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_skip_noops(start, end);
  36              k != end;
  37              p = k, k = bkey_next_skip_noops(k, end)) {
  38                 struct bkey l = bkey_unpack_key(b, p);
  39                 struct bkey r = bkey_unpack_key(b, k);
  40
  41                 BUG_ON(bkey_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;
  51              k != vstruct_last(i);
  52              k = bkey_next_skip_noops(k, vstruct_last(i)))
  53                 k->needs_whiteout = v;
  54 }
  55
  56 static void btree_bounce_free(struct bch_fs *c, size_t size,
  57                               bool used_mempool, void *p)
  58 {
  59         if (used_mempool)
  60                 mempool_free(p, &c->btree_bounce_pool);
  61         else
  62                 vpfree(p, size);
  63 }
  64
  65 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
  66                                 bool *used_mempool)
  67 {
  68         unsigned flags = memalloc_nofs_save();
  69         void *p;
  70
  71         BUG_ON(size > btree_bytes(c));
  72
  73         *used_mempool = false;
  74         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
  75         if (!p) {
  76                 *used_mempool = true;
  77                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
  78         }
  79         memalloc_nofs_restore(flags);
  80         return p;
  81 }
  82
  83 static void sort_bkey_ptrs(const struct btree *bt,
  84                            struct bkey_packed **ptrs, unsigned nr)
  85 {
  86         unsigned n = nr, a = nr / 2, b, c, d;
  87
  88         if (!a)
  89                 return;
  90
  91         /* Heap sort: see lib/sort.c: */
  92         while (1) {
  93                 if (a)
  94                         a--;
  95                 else if (--n)
  96                         swap(ptrs[0], ptrs[n]);
  97                 else
  98                         break;
  99
 100                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 101                         b = bch2_bkey_cmp_packed(bt,
 102                                             ptrs[c],
 103                                             ptrs[d]) >= 0 ? c : d;
 104                 if (d == n)
 105                         b = c;
 106
 107                 while (b != a &&
 108                        bch2_bkey_cmp_packed(bt,
 109                                        ptrs[a],
 110                                        ptrs[b]) >= 0)
 111                         b = (b - 1) / 2;
 112                 c = b;
 113                 while (b != a) {
 114                         b = (b - 1) / 2;
 115                         swap(ptrs[b], ptrs[c]);
 116                 }
 117         }
 118 }
 119
 120 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 121 {
 122         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 123         bool used_mempool = false;
 124         size_t bytes = b->whiteout_u64s * sizeof(u64);
 125
 126         if (!b->whiteout_u64s)
 127                 return;
 128
 129         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 130
 131         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 132
 133         for (k = unwritten_whiteouts_start(c, b);
 134              k != unwritten_whiteouts_end(c, b);
 135              k = bkey_next(k))
 136                 *--ptrs = k;
 137
 138         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 139
 140         k = new_whiteouts;
 141
 142         while (ptrs != ptrs_end) {
 143                 bkey_copy(k, *ptrs);
 144                 k = bkey_next(k);
 145                 ptrs++;
 146         }
 147
 148         verify_no_dups(b, new_whiteouts,
 149                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 150
 151         memcpy_u64s(unwritten_whiteouts_start(c, b),
 152                     new_whiteouts, b->whiteout_u64s);
 153
 154         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 155 }
 156
 157 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 158                                 bool compacting, enum compact_mode mode)
 159 {
 160         if (!bset_dead_u64s(b, t))
 161                 return false;
 162
 163         switch (mode) {
 164         case COMPACT_LAZY:
 165                 return should_compact_bset_lazy(b, t) ||
 166                         (compacting && !bset_written(b, bset(b, t)));
 167         case COMPACT_ALL:
 168                 return true;
 169         default:
 170                 BUG();
 171         }
 172 }
 173
 174 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 175 {
 176         struct bset_tree *t;
 177         bool ret = false;
 178
 179         for_each_bset(b, t) {
 180                 struct bset *i = bset(b, t);
 181                 struct bkey_packed *k, *n, *out, *start, *end;
 182                 struct btree_node_entry *src = NULL, *dst = NULL;
 183
 184                 if (t != b->set && !bset_written(b, i)) {
 185                         src = container_of(i, struct btree_node_entry, keys);
 186                         dst = max(write_block(b),
 187                                   (void *) btree_bkey_last(b, t - 1));
 188                 }
 189
 190                 if (src != dst)
 191                         ret = true;
 192
 193                 if (!should_compact_bset(b, t, ret, mode)) {
 194                         if (src != dst) {
 195                                 memmove(dst, src, sizeof(*src) +
 196                                         le16_to_cpu(src->keys.u64s) *
 197                                         sizeof(u64));
 198                                 i = &dst->keys;
 199                                 set_btree_bset(b, t, i);
 200                         }
 201                         continue;
 202                 }
 203
 204                 start   = btree_bkey_first(b, t);
 205                 end     = btree_bkey_last(b, t);
 206
 207                 if (src != dst) {
 208                         memmove(dst, src, sizeof(*src));
 209                         i = &dst->keys;
 210                         set_btree_bset(b, t, i);
 211                 }
 212
 213                 out = i->start;
 214
 215                 for (k = start; k != end; k = n) {
 216                         n = bkey_next_skip_noops(k, end);
 217
 218                         if (!bkey_deleted(k)) {
 219                                 bkey_copy(out, k);
 220                                 out = bkey_next(out);
 221                         } else {
 222                                 BUG_ON(k->needs_whiteout);
 223                         }
 224                 }
 225
 226                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 227                 set_btree_bset_end(b, t);
 228                 bch2_bset_set_no_aux_tree(b, t);
 229                 ret = true;
 230         }
 231
 232         bch2_verify_btree_nr_keys(b);
 233
 234         bch2_btree_build_aux_trees(b);
 235
 236         return ret;
 237 }
 238
 239 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 240                             enum compact_mode mode)
 241 {
 242         return bch2_drop_whiteouts(b, mode);
 243 }
 244
 245 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 246                             struct btree_iter *iter,
 247                             unsigned start_idx,
 248                             unsigned end_idx,
 249                             bool filter_whiteouts)
 250 {
 251         struct btree_node *out;
 252         struct sort_iter sort_iter;
 253         struct bset_tree *t;
 254         struct bset *start_bset = bset(b, &b->set[start_idx]);
 255         bool used_mempool = false;
 256         u64 start_time, seq = 0;
 257         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 258         bool sorting_entire_node = start_idx == 0 &&
 259                 end_idx == b->nsets;
 260
 261         sort_iter_init(&sort_iter, b);
 262
 263         for (t = b->set + start_idx;
 264              t < b->set + end_idx;
 265              t++) {
 266                 u64s += le16_to_cpu(bset(b, t)->u64s);
 267                 sort_iter_add(&sort_iter,
 268                               btree_bkey_first(b, t),
 269                               btree_bkey_last(b, t));
 270         }
 271
 272         bytes = sorting_entire_node
 273                 ? btree_bytes(c)
 274                 : __vstruct_bytes(struct btree_node, u64s);
 275
 276         out = btree_bounce_alloc(c, bytes, &used_mempool);
 277
 278         start_time = local_clock();
 279
 280         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 281
 282         out->keys.u64s = cpu_to_le16(u64s);
 283
 284         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 285
 286         if (sorting_entire_node)
 287                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 288                                        start_time);
 289
 290         /* Make sure we preserve bset journal_seq: */
 291         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 292                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 293         start_bset->journal_seq = cpu_to_le64(seq);
 294
 295         if (sorting_entire_node) {
 296                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 297
 298                 BUG_ON(bytes != btree_bytes(c));
 299
 300                 /*
 301                  * Our temporary buffer is the same size as the btree node's
 302                  * buffer, we can just swap buffers instead of doing a big
 303                  * memcpy()
 304                  */
 305                 *out = *b->data;
 306                 out->keys.u64s = cpu_to_le16(u64s);
 307                 swap(out, b->data);
 308                 set_btree_bset(b, b->set, &b->data->keys);
 309         } else {
 310                 start_bset->u64s = out->keys.u64s;
 311                 memcpy_u64s(start_bset->start,
 312                             out->keys.start,
 313                             le16_to_cpu(out->keys.u64s));
 314         }
 315
 316         for (i = start_idx + 1; i < end_idx; i++)
 317                 b->nr.bset_u64s[start_idx] +=
 318                         b->nr.bset_u64s[i];
 319
 320         b->nsets -= shift;
 321
 322         for (i = start_idx + 1; i < b->nsets; i++) {
 323                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 324                 b->set[i]               = b->set[i + shift];
 325         }
 326
 327         for (i = b->nsets; i < MAX_BSETS; i++)
 328                 b->nr.bset_u64s[i] = 0;
 329
 330         set_btree_bset_end(b, &b->set[start_idx]);
 331         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 332
 333         btree_bounce_free(c, bytes, used_mempool, out);
 334
 335         bch2_verify_btree_nr_keys(b);
 336 }
 337
 338 void bch2_btree_sort_into(struct bch_fs *c,
 339                          struct btree *dst,
 340                          struct btree *src)
 341 {
 342         struct btree_nr_keys nr;
 343         struct btree_node_iter src_iter;
 344         u64 start_time = local_clock();
 345
 346         BUG_ON(dst->nsets != 1);
 347
 348         bch2_bset_set_no_aux_tree(dst, dst->set);
 349
 350         bch2_btree_node_iter_init_from_start(&src_iter, src);
 351
 352         if (btree_node_is_extents(src))
 353                 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
 354                                 src, &src_iter,
 355                                 &dst->format,
 356                                 true);
 357         else
 358                 nr = bch2_sort_repack(btree_bset_first(dst),
 359                                 src, &src_iter,
 360                                 &dst->format,
 361                                 true);
 362
 363         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 364                                start_time);
 365
 366         set_btree_bset_end(dst, dst->set);
 367
 368         dst->nr.live_u64s       += nr.live_u64s;
 369         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 370         dst->nr.packed_keys     += nr.packed_keys;
 371         dst->nr.unpacked_keys   += nr.unpacked_keys;
 372
 373         bch2_verify_btree_nr_keys(dst);
 374 }
 375
 376 #define SORT_CRIT       (4096 / sizeof(u64))
 377
 378 /*
 379  * We're about to add another bset to the btree node, so if there's currently
 380  * too many bsets - sort some of them together:
 381  */
 382 static bool btree_node_compact(struct bch_fs *c, struct btree *b,
 383                                struct btree_iter *iter)
 384 {
 385         unsigned unwritten_idx;
 386         bool ret = false;
 387
 388         for (unwritten_idx = 0;
 389              unwritten_idx < b->nsets;
 390              unwritten_idx++)
 391                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 392                         break;
 393
 394         if (b->nsets - unwritten_idx > 1) {
 395                 btree_node_sort(c, b, iter, unwritten_idx,
 396                                 b->nsets, false);
 397                 ret = true;
 398         }
 399
 400         if (unwritten_idx > 1) {
 401                 btree_node_sort(c, b, iter, 0, unwritten_idx, false);
 402                 ret = true;
 403         }
 404
 405         return ret;
 406 }
 407
 408 void bch2_btree_build_aux_trees(struct btree *b)
 409 {
 410         struct bset_tree *t;
 411
 412         for_each_bset(b, t)
 413                 bch2_bset_build_aux_tree(b, t,
 414                                 !bset_written(b, bset(b, t)) &&
 415                                 t == bset_tree_last(b));
 416 }
 417
 418 /*
 419  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 420  * inserted into
 421  *
 422  * Safe to call if there already is an unwritten bset - will only add a new bset
 423  * if @b doesn't already have one.
 424  *
 425  * Returns true if we sorted (i.e. invalidated iterators
 426  */
 427 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
 428                           struct btree_iter *iter)
 429 {
 430         struct btree_node_entry *bne;
 431         bool did_sort;
 432
 433         EBUG_ON(!(b->c.lock.state.seq & 1));
 434         EBUG_ON(iter && iter->l[b->c.level].b != b);
 435
 436         did_sort = btree_node_compact(c, b, iter);
 437
 438         bne = want_new_bset(c, b);
 439         if (bne)
 440                 bch2_bset_init_next(c, b, bne);
 441
 442         bch2_btree_build_aux_trees(b);
 443
 444         if (iter && did_sort)
 445                 bch2_btree_iter_reinit_node(iter, b);
 446 }
 447
 448 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 449                           struct btree *b)
 450 {
 451         pr_buf(out, "%s level %u/%u\n  ",
 452                bch2_btree_ids[b->c.btree_id],
 453                b->c.level,
 454                c->btree_roots[b->c.btree_id].level);
 455         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 456 }
 457
 458 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 459                           struct bch_dev *ca,
 460                           struct btree *b, struct bset *i,
 461                           unsigned offset, int write)
 462 {
 463         pr_buf(out, "error validating btree node ");
 464         if (write)
 465                 pr_buf(out, "before write ");
 466         if (ca)
 467                 pr_buf(out, "on %s ", ca->name);
 468         pr_buf(out, "at btree ");
 469         btree_pos_to_text(out, c, b);
 470
 471         pr_buf(out, "\n  node offset %u", b->written);
 472         if (i)
 473                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 474 }
 475
 476 enum btree_err_type {
 477         BTREE_ERR_FIXABLE,
 478         BTREE_ERR_WANT_RETRY,
 479         BTREE_ERR_MUST_RETRY,
 480         BTREE_ERR_FATAL,
 481 };
 482
 483 enum btree_validate_ret {
 484         BTREE_RETRY_READ = 64,
 485 };
 486
 487 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 488 ({                                                                      \
 489         __label__ out;                                                  \
 490         char _buf[300];                                                 \
 491         char *_buf2 = _buf;                                             \
 492         struct printbuf out = PBUF(_buf);                               \
 493                                                                         \
 494         _buf2 = kmalloc(4096, GFP_ATOMIC);                              \
 495         if (_buf2)                                                      \
 496                 out = _PBUF(_buf2, 4986);                               \
 497                                                                         \
 498         btree_err_msg(&out, c, ca, b, i, b->written, write);            \
 499         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 500                                                                         \
 501         if (type == BTREE_ERR_FIXABLE &&                                \
 502             write == READ &&                                            \
 503             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 504                 mustfix_fsck_err(c, "%s", _buf2);                       \
 505                 goto out;                                               \
 506         }                                                               \
 507                                                                         \
 508         switch (write) {                                                \
 509         case READ:                                                      \
 510                 bch_err(c, "%s", _buf2);                                        \
 511                                                                         \
 512                 switch (type) {                                         \
 513                 case BTREE_ERR_FIXABLE:                                 \
 514                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 515                         goto fsck_err;                                  \
 516                 case BTREE_ERR_WANT_RETRY:                              \
 517                         if (have_retry) {                               \
 518                                 ret = BTREE_RETRY_READ;                 \
 519                                 goto fsck_err;                          \
 520                         }                                               \
 521                         break;                                          \
 522                 case BTREE_ERR_MUST_RETRY:                              \
 523                         ret = BTREE_RETRY_READ;                         \
 524                         goto fsck_err;                                  \
 525                 case BTREE_ERR_FATAL:                                   \
 526                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 527                         goto fsck_err;                                  \
 528                 }                                                       \
 529                 break;                                                  \
 530         case WRITE:                                                     \
 531                 bch_err(c, "corrupt metadata before write: %s", _buf2); \
 532                                                                         \
 533                 if (bch2_fs_inconsistent(c)) {                          \
 534                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 535                         goto fsck_err;                                  \
 536                 }                                                       \
 537                 break;                                                  \
 538         }                                                               \
 539 out:                                                                    \
 540         if (_buf2 != _buf)                                              \
 541                 kfree(_buf2);                                           \
 542         true;                                                           \
 543 })
 544
 545 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 546
 547 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 548                          struct btree *b, struct bset *i,
 549                          unsigned sectors, int write, bool have_retry)
 550 {
 551         unsigned version = le16_to_cpu(i->version);
 552         const char *err;
 553         char buf1[100];
 554         char buf2[100];
 555         int ret = 0;
 556
 557         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 558                       version < bcachefs_metadata_version_min) ||
 559                      version >= bcachefs_metadata_version_max,
 560                      BTREE_ERR_FATAL, c, ca, b, i,
 561                      "unsupported bset version");
 562
 563         if (btree_err_on(version < c->sb.version_min,
 564                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 565                          "bset version %u older than superblock version_min %u",
 566                          version, c->sb.version_min)) {
 567                 mutex_lock(&c->sb_lock);
 568                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 569                 bch2_write_super(c);
 570                 mutex_unlock(&c->sb_lock);
 571         }
 572
 573         if (btree_err_on(version > c->sb.version,
 574                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 575                          "bset version %u newer than superblock version %u",
 576                          version, c->sb.version)) {
 577                 mutex_lock(&c->sb_lock);
 578                 c->disk_sb.sb->version = cpu_to_le16(version);
 579                 bch2_write_super(c);
 580                 mutex_unlock(&c->sb_lock);
 581         }
 582
 583         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 584                          BTREE_ERR_FIXABLE, c, ca, b, i,
 585                          "bset past end of btree node")) {
 586                 i->u64s = 0;
 587                 return 0;
 588         }
 589
 590         btree_err_on(b->written && !i->u64s,
 591                      BTREE_ERR_FIXABLE, c, ca, b, i,
 592                      "empty bset");
 593
 594         if (!b->written) {
 595                 struct btree_node *bn =
 596                         container_of(i, struct btree_node, keys);
 597                 /* These indicate that we read the wrong btree node: */
 598
 599                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 600                         struct bch_btree_ptr_v2 *bp =
 601                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 602
 603                         /* XXX endianness */
 604                         btree_err_on(bp->seq != bn->keys.seq,
 605                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 606                                      "incorrect sequence number (wrong btree node)");
 607                 }
 608
 609                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 610                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 611                              "incorrect btree id");
 612
 613                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 614                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 615                              "incorrect level");
 616
 617                 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
 618                         u64 *p = (u64 *) &bn->ptr;
 619
 620                         *p = swab64(*p);
 621                 }
 622
 623                 if (!write)
 624                         compat_btree_node(b->c.level, b->c.btree_id, version,
 625                                           BSET_BIG_ENDIAN(i), write, bn);
 626
 627                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 628                         struct bch_btree_ptr_v2 *bp =
 629                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 630
 631                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 632                                 b->data->min_key = bp->min_key;
 633                                 b->data->max_key = b->key.k.p;
 634                         }
 635
 636                         btree_err_on(bkey_cmp(b->data->min_key, bp->min_key),
 637                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 638                                      "incorrect min_key: got %s should be %s",
 639                                      (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
 640                                      (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
 641                 }
 642
 643                 btree_err_on(bkey_cmp(bn->max_key, b->key.k.p),
 644                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 645                              "incorrect max key %s",
 646                              (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
 647
 648                 if (write)
 649                         compat_btree_node(b->c.level, b->c.btree_id, version,
 650                                           BSET_BIG_ENDIAN(i), write, bn);
 651
 652                 err = bch2_bkey_format_validate(&bn->format);
 653                 btree_err_on(err,
 654                              BTREE_ERR_FATAL, c, ca, b, i,
 655                              "invalid bkey format: %s", err);
 656
 657                 compat_bformat(b->c.level, b->c.btree_id, version,
 658                                BSET_BIG_ENDIAN(i), write,
 659                                &bn->format);
 660         }
 661 fsck_err:
 662         return ret;
 663 }
 664
 665 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 666                          struct bset *i, unsigned *whiteout_u64s,
 667                          int write, bool have_retry)
 668 {
 669         unsigned version = le16_to_cpu(i->version);
 670         struct bkey_packed *k, *prev = NULL;
 671         bool seen_non_whiteout = false;
 672         int ret = 0;
 673
 674         if (!BSET_SEPARATE_WHITEOUTS(i)) {
 675                 seen_non_whiteout = true;
 676                 *whiteout_u64s = 0;
 677         }
 678
 679         for (k = i->start;
 680              k != vstruct_last(i);) {
 681                 struct bkey_s u;
 682                 struct bkey tmp;
 683                 const char *invalid;
 684
 685                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 686                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 687                                  "key extends past end of bset")) {
 688                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 689                         break;
 690                 }
 691
 692                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 693                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 694                                  "invalid bkey format %u", k->format)) {
 695                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 696                         memmove_u64s_down(k, bkey_next(k),
 697                                           (u64 *) vstruct_end(i) - (u64 *) k);
 698                         continue;
 699                 }
 700
 701                 /* XXX: validate k->u64s */
 702                 if (!write)
 703                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 704                                     BSET_BIG_ENDIAN(i), write,
 705                                     &b->format, k);
 706
 707                 u = __bkey_disassemble(b, k, &tmp);
 708
 709                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 710                         bch2_bkey_in_btree_node(b, u.s_c) ?:
 711                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 712                 if (invalid) {
 713                         char buf[160];
 714
 715                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 716                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 717                                   "invalid bkey: %s\n%s", invalid, buf);
 718
 719                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 720                         memmove_u64s_down(k, bkey_next(k),
 721                                           (u64 *) vstruct_end(i) - (u64 *) k);
 722                         continue;
 723                 }
 724
 725                 if (write)
 726                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 727                                     BSET_BIG_ENDIAN(i), write,
 728                                     &b->format, k);
 729
 730                 /*
 731                  * with the separate whiteouts thing (used for extents), the
 732                  * second set of keys actually can have whiteouts too, so we
 733                  * can't solely go off bkey_deleted()...
 734                  */
 735
 736                 if (!seen_non_whiteout &&
 737                     (!bkey_deleted(k) ||
 738                      (prev && bkey_iter_cmp(b, prev, k) > 0))) {
 739                         *whiteout_u64s = k->_data - i->_data;
 740                         seen_non_whiteout = true;
 741                 } else if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 742                         char buf1[80];
 743                         char buf2[80];
 744                         struct bkey up = bkey_unpack_key(b, prev);
 745
 746                         bch2_bkey_to_text(&PBUF(buf1), &up);
 747                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 748
 749                         bch2_dump_bset(c, b, i, 0);
 750                         btree_err(BTREE_ERR_FATAL, c, NULL, b, i,
 751                                   "keys out of order: %s > %s",
 752                                   buf1, buf2);
 753                         /* XXX: repair this */
 754                 }
 755
 756                 prev = k;
 757                 k = bkey_next_skip_noops(k, vstruct_last(i));
 758         }
 759 fsck_err:
 760         return ret;
 761 }
 762
 763 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 764                               struct btree *b, bool have_retry)
 765 {
 766         struct btree_node_entry *bne;
 767         struct sort_iter *iter;
 768         struct btree_node *sorted;
 769         struct bkey_packed *k;
 770         struct bch_extent_ptr *ptr;
 771         struct bset *i;
 772         bool used_mempool, blacklisted;
 773         unsigned u64s;
 774         int ret, retry_read = 0, write = READ;
 775
 776         b->version_ondisk = U16_MAX;
 777
 778         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 779         sort_iter_init(iter, b);
 780         iter->size = (btree_blocks(c) + 1) * 2;
 781
 782         if (bch2_meta_read_fault("btree"))
 783                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 784                           "dynamic fault");
 785
 786         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 787                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 788                      "bad magic");
 789
 790         btree_err_on(!b->data->keys.seq,
 791                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 792                      "bad btree header");
 793
 794         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 795                 struct bch_btree_ptr_v2 *bp =
 796                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 797
 798                 btree_err_on(b->data->keys.seq != bp->seq,
 799                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 800                              "got wrong btree node (seq %llx want %llx)",
 801                              b->data->keys.seq, bp->seq);
 802         }
 803
 804         while (b->written < c->opts.btree_node_size) {
 805                 unsigned sectors, whiteout_u64s = 0;
 806                 struct nonce nonce;
 807                 struct bch_csum csum;
 808                 bool first = !b->written;
 809
 810                 if (!b->written) {
 811                         i = &b->data->keys;
 812
 813                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 814                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 815                                      "unknown checksum type %llu",
 816                                      BSET_CSUM_TYPE(i));
 817
 818                         nonce = btree_nonce(i, b->written << 9);
 819                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 820
 821                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 822                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 823                                      "invalid checksum");
 824
 825                         bset_encrypt(c, i, b->written << 9);
 826
 827                         btree_err_on(btree_node_is_extents(b) &&
 828                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 829                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 830                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 831
 832                         sectors = vstruct_sectors(b->data, c->block_bits);
 833                 } else {
 834                         bne = write_block(b);
 835                         i = &bne->keys;
 836
 837                         if (i->seq != b->data->keys.seq)
 838                                 break;
 839
 840                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 841                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 842                                      "unknown checksum type %llu",
 843                                      BSET_CSUM_TYPE(i));
 844
 845                         nonce = btree_nonce(i, b->written << 9);
 846                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 847
 848                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 849                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 850                                      "invalid checksum");
 851
 852                         bset_encrypt(c, i, b->written << 9);
 853
 854                         sectors = vstruct_sectors(bne, c->block_bits);
 855                 }
 856
 857                 b->version_ondisk = min(b->version_ondisk,
 858                                         le16_to_cpu(i->version));
 859
 860                 ret = validate_bset(c, ca, b, i, sectors,
 861                                     READ, have_retry);
 862                 if (ret)
 863                         goto fsck_err;
 864
 865                 if (!b->written)
 866                         btree_node_set_format(b, b->data->format);
 867
 868                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 869                                     READ, have_retry);
 870                 if (ret)
 871                         goto fsck_err;
 872
 873                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 874
 875                 b->written += sectors;
 876
 877                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 878                                         le64_to_cpu(i->journal_seq),
 879                                         true);
 880
 881                 btree_err_on(blacklisted && first,
 882                              BTREE_ERR_FIXABLE, c, ca, b, i,
 883                              "first btree node bset has blacklisted journal seq");
 884                 if (blacklisted && !first)
 885                         continue;
 886
 887                 sort_iter_add(iter, i->start,
 888                               vstruct_idx(i, whiteout_u64s));
 889
 890                 sort_iter_add(iter,
 891                               vstruct_idx(i, whiteout_u64s),
 892                               vstruct_last(i));
 893         }
 894
 895         for (bne = write_block(b);
 896              bset_byte_offset(b, bne) < btree_bytes(c);
 897              bne = (void *) bne + block_bytes(c))
 898                 btree_err_on(bne->keys.seq == b->data->keys.seq,
 899                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
 900                              "found bset signature after last bset");
 901
 902         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
 903         sorted->keys.u64s = 0;
 904
 905         set_btree_bset(b, b->set, &b->data->keys);
 906
 907         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
 908
 909         u64s = le16_to_cpu(sorted->keys.u64s);
 910         *sorted = *b->data;
 911         sorted->keys.u64s = cpu_to_le16(u64s);
 912         swap(sorted, b->data);
 913         set_btree_bset(b, b->set, &b->data->keys);
 914         b->nsets = 1;
 915
 916         BUG_ON(b->nr.live_u64s != u64s);
 917
 918         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
 919
 920         i = &b->data->keys;
 921         for (k = i->start; k != vstruct_last(i);) {
 922                 struct bkey tmp;
 923                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 924                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
 925
 926                 if (invalid ||
 927                     (bch2_inject_invalid_keys &&
 928                      !bversion_cmp(u.k->version, MAX_VERSION))) {
 929                         char buf[160];
 930
 931                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 932                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 933                                   "invalid bkey %s: %s", buf, invalid);
 934
 935                         btree_keys_account_key_drop(&b->nr, 0, k);
 936
 937                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 938                         memmove_u64s_down(k, bkey_next(k),
 939                                           (u64 *) vstruct_end(i) - (u64 *) k);
 940                         set_btree_bset_end(b, b->set);
 941                         continue;
 942                 }
 943
 944                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
 945                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
 946
 947                         bp.v->mem_ptr = 0;
 948                 }
 949
 950                 k = bkey_next_skip_noops(k, vstruct_last(i));
 951         }
 952
 953         bch2_bset_build_aux_tree(b, b->set, false);
 954
 955         set_needs_whiteout(btree_bset_first(b), true);
 956
 957         btree_node_reset_sib_u64s(b);
 958
 959         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
 960                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 961
 962                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
 963                         set_btree_node_need_rewrite(b);
 964         }
 965 out:
 966         mempool_free(iter, &c->fill_iter);
 967         return retry_read;
 968 fsck_err:
 969         if (ret == BTREE_RETRY_READ) {
 970                 retry_read = 1;
 971         } else {
 972                 bch2_inconsistent_error(c);
 973                 set_btree_node_read_error(b);
 974         }
 975         goto out;
 976 }
 977
 978 static void btree_node_read_work(struct work_struct *work)
 979 {
 980         struct btree_read_bio *rb =
 981                 container_of(work, struct btree_read_bio, work);
 982         struct bch_fs *c        = rb->c;
 983         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
 984         struct btree *b         = rb->bio.bi_private;
 985         struct bio *bio         = &rb->bio;
 986         struct bch_io_failures failed = { .nr = 0 };
 987         char buf[200];
 988         struct printbuf out;
 989         bool can_retry;
 990
 991         goto start;
 992         while (1) {
 993                 bch_info(c, "retrying read");
 994                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
 995                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
 996                 bio_reset(bio);
 997                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
 998                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
 999                 bio->bi_iter.bi_size    = btree_bytes(c);
1000
1001                 if (rb->have_ioref) {
1002                         bio_set_dev(bio, ca->disk_sb.bdev);
1003                         submit_bio_wait(bio);
1004                 } else {
1005                         bio->bi_status = BLK_STS_REMOVED;
1006                 }
1007 start:
1008                 out = PBUF(buf);
1009                 btree_pos_to_text(&out, c, b);
1010                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1011                                    bch2_blk_status_to_str(bio->bi_status), buf);
1012                 if (rb->have_ioref)
1013                         percpu_ref_put(&ca->io_ref);
1014                 rb->have_ioref = false;
1015
1016                 bch2_mark_io_failure(&failed, &rb->pick);
1017
1018                 can_retry = bch2_bkey_pick_read_device(c,
1019                                 bkey_i_to_s_c(&b->key),
1020                                 &failed, &rb->pick) > 0;
1021
1022                 if (!bio->bi_status &&
1023                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1024                         break;
1025
1026                 if (!can_retry) {
1027                         set_btree_node_read_error(b);
1028                         break;
1029                 }
1030         }
1031
1032         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1033                                rb->start_time);
1034         bio_put(&rb->bio);
1035         clear_btree_node_read_in_flight(b);
1036         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1037 }
1038
1039 static void btree_node_read_endio(struct bio *bio)
1040 {
1041         struct btree_read_bio *rb =
1042                 container_of(bio, struct btree_read_bio, bio);
1043         struct bch_fs *c        = rb->c;
1044
1045         if (rb->have_ioref) {
1046                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1047                 bch2_latency_acct(ca, rb->start_time, READ);
1048         }
1049
1050         queue_work(system_unbound_wq, &rb->work);
1051 }
1052
1053 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1054                           bool sync)
1055 {
1056         struct extent_ptr_decoded pick;
1057         struct btree_read_bio *rb;
1058         struct bch_dev *ca;
1059         struct bio *bio;
1060         int ret;
1061
1062         trace_btree_read(c, b);
1063
1064         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1065                                          NULL, &pick);
1066         if (bch2_fs_fatal_err_on(ret <= 0, c,
1067                         "btree node read error: no device to read from")) {
1068                 set_btree_node_read_error(b);
1069                 return;
1070         }
1071
1072         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1073
1074         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1075                                                    btree_bytes(c)),
1076                                &c->btree_bio);
1077         rb = container_of(bio, struct btree_read_bio, bio);
1078         rb->c                   = c;
1079         rb->start_time          = local_clock();
1080         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1081         rb->pick                = pick;
1082         INIT_WORK(&rb->work, btree_node_read_work);
1083         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1084         bio->bi_iter.bi_sector  = pick.ptr.offset;
1085         bio->bi_end_io          = btree_node_read_endio;
1086         bio->bi_private         = b;
1087         bch2_bio_map(bio, b->data, btree_bytes(c));
1088
1089         set_btree_node_read_in_flight(b);
1090
1091         if (rb->have_ioref) {
1092                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1093                              bio_sectors(bio));
1094                 bio_set_dev(bio, ca->disk_sb.bdev);
1095
1096                 if (sync) {
1097                         submit_bio_wait(bio);
1098
1099                         bio->bi_private = b;
1100                         btree_node_read_work(&rb->work);
1101                 } else {
1102                         submit_bio(bio);
1103                 }
1104         } else {
1105                 bio->bi_status = BLK_STS_REMOVED;
1106
1107                 if (sync)
1108                         btree_node_read_work(&rb->work);
1109                 else
1110                         queue_work(system_unbound_wq, &rb->work);
1111
1112         }
1113 }
1114
1115 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1116                         const struct bkey_i *k, unsigned level)
1117 {
1118         struct closure cl;
1119         struct btree *b;
1120         int ret;
1121
1122         closure_init_stack(&cl);
1123
1124         do {
1125                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1126                 closure_sync(&cl);
1127         } while (ret);
1128
1129         b = bch2_btree_node_mem_alloc(c);
1130         bch2_btree_cache_cannibalize_unlock(c);
1131
1132         BUG_ON(IS_ERR(b));
1133
1134         bkey_copy(&b->key, k);
1135         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1136
1137         bch2_btree_node_read(c, b, true);
1138
1139         if (btree_node_read_error(b)) {
1140                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1141
1142                 mutex_lock(&c->btree_cache.lock);
1143                 list_move(&b->list, &c->btree_cache.freeable);
1144                 mutex_unlock(&c->btree_cache.lock);
1145
1146                 ret = -EIO;
1147                 goto err;
1148         }
1149
1150         bch2_btree_set_root_for_read(c, b);
1151 err:
1152         six_unlock_write(&b->c.lock);
1153         six_unlock_intent(&b->c.lock);
1154
1155         return ret;
1156 }
1157
1158 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1159                               struct btree_write *w)
1160 {
1161         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1162
1163         do {
1164                 old = new = v;
1165                 if (!(old & 1))
1166                         break;
1167
1168                 new &= ~1UL;
1169         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1170
1171         if (old & 1)
1172                 closure_put(&((struct btree_update *) new)->cl);
1173
1174         bch2_journal_pin_drop(&c->journal, &w->journal);
1175 }
1176
1177 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1178 {
1179         struct btree_write *w = btree_prev_write(b);
1180
1181         bch2_btree_complete_write(c, b, w);
1182         btree_node_io_unlock(b);
1183 }
1184
1185 static void bch2_btree_node_write_error(struct bch_fs *c,
1186                                         struct btree_write_bio *wbio)
1187 {
1188         struct btree *b         = wbio->wbio.bio.bi_private;
1189         struct bkey_buf k;
1190         struct bch_extent_ptr *ptr;
1191         struct btree_trans trans;
1192         struct btree_iter *iter;
1193         int ret;
1194
1195         bch2_bkey_buf_init(&k);
1196         bch2_trans_init(&trans, c, 0, 0);
1197
1198         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1199                                         BTREE_MAX_DEPTH, b->c.level, 0);
1200 retry:
1201         ret = bch2_btree_iter_traverse(iter);
1202         if (ret)
1203                 goto err;
1204
1205         /* has node been freed? */
1206         if (iter->l[b->c.level].b != b) {
1207                 /* node has been freed: */
1208                 BUG_ON(!btree_node_dying(b));
1209                 goto out;
1210         }
1211
1212         BUG_ON(!btree_node_hashed(b));
1213
1214         bch2_bkey_buf_copy(&k, c, &b->key);
1215
1216         bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1217                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1218
1219         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1220                 goto err;
1221
1222         ret = bch2_btree_node_update_key(c, iter, b, k.k);
1223         if (ret == -EINTR)
1224                 goto retry;
1225         if (ret)
1226                 goto err;
1227 out:
1228         bch2_trans_iter_put(&trans, iter);
1229         bch2_trans_exit(&trans);
1230         bch2_bkey_buf_exit(&k, c);
1231         bio_put(&wbio->wbio.bio);
1232         btree_node_write_done(c, b);
1233         return;
1234 err:
1235         set_btree_node_noevict(b);
1236         bch2_fs_fatal_error(c, "fatal error writing btree node");
1237         goto out;
1238 }
1239
1240 void bch2_btree_write_error_work(struct work_struct *work)
1241 {
1242         struct bch_fs *c = container_of(work, struct bch_fs,
1243                                         btree_write_error_work);
1244         struct bio *bio;
1245
1246         while (1) {
1247                 spin_lock_irq(&c->btree_write_error_lock);
1248                 bio = bio_list_pop(&c->btree_write_error_list);
1249                 spin_unlock_irq(&c->btree_write_error_lock);
1250
1251                 if (!bio)
1252                         break;
1253
1254                 bch2_btree_node_write_error(c,
1255                         container_of(bio, struct btree_write_bio, wbio.bio));
1256         }
1257 }
1258
1259 static void btree_node_write_work(struct work_struct *work)
1260 {
1261         struct btree_write_bio *wbio =
1262                 container_of(work, struct btree_write_bio, work);
1263         struct bch_fs *c        = wbio->wbio.c;
1264         struct btree *b         = wbio->wbio.bio.bi_private;
1265
1266         btree_bounce_free(c,
1267                 wbio->bytes,
1268                 wbio->wbio.used_mempool,
1269                 wbio->data);
1270
1271         if (wbio->wbio.failed.nr) {
1272                 unsigned long flags;
1273
1274                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1275                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1276                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1277
1278                 queue_work(c->wq, &c->btree_write_error_work);
1279                 return;
1280         }
1281
1282         bio_put(&wbio->wbio.bio);
1283         btree_node_write_done(c, b);
1284 }
1285
1286 static void btree_node_write_endio(struct bio *bio)
1287 {
1288         struct bch_write_bio *wbio      = to_wbio(bio);
1289         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1290         struct bch_write_bio *orig      = parent ?: wbio;
1291         struct bch_fs *c                = wbio->c;
1292         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1293         unsigned long flags;
1294
1295         if (wbio->have_ioref)
1296                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1297
1298         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1299                                bch2_blk_status_to_str(bio->bi_status)) ||
1300             bch2_meta_write_fault("btree")) {
1301                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1302                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1303                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1304         }
1305
1306         if (wbio->have_ioref)
1307                 percpu_ref_put(&ca->io_ref);
1308
1309         if (parent) {
1310                 bio_put(bio);
1311                 bio_endio(&parent->bio);
1312         } else {
1313                 struct btree_write_bio *wb =
1314                         container_of(orig, struct btree_write_bio, wbio);
1315
1316                 INIT_WORK(&wb->work, btree_node_write_work);
1317                 queue_work(system_unbound_wq, &wb->work);
1318         }
1319 }
1320
1321 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1322                                    struct bset *i, unsigned sectors)
1323 {
1324         unsigned whiteout_u64s = 0;
1325         int ret;
1326
1327         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1328                 return -1;
1329
1330         ret = validate_bset(c, NULL, b, i, sectors, WRITE, false) ?:
1331                 validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false);
1332         if (ret) {
1333                 bch2_inconsistent_error(c);
1334                 dump_stack();
1335         }
1336
1337         return ret;
1338 }
1339
1340 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1341                             enum six_lock_type lock_type_held)
1342 {
1343         struct btree_write_bio *wbio;
1344         struct bset_tree *t;
1345         struct bset *i;
1346         struct btree_node *bn = NULL;
1347         struct btree_node_entry *bne = NULL;
1348         struct bkey_buf k;
1349         struct bch_extent_ptr *ptr;
1350         struct sort_iter sort_iter;
1351         struct nonce nonce;
1352         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1353         u64 seq = 0;
1354         bool used_mempool;
1355         unsigned long old, new;
1356         bool validate_before_checksum = false;
1357         void *data;
1358
1359         bch2_bkey_buf_init(&k);
1360
1361         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1362                 return;
1363
1364         /*
1365          * We may only have a read lock on the btree node - the dirty bit is our
1366          * "lock" against racing with other threads that may be trying to start
1367          * a write, we do a write iff we clear the dirty bit. Since setting the
1368          * dirty bit requires a write lock, we can't race with other threads
1369          * redirtying it:
1370          */
1371         do {
1372                 old = new = READ_ONCE(b->flags);
1373
1374                 if (!(old & (1 << BTREE_NODE_dirty)))
1375                         return;
1376
1377                 if (!btree_node_may_write(b))
1378                         return;
1379
1380                 if (old & (1 << BTREE_NODE_never_write))
1381                         return;
1382
1383                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1384                         btree_node_wait_on_io(b);
1385                         continue;
1386                 }
1387
1388                 new &= ~(1 << BTREE_NODE_dirty);
1389                 new &= ~(1 << BTREE_NODE_need_write);
1390                 new |=  (1 << BTREE_NODE_write_in_flight);
1391                 new |=  (1 << BTREE_NODE_just_written);
1392                 new ^=  (1 << BTREE_NODE_write_idx);
1393         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1394
1395         atomic_dec(&c->btree_cache.dirty);
1396
1397         BUG_ON(btree_node_fake(b));
1398         BUG_ON((b->will_make_reachable != 0) != !b->written);
1399
1400         BUG_ON(b->written >= c->opts.btree_node_size);
1401         BUG_ON(b->written & (c->opts.block_size - 1));
1402         BUG_ON(bset_written(b, btree_bset_last(b)));
1403         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1404         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1405
1406         bch2_sort_whiteouts(c, b);
1407
1408         sort_iter_init(&sort_iter, b);
1409
1410         bytes = !b->written
1411                 ? sizeof(struct btree_node)
1412                 : sizeof(struct btree_node_entry);
1413
1414         bytes += b->whiteout_u64s * sizeof(u64);
1415
1416         for_each_bset(b, t) {
1417                 i = bset(b, t);
1418
1419                 if (bset_written(b, i))
1420                         continue;
1421
1422                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1423                 sort_iter_add(&sort_iter,
1424                               btree_bkey_first(b, t),
1425                               btree_bkey_last(b, t));
1426                 seq = max(seq, le64_to_cpu(i->journal_seq));
1427         }
1428
1429         BUG_ON(b->written && !seq);
1430
1431         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1432         bytes += 8;
1433
1434         data = btree_bounce_alloc(c, bytes, &used_mempool);
1435
1436         if (!b->written) {
1437                 bn = data;
1438                 *bn = *b->data;
1439                 i = &bn->keys;
1440         } else {
1441                 bne = data;
1442                 bne->keys = b->data->keys;
1443                 i = &bne->keys;
1444         }
1445
1446         i->journal_seq  = cpu_to_le64(seq);
1447         i->u64s         = 0;
1448
1449         sort_iter_add(&sort_iter,
1450                       unwritten_whiteouts_start(c, b),
1451                       unwritten_whiteouts_end(c, b));
1452         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1453
1454         b->whiteout_u64s = 0;
1455
1456         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1457         le16_add_cpu(&i->u64s, u64s);
1458
1459         set_needs_whiteout(i, false);
1460
1461         /* do we have data to write? */
1462         if (b->written && !i->u64s)
1463                 goto nowrite;
1464
1465         bytes_to_write = vstruct_end(i) - data;
1466         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1467
1468         memset(data + bytes_to_write, 0,
1469                (sectors_to_write << 9) - bytes_to_write);
1470
1471         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1472         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1473         BUG_ON(i->seq != b->data->keys.seq);
1474
1475         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1476                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1477                 : cpu_to_le16(c->sb.version);
1478         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1479
1480         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1481                 validate_before_checksum = true;
1482
1483         /* validate_bset will be modifying: */
1484         if (le16_to_cpu(i->version) <= bcachefs_metadata_version_inode_btree_change)
1485                 validate_before_checksum = true;
1486
1487         /* if we're going to be encrypting, check metadata validity first: */
1488         if (validate_before_checksum &&
1489             validate_bset_for_write(c, b, i, sectors_to_write))
1490                 goto err;
1491
1492         bset_encrypt(c, i, b->written << 9);
1493
1494         nonce = btree_nonce(i, b->written << 9);
1495
1496         if (bn)
1497                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1498         else
1499                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1500
1501         /* if we're not encrypting, check metadata after checksumming: */
1502         if (!validate_before_checksum &&
1503             validate_bset_for_write(c, b, i, sectors_to_write))
1504                 goto err;
1505
1506         /*
1507          * We handle btree write errors by immediately halting the journal -
1508          * after we've done that, we can't issue any subsequent btree writes
1509          * because they might have pointers to new nodes that failed to write.
1510          *
1511          * Furthermore, there's no point in doing any more btree writes because
1512          * with the journal stopped, we're never going to update the journal to
1513          * reflect that those writes were done and the data flushed from the
1514          * journal:
1515          *
1516          * Also on journal error, the pending write may have updates that were
1517          * never journalled (interior nodes, see btree_update_nodes_written()) -
1518          * it's critical that we don't do the write in that case otherwise we
1519          * will have updates visible that weren't in the journal:
1520          *
1521          * Make sure to update b->written so bch2_btree_init_next() doesn't
1522          * break:
1523          */
1524         if (bch2_journal_error(&c->journal) ||
1525             c->opts.nochanges)
1526                 goto err;
1527
1528         trace_btree_write(b, bytes_to_write, sectors_to_write);
1529
1530         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1531                                 buf_pages(data, sectors_to_write << 9),
1532                                 &c->btree_bio),
1533                             struct btree_write_bio, wbio.bio);
1534         wbio_init(&wbio->wbio.bio);
1535         wbio->data                      = data;
1536         wbio->bytes                     = bytes;
1537         wbio->wbio.used_mempool         = used_mempool;
1538         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1539         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1540         wbio->wbio.bio.bi_private       = b;
1541
1542         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1543
1544         /*
1545          * If we're appending to a leaf node, we don't technically need FUA -
1546          * this write just needs to be persisted before the next journal write,
1547          * which will be marked FLUSH|FUA.
1548          *
1549          * Similarly if we're writing a new btree root - the pointer is going to
1550          * be in the next journal entry.
1551          *
1552          * But if we're writing a new btree node (that isn't a root) or
1553          * appending to a non leaf btree node, we need either FUA or a flush
1554          * when we write the parent with the new pointer. FUA is cheaper than a
1555          * flush, and writes appending to leaf nodes aren't blocking anything so
1556          * just make all btree node writes FUA to keep things sane.
1557          */
1558
1559         bch2_bkey_buf_copy(&k, c, &b->key);
1560
1561         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(k.k)), ptr)
1562                 ptr->offset += b->written;
1563
1564         b->written += sectors_to_write;
1565
1566         /* XXX: submitting IO with btree locks held: */
1567         bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_btree, k.k);
1568         bch2_bkey_buf_exit(&k, c);
1569         return;
1570 err:
1571         set_btree_node_noevict(b);
1572         b->written += sectors_to_write;
1573 nowrite:
1574         btree_bounce_free(c, bytes, used_mempool, data);
1575         btree_node_write_done(c, b);
1576 }
1577
1578 /*
1579  * Work that must be done with write lock held:
1580  */
1581 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1582 {
1583         bool invalidated_iter = false;
1584         struct btree_node_entry *bne;
1585         struct bset_tree *t;
1586
1587         if (!btree_node_just_written(b))
1588                 return false;
1589
1590         BUG_ON(b->whiteout_u64s);
1591
1592         clear_btree_node_just_written(b);
1593
1594         /*
1595          * Note: immediately after write, bset_written() doesn't work - the
1596          * amount of data we had to write after compaction might have been
1597          * smaller than the offset of the last bset.
1598          *
1599          * However, we know that all bsets have been written here, as long as
1600          * we're still holding the write lock:
1601          */
1602
1603         /*
1604          * XXX: decide if we really want to unconditionally sort down to a
1605          * single bset:
1606          */
1607         if (b->nsets > 1) {
1608                 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1609                 invalidated_iter = true;
1610         } else {
1611                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1612         }
1613
1614         for_each_bset(b, t)
1615                 set_needs_whiteout(bset(b, t), true);
1616
1617         bch2_btree_verify(c, b);
1618
1619         /*
1620          * If later we don't unconditionally sort down to a single bset, we have
1621          * to ensure this is still true:
1622          */
1623         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1624
1625         bne = want_new_bset(c, b);
1626         if (bne)
1627                 bch2_bset_init_next(c, b, bne);
1628
1629         bch2_btree_build_aux_trees(b);
1630
1631         return invalidated_iter;
1632 }
1633
1634 /*
1635  * Use this one if the node is intent locked:
1636  */
1637 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1638                           enum six_lock_type lock_type_held)
1639 {
1640         BUG_ON(lock_type_held == SIX_LOCK_write);
1641
1642         if (lock_type_held == SIX_LOCK_intent ||
1643             six_lock_tryupgrade(&b->c.lock)) {
1644                 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1645
1646                 /* don't cycle lock unnecessarily: */
1647                 if (btree_node_just_written(b) &&
1648                     six_trylock_write(&b->c.lock)) {
1649                         bch2_btree_post_write_cleanup(c, b);
1650                         six_unlock_write(&b->c.lock);
1651                 }
1652
1653                 if (lock_type_held == SIX_LOCK_read)
1654                         six_lock_downgrade(&b->c.lock);
1655         } else {
1656                 __bch2_btree_node_write(c, b, SIX_LOCK_read);
1657         }
1658 }
1659
1660 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
1661 {
1662         struct bucket_table *tbl;
1663         struct rhash_head *pos;
1664         struct btree *b;
1665         unsigned i;
1666 restart:
1667         rcu_read_lock();
1668         for_each_cached_btree(b, c, tbl, i, pos)
1669                 if (test_bit(flag, &b->flags)) {
1670                         rcu_read_unlock();
1671                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
1672                         goto restart;
1673
1674                 }
1675         rcu_read_unlock();
1676 }
1677
1678 void bch2_btree_flush_all_reads(struct bch_fs *c)
1679 {
1680         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
1681 }
1682
1683 void bch2_btree_flush_all_writes(struct bch_fs *c)
1684 {
1685         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
1686 }
1687
1688 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
1689 {
1690         struct bucket_table *tbl;
1691         struct rhash_head *pos;
1692         struct btree *b;
1693         unsigned i;
1694
1695         rcu_read_lock();
1696         for_each_cached_btree(b, c, tbl, i, pos) {
1697                 unsigned long flags = READ_ONCE(b->flags);
1698
1699                 if (!(flags & (1 << BTREE_NODE_dirty)))
1700                         continue;
1701
1702                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
1703                        b,
1704                        (flags & (1 << BTREE_NODE_dirty)) != 0,
1705                        (flags & (1 << BTREE_NODE_need_write)) != 0,
1706                        b->c.level,
1707                        b->written,
1708                        !list_empty_careful(&b->write_blocked),
1709                        b->will_make_reachable != 0,
1710                        b->will_make_reachable & 1);
1711         }
1712         rcu_read_unlock();
1713 }