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         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 582                          BTREE_ERR_FIXABLE, c, ca, b, i,
 583                          "bset past end of btree node")) {
 584                 i->u64s = 0;
 585                 return 0;
 586         }
 587
 588         btree_err_on(b->written && !i->u64s,
 589                      BTREE_ERR_FIXABLE, c, ca, b, i,
 590                      "empty bset");
 591
 592         if (!b->written) {
 593                 struct btree_node *bn =
 594                         container_of(i, struct btree_node, keys);
 595                 /* These indicate that we read the wrong btree node: */
 596
 597                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 598                         struct bch_btree_ptr_v2 *bp =
 599                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 600
 601                         /* XXX endianness */
 602                         btree_err_on(bp->seq != bn->keys.seq,
 603                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 604                                      "incorrect sequence number (wrong btree node)");
 605                 }
 606
 607                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 608                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 609                              "incorrect btree id");
 610
 611                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 612                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 613                              "incorrect level");
 614
 615                 if (!write)
 616                         compat_btree_node(b->c.level, b->c.btree_id, version,
 617                                           BSET_BIG_ENDIAN(i), write, bn);
 618
 619                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 620                         struct bch_btree_ptr_v2 *bp =
 621                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 622
 623                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 624                                 b->data->min_key = bp->min_key;
 625                                 b->data->max_key = b->key.k.p;
 626                         }
 627
 628                         btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
 629                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 630                                      "incorrect min_key: got %s should be %s",
 631                                      (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
 632                                      (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
 633                 }
 634
 635                 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
 636                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 637                              "incorrect max key %s",
 638                              (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
 639
 640                 if (write)
 641                         compat_btree_node(b->c.level, b->c.btree_id, version,
 642                                           BSET_BIG_ENDIAN(i), write, bn);
 643
 644                 err = bch2_bkey_format_validate(&bn->format);
 645                 btree_err_on(err,
 646                              BTREE_ERR_FATAL, c, ca, b, i,
 647                              "invalid bkey format: %s", err);
 648
 649                 compat_bformat(b->c.level, b->c.btree_id, version,
 650                                BSET_BIG_ENDIAN(i), write,
 651                                &bn->format);
 652         }
 653 fsck_err:
 654         return ret;
 655 }
 656
 657 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 658                          struct bset *i, unsigned *whiteout_u64s,
 659                          int write, bool have_retry)
 660 {
 661         unsigned version = le16_to_cpu(i->version);
 662         struct bkey_packed *k, *prev = NULL;
 663         bool seen_non_whiteout = false;
 664         int ret = 0;
 665
 666         if (!BSET_SEPARATE_WHITEOUTS(i)) {
 667                 seen_non_whiteout = true;
 668                 *whiteout_u64s = 0;
 669         }
 670
 671         for (k = i->start;
 672              k != vstruct_last(i);) {
 673                 struct bkey_s u;
 674                 struct bkey tmp;
 675                 const char *invalid;
 676
 677                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 678                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 679                                  "key extends past end of bset")) {
 680                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 681                         break;
 682                 }
 683
 684                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 685                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 686                                  "invalid bkey format %u", k->format)) {
 687                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 688                         memmove_u64s_down(k, bkey_next(k),
 689                                           (u64 *) vstruct_end(i) - (u64 *) k);
 690                         continue;
 691                 }
 692
 693                 /* XXX: validate k->u64s */
 694                 if (!write)
 695                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 696                                     BSET_BIG_ENDIAN(i), write,
 697                                     &b->format, k);
 698
 699                 u = __bkey_disassemble(b, k, &tmp);
 700
 701                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 702                         bch2_bkey_in_btree_node(b, u.s_c) ?:
 703                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 704                 if (invalid) {
 705                         char buf[160];
 706
 707                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 708                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 709                                   "invalid bkey: %s\n%s", invalid, buf);
 710
 711                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 712                         memmove_u64s_down(k, bkey_next(k),
 713                                           (u64 *) vstruct_end(i) - (u64 *) k);
 714                         continue;
 715                 }
 716
 717                 if (write)
 718                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 719                                     BSET_BIG_ENDIAN(i), write,
 720                                     &b->format, k);
 721
 722                 /*
 723                  * with the separate whiteouts thing (used for extents), the
 724                  * second set of keys actually can have whiteouts too, so we
 725                  * can't solely go off bkey_deleted()...
 726                  */
 727
 728                 if (!seen_non_whiteout &&
 729                     (!bkey_deleted(k) ||
 730                      (prev && bkey_iter_cmp(b, prev, k) > 0))) {
 731                         *whiteout_u64s = k->_data - i->_data;
 732                         seen_non_whiteout = true;
 733                 } else if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 734                         char buf1[80];
 735                         char buf2[80];
 736                         struct bkey up = bkey_unpack_key(b, prev);
 737
 738                         bch2_bkey_to_text(&PBUF(buf1), &up);
 739                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 740
 741                         bch2_dump_bset(c, b, i, 0);
 742                         btree_err(BTREE_ERR_FATAL, c, NULL, b, i,
 743                                   "keys out of order: %s > %s",
 744                                   buf1, buf2);
 745                         /* XXX: repair this */
 746                 }
 747
 748                 prev = k;
 749                 k = bkey_next(k);
 750         }
 751 fsck_err:
 752         return ret;
 753 }
 754
 755 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 756                               struct btree *b, bool have_retry)
 757 {
 758         struct btree_node_entry *bne;
 759         struct sort_iter *iter;
 760         struct btree_node *sorted;
 761         struct bkey_packed *k;
 762         struct bch_extent_ptr *ptr;
 763         struct bset *i;
 764         bool used_mempool, blacklisted;
 765         unsigned u64s;
 766         int ret, retry_read = 0, write = READ;
 767
 768         b->version_ondisk = U16_MAX;
 769
 770         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 771         sort_iter_init(iter, b);
 772         iter->size = (btree_blocks(c) + 1) * 2;
 773
 774         if (bch2_meta_read_fault("btree"))
 775                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 776                           "dynamic fault");
 777
 778         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 779                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 780                      "bad magic");
 781
 782         btree_err_on(!b->data->keys.seq,
 783                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 784                      "bad btree header");
 785
 786         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 787                 struct bch_btree_ptr_v2 *bp =
 788                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 789
 790                 btree_err_on(b->data->keys.seq != bp->seq,
 791                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 792                              "got wrong btree node (seq %llx want %llx)",
 793                              b->data->keys.seq, bp->seq);
 794         }
 795
 796         while (b->written < c->opts.btree_node_size) {
 797                 unsigned sectors, whiteout_u64s = 0;
 798                 struct nonce nonce;
 799                 struct bch_csum csum;
 800                 bool first = !b->written;
 801
 802                 if (!b->written) {
 803                         i = &b->data->keys;
 804
 805                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 806                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 807                                      "unknown checksum type %llu",
 808                                      BSET_CSUM_TYPE(i));
 809
 810                         nonce = btree_nonce(i, b->written << 9);
 811                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 812
 813                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 814                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 815                                      "invalid checksum");
 816
 817                         bset_encrypt(c, i, b->written << 9);
 818
 819                         btree_err_on(btree_node_is_extents(b) &&
 820                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 821                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 822                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 823
 824                         sectors = vstruct_sectors(b->data, c->block_bits);
 825                 } else {
 826                         bne = write_block(b);
 827                         i = &bne->keys;
 828
 829                         if (i->seq != b->data->keys.seq)
 830                                 break;
 831
 832                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 833                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 834                                      "unknown checksum type %llu",
 835                                      BSET_CSUM_TYPE(i));
 836
 837                         nonce = btree_nonce(i, b->written << 9);
 838                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 839
 840                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 841                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 842                                      "invalid checksum");
 843
 844                         bset_encrypt(c, i, b->written << 9);
 845
 846                         sectors = vstruct_sectors(bne, c->block_bits);
 847                 }
 848
 849                 b->version_ondisk = min(b->version_ondisk,
 850                                         le16_to_cpu(i->version));
 851
 852                 ret = validate_bset(c, ca, b, i, sectors,
 853                                     READ, have_retry);
 854                 if (ret)
 855                         goto fsck_err;
 856
 857                 if (!b->written)
 858                         btree_node_set_format(b, b->data->format);
 859
 860                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 861                                     READ, have_retry);
 862                 if (ret)
 863                         goto fsck_err;
 864
 865                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 866
 867                 b->written += sectors;
 868
 869                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 870                                         le64_to_cpu(i->journal_seq),
 871                                         true);
 872
 873                 btree_err_on(blacklisted && first,
 874                              BTREE_ERR_FIXABLE, c, ca, b, i,
 875                              "first btree node bset has blacklisted journal seq");
 876                 if (blacklisted && !first)
 877                         continue;
 878
 879                 sort_iter_add(iter, i->start,
 880                               vstruct_idx(i, whiteout_u64s));
 881
 882                 sort_iter_add(iter,
 883                               vstruct_idx(i, whiteout_u64s),
 884                               vstruct_last(i));
 885         }
 886
 887         for (bne = write_block(b);
 888              bset_byte_offset(b, bne) < btree_bytes(c);
 889              bne = (void *) bne + block_bytes(c))
 890                 btree_err_on(bne->keys.seq == b->data->keys.seq,
 891                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
 892                              "found bset signature after last bset");
 893
 894         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
 895         sorted->keys.u64s = 0;
 896
 897         set_btree_bset(b, b->set, &b->data->keys);
 898
 899         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
 900
 901         u64s = le16_to_cpu(sorted->keys.u64s);
 902         *sorted = *b->data;
 903         sorted->keys.u64s = cpu_to_le16(u64s);
 904         swap(sorted, b->data);
 905         set_btree_bset(b, b->set, &b->data->keys);
 906         b->nsets = 1;
 907
 908         BUG_ON(b->nr.live_u64s != u64s);
 909
 910         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
 911
 912         i = &b->data->keys;
 913         for (k = i->start; k != vstruct_last(i);) {
 914                 struct bkey tmp;
 915                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 916                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
 917
 918                 if (invalid ||
 919                     (bch2_inject_invalid_keys &&
 920                      !bversion_cmp(u.k->version, MAX_VERSION))) {
 921                         char buf[160];
 922
 923                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 924                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 925                                   "invalid bkey %s: %s", buf, invalid);
 926
 927                         btree_keys_account_key_drop(&b->nr, 0, k);
 928
 929                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 930                         memmove_u64s_down(k, bkey_next(k),
 931                                           (u64 *) vstruct_end(i) - (u64 *) k);
 932                         set_btree_bset_end(b, b->set);
 933                         continue;
 934                 }
 935
 936                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
 937                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
 938
 939                         bp.v->mem_ptr = 0;
 940                 }
 941
 942                 k = bkey_next(k);
 943         }
 944
 945         bch2_bset_build_aux_tree(b, b->set, false);
 946
 947         set_needs_whiteout(btree_bset_first(b), true);
 948
 949         btree_node_reset_sib_u64s(b);
 950
 951         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
 952                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 953
 954                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
 955                         set_btree_node_need_rewrite(b);
 956         }
 957 out:
 958         mempool_free(iter, &c->fill_iter);
 959         return retry_read;
 960 fsck_err:
 961         if (ret == BTREE_RETRY_READ) {
 962                 retry_read = 1;
 963         } else {
 964                 bch2_inconsistent_error(c);
 965                 set_btree_node_read_error(b);
 966         }
 967         goto out;
 968 }
 969
 970 static void btree_node_read_work(struct work_struct *work)
 971 {
 972         struct btree_read_bio *rb =
 973                 container_of(work, struct btree_read_bio, work);
 974         struct bch_fs *c        = rb->c;
 975         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
 976         struct btree *b         = rb->bio.bi_private;
 977         struct bio *bio         = &rb->bio;
 978         struct bch_io_failures failed = { .nr = 0 };
 979         char buf[200];
 980         struct printbuf out;
 981         bool can_retry;
 982
 983         goto start;
 984         while (1) {
 985                 bch_info(c, "retrying read");
 986                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
 987                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
 988                 bio_reset(bio);
 989                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
 990                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
 991                 bio->bi_iter.bi_size    = btree_bytes(c);
 992
 993                 if (rb->have_ioref) {
 994                         bio_set_dev(bio, ca->disk_sb.bdev);
 995                         submit_bio_wait(bio);
 996                 } else {
 997                         bio->bi_status = BLK_STS_REMOVED;
 998                 }
 999 start:
1000                 out = PBUF(buf);
1001                 btree_pos_to_text(&out, c, b);
1002                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1003                                    bch2_blk_status_to_str(bio->bi_status), buf);
1004                 if (rb->have_ioref)
1005                         percpu_ref_put(&ca->io_ref);
1006                 rb->have_ioref = false;
1007
1008                 bch2_mark_io_failure(&failed, &rb->pick);
1009
1010                 can_retry = bch2_bkey_pick_read_device(c,
1011                                 bkey_i_to_s_c(&b->key),
1012                                 &failed, &rb->pick) > 0;
1013
1014                 if (!bio->bi_status &&
1015                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1016                         break;
1017
1018                 if (!can_retry) {
1019                         set_btree_node_read_error(b);
1020                         break;
1021                 }
1022         }
1023
1024         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1025                                rb->start_time);
1026         bio_put(&rb->bio);
1027         clear_btree_node_read_in_flight(b);
1028         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1029 }
1030
1031 static void btree_node_read_endio(struct bio *bio)
1032 {
1033         struct btree_read_bio *rb =
1034                 container_of(bio, struct btree_read_bio, bio);
1035         struct bch_fs *c        = rb->c;
1036
1037         if (rb->have_ioref) {
1038                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1039                 bch2_latency_acct(ca, rb->start_time, READ);
1040         }
1041
1042         queue_work(system_unbound_wq, &rb->work);
1043 }
1044
1045 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1046                           bool sync)
1047 {
1048         struct extent_ptr_decoded pick;
1049         struct btree_read_bio *rb;
1050         struct bch_dev *ca;
1051         struct bio *bio;
1052         int ret;
1053
1054         trace_btree_read(c, b);
1055
1056         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1057                                          NULL, &pick);
1058         if (bch2_fs_fatal_err_on(ret <= 0, c,
1059                         "btree node read error: no device to read from")) {
1060                 set_btree_node_read_error(b);
1061                 return;
1062         }
1063
1064         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1065
1066         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1067                                                    btree_bytes(c)),
1068                                &c->btree_bio);
1069         rb = container_of(bio, struct btree_read_bio, bio);
1070         rb->c                   = c;
1071         rb->start_time          = local_clock();
1072         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1073         rb->pick                = pick;
1074         INIT_WORK(&rb->work, btree_node_read_work);
1075         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1076         bio->bi_iter.bi_sector  = pick.ptr.offset;
1077         bio->bi_end_io          = btree_node_read_endio;
1078         bio->bi_private         = b;
1079         bch2_bio_map(bio, b->data, btree_bytes(c));
1080
1081         set_btree_node_read_in_flight(b);
1082
1083         if (rb->have_ioref) {
1084                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1085                              bio_sectors(bio));
1086                 bio_set_dev(bio, ca->disk_sb.bdev);
1087
1088                 if (sync) {
1089                         submit_bio_wait(bio);
1090
1091                         bio->bi_private = b;
1092                         btree_node_read_work(&rb->work);
1093                 } else {
1094                         submit_bio(bio);
1095                 }
1096         } else {
1097                 bio->bi_status = BLK_STS_REMOVED;
1098
1099                 if (sync)
1100                         btree_node_read_work(&rb->work);
1101                 else
1102                         queue_work(system_unbound_wq, &rb->work);
1103
1104         }
1105 }
1106
1107 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1108                         const struct bkey_i *k, unsigned level)
1109 {
1110         struct closure cl;
1111         struct btree *b;
1112         int ret;
1113
1114         closure_init_stack(&cl);
1115
1116         do {
1117                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1118                 closure_sync(&cl);
1119         } while (ret);
1120
1121         b = bch2_btree_node_mem_alloc(c);
1122         bch2_btree_cache_cannibalize_unlock(c);
1123
1124         BUG_ON(IS_ERR(b));
1125
1126         bkey_copy(&b->key, k);
1127         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1128
1129         bch2_btree_node_read(c, b, true);
1130
1131         if (btree_node_read_error(b)) {
1132                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1133
1134                 mutex_lock(&c->btree_cache.lock);
1135                 list_move(&b->list, &c->btree_cache.freeable);
1136                 mutex_unlock(&c->btree_cache.lock);
1137
1138                 ret = -EIO;
1139                 goto err;
1140         }
1141
1142         bch2_btree_set_root_for_read(c, b);
1143 err:
1144         six_unlock_write(&b->c.lock);
1145         six_unlock_intent(&b->c.lock);
1146
1147         return ret;
1148 }
1149
1150 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1151                               struct btree_write *w)
1152 {
1153         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1154
1155         do {
1156                 old = new = v;
1157                 if (!(old & 1))
1158                         break;
1159
1160                 new &= ~1UL;
1161         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1162
1163         if (old & 1)
1164                 closure_put(&((struct btree_update *) new)->cl);
1165
1166         bch2_journal_pin_drop(&c->journal, &w->journal);
1167 }
1168
1169 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1170 {
1171         struct btree_write *w = btree_prev_write(b);
1172
1173         bch2_btree_complete_write(c, b, w);
1174         btree_node_io_unlock(b);
1175 }
1176
1177 static void bch2_btree_node_write_error(struct bch_fs *c,
1178                                         struct btree_write_bio *wbio)
1179 {
1180         struct btree *b         = wbio->wbio.bio.bi_private;
1181         struct bkey_buf k;
1182         struct bch_extent_ptr *ptr;
1183         struct btree_trans trans;
1184         struct btree_iter *iter;
1185         int ret;
1186
1187         bch2_bkey_buf_init(&k);
1188         bch2_trans_init(&trans, c, 0, 0);
1189
1190         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1191                                         BTREE_MAX_DEPTH, b->c.level, 0);
1192 retry:
1193         ret = bch2_btree_iter_traverse(iter);
1194         if (ret)
1195                 goto err;
1196
1197         /* has node been freed? */
1198         if (iter->l[b->c.level].b != b) {
1199                 /* node has been freed: */
1200                 BUG_ON(!btree_node_dying(b));
1201                 goto out;
1202         }
1203
1204         BUG_ON(!btree_node_hashed(b));
1205
1206         bch2_bkey_buf_copy(&k, c, &b->key);
1207
1208         bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1209                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1210
1211         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1212                 goto err;
1213
1214         ret = bch2_btree_node_update_key(c, iter, b, k.k);
1215         if (ret == -EINTR)
1216                 goto retry;
1217         if (ret)
1218                 goto err;
1219 out:
1220         bch2_trans_iter_put(&trans, iter);
1221         bch2_trans_exit(&trans);
1222         bch2_bkey_buf_exit(&k, c);
1223         bio_put(&wbio->wbio.bio);
1224         btree_node_write_done(c, b);
1225         return;
1226 err:
1227         set_btree_node_noevict(b);
1228         bch2_fs_fatal_error(c, "fatal error writing btree node");
1229         goto out;
1230 }
1231
1232 void bch2_btree_write_error_work(struct work_struct *work)
1233 {
1234         struct bch_fs *c = container_of(work, struct bch_fs,
1235                                         btree_write_error_work);
1236         struct bio *bio;
1237
1238         while (1) {
1239                 spin_lock_irq(&c->btree_write_error_lock);
1240                 bio = bio_list_pop(&c->btree_write_error_list);
1241                 spin_unlock_irq(&c->btree_write_error_lock);
1242
1243                 if (!bio)
1244                         break;
1245
1246                 bch2_btree_node_write_error(c,
1247                         container_of(bio, struct btree_write_bio, wbio.bio));
1248         }
1249 }
1250
1251 static void btree_node_write_work(struct work_struct *work)
1252 {
1253         struct btree_write_bio *wbio =
1254                 container_of(work, struct btree_write_bio, work);
1255         struct bch_fs *c        = wbio->wbio.c;
1256         struct btree *b         = wbio->wbio.bio.bi_private;
1257
1258         btree_bounce_free(c,
1259                 wbio->bytes,
1260                 wbio->wbio.used_mempool,
1261                 wbio->data);
1262
1263         if (wbio->wbio.failed.nr) {
1264                 unsigned long flags;
1265
1266                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1267                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1268                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1269
1270                 queue_work(c->wq, &c->btree_write_error_work);
1271                 return;
1272         }
1273
1274         bio_put(&wbio->wbio.bio);
1275         btree_node_write_done(c, b);
1276 }
1277
1278 static void btree_node_write_endio(struct bio *bio)
1279 {
1280         struct bch_write_bio *wbio      = to_wbio(bio);
1281         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1282         struct bch_write_bio *orig      = parent ?: wbio;
1283         struct bch_fs *c                = wbio->c;
1284         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1285         unsigned long flags;
1286
1287         if (wbio->have_ioref)
1288                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1289
1290         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1291                                bch2_blk_status_to_str(bio->bi_status)) ||
1292             bch2_meta_write_fault("btree")) {
1293                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1294                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1295                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1296         }
1297
1298         if (wbio->have_ioref)
1299                 percpu_ref_put(&ca->io_ref);
1300
1301         if (parent) {
1302                 bio_put(bio);
1303                 bio_endio(&parent->bio);
1304         } else {
1305                 struct btree_write_bio *wb =
1306                         container_of(orig, struct btree_write_bio, wbio);
1307
1308                 INIT_WORK(&wb->work, btree_node_write_work);
1309                 queue_work(system_unbound_wq, &wb->work);
1310         }
1311 }
1312
1313 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1314                                    struct bset *i, unsigned sectors)
1315 {
1316         unsigned whiteout_u64s = 0;
1317         int ret;
1318
1319         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1320                 return -1;
1321
1322         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1323                 validate_bset(c, NULL, b, i, sectors, WRITE, false);
1324         if (ret) {
1325                 bch2_inconsistent_error(c);
1326                 dump_stack();
1327         }
1328
1329         return ret;
1330 }
1331
1332 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1333                             enum six_lock_type lock_type_held)
1334 {
1335         struct btree_write_bio *wbio;
1336         struct bset_tree *t;
1337         struct bset *i;
1338         struct btree_node *bn = NULL;
1339         struct btree_node_entry *bne = NULL;
1340         struct bkey_buf k;
1341         struct bch_extent_ptr *ptr;
1342         struct sort_iter sort_iter;
1343         struct nonce nonce;
1344         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1345         u64 seq = 0;
1346         bool used_mempool;
1347         unsigned long old, new;
1348         bool validate_before_checksum = false;
1349         void *data;
1350
1351         bch2_bkey_buf_init(&k);
1352
1353         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1354                 return;
1355
1356         /*
1357          * We may only have a read lock on the btree node - the dirty bit is our
1358          * "lock" against racing with other threads that may be trying to start
1359          * a write, we do a write iff we clear the dirty bit. Since setting the
1360          * dirty bit requires a write lock, we can't race with other threads
1361          * redirtying it:
1362          */
1363         do {
1364                 old = new = READ_ONCE(b->flags);
1365
1366                 if (!(old & (1 << BTREE_NODE_dirty)))
1367                         return;
1368
1369                 if (!btree_node_may_write(b))
1370                         return;
1371
1372                 if (old & (1 << BTREE_NODE_never_write))
1373                         return;
1374
1375                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1376                         btree_node_wait_on_io(b);
1377                         continue;
1378                 }
1379
1380                 new &= ~(1 << BTREE_NODE_dirty);
1381                 new &= ~(1 << BTREE_NODE_need_write);
1382                 new |=  (1 << BTREE_NODE_write_in_flight);
1383                 new |=  (1 << BTREE_NODE_just_written);
1384                 new ^=  (1 << BTREE_NODE_write_idx);
1385         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1386
1387         atomic_dec(&c->btree_cache.dirty);
1388
1389         BUG_ON(btree_node_fake(b));
1390         BUG_ON((b->will_make_reachable != 0) != !b->written);
1391
1392         BUG_ON(b->written >= c->opts.btree_node_size);
1393         BUG_ON(b->written & (c->opts.block_size - 1));
1394         BUG_ON(bset_written(b, btree_bset_last(b)));
1395         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1396         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1397
1398         bch2_sort_whiteouts(c, b);
1399
1400         sort_iter_init(&sort_iter, b);
1401
1402         bytes = !b->written
1403                 ? sizeof(struct btree_node)
1404                 : sizeof(struct btree_node_entry);
1405
1406         bytes += b->whiteout_u64s * sizeof(u64);
1407
1408         for_each_bset(b, t) {
1409                 i = bset(b, t);
1410
1411                 if (bset_written(b, i))
1412                         continue;
1413
1414                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1415                 sort_iter_add(&sort_iter,
1416                               btree_bkey_first(b, t),
1417                               btree_bkey_last(b, t));
1418                 seq = max(seq, le64_to_cpu(i->journal_seq));
1419         }
1420
1421         BUG_ON(b->written && !seq);
1422
1423         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1424         bytes += 8;
1425
1426         data = btree_bounce_alloc(c, bytes, &used_mempool);
1427
1428         if (!b->written) {
1429                 bn = data;
1430                 *bn = *b->data;
1431                 i = &bn->keys;
1432         } else {
1433                 bne = data;
1434                 bne->keys = b->data->keys;
1435                 i = &bne->keys;
1436         }
1437
1438         i->journal_seq  = cpu_to_le64(seq);
1439         i->u64s         = 0;
1440
1441         sort_iter_add(&sort_iter,
1442                       unwritten_whiteouts_start(c, b),
1443                       unwritten_whiteouts_end(c, b));
1444         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1445
1446         b->whiteout_u64s = 0;
1447
1448         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1449         le16_add_cpu(&i->u64s, u64s);
1450
1451         set_needs_whiteout(i, false);
1452
1453         /* do we have data to write? */
1454         if (b->written && !i->u64s)
1455                 goto nowrite;
1456
1457         bytes_to_write = vstruct_end(i) - data;
1458         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1459
1460         memset(data + bytes_to_write, 0,
1461                (sectors_to_write << 9) - bytes_to_write);
1462
1463         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1464         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1465         BUG_ON(i->seq != b->data->keys.seq);
1466
1467         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1468                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1469                 : cpu_to_le16(c->sb.version);
1470         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1471
1472         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1473                 validate_before_checksum = true;
1474
1475         /* validate_bset will be modifying: */
1476         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1477                 validate_before_checksum = true;
1478
1479         /* if we're going to be encrypting, check metadata validity first: */
1480         if (validate_before_checksum &&
1481             validate_bset_for_write(c, b, i, sectors_to_write))
1482                 goto err;
1483
1484         bset_encrypt(c, i, b->written << 9);
1485
1486         nonce = btree_nonce(i, b->written << 9);
1487
1488         if (bn)
1489                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1490         else
1491                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1492
1493         /* if we're not encrypting, check metadata after checksumming: */
1494         if (!validate_before_checksum &&
1495             validate_bset_for_write(c, b, i, sectors_to_write))
1496                 goto err;
1497
1498         /*
1499          * We handle btree write errors by immediately halting the journal -
1500          * after we've done that, we can't issue any subsequent btree writes
1501          * because they might have pointers to new nodes that failed to write.
1502          *
1503          * Furthermore, there's no point in doing any more btree writes because
1504          * with the journal stopped, we're never going to update the journal to
1505          * reflect that those writes were done and the data flushed from the
1506          * journal:
1507          *
1508          * Also on journal error, the pending write may have updates that were
1509          * never journalled (interior nodes, see btree_update_nodes_written()) -
1510          * it's critical that we don't do the write in that case otherwise we
1511          * will have updates visible that weren't in the journal:
1512          *
1513          * Make sure to update b->written so bch2_btree_init_next() doesn't
1514          * break:
1515          */
1516         if (bch2_journal_error(&c->journal) ||
1517             c->opts.nochanges)
1518                 goto err;
1519
1520         trace_btree_write(b, bytes_to_write, sectors_to_write);
1521
1522         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1523                                 buf_pages(data, sectors_to_write << 9),
1524                                 &c->btree_bio),
1525                             struct btree_write_bio, wbio.bio);
1526         wbio_init(&wbio->wbio.bio);
1527         wbio->data                      = data;
1528         wbio->bytes                     = bytes;
1529         wbio->wbio.used_mempool         = used_mempool;
1530         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1531         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1532         wbio->wbio.bio.bi_private       = b;
1533
1534         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1535
1536         /*
1537          * If we're appending to a leaf node, we don't technically need FUA -
1538          * this write just needs to be persisted before the next journal write,
1539          * which will be marked FLUSH|FUA.
1540          *
1541          * Similarly if we're writing a new btree root - the pointer is going to
1542          * be in the next journal entry.
1543          *
1544          * But if we're writing a new btree node (that isn't a root) or
1545          * appending to a non leaf btree node, we need either FUA or a flush
1546          * when we write the parent with the new pointer. FUA is cheaper than a
1547          * flush, and writes appending to leaf nodes aren't blocking anything so
1548          * just make all btree node writes FUA to keep things sane.
1549          */
1550
1551         bch2_bkey_buf_copy(&k, c, &b->key);
1552
1553         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(k.k)), ptr)
1554                 ptr->offset += b->written;
1555
1556         b->written += sectors_to_write;
1557
1558         /* XXX: submitting IO with btree locks held: */
1559         bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_btree, k.k);
1560         bch2_bkey_buf_exit(&k, c);
1561         return;
1562 err:
1563         set_btree_node_noevict(b);
1564         b->written += sectors_to_write;
1565 nowrite:
1566         btree_bounce_free(c, bytes, used_mempool, data);
1567         btree_node_write_done(c, b);
1568 }
1569
1570 /*
1571  * Work that must be done with write lock held:
1572  */
1573 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1574 {
1575         bool invalidated_iter = false;
1576         struct btree_node_entry *bne;
1577         struct bset_tree *t;
1578
1579         if (!btree_node_just_written(b))
1580                 return false;
1581
1582         BUG_ON(b->whiteout_u64s);
1583
1584         clear_btree_node_just_written(b);
1585
1586         /*
1587          * Note: immediately after write, bset_written() doesn't work - the
1588          * amount of data we had to write after compaction might have been
1589          * smaller than the offset of the last bset.
1590          *
1591          * However, we know that all bsets have been written here, as long as
1592          * we're still holding the write lock:
1593          */
1594
1595         /*
1596          * XXX: decide if we really want to unconditionally sort down to a
1597          * single bset:
1598          */
1599         if (b->nsets > 1) {
1600                 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1601                 invalidated_iter = true;
1602         } else {
1603                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1604         }
1605
1606         for_each_bset(b, t)
1607                 set_needs_whiteout(bset(b, t), true);
1608
1609         bch2_btree_verify(c, b);
1610
1611         /*
1612          * If later we don't unconditionally sort down to a single bset, we have
1613          * to ensure this is still true:
1614          */
1615         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1616
1617         bne = want_new_bset(c, b);
1618         if (bne)
1619                 bch2_bset_init_next(c, b, bne);
1620
1621         bch2_btree_build_aux_trees(b);
1622
1623         return invalidated_iter;
1624 }
1625
1626 /*
1627  * Use this one if the node is intent locked:
1628  */
1629 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1630                           enum six_lock_type lock_type_held)
1631 {
1632         BUG_ON(lock_type_held == SIX_LOCK_write);
1633
1634         if (lock_type_held == SIX_LOCK_intent ||
1635             six_lock_tryupgrade(&b->c.lock)) {
1636                 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1637
1638                 /* don't cycle lock unnecessarily: */
1639                 if (btree_node_just_written(b) &&
1640                     six_trylock_write(&b->c.lock)) {
1641                         bch2_btree_post_write_cleanup(c, b);
1642                         six_unlock_write(&b->c.lock);
1643                 }
1644
1645                 if (lock_type_held == SIX_LOCK_read)
1646                         six_lock_downgrade(&b->c.lock);
1647         } else {
1648                 __bch2_btree_node_write(c, b, SIX_LOCK_read);
1649         }
1650 }
1651
1652 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
1653 {
1654         struct bucket_table *tbl;
1655         struct rhash_head *pos;
1656         struct btree *b;
1657         unsigned i;
1658 restart:
1659         rcu_read_lock();
1660         for_each_cached_btree(b, c, tbl, i, pos)
1661                 if (test_bit(flag, &b->flags)) {
1662                         rcu_read_unlock();
1663                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
1664                         goto restart;
1665
1666                 }
1667         rcu_read_unlock();
1668 }
1669
1670 void bch2_btree_flush_all_reads(struct bch_fs *c)
1671 {
1672         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
1673 }
1674
1675 void bch2_btree_flush_all_writes(struct bch_fs *c)
1676 {
1677         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
1678 }
1679
1680 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
1681 {
1682         struct bucket_table *tbl;
1683         struct rhash_head *pos;
1684         struct btree *b;
1685         unsigned i;
1686
1687         rcu_read_lock();
1688         for_each_cached_btree(b, c, tbl, i, pos) {
1689                 unsigned long flags = READ_ONCE(b->flags);
1690
1691                 if (!(flags & (1 << BTREE_NODE_dirty)))
1692                         continue;
1693
1694                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
1695                        b,
1696                        (flags & (1 << BTREE_NODE_dirty)) != 0,
1697                        (flags & (1 << BTREE_NODE_need_write)) != 0,
1698                        b->c.level,
1699                        b->written,
1700                        !list_empty_careful(&b->write_blocked),
1701                        b->will_make_reachable != 0,
1702                        b->will_make_reachable & 1);
1703         }
1704         rcu_read_unlock();
1705 }