fatal_error(c);
1885         continue_at(cl, journal_write_done, c->io_complete_wq);
1886 }