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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_sort.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "buckets.h"
  13 #include "checksum.h"
  14 #include "debug.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "io.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_seq_blacklist.h"
  20 #include "super-io.h"
  21
  22 #include <linux/sched/mm.h>
  23 #include <trace/events/bcachefs.h>
  24
  25 static void verify_no_dups(struct btree *b,
  26                            struct bkey_packed *start,
  27                            struct bkey_packed *end)
  28 {
  29 #ifdef CONFIG_BCACHEFS_DEBUG
  30         struct bkey_packed *k, *p;
  31
  32         if (start == end)
  33                 return;
  34
  35         for (p = start, k = bkey_next(start);
  36              k != end;
  37              p = k, k = bkey_next(k)) {
  38                 struct bkey l = bkey_unpack_key(b, p);
  39                 struct bkey r = bkey_unpack_key(b, k);
  40
  41                 BUG_ON(bpos_cmp(l.p, bkey_start_pos(&r)) >= 0);
  42         }
  43 #endif
  44 }
  45
  46 static void set_needs_whiteout(struct bset *i, int v)
  47 {
  48         struct bkey_packed *k;
  49
  50         for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
  51                 k->needs_whiteout = v;
  52 }
  53
  54 static void btree_bounce_free(struct bch_fs *c, size_t size,
  55                               bool used_mempool, void *p)
  56 {
  57         if (used_mempool)
  58                 mempool_free(p, &c->btree_bounce_pool);
  59         else
  60                 vpfree(p, size);
  61 }
  62
  63 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
  64                                 bool *used_mempool)
  65 {
  66         unsigned flags = memalloc_nofs_save();
  67         void *p;
  68
  69         BUG_ON(size > btree_bytes(c));
  70
  71         *used_mempool = false;
  72         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
  73         if (!p) {
  74                 *used_mempool = true;
  75                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
  76         }
  77         memalloc_nofs_restore(flags);
  78         return p;
  79 }
  80
  81 static void sort_bkey_ptrs(const struct btree *bt,
  82                            struct bkey_packed **ptrs, unsigned nr)
  83 {
  84         unsigned n = nr, a = nr / 2, b, c, d;
  85
  86         if (!a)
  87                 return;
  88
  89         /* Heap sort: see lib/sort.c: */
  90         while (1) {
  91                 if (a)
  92                         a--;
  93                 else if (--n)
  94                         swap(ptrs[0], ptrs[n]);
  95                 else
  96                         break;
  97
  98                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
  99                         b = bch2_bkey_cmp_packed(bt,
 100                                             ptrs[c],
 101                                             ptrs[d]) >= 0 ? c : d;
 102                 if (d == n)
 103                         b = c;
 104
 105                 while (b != a &&
 106                        bch2_bkey_cmp_packed(bt,
 107                                        ptrs[a],
 108                                        ptrs[b]) >= 0)
 109                         b = (b - 1) / 2;
 110                 c = b;
 111                 while (b != a) {
 112                         b = (b - 1) / 2;
 113                         swap(ptrs[b], ptrs[c]);
 114                 }
 115         }
 116 }
 117
 118 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 119 {
 120         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 121         bool used_mempool = false;
 122         size_t bytes = b->whiteout_u64s * sizeof(u64);
 123
 124         if (!b->whiteout_u64s)
 125                 return;
 126
 127         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 128
 129         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 130
 131         for (k = unwritten_whiteouts_start(c, b);
 132              k != unwritten_whiteouts_end(c, b);
 133              k = bkey_next(k))
 134                 *--ptrs = k;
 135
 136         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 137
 138         k = new_whiteouts;
 139
 140         while (ptrs != ptrs_end) {
 141                 bkey_copy(k, *ptrs);
 142                 k = bkey_next(k);
 143                 ptrs++;
 144         }
 145
 146         verify_no_dups(b, new_whiteouts,
 147                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 148
 149         memcpy_u64s(unwritten_whiteouts_start(c, b),
 150                     new_whiteouts, b->whiteout_u64s);
 151
 152         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 153 }
 154
 155 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 156                                 bool compacting, enum compact_mode mode)
 157 {
 158         if (!bset_dead_u64s(b, t))
 159                 return false;
 160
 161         switch (mode) {
 162         case COMPACT_LAZY:
 163                 return should_compact_bset_lazy(b, t) ||
 164                         (compacting && !bset_written(b, bset(b, t)));
 165         case COMPACT_ALL:
 166                 return true;
 167         default:
 168                 BUG();
 169         }
 170 }
 171
 172 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 173 {
 174         struct bset_tree *t;
 175         bool ret = false;
 176
 177         for_each_bset(b, t) {
 178                 struct bset *i = bset(b, t);
 179                 struct bkey_packed *k, *n, *out, *start, *end;
 180                 struct btree_node_entry *src = NULL, *dst = NULL;
 181
 182                 if (t != b->set && !bset_written(b, i)) {
 183                         src = container_of(i, struct btree_node_entry, keys);
 184                         dst = max(write_block(b),
 185                                   (void *) btree_bkey_last(b, t - 1));
 186                 }
 187
 188                 if (src != dst)
 189                         ret = true;
 190
 191                 if (!should_compact_bset(b, t, ret, mode)) {
 192                         if (src != dst) {
 193                                 memmove(dst, src, sizeof(*src) +
 194                                         le16_to_cpu(src->keys.u64s) *
 195                                         sizeof(u64));
 196                                 i = &dst->keys;
 197                                 set_btree_bset(b, t, i);
 198                         }
 199                         continue;
 200                 }
 201
 202                 start   = btree_bkey_first(b, t);
 203                 end     = btree_bkey_last(b, t);
 204
 205                 if (src != dst) {
 206                         memmove(dst, src, sizeof(*src));
 207                         i = &dst->keys;
 208                         set_btree_bset(b, t, i);
 209                 }
 210
 211                 out = i->start;
 212
 213                 for (k = start; k != end; k = n) {
 214                         n = bkey_next(k);
 215
 216                         if (!bkey_deleted(k)) {
 217                                 bkey_copy(out, k);
 218                                 out = bkey_next(out);
 219                         } else {
 220                                 BUG_ON(k->needs_whiteout);
 221                         }
 222                 }
 223
 224                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 225                 set_btree_bset_end(b, t);
 226                 bch2_bset_set_no_aux_tree(b, t);
 227                 ret = true;
 228         }
 229
 230         bch2_verify_btree_nr_keys(b);
 231
 232         bch2_btree_build_aux_trees(b);
 233
 234         return ret;
 235 }
 236
 237 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 238                             enum compact_mode mode)
 239 {
 240         return bch2_drop_whiteouts(b, mode);
 241 }
 242
 243 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 244                             struct btree_iter *iter,
 245                             unsigned start_idx,
 246                             unsigned end_idx,
 247                             bool filter_whiteouts)
 248 {
 249         struct btree_node *out;
 250         struct sort_iter sort_iter;
 251         struct bset_tree *t;
 252         struct bset *start_bset = bset(b, &b->set[start_idx]);
 253         bool used_mempool = false;
 254         u64 start_time, seq = 0;
 255         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 256         bool sorting_entire_node = start_idx == 0 &&
 257                 end_idx == b->nsets;
 258
 259         sort_iter_init(&sort_iter, b);
 260
 261         for (t = b->set + start_idx;
 262              t < b->set + end_idx;
 263              t++) {
 264                 u64s += le16_to_cpu(bset(b, t)->u64s);
 265                 sort_iter_add(&sort_iter,
 266                               btree_bkey_first(b, t),
 267                               btree_bkey_last(b, t));
 268         }
 269
 270         bytes = sorting_entire_node
 271                 ? btree_bytes(c)
 272                 : __vstruct_bytes(struct btree_node, u64s);
 273
 274         out = btree_bounce_alloc(c, bytes, &used_mempool);
 275
 276         start_time = local_clock();
 277
 278         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 279
 280         out->keys.u64s = cpu_to_le16(u64s);
 281
 282         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 283
 284         if (sorting_entire_node)
 285                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 286                                        start_time);
 287
 288         /* Make sure we preserve bset journal_seq: */
 289         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 290                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 291         start_bset->journal_seq = cpu_to_le64(seq);
 292
 293         if (sorting_entire_node) {
 294                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 295
 296                 BUG_ON(bytes != btree_bytes(c));
 297
 298                 /*
 299                  * Our temporary buffer is the same size as the btree node's
 300                  * buffer, we can just swap buffers instead of doing a big
 301                  * memcpy()
 302                  */
 303                 *out = *b->data;
 304                 out->keys.u64s = cpu_to_le16(u64s);
 305                 swap(out, b->data);
 306                 set_btree_bset(b, b->set, &b->data->keys);
 307         } else {
 308                 start_bset->u64s = out->keys.u64s;
 309                 memcpy_u64s(start_bset->start,
 310                             out->keys.start,
 311                             le16_to_cpu(out->keys.u64s));
 312         }
 313
 314         for (i = start_idx + 1; i < end_idx; i++)
 315                 b->nr.bset_u64s[start_idx] +=
 316                         b->nr.bset_u64s[i];
 317
 318         b->nsets -= shift;
 319
 320         for (i = start_idx + 1; i < b->nsets; i++) {
 321                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 322                 b->set[i]               = b->set[i + shift];
 323         }
 324
 325         for (i = b->nsets; i < MAX_BSETS; i++)
 326                 b->nr.bset_u64s[i] = 0;
 327
 328         set_btree_bset_end(b, &b->set[start_idx]);
 329         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 330
 331         btree_bounce_free(c, bytes, used_mempool, out);
 332
 333         bch2_verify_btree_nr_keys(b);
 334 }
 335
 336 void bch2_btree_sort_into(struct bch_fs *c,
 337                          struct btree *dst,
 338                          struct btree *src)
 339 {
 340         struct btree_nr_keys nr;
 341         struct btree_node_iter src_iter;
 342         u64 start_time = local_clock();
 343
 344         BUG_ON(dst->nsets != 1);
 345
 346         bch2_bset_set_no_aux_tree(dst, dst->set);
 347
 348         bch2_btree_node_iter_init_from_start(&src_iter, src);
 349
 350         if (btree_node_is_extents(src))
 351                 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
 352                                 src, &src_iter,
 353                                 &dst->format,
 354                                 true);
 355         else
 356                 nr = bch2_sort_repack(btree_bset_first(dst),
 357                                 src, &src_iter,
 358                                 &dst->format,
 359                                 true);
 360
 361         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 362                                start_time);
 363
 364         set_btree_bset_end(dst, dst->set);
 365
 366         dst->nr.live_u64s       += nr.live_u64s;
 367         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 368         dst->nr.packed_keys     += nr.packed_keys;
 369         dst->nr.unpacked_keys   += nr.unpacked_keys;
 370
 371         bch2_verify_btree_nr_keys(dst);
 372 }
 373
 374 #define SORT_CRIT       (4096 / sizeof(u64))
 375
 376 /*
 377  * We're about to add another bset to the btree node, so if there's currently
 378  * too many bsets - sort some of them together:
 379  */
 380 static bool btree_node_compact(struct bch_fs *c, struct btree *b,
 381                                struct btree_iter *iter)
 382 {
 383         unsigned unwritten_idx;
 384         bool ret = false;
 385
 386         for (unwritten_idx = 0;
 387              unwritten_idx < b->nsets;
 388              unwritten_idx++)
 389                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 390                         break;
 391
 392         if (b->nsets - unwritten_idx > 1) {
 393                 btree_node_sort(c, b, iter, unwritten_idx,
 394                                 b->nsets, false);
 395                 ret = true;
 396         }
 397
 398         if (unwritten_idx > 1) {
 399                 btree_node_sort(c, b, iter, 0, unwritten_idx, false);
 400                 ret = true;
 401         }
 402
 403         return ret;
 404 }
 405
 406 void bch2_btree_build_aux_trees(struct btree *b)
 407 {
 408         struct bset_tree *t;
 409
 410         for_each_bset(b, t)
 411                 bch2_bset_build_aux_tree(b, t,
 412                                 !bset_written(b, bset(b, t)) &&
 413                                 t == bset_tree_last(b));
 414 }
 415
 416 /*
 417  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 418  * inserted into
 419  *
 420  * Safe to call if there already is an unwritten bset - will only add a new bset
 421  * if @b doesn't already have one.
 422  *
 423  * Returns true if we sorted (i.e. invalidated iterators
 424  */
 425 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
 426                           struct btree_iter *iter)
 427 {
 428         struct btree_node_entry *bne;
 429         bool did_sort;
 430
 431         EBUG_ON(!(b->c.lock.state.seq & 1));
 432         EBUG_ON(iter && iter->l[b->c.level].b != b);
 433
 434         did_sort = btree_node_compact(c, b, iter);
 435
 436         bne = want_new_bset(c, b);
 437         if (bne)
 438                 bch2_bset_init_next(c, b, bne);
 439
 440         bch2_btree_build_aux_trees(b);
 441
 442         if (iter && did_sort)
 443                 bch2_btree_iter_reinit_node(iter, b);
 444 }
 445
 446 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 447                           struct btree *b)
 448 {
 449         pr_buf(out, "%s level %u/%u\n  ",
 450                bch2_btree_ids[b->c.btree_id],
 451                b->c.level,
 452                c->btree_roots[b->c.btree_id].level);
 453         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 454 }
 455
 456 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 457                           struct bch_dev *ca,
 458                           struct btree *b, struct bset *i,
 459                           unsigned offset, int write)
 460 {
 461         pr_buf(out, "error validating btree node ");
 462         if (write)
 463                 pr_buf(out, "before write ");
 464         if (ca)
 465                 pr_buf(out, "on %s ", ca->name);
 466         pr_buf(out, "at btree ");
 467         btree_pos_to_text(out, c, b);
 468
 469         pr_buf(out, "\n  node offset %u", b->written);
 470         if (i)
 471                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 472 }
 473
 474 enum btree_err_type {
 475         BTREE_ERR_FIXABLE,
 476         BTREE_ERR_WANT_RETRY,
 477         BTREE_ERR_MUST_RETRY,
 478         BTREE_ERR_FATAL,
 479 };
 480
 481 enum btree_validate_ret {
 482         BTREE_RETRY_READ = 64,
 483 };
 484
 485 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 486 ({                                                                      \
 487         __label__ out;                                                  \
 488         char _buf[300];                                                 \
 489         char *_buf2 = _buf;                                             \
 490         struct printbuf out = PBUF(_buf);                               \
 491                                                                         \
 492         _buf2 = kmalloc(4096, GFP_ATOMIC);                              \
 493         if (_buf2)                                                      \
 494                 out = _PBUF(_buf2, 4986);                               \
 495                                                                         \
 496         btree_err_msg(&out, c, ca, b, i, b->written, write);            \
 497         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 498                                                                         \
 499         if (type == BTREE_ERR_FIXABLE &&                                \
 500             write == READ &&                                            \
 501             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 502                 mustfix_fsck_err(c, "%s", _buf2);                       \
 503                 goto out;                                               \
 504         }                                                               \
 505                                                                         \
 506         switch (write) {                                                \
 507         case READ:                                                      \
 508                 bch_err(c, "%s", _buf2);                                        \
 509                                                                         \
 510                 switch (type) {                                         \
 511                 case BTREE_ERR_FIXABLE:                                 \
 512                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 513                         goto fsck_err;                                  \
 514                 case BTREE_ERR_WANT_RETRY:                              \
 515                         if (have_retry) {                               \
 516                                 ret = BTREE_RETRY_READ;                 \
 517                                 goto fsck_err;                          \
 518                         }                                               \
 519                         break;                                          \
 520                 case BTREE_ERR_MUST_RETRY:                              \
 521                         ret = BTREE_RETRY_READ;                         \
 522                         goto fsck_err;                                  \
 523                 case BTREE_ERR_FATAL:                                   \
 524                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 525                         goto fsck_err;                                  \
 526                 }                                                       \
 527                 break;                                                  \
 528         case WRITE:                                                     \
 529                 bch_err(c, "corrupt metadata before write: %s", _buf2); \
 530                                                                         \
 531                 if (bch2_fs_inconsistent(c)) {                          \
 532                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 533                         goto fsck_err;                                  \
 534                 }                                                       \
 535                 break;                                                  \
 536         }                                                               \
 537 out:                                                                    \
 538         if (_buf2 != _buf)                                              \
 539                 kfree(_buf2);                                           \
 540         true;                                                           \
 541 })
 542
 543 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 544
 545 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 546                          struct btree *b, struct bset *i,
 547                          unsigned sectors, int write, bool have_retry)
 548 {
 549         unsigned version = le16_to_cpu(i->version);
 550         const char *err;
 551         char buf1[100];
 552         char buf2[100];
 553         int ret = 0;
 554
 555         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 556                       version < bcachefs_metadata_version_min) ||
 557                      version >= bcachefs_metadata_version_max,
 558                      BTREE_ERR_FATAL, c, ca, b, i,
 559                      "unsupported bset version");
 560
 561         if (btree_err_on(version < c->sb.version_min,
 562                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 563                          "bset version %u older than superblock version_min %u",
 564                          version, c->sb.version_min)) {
 565                 mutex_lock(&c->sb_lock);
 566                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 567                 bch2_write_super(c);
 568                 mutex_unlock(&c->sb_lock);
 569         }
 570
 571         if (btree_err_on(version > c->sb.version,
 572                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 573                          "bset version %u newer than superblock version %u",
 574                          version, c->sb.version)) {
 575                 mutex_lock(&c->sb_lock);
 576                 c->disk_sb.sb->version = cpu_to_le16(version);
 577                 bch2_write_super(c);
 578                 mutex_unlock(&c->sb_lock);
 579         }
 580
 581         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 582                      BTREE_ERR_FATAL, c, ca, b, i,
 583                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 584
 585         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 586                          BTREE_ERR_FIXABLE, c, ca, b, i,
 587                          "bset past end of btree node")) {
 588                 i->u64s = 0;
 589                 return 0;
 590         }
 591
 592         btree_err_on(b->written && !i->u64s,
 593                      BTREE_ERR_FIXABLE, c, ca, b, i,
 594                      "empty bset");
 595
 596         if (!b->written) {
 597                 struct btree_node *bn =
 598                         container_of(i, struct btree_node, keys);
 599                 /* These indicate that we read the wrong btree node: */
 600
 601                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 602                         struct bch_btree_ptr_v2 *bp =
 603                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 604
 605                         /* XXX endianness */
 606                         btree_err_on(bp->seq != bn->keys.seq,
 607                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 608                                      "incorrect sequence number (wrong btree node)");
 609                 }
 610
 611                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 612                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 613                              "incorrect btree id");
 614
 615                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 616                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 617                              "incorrect level");
 618
 619                 if (!write)
 620                         compat_btree_node(b->c.level, b->c.btree_id, version,
 621                                           BSET_BIG_ENDIAN(i), write, bn);
 622
 623                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 624                         struct bch_btree_ptr_v2 *bp =
 625                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 626
 627                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 628                                 b->data->min_key = bp->min_key;
 629                                 b->data->max_key = b->key.k.p;
 630                         }
 631
 632                         btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
 633                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 634                                      "incorrect min_key: got %s should be %s",
 635                                      (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
 636                                      (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
 637                 }
 638
 639                 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
 640                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 641                              "incorrect max key %s",
 642                              (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
 643
 644                 if (write)
 645                         compat_btree_node(b->c.level, b->c.btree_id, version,
 646                                           BSET_BIG_ENDIAN(i), write, bn);
 647
 648                 err = bch2_bkey_format_validate(&bn->format);
 649                 btree_err_on(err,
 650                              BTREE_ERR_FATAL, c, ca, b, i,
 651                              "invalid bkey format: %s", err);
 652
 653                 compat_bformat(b->c.level, b->c.btree_id, version,
 654                                BSET_BIG_ENDIAN(i), write,
 655                                &bn->format);
 656         }
 657 fsck_err:
 658         return ret;
 659 }
 660
 661 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 662                          struct bset *i, unsigned *whiteout_u64s,
 663                          int write, bool have_retry)
 664 {
 665         unsigned version = le16_to_cpu(i->version);
 666         struct bkey_packed *k, *prev = NULL;
 667         int ret = 0;
 668
 669         for (k = i->start;
 670              k != vstruct_last(i);) {
 671                 struct bkey_s u;
 672                 struct bkey tmp;
 673                 const char *invalid;
 674
 675                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 676                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 677                                  "key extends past end of bset")) {
 678                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 679                         break;
 680                 }
 681
 682                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 683                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 684                                  "invalid bkey format %u", k->format)) {
 685                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 686                         memmove_u64s_down(k, bkey_next(k),
 687                                           (u64 *) vstruct_end(i) - (u64 *) k);
 688                         continue;
 689                 }
 690
 691                 /* XXX: validate k->u64s */
 692                 if (!write)
 693                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 694                                     BSET_BIG_ENDIAN(i), write,
 695                                     &b->format, k);
 696
 697                 u = __bkey_disassemble(b, k, &tmp);
 698
 699                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 700                         bch2_bkey_in_btree_node(b, u.s_c) ?:
 701                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 702                 if (invalid) {
 703                         char buf[160];
 704
 705                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 706                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 707                                   "invalid bkey: %s\n%s", invalid, buf);
 708
 709                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 710                         memmove_u64s_down(k, bkey_next(k),
 711                                           (u64 *) vstruct_end(i) - (u64 *) k);
 712                         continue;
 713                 }
 714
 715                 if (write)
 716                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 717                                     BSET_BIG_ENDIAN(i), write,
 718                                     &b->format, k);
 719
 720                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 721                         char buf1[80];
 722                         char buf2[80];
 723                         struct bkey up = bkey_unpack_key(b, prev);
 724
 725                         bch2_bkey_to_text(&PBUF(buf1), &up);
 726                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 727
 728                         bch2_dump_bset(c, b, i, 0);
 729
 730                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 731                                       "keys out of order: %s > %s",
 732                                       buf1, buf2)) {
 733                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 734                                 memmove_u64s_down(k, bkey_next(k),
 735                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 736                                 continue;
 737                         }
 738                 }
 739
 740                 prev = k;
 741                 k = bkey_next(k);
 742         }
 743 fsck_err:
 744         return ret;
 745 }
 746
 747 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 748                               struct btree *b, bool have_retry)
 749 {
 750         struct btree_node_entry *bne;
 751         struct sort_iter *iter;
 752         struct btree_node *sorted;
 753         struct bkey_packed *k;
 754         struct bch_extent_ptr *ptr;
 755         struct bset *i;
 756         bool used_mempool, blacklisted;
 757         unsigned u64s;
 758         int ret, retry_read = 0, write = READ;
 759
 760         b->version_ondisk = U16_MAX;
 761
 762         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 763         sort_iter_init(iter, b);
 764         iter->size = (btree_blocks(c) + 1) * 2;
 765
 766         if (bch2_meta_read_fault("btree"))
 767                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 768                           "dynamic fault");
 769
 770         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 771                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 772                      "bad magic");
 773
 774         btree_err_on(!b->data->keys.seq,
 775                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 776                      "bad btree header");
 777
 778         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 779                 struct bch_btree_ptr_v2 *bp =
 780                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 781
 782                 btree_err_on(b->data->keys.seq != bp->seq,
 783                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 784                              "got wrong btree node (seq %llx want %llx)",
 785                              b->data->keys.seq, bp->seq);
 786         }
 787
 788         while (b->written < c->opts.btree_node_size) {
 789                 unsigned sectors, whiteout_u64s = 0;
 790                 struct nonce nonce;
 791                 struct bch_csum csum;
 792                 bool first = !b->written;
 793
 794                 if (!b->written) {
 795                         i = &b->data->keys;
 796
 797                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 798                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 799                                      "unknown checksum type %llu",
 800                                      BSET_CSUM_TYPE(i));
 801
 802                         nonce = btree_nonce(i, b->written << 9);
 803                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 804
 805                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 806                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 807                                      "invalid checksum");
 808
 809                         bset_encrypt(c, i, b->written << 9);
 810
 811                         btree_err_on(btree_node_is_extents(b) &&
 812                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 813                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 814                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 815
 816                         sectors = vstruct_sectors(b->data, c->block_bits);
 817                 } else {
 818                         bne = write_block(b);
 819                         i = &bne->keys;
 820
 821                         if (i->seq != b->data->keys.seq)
 822                                 break;
 823
 824                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 825                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 826                                      "unknown checksum type %llu",
 827                                      BSET_CSUM_TYPE(i));
 828
 829                         nonce = btree_nonce(i, b->written << 9);
 830                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 831
 832                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 833                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 834                                      "invalid checksum");
 835
 836                         bset_encrypt(c, i, b->written << 9);
 837
 838                         sectors = vstruct_sectors(bne, c->block_bits);
 839                 }
 840
 841                 b->version_ondisk = min(b->version_ondisk,
 842                                         le16_to_cpu(i->version));
 843
 844                 ret = validate_bset(c, ca, b, i, sectors,
 845                                     READ, have_retry);
 846                 if (ret)
 847                         goto fsck_err;
 848
 849                 if (!b->written)
 850                         btree_node_set_format(b, b->data->format);
 851
 852                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 853                                     READ, have_retry);
 854                 if (ret)
 855                         goto fsck_err;
 856
 857                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 858
 859                 b->written += sectors;
 860
 861                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 862                                         le64_to_cpu(i->journal_seq),
 863                                         true);
 864
 865                 btree_err_on(blacklisted && first,
 866                              BTREE_ERR_FIXABLE, c, ca, b, i,
 867                              "first btree node bset has blacklisted journal seq");
 868                 if (blacklisted && !first)
 869                         continue;
 870
 871                 sort_iter_add(iter, i->start,
 872                               vstruct_idx(i, whiteout_u64s));
 873
 874                 sort_iter_add(iter,
 875                               vstruct_idx(i, whiteout_u64s),
 876                               vstruct_last(i));
 877         }
 878
 879         for (bne = write_block(b);
 880              bset_byte_offset(b, bne) < btree_bytes(c);
 881              bne = (void *) bne + block_bytes(c))
 882                 btree_err_on(bne->keys.seq == b->data->keys.seq,
 883                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
 884                              "found bset signature after last bset");
 885
 886         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
 887         sorted->keys.u64s = 0;
 888
 889         set_btree_bset(b, b->set, &b->data->keys);
 890
 891         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
 892
 893         u64s = le16_to_cpu(sorted->keys.u64s);
 894         *sorted = *b->data;
 895         sorted->keys.u64s = cpu_to_le16(u64s);
 896         swap(sorted, b->data);
 897         set_btree_bset(b, b->set, &b->data->keys);
 898         b->nsets = 1;
 899
 900         BUG_ON(b->nr.live_u64s != u64s);
 901
 902         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
 903
 904         i = &b->data->keys;
 905         for (k = i->start; k != vstruct_last(i);) {
 906                 struct bkey tmp;
 907                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 908                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
 909
 910                 if (invalid ||
 911                     (bch2_inject_invalid_keys &&
 912                      !bversion_cmp(u.k->version, MAX_VERSION))) {
 913                         char buf[160];
 914
 915                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 916                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 917                                   "invalid bkey %s: %s", buf, invalid);
 918
 919                         btree_keys_account_key_drop(&b->nr, 0, k);
 920
 921                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 922                         memmove_u64s_down(k, bkey_next(k),
 923                                           (u64 *) vstruct_end(i) - (u64 *) k);
 924                         set_btree_bset_end(b, b->set);
 925                         continue;
 926                 }
 927
 928                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
 929                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
 930
 931                         bp.v->mem_ptr = 0;
 932                 }
 933
 934                 k = bkey_next(k);
 935         }
 936
 937         bch2_bset_build_aux_tree(b, b->set, false);
 938
 939         set_needs_whiteout(btree_bset_first(b), true);
 940
 941         btree_node_reset_sib_u64s(b);
 942
 943         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
 944                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 945
 946                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
 947                         set_btree_node_need_rewrite(b);
 948         }
 949 out:
 950         mempool_free(iter, &c->fill_iter);
 951         return retry_read;
 952 fsck_err:
 953         if (ret == BTREE_RETRY_READ) {
 954                 retry_read = 1;
 955         } else {
 956                 bch2_inconsistent_error(c);
 957                 set_btree_node_read_error(b);
 958         }
 959         goto out;
 960 }
 961
 962 static void btree_node_read_work(struct work_struct *work)
 963 {
 964         struct btree_read_bio *rb =
 965                 container_of(work, struct btree_read_bio, work);
 966         struct bch_fs *c        = rb->c;
 967         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
 968         struct btree *b         = rb->bio.bi_private;
 969         struct bio *bio         = &rb->bio;
 970         struct bch_io_failures failed = { .nr = 0 };
 971         char buf[200];
 972         struct printbuf out;
 973         bool can_retry;
 974
 975         goto start;
 976         while (1) {
 977                 bch_info(c, "retrying read");
 978                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
 979                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
 980                 bio_reset(bio);
 981                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
 982                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
 983                 bio->bi_iter.bi_size    = btree_bytes(c);
 984
 985                 if (rb->have_ioref) {
 986                         bio_set_dev(bio, ca->disk_sb.bdev);
 987                         submit_bio_wait(bio);
 988                 } else {
 989                         bio->bi_status = BLK_STS_REMOVED;
 990                 }
 991 start:
 992                 out = PBUF(buf);
 993                 btree_pos_to_text(&out, c, b);
 994                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
 995                                    bch2_blk_status_to_str(bio->bi_status), buf);
 996                 if (rb->have_ioref)
 997                         percpu_ref_put(&ca->io_ref);
 998                 rb->have_ioref = false;
 999
1000                 bch2_mark_io_failure(&failed, &rb->pick);
1001
1002                 can_retry = bch2_bkey_pick_read_device(c,
1003                                 bkey_i_to_s_c(&b->key),
1004                                 &failed, &rb->pick) > 0;
1005
1006                 if (!bio->bi_status &&
1007                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1008                         break;
1009
1010                 if (!can_retry) {
1011                         set_btree_node_read_error(b);
1012                         break;
1013                 }
1014         }
1015
1016         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1017                                rb->start_time);
1018         bio_put(&rb->bio);
1019         clear_btree_node_read_in_flight(b);
1020         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1021 }
1022
1023 static void btree_node_read_endio(struct bio *bio)
1024 {
1025         struct btree_read_bio *rb =
1026                 container_of(bio, struct btree_read_bio, bio);
1027         struct bch_fs *c        = rb->c;
1028
1029         if (rb->have_ioref) {
1030                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1031                 bch2_latency_acct(ca, rb->start_time, READ);
1032         }
1033
1034         queue_work(system_unbound_wq, &rb->work);
1035 }
1036
1037 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1038                           bool sync)
1039 {
1040         struct extent_ptr_decoded pick;
1041         struct btree_read_bio *rb;
1042         struct bch_dev *ca;
1043         struct bio *bio;
1044         int ret;
1045
1046         trace_btree_read(c, b);
1047
1048         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1049                                          NULL, &pick);
1050         if (bch2_fs_fatal_err_on(ret <= 0, c,
1051                         "btree node read error: no device to read from")) {
1052                 set_btree_node_read_error(b);
1053                 return;
1054         }
1055
1056         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1057
1058         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1059                                                    btree_bytes(c)),
1060                                &c->btree_bio);
1061         rb = container_of(bio, struct btree_read_bio, bio);
1062         rb->c                   = c;
1063         rb->start_time          = local_clock();
1064         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1065         rb->pick                = pick;
1066         INIT_WORK(&rb->work, btree_node_read_work);
1067         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1068         bio->bi_iter.bi_sector  = pick.ptr.offset;
1069         bio->bi_end_io          = btree_node_read_endio;
1070         bio->bi_private         = b;
1071         bch2_bio_map(bio, b->data, btree_bytes(c));
1072
1073         set_btree_node_read_in_flight(b);
1074
1075         if (rb->have_ioref) {
1076                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1077                              bio_sectors(bio));
1078                 bio_set_dev(bio, ca->disk_sb.bdev);
1079
1080                 if (sync) {
1081                         submit_bio_wait(bio);
1082
1083                         bio->bi_private = b;
1084                         btree_node_read_work(&rb->work);
1085                 } else {
1086                         submit_bio(bio);
1087                 }
1088         } else {
1089                 bio->bi_status = BLK_STS_REMOVED;
1090
1091                 if (sync)
1092                         btree_node_read_work(&rb->work);
1093                 else
1094                         queue_work(system_unbound_wq, &rb->work);
1095
1096         }
1097 }
1098
1099 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1100                         const struct bkey_i *k, unsigned level)
1101 {
1102         struct closure cl;
1103         struct btree *b;
1104         int ret;
1105
1106         closure_init_stack(&cl);
1107
1108         do {
1109                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1110                 closure_sync(&cl);
1111         } while (ret);
1112
1113         b = bch2_btree_node_mem_alloc(c);
1114         bch2_btree_cache_cannibalize_unlock(c);
1115
1116         BUG_ON(IS_ERR(b));
1117
1118         bkey_copy(&b->key, k);
1119         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1120
1121         bch2_btree_node_read(c, b, true);
1122
1123         if (btree_node_read_error(b)) {
1124                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1125
1126                 mutex_lock(&c->btree_cache.lock);
1127                 list_move(&b->list, &c->btree_cache.freeable);
1128                 mutex_unlock(&c->btree_cache.lock);
1129
1130                 ret = -EIO;
1131                 goto err;
1132         }
1133
1134         bch2_btree_set_root_for_read(c, b);
1135 err:
1136         six_unlock_write(&b->c.lock);
1137         six_unlock_intent(&b->c.lock);
1138
1139         return ret;
1140 }
1141
1142 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1143                               struct btree_write *w)
1144 {
1145         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1146
1147         do {
1148                 old = new = v;
1149                 if (!(old & 1))
1150                         break;
1151
1152                 new &= ~1UL;
1153         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1154
1155         if (old & 1)
1156                 closure_put(&((struct btree_update *) new)->cl);
1157
1158         bch2_journal_pin_drop(&c->journal, &w->journal);
1159 }
1160
1161 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1162 {
1163         struct btree_write *w = btree_prev_write(b);
1164
1165         bch2_btree_complete_write(c, b, w);
1166         btree_node_io_unlock(b);
1167 }
1168
1169 static void bch2_btree_node_write_error(struct bch_fs *c,
1170                                         struct btree_write_bio *wbio)
1171 {
1172         struct btree *b         = wbio->wbio.bio.bi_private;
1173         struct bkey_buf k;
1174         struct bch_extent_ptr *ptr;
1175         struct btree_trans trans;
1176         struct btree_iter *iter;
1177         int ret;
1178
1179         bch2_bkey_buf_init(&k);
1180         bch2_trans_init(&trans, c, 0, 0);
1181
1182         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1183                                         BTREE_MAX_DEPTH, b->c.level, 0);
1184 retry:
1185         ret = bch2_btree_iter_traverse(iter);
1186         if (ret)
1187                 goto err;
1188
1189         /* has node been freed? */
1190         if (iter->l[b->c.level].b != b) {
1191                 /* node has been freed: */
1192                 BUG_ON(!btree_node_dying(b));
1193                 goto out;
1194         }
1195
1196         BUG_ON(!btree_node_hashed(b));
1197
1198         bch2_bkey_buf_copy(&k, c, &b->key);
1199
1200         bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1201                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1202
1203         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1204                 goto err;
1205
1206         ret = bch2_btree_node_update_key(c, iter, b, k.k);
1207         if (ret == -EINTR)
1208                 goto retry;
1209         if (ret)
1210                 goto err;
1211 out:
1212         bch2_trans_iter_put(&trans, iter);
1213         bch2_trans_exit(&trans);
1214         bch2_bkey_buf_exit(&k, c);
1215         bio_put(&wbio->wbio.bio);
1216         btree_node_write_done(c, b);
1217         return;
1218 err:
1219         set_btree_node_noevict(b);
1220         bch2_fs_fatal_error(c, "fatal error writing btree node");
1221         goto out;
1222 }
1223
1224 void bch2_btree_write_error_work(struct work_struct *work)
1225 {
1226         struct bch_fs *c = container_of(work, struct bch_fs,
1227                                         btree_write_error_work);
1228         struct bio *bio;
1229
1230         while (1) {
1231                 spin_lock_irq(&c->btree_write_error_lock);
1232                 bio = bio_list_pop(&c->btree_write_error_list);
1233                 spin_unlock_irq(&c->btree_write_error_lock);
1234
1235                 if (!bio)
1236                         break;
1237
1238                 bch2_btree_node_write_error(c,
1239                         container_of(bio, struct btree_write_bio, wbio.bio));
1240         }
1241 }
1242
1243 static void btree_node_write_work(struct work_struct *work)
1244 {
1245         struct btree_write_bio *wbio =
1246                 container_of(work, struct btree_write_bio, work);
1247         struct bch_fs *c        = wbio->wbio.c;
1248         struct btree *b         = wbio->wbio.bio.bi_private;
1249
1250         btree_bounce_free(c,
1251                 wbio->bytes,
1252                 wbio->wbio.used_mempool,
1253                 wbio->data);
1254
1255         if (wbio->wbio.failed.nr) {
1256                 unsigned long flags;
1257
1258                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1259                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1260                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1261
1262                 queue_work(c->wq, &c->btree_write_error_work);
1263                 return;
1264         }
1265
1266         bio_put(&wbio->wbio.bio);
1267         btree_node_write_done(c, b);
1268 }
1269
1270 static void btree_node_write_endio(struct bio *bio)
1271 {
1272         struct bch_write_bio *wbio      = to_wbio(bio);
1273         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1274         struct bch_write_bio *orig      = parent ?: wbio;
1275         struct bch_fs *c                = wbio->c;
1276         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1277         unsigned long flags;
1278
1279         if (wbio->have_ioref)
1280                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1281
1282         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1283                                bch2_blk_status_to_str(bio->bi_status)) ||
1284             bch2_meta_write_fault("btree")) {
1285                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1286                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1287                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1288         }
1289
1290         if (wbio->have_ioref)
1291                 percpu_ref_put(&ca->io_ref);
1292
1293         if (parent) {
1294                 bio_put(bio);
1295                 bio_endio(&parent->bio);
1296         } else {
1297                 struct btree_write_bio *wb =
1298                         container_of(orig, struct btree_write_bio, wbio);
1299
1300                 INIT_WORK(&wb->work, btree_node_write_work);
1301                 queue_work(system_unbound_wq, &wb->work);
1302         }
1303 }
1304
1305 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1306                                    struct bset *i, unsigned sectors)
1307 {
1308         unsigned whiteout_u64s = 0;
1309         int ret;
1310
1311         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1312                 return -1;
1313
1314         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1315                 validate_bset(c, NULL, b, i, sectors, WRITE, false);
1316         if (ret) {
1317                 bch2_inconsistent_error(c);
1318                 dump_stack();
1319         }
1320
1321         return ret;
1322 }
1323
1324 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1325                             enum six_lock_type lock_type_held)
1326 {
1327         struct btree_write_bio *wbio;
1328         struct bset_tree *t;
1329         struct bset *i;
1330         struct btree_node *bn = NULL;
1331         struct btree_node_entry *bne = NULL;
1332         struct bkey_buf k;
1333         struct bch_extent_ptr *ptr;
1334         struct sort_iter sort_iter;
1335         struct nonce nonce;
1336         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1337         u64 seq = 0;
1338         bool used_mempool;
1339         unsigned long old, new;
1340         bool validate_before_checksum = false;
1341         void *data;
1342
1343         bch2_bkey_buf_init(&k);
1344
1345         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1346                 return;
1347
1348         /*
1349          * We may only have a read lock on the btree node - the dirty bit is our
1350          * "lock" against racing with other threads that may be trying to start
1351          * a write, we do a write iff we clear the dirty bit. Since setting the
1352          * dirty bit requires a write lock, we can't race with other threads
1353          * redirtying it:
1354          */
1355         do {
1356                 old = new = READ_ONCE(b->flags);
1357
1358                 if (!(old & (1 << BTREE_NODE_dirty)))
1359                         return;
1360
1361                 if (!btree_node_may_write(b))
1362                         return;
1363
1364                 if (old & (1 << BTREE_NODE_never_write))
1365                         return;
1366
1367                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1368                         btree_node_wait_on_io(b);
1369                         continue;
1370                 }
1371
1372                 new &= ~(1 << BTREE_NODE_dirty);
1373                 new &= ~(1 << BTREE_NODE_need_write);
1374                 new |=  (1 << BTREE_NODE_write_in_flight);
1375                 new |=  (1 << BTREE_NODE_just_written);
1376                 new ^=  (1 << BTREE_NODE_write_idx);
1377         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1378
1379         atomic_dec(&c->btree_cache.dirty);
1380
1381         BUG_ON(btree_node_fake(b));
1382         BUG_ON((b->will_make_reachable != 0) != !b->written);
1383
1384         BUG_ON(b->written >= c->opts.btree_node_size);
1385         BUG_ON(b->written & (c->opts.block_size - 1));
1386         BUG_ON(bset_written(b, btree_bset_last(b)));
1387         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1388         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1389
1390         bch2_sort_whiteouts(c, b);
1391
1392         sort_iter_init(&sort_iter, b);
1393
1394         bytes = !b->written
1395                 ? sizeof(struct btree_node)
1396                 : sizeof(struct btree_node_entry);
1397
1398         bytes += b->whiteout_u64s * sizeof(u64);
1399
1400         for_each_bset(b, t) {
1401                 i = bset(b, t);
1402
1403                 if (bset_written(b, i))
1404                         continue;
1405
1406                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1407                 sort_iter_add(&sort_iter,
1408                               btree_bkey_first(b, t),
1409                               btree_bkey_last(b, t));
1410                 seq = max(seq, le64_to_cpu(i->journal_seq));
1411         }
1412
1413         BUG_ON(b->written && !seq);
1414
1415         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1416         bytes += 8;
1417
1418         data = btree_bounce_alloc(c, bytes, &used_mempool);
1419
1420         if (!b->written) {
1421                 bn = data;
1422                 *bn = *b->data;
1423                 i = &bn->keys;
1424         } else {
1425                 bne = data;
1426                 bne->keys = b->data->keys;
1427                 i = &bne->keys;
1428         }
1429
1430         i->journal_seq  = cpu_to_le64(seq);
1431         i->u64s         = 0;
1432
1433         sort_iter_add(&sort_iter,
1434                       unwritten_whiteouts_start(c, b),
1435                       unwritten_whiteouts_end(c, b));
1436         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1437
1438         b->whiteout_u64s = 0;
1439
1440         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1441         le16_add_cpu(&i->u64s, u64s);
1442
1443         set_needs_whiteout(i, false);
1444
1445         /* do we have data to write? */
1446         if (b->written && !i->u64s)
1447                 goto nowrite;
1448
1449         bytes_to_write = vstruct_end(i) - data;
1450         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1451
1452         memset(data + bytes_to_write, 0,
1453                (sectors_to_write << 9) - bytes_to_write);
1454
1455         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1456         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1457         BUG_ON(i->seq != b->data->keys.seq);
1458
1459         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1460                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1461                 : cpu_to_le16(c->sb.version);
1462         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1463
1464         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1465                 validate_before_checksum = true;
1466
1467         /* validate_bset will be modifying: */
1468         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1469                 validate_before_checksum = true;
1470
1471         /* if we're going to be encrypting, check metadata validity first: */
1472         if (validate_before_checksum &&
1473             validate_bset_for_write(c, b, i, sectors_to_write))
1474                 goto err;
1475
1476         bset_encrypt(c, i, b->written << 9);
1477
1478         nonce = btree_nonce(i, b->written << 9);
1479
1480         if (bn)
1481                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1482         else
1483                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1484
1485         /* if we're not encrypting, check metadata after checksumming: */
1486         if (!validate_before_checksum &&
1487             validate_bset_for_write(c, b, i, sectors_to_write))
1488                 goto err;
1489
1490         /*
1491          * We handle btree write errors by immediately halting the journal -
1492          * after we've done that, we can't issue any subsequent btree writes
1493          * because they might have pointers to new nodes that failed to write.
1494          *
1495          * Furthermore, there's no point in doing any more btree writes because
1496          * with the journal stopped, we're never going to update the journal to
1497          * reflect that those writes were done and the data flushed from the
1498          * journal:
1499          *
1500          * Also on journal error, the pending write may have updates that were
1501          * never journalled (interior nodes, see btree_update_nodes_written()) -
1502          * it's critical that we don't do the write in that case otherwise we
1503          * will have updates visible that weren't in the journal:
1504          *
1505          * Make sure to update b->written so bch2_btree_init_next() doesn't
1506          * break:
1507          */
1508         if (bch2_journal_error(&c->journal) ||
1509             c->opts.nochanges)
1510                 goto err;
1511
1512         trace_btree_write(b, bytes_to_write, sectors_to_write);
1513
1514         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1515                                 buf_pages(data, sectors_to_write << 9),
1516                                 &c->btree_bio),
1517                             struct btree_write_bio, wbio.bio);
1518         wbio_init(&wbio->wbio.bio);
1519         wbio->data                      = data;
1520         wbio->bytes                     = bytes;
1521         wbio->wbio.used_mempool         = used_mempool;
1522         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1523         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1524         wbio->wbio.bio.bi_private       = b;
1525
1526         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1527
1528         /*
1529          * If we're appending to a leaf node, we don't technically need FUA -
1530          * this write just needs to be persisted before the next journal write,
1531          * which will be marked FLUSH|FUA.
1532          *
1533          * Similarly if we're writing a new btree root - the pointer is going to
1534          * be in the next journal entry.
1535          *
1536          * But if we're writing a new btree node (that isn't a root) or
1537          * appending to a non leaf btree node, we need either FUA or a flush
1538          * when we write the parent with the new pointer. FUA is cheaper than a
1539          * flush, and writes appending to leaf nodes aren't blocking anything so
1540          * just make all btree node writes FUA to keep things sane.
1541          */
1542
1543         bch2_bkey_buf_copy(&k, c, &b->key);
1544
1545         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(k.k)), ptr)
1546                 ptr->offset += b->written;
1547
1548         b->written += sectors_to_write;
1549
1550         atomic64_inc(&c->btree_writes_nr);
1551         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
1552
1553         /* XXX: submitting IO with btree locks held: */
1554         bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_btree, k.k);
1555         bch2_bkey_buf_exit(&k, c);
1556         return;
1557 err:
1558         set_btree_node_noevict(b);
1559         b->written += sectors_to_write;
1560 nowrite:
1561         btree_bounce_free(c, bytes, used_mempool, data);
1562         btree_node_write_done(c, b);
1563 }
1564
1565 /*
1566  * Work that must be done with write lock held:
1567  */
1568 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1569 {
1570         bool invalidated_iter = false;
1571         struct btree_node_entry *bne;
1572         struct bset_tree *t;
1573
1574         if (!btree_node_just_written(b))
1575                 return false;
1576
1577         BUG_ON(b->whiteout_u64s);
1578
1579         clear_btree_node_just_written(b);
1580
1581         /*
1582          * Note: immediately after write, bset_written() doesn't work - the
1583          * amount of data we had to write after compaction might have been
1584          * smaller than the offset of the last bset.
1585          *
1586          * However, we know that all bsets have been written here, as long as
1587          * we're still holding the write lock:
1588          */
1589
1590         /*
1591          * XXX: decide if we really want to unconditionally sort down to a
1592          * single bset:
1593          */
1594         if (b->nsets > 1) {
1595                 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1596                 invalidated_iter = true;
1597         } else {
1598                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1599         }
1600
1601         for_each_bset(b, t)
1602                 set_needs_whiteout(bset(b, t), true);
1603
1604         bch2_btree_verify(c, b);
1605
1606         /*
1607          * If later we don't unconditionally sort down to a single bset, we have
1608          * to ensure this is still true:
1609          */
1610         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1611
1612         bne = want_new_bset(c, b);
1613         if (bne)
1614                 bch2_bset_init_next(c, b, bne);
1615
1616         bch2_btree_build_aux_trees(b);
1617
1618         return invalidated_iter;
1619 }
1620
1621 /*
1622  * Use this one if the node is intent locked:
1623  */
1624 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1625                           enum six_lock_type lock_type_held)
1626 {
1627         BUG_ON(lock_type_held == SIX_LOCK_write);
1628
1629         if (lock_type_held == SIX_LOCK_intent ||
1630             six_lock_tryupgrade(&b->c.lock)) {
1631                 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1632
1633                 /* don't cycle lock unnecessarily: */
1634                 if (btree_node_just_written(b) &&
1635                     six_trylock_write(&b->c.lock)) {
1636                         bch2_btree_post_write_cleanup(c, b);
1637                         six_unlock_write(&b->c.lock);
1638                 }
1639
1640                 if (lock_type_held == SIX_LOCK_read)
1641                         six_lock_downgrade(&b->c.lock);
1642         } else {
1643                 __bch2_btree_node_write(c, b, SIX_LOCK_read);
1644         }
1645 }
1646
1647 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
1648 {
1649         struct bucket_table *tbl;
1650         struct rhash_head *pos;
1651         struct btree *b;
1652         unsigned i;
1653 restart:
1654         rcu_read_lock();
1655         for_each_cached_btree(b, c, tbl, i, pos)
1656                 if (test_bit(flag, &b->flags)) {
1657                         rcu_read_unlock();
1658                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
1659                         goto restart;
1660
1661                 }
1662         rcu_read_unlock();
1663 }
1664
1665 void bch2_btree_flush_all_reads(struct bch_fs *c)
1666 {
1667         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
1668 }
1669
1670 void bch2_btree_flush_all_writes(struct bch_fs *c)
1671 {
1672         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
1673 }
1674
1675 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
1676 {
1677         struct bucket_table *tbl;
1678         struct rhash_head *pos;
1679         struct btree *b;
1680         unsigned i;
1681
1682         rcu_read_lock();
1683         for_each_cached_btree(b, c, tbl, i, pos) {
1684                 unsigned long flags = READ_ONCE(b->flags);
1685
1686                 if (!(flags & (1 << BTREE_NODE_dirty)))
1687                         continue;
1688
1689                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
1690                        b,
1691                        (flags & (1 << BTREE_NODE_dirty)) != 0,
1692                        (flags & (1 << BTREE_NODE_need_write)) != 0,
1693                        b->c.level,
1694                        b->written,
1695                        !list_empty_careful(&b->write_blocked),
1696                        b->will_make_reachable != 0,
1697                        b->will_make_reachable & 1);
1698         }
1699         rcu_read_unlock();
1700 }