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 void bch2_btree_node_io_unlock(struct btree *b)
  26 {
  27         EBUG_ON(!btree_node_write_in_flight(b));
  28
  29         clear_btree_node_write_in_flight(b);
  30         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
  31 }
  32
  33 void bch2_btree_node_io_lock(struct btree *b)
  34 {
  35         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
  36
  37         wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
  38                             TASK_UNINTERRUPTIBLE);
  39 }
  40
  41 void __bch2_btree_node_wait_on_read(struct btree *b)
  42 {
  43         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  44                        TASK_UNINTERRUPTIBLE);
  45 }
  46
  47 void __bch2_btree_node_wait_on_write(struct btree *b)
  48 {
  49         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  50                        TASK_UNINTERRUPTIBLE);
  51 }
  52
  53 void bch2_btree_node_wait_on_read(struct btree *b)
  54 {
  55         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
  56
  57         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  58                        TASK_UNINTERRUPTIBLE);
  59 }
  60
  61 void bch2_btree_node_wait_on_write(struct btree *b)
  62 {
  63         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
  64
  65         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  66                        TASK_UNINTERRUPTIBLE);
  67 }
  68
  69 static void verify_no_dups(struct btree *b,
  70                            struct bkey_packed *start,
  71                            struct bkey_packed *end)
  72 {
  73 #ifdef CONFIG_BCACHEFS_DEBUG
  74         struct bkey_packed *k, *p;
  75
  76         if (start == end)
  77                 return;
  78
  79         for (p = start, k = bkey_next(start);
  80              k != end;
  81              p = k, k = bkey_next(k)) {
  82                 struct bkey l = bkey_unpack_key(b, p);
  83                 struct bkey r = bkey_unpack_key(b, k);
  84
  85                 BUG_ON(bpos_cmp(l.p, bkey_start_pos(&r)) >= 0);
  86         }
  87 #endif
  88 }
  89
  90 static void set_needs_whiteout(struct bset *i, int v)
  91 {
  92         struct bkey_packed *k;
  93
  94         for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
  95                 k->needs_whiteout = v;
  96 }
  97
  98 static void btree_bounce_free(struct bch_fs *c, size_t size,
  99                               bool used_mempool, void *p)
 100 {
 101         if (used_mempool)
 102                 mempool_free(p, &c->btree_bounce_pool);
 103         else
 104                 vpfree(p, size);
 105 }
 106
 107 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
 108                                 bool *used_mempool)
 109 {
 110         unsigned flags = memalloc_nofs_save();
 111         void *p;
 112
 113         BUG_ON(size > btree_bytes(c));
 114
 115         *used_mempool = false;
 116         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
 117         if (!p) {
 118                 *used_mempool = true;
 119                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
 120         }
 121         memalloc_nofs_restore(flags);
 122         return p;
 123 }
 124
 125 static void sort_bkey_ptrs(const struct btree *bt,
 126                            struct bkey_packed **ptrs, unsigned nr)
 127 {
 128         unsigned n = nr, a = nr / 2, b, c, d;
 129
 130         if (!a)
 131                 return;
 132
 133         /* Heap sort: see lib/sort.c: */
 134         while (1) {
 135                 if (a)
 136                         a--;
 137                 else if (--n)
 138                         swap(ptrs[0], ptrs[n]);
 139                 else
 140                         break;
 141
 142                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 143                         b = bch2_bkey_cmp_packed(bt,
 144                                             ptrs[c],
 145                                             ptrs[d]) >= 0 ? c : d;
 146                 if (d == n)
 147                         b = c;
 148
 149                 while (b != a &&
 150                        bch2_bkey_cmp_packed(bt,
 151                                        ptrs[a],
 152                                        ptrs[b]) >= 0)
 153                         b = (b - 1) / 2;
 154                 c = b;
 155                 while (b != a) {
 156                         b = (b - 1) / 2;
 157                         swap(ptrs[b], ptrs[c]);
 158                 }
 159         }
 160 }
 161
 162 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 163 {
 164         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 165         bool used_mempool = false;
 166         size_t bytes = b->whiteout_u64s * sizeof(u64);
 167
 168         if (!b->whiteout_u64s)
 169                 return;
 170
 171         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 172
 173         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 174
 175         for (k = unwritten_whiteouts_start(c, b);
 176              k != unwritten_whiteouts_end(c, b);
 177              k = bkey_next(k))
 178                 *--ptrs = k;
 179
 180         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 181
 182         k = new_whiteouts;
 183
 184         while (ptrs != ptrs_end) {
 185                 bkey_copy(k, *ptrs);
 186                 k = bkey_next(k);
 187                 ptrs++;
 188         }
 189
 190         verify_no_dups(b, new_whiteouts,
 191                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 192
 193         memcpy_u64s(unwritten_whiteouts_start(c, b),
 194                     new_whiteouts, b->whiteout_u64s);
 195
 196         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 197 }
 198
 199 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 200                                 bool compacting, enum compact_mode mode)
 201 {
 202         if (!bset_dead_u64s(b, t))
 203                 return false;
 204
 205         switch (mode) {
 206         case COMPACT_LAZY:
 207                 return should_compact_bset_lazy(b, t) ||
 208                         (compacting && !bset_written(b, bset(b, t)));
 209         case COMPACT_ALL:
 210                 return true;
 211         default:
 212                 BUG();
 213         }
 214 }
 215
 216 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 217 {
 218         struct bset_tree *t;
 219         bool ret = false;
 220
 221         for_each_bset(b, t) {
 222                 struct bset *i = bset(b, t);
 223                 struct bkey_packed *k, *n, *out, *start, *end;
 224                 struct btree_node_entry *src = NULL, *dst = NULL;
 225
 226                 if (t != b->set && !bset_written(b, i)) {
 227                         src = container_of(i, struct btree_node_entry, keys);
 228                         dst = max(write_block(b),
 229                                   (void *) btree_bkey_last(b, t - 1));
 230                 }
 231
 232                 if (src != dst)
 233                         ret = true;
 234
 235                 if (!should_compact_bset(b, t, ret, mode)) {
 236                         if (src != dst) {
 237                                 memmove(dst, src, sizeof(*src) +
 238                                         le16_to_cpu(src->keys.u64s) *
 239                                         sizeof(u64));
 240                                 i = &dst->keys;
 241                                 set_btree_bset(b, t, i);
 242                         }
 243                         continue;
 244                 }
 245
 246                 start   = btree_bkey_first(b, t);
 247                 end     = btree_bkey_last(b, t);
 248
 249                 if (src != dst) {
 250                         memmove(dst, src, sizeof(*src));
 251                         i = &dst->keys;
 252                         set_btree_bset(b, t, i);
 253                 }
 254
 255                 out = i->start;
 256
 257                 for (k = start; k != end; k = n) {
 258                         n = bkey_next(k);
 259
 260                         if (!bkey_deleted(k)) {
 261                                 bkey_copy(out, k);
 262                                 out = bkey_next(out);
 263                         } else {
 264                                 BUG_ON(k->needs_whiteout);
 265                         }
 266                 }
 267
 268                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 269                 set_btree_bset_end(b, t);
 270                 bch2_bset_set_no_aux_tree(b, t);
 271                 ret = true;
 272         }
 273
 274         bch2_verify_btree_nr_keys(b);
 275
 276         bch2_btree_build_aux_trees(b);
 277
 278         return ret;
 279 }
 280
 281 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 282                             enum compact_mode mode)
 283 {
 284         return bch2_drop_whiteouts(b, mode);
 285 }
 286
 287 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 288                             unsigned start_idx,
 289                             unsigned end_idx,
 290                             bool filter_whiteouts)
 291 {
 292         struct btree_node *out;
 293         struct sort_iter sort_iter;
 294         struct bset_tree *t;
 295         struct bset *start_bset = bset(b, &b->set[start_idx]);
 296         bool used_mempool = false;
 297         u64 start_time, seq = 0;
 298         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 299         bool sorting_entire_node = start_idx == 0 &&
 300                 end_idx == b->nsets;
 301
 302         sort_iter_init(&sort_iter, b);
 303
 304         for (t = b->set + start_idx;
 305              t < b->set + end_idx;
 306              t++) {
 307                 u64s += le16_to_cpu(bset(b, t)->u64s);
 308                 sort_iter_add(&sort_iter,
 309                               btree_bkey_first(b, t),
 310                               btree_bkey_last(b, t));
 311         }
 312
 313         bytes = sorting_entire_node
 314                 ? btree_bytes(c)
 315                 : __vstruct_bytes(struct btree_node, u64s);
 316
 317         out = btree_bounce_alloc(c, bytes, &used_mempool);
 318
 319         start_time = local_clock();
 320
 321         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 322
 323         out->keys.u64s = cpu_to_le16(u64s);
 324
 325         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 326
 327         if (sorting_entire_node)
 328                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 329                                        start_time);
 330
 331         /* Make sure we preserve bset journal_seq: */
 332         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 333                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 334         start_bset->journal_seq = cpu_to_le64(seq);
 335
 336         if (sorting_entire_node) {
 337                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 338
 339                 BUG_ON(bytes != btree_bytes(c));
 340
 341                 /*
 342                  * Our temporary buffer is the same size as the btree node's
 343                  * buffer, we can just swap buffers instead of doing a big
 344                  * memcpy()
 345                  */
 346                 *out = *b->data;
 347                 out->keys.u64s = cpu_to_le16(u64s);
 348                 swap(out, b->data);
 349                 set_btree_bset(b, b->set, &b->data->keys);
 350         } else {
 351                 start_bset->u64s = out->keys.u64s;
 352                 memcpy_u64s(start_bset->start,
 353                             out->keys.start,
 354                             le16_to_cpu(out->keys.u64s));
 355         }
 356
 357         for (i = start_idx + 1; i < end_idx; i++)
 358                 b->nr.bset_u64s[start_idx] +=
 359                         b->nr.bset_u64s[i];
 360
 361         b->nsets -= shift;
 362
 363         for (i = start_idx + 1; i < b->nsets; i++) {
 364                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 365                 b->set[i]               = b->set[i + shift];
 366         }
 367
 368         for (i = b->nsets; i < MAX_BSETS; i++)
 369                 b->nr.bset_u64s[i] = 0;
 370
 371         set_btree_bset_end(b, &b->set[start_idx]);
 372         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 373
 374         btree_bounce_free(c, bytes, used_mempool, out);
 375
 376         bch2_verify_btree_nr_keys(b);
 377 }
 378
 379 void bch2_btree_sort_into(struct bch_fs *c,
 380                          struct btree *dst,
 381                          struct btree *src)
 382 {
 383         struct btree_nr_keys nr;
 384         struct btree_node_iter src_iter;
 385         u64 start_time = local_clock();
 386
 387         BUG_ON(dst->nsets != 1);
 388
 389         bch2_bset_set_no_aux_tree(dst, dst->set);
 390
 391         bch2_btree_node_iter_init_from_start(&src_iter, src);
 392
 393         if (btree_node_is_extents(src))
 394                 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
 395                                 src, &src_iter,
 396                                 &dst->format,
 397                                 true);
 398         else
 399                 nr = bch2_sort_repack(btree_bset_first(dst),
 400                                 src, &src_iter,
 401                                 &dst->format,
 402                                 true);
 403
 404         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 405                                start_time);
 406
 407         set_btree_bset_end(dst, dst->set);
 408
 409         dst->nr.live_u64s       += nr.live_u64s;
 410         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 411         dst->nr.packed_keys     += nr.packed_keys;
 412         dst->nr.unpacked_keys   += nr.unpacked_keys;
 413
 414         bch2_verify_btree_nr_keys(dst);
 415 }
 416
 417 #define SORT_CRIT       (4096 / sizeof(u64))
 418
 419 /*
 420  * We're about to add another bset to the btree node, so if there's currently
 421  * too many bsets - sort some of them together:
 422  */
 423 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 424 {
 425         unsigned unwritten_idx;
 426         bool ret = false;
 427
 428         for (unwritten_idx = 0;
 429              unwritten_idx < b->nsets;
 430              unwritten_idx++)
 431                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 432                         break;
 433
 434         if (b->nsets - unwritten_idx > 1) {
 435                 btree_node_sort(c, b, unwritten_idx,
 436                                 b->nsets, false);
 437                 ret = true;
 438         }
 439
 440         if (unwritten_idx > 1) {
 441                 btree_node_sort(c, b, 0, unwritten_idx, false);
 442                 ret = true;
 443         }
 444
 445         return ret;
 446 }
 447
 448 void bch2_btree_build_aux_trees(struct btree *b)
 449 {
 450         struct bset_tree *t;
 451
 452         for_each_bset(b, t)
 453                 bch2_bset_build_aux_tree(b, t,
 454                                 !bset_written(b, bset(b, t)) &&
 455                                 t == bset_tree_last(b));
 456 }
 457
 458 /*
 459  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 460  * inserted into
 461  *
 462  * Safe to call if there already is an unwritten bset - will only add a new bset
 463  * if @b doesn't already have one.
 464  *
 465  * Returns true if we sorted (i.e. invalidated iterators
 466  */
 467 void bch2_btree_init_next(struct btree_trans *trans,
 468                           struct btree_iter *iter,
 469                           struct btree *b)
 470 {
 471         struct bch_fs *c = trans->c;
 472         struct btree_node_entry *bne;
 473         bool reinit_iter = false;
 474
 475         EBUG_ON(!(b->c.lock.state.seq & 1));
 476         EBUG_ON(iter && iter->l[b->c.level].b != b);
 477         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 478
 479         if (b->nsets == MAX_BSETS &&
 480             !btree_node_write_in_flight(b)) {
 481                 unsigned log_u64s[] = {
 482                         ilog2(bset_u64s(&b->set[0])),
 483                         ilog2(bset_u64s(&b->set[1])),
 484                         ilog2(bset_u64s(&b->set[2])),
 485                 };
 486
 487                 if (log_u64s[1] >= (log_u64s[0] + log_u64s[2]) / 2) {
 488                         bch2_btree_node_write(c, b, SIX_LOCK_write);
 489                         reinit_iter = true;
 490                 }
 491         }
 492
 493         if (b->nsets == MAX_BSETS &&
 494             btree_node_compact(c, b))
 495                 reinit_iter = true;
 496
 497         BUG_ON(b->nsets >= MAX_BSETS);
 498
 499         bne = want_new_bset(c, b);
 500         if (bne)
 501                 bch2_bset_init_next(c, b, bne);
 502
 503         bch2_btree_build_aux_trees(b);
 504
 505         if (iter && reinit_iter)
 506                 bch2_btree_iter_reinit_node(iter, b);
 507 }
 508
 509 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 510                           struct btree *b)
 511 {
 512         pr_buf(out, "%s level %u/%u\n  ",
 513                bch2_btree_ids[b->c.btree_id],
 514                b->c.level,
 515                c->btree_roots[b->c.btree_id].level);
 516         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 517 }
 518
 519 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 520                           struct bch_dev *ca,
 521                           struct btree *b, struct bset *i,
 522                           unsigned offset, int write)
 523 {
 524         pr_buf(out, "error validating btree node ");
 525         if (write)
 526                 pr_buf(out, "before write ");
 527         if (ca)
 528                 pr_buf(out, "on %s ", ca->name);
 529         pr_buf(out, "at btree ");
 530         btree_pos_to_text(out, c, b);
 531
 532         pr_buf(out, "\n  node offset %u", b->written);
 533         if (i)
 534                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 535 }
 536
 537 enum btree_err_type {
 538         BTREE_ERR_FIXABLE,
 539         BTREE_ERR_WANT_RETRY,
 540         BTREE_ERR_MUST_RETRY,
 541         BTREE_ERR_FATAL,
 542 };
 543
 544 enum btree_validate_ret {
 545         BTREE_RETRY_READ = 64,
 546 };
 547
 548 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 549 ({                                                                      \
 550         __label__ out;                                                  \
 551         char _buf[300];                                                 \
 552         char *_buf2 = _buf;                                             \
 553         struct printbuf out = PBUF(_buf);                               \
 554                                                                         \
 555         _buf2 = kmalloc(4096, GFP_ATOMIC);                              \
 556         if (_buf2)                                                      \
 557                 out = _PBUF(_buf2, 4986);                               \
 558                                                                         \
 559         btree_err_msg(&out, c, ca, b, i, b->written, write);            \
 560         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 561                                                                         \
 562         if (type == BTREE_ERR_FIXABLE &&                                \
 563             write == READ &&                                            \
 564             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 565                 mustfix_fsck_err(c, "%s", _buf2);                       \
 566                 goto out;                                               \
 567         }                                                               \
 568                                                                         \
 569         switch (write) {                                                \
 570         case READ:                                                      \
 571                 bch_err(c, "%s", _buf2);                                \
 572                                                                         \
 573                 switch (type) {                                         \
 574                 case BTREE_ERR_FIXABLE:                                 \
 575                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 576                         goto fsck_err;                                  \
 577                 case BTREE_ERR_WANT_RETRY:                              \
 578                         if (have_retry) {                               \
 579                                 ret = BTREE_RETRY_READ;                 \
 580                                 goto fsck_err;                          \
 581                         }                                               \
 582                         break;                                          \
 583                 case BTREE_ERR_MUST_RETRY:                              \
 584                         ret = BTREE_RETRY_READ;                         \
 585                         goto fsck_err;                                  \
 586                 case BTREE_ERR_FATAL:                                   \
 587                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 588                         goto fsck_err;                                  \
 589                 }                                                       \
 590                 break;                                                  \
 591         case WRITE:                                                     \
 592                 bch_err(c, "corrupt metadata before write: %s", _buf2); \
 593                                                                         \
 594                 if (bch2_fs_inconsistent(c)) {                          \
 595                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 596                         goto fsck_err;                                  \
 597                 }                                                       \
 598                 break;                                                  \
 599         }                                                               \
 600 out:                                                                    \
 601         if (_buf2 != _buf)                                              \
 602                 kfree(_buf2);                                           \
 603         true;                                                           \
 604 })
 605
 606 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 607
 608 /*
 609  * When btree topology repair changes the start or end of a node, that might
 610  * mean we have to drop keys that are no longer inside the node:
 611  */
 612 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 613 {
 614         struct bset_tree *t;
 615         struct bkey_s_c k;
 616         struct bkey unpacked;
 617         struct btree_node_iter iter;
 618
 619         for_each_bset(b, t) {
 620                 struct bset *i = bset(b, t);
 621                 struct bkey_packed *k;
 622
 623                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 624                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 625                                 break;
 626
 627                 if (k != i->start) {
 628                         unsigned shift = (u64 *) k - (u64 *) i->start;
 629
 630                         memmove_u64s_down(i->start, k,
 631                                           (u64 *) vstruct_end(i) - (u64 *) k);
 632                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 633                         set_btree_bset_end(b, t);
 634                         bch2_bset_set_no_aux_tree(b, t);
 635                 }
 636
 637                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 638                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 639                                 break;
 640
 641                 if (k != vstruct_last(i)) {
 642                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 643                         set_btree_bset_end(b, t);
 644                         bch2_bset_set_no_aux_tree(b, t);
 645                 }
 646         }
 647
 648         bch2_btree_build_aux_trees(b);
 649
 650         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 651                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 652                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 653         }
 654 }
 655
 656 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 657                          struct btree *b, struct bset *i,
 658                          unsigned sectors, int write, bool have_retry)
 659 {
 660         unsigned version = le16_to_cpu(i->version);
 661         const char *err;
 662         char buf1[100];
 663         char buf2[100];
 664         int ret = 0;
 665
 666         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 667                       version < bcachefs_metadata_version_min) ||
 668                      version >= bcachefs_metadata_version_max,
 669                      BTREE_ERR_FATAL, c, ca, b, i,
 670                      "unsupported bset version");
 671
 672         if (btree_err_on(version < c->sb.version_min,
 673                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 674                          "bset version %u older than superblock version_min %u",
 675                          version, c->sb.version_min)) {
 676                 mutex_lock(&c->sb_lock);
 677                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 678                 bch2_write_super(c);
 679                 mutex_unlock(&c->sb_lock);
 680         }
 681
 682         if (btree_err_on(version > c->sb.version,
 683                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 684                          "bset version %u newer than superblock version %u",
 685                          version, c->sb.version)) {
 686                 mutex_lock(&c->sb_lock);
 687                 c->disk_sb.sb->version = cpu_to_le16(version);
 688                 bch2_write_super(c);
 689                 mutex_unlock(&c->sb_lock);
 690         }
 691
 692         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 693                      BTREE_ERR_FATAL, c, ca, b, i,
 694                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 695
 696         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 697                          BTREE_ERR_FIXABLE, c, ca, b, i,
 698                          "bset past end of btree node")) {
 699                 i->u64s = 0;
 700                 return 0;
 701         }
 702
 703         btree_err_on(b->written && !i->u64s,
 704                      BTREE_ERR_FIXABLE, c, ca, b, i,
 705                      "empty bset");
 706
 707         if (!b->written) {
 708                 struct btree_node *bn =
 709                         container_of(i, struct btree_node, keys);
 710                 /* These indicate that we read the wrong btree node: */
 711
 712                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 713                         struct bch_btree_ptr_v2 *bp =
 714                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 715
 716                         /* XXX endianness */
 717                         btree_err_on(bp->seq != bn->keys.seq,
 718                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 719                                      "incorrect sequence number (wrong btree node)");
 720                 }
 721
 722                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 723                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 724                              "incorrect btree id");
 725
 726                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 727                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 728                              "incorrect level");
 729
 730                 if (!write)
 731                         compat_btree_node(b->c.level, b->c.btree_id, version,
 732                                           BSET_BIG_ENDIAN(i), write, bn);
 733
 734                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 735                         struct bch_btree_ptr_v2 *bp =
 736                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 737
 738                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 739                                 b->data->min_key = bp->min_key;
 740                                 b->data->max_key = b->key.k.p;
 741                         }
 742
 743                         btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
 744                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 745                                      "incorrect min_key: got %s should be %s",
 746                                      (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
 747                                      (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
 748                 }
 749
 750                 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
 751                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 752                              "incorrect max key %s",
 753                              (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
 754
 755                 if (write)
 756                         compat_btree_node(b->c.level, b->c.btree_id, version,
 757                                           BSET_BIG_ENDIAN(i), write, bn);
 758
 759                 err = bch2_bkey_format_validate(&bn->format);
 760                 btree_err_on(err,
 761                              BTREE_ERR_FATAL, c, ca, b, i,
 762                              "invalid bkey format: %s", err);
 763
 764                 compat_bformat(b->c.level, b->c.btree_id, version,
 765                                BSET_BIG_ENDIAN(i), write,
 766                                &bn->format);
 767         }
 768 fsck_err:
 769         return ret;
 770 }
 771
 772 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 773                          struct bset *i, unsigned *whiteout_u64s,
 774                          int write, bool have_retry)
 775 {
 776         unsigned version = le16_to_cpu(i->version);
 777         struct bkey_packed *k, *prev = NULL;
 778         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 779                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 780         int ret = 0;
 781
 782         for (k = i->start;
 783              k != vstruct_last(i);) {
 784                 struct bkey_s u;
 785                 struct bkey tmp;
 786                 const char *invalid;
 787
 788                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 789                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 790                                  "key extends past end of bset")) {
 791                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 792                         break;
 793                 }
 794
 795                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 796                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 797                                  "invalid bkey format %u", k->format)) {
 798                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 799                         memmove_u64s_down(k, bkey_next(k),
 800                                           (u64 *) vstruct_end(i) - (u64 *) k);
 801                         continue;
 802                 }
 803
 804                 /* XXX: validate k->u64s */
 805                 if (!write)
 806                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 807                                     BSET_BIG_ENDIAN(i), write,
 808                                     &b->format, k);
 809
 810                 u = __bkey_disassemble(b, k, &tmp);
 811
 812                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 813                         (!updated_range ?  bch2_bkey_in_btree_node(b, u.s_c) : NULL) ?:
 814                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 815                 if (invalid) {
 816                         char buf[160];
 817
 818                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 819                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 820                                   "invalid bkey: %s\n%s", invalid, buf);
 821
 822                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 823                         memmove_u64s_down(k, bkey_next(k),
 824                                           (u64 *) vstruct_end(i) - (u64 *) k);
 825                         continue;
 826                 }
 827
 828                 if (write)
 829                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 830                                     BSET_BIG_ENDIAN(i), write,
 831                                     &b->format, k);
 832
 833                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 834                         char buf1[80];
 835                         char buf2[80];
 836                         struct bkey up = bkey_unpack_key(b, prev);
 837
 838                         bch2_bkey_to_text(&PBUF(buf1), &up);
 839                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 840
 841                         bch2_dump_bset(c, b, i, 0);
 842
 843                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 844                                       "keys out of order: %s > %s",
 845                                       buf1, buf2)) {
 846                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 847                                 memmove_u64s_down(k, bkey_next(k),
 848                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 849                                 continue;
 850                         }
 851                 }
 852
 853                 prev = k;
 854                 k = bkey_next(k);
 855         }
 856 fsck_err:
 857         return ret;
 858 }
 859
 860 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 861                               struct btree *b, bool have_retry)
 862 {
 863         struct btree_node_entry *bne;
 864         struct sort_iter *iter;
 865         struct btree_node *sorted;
 866         struct bkey_packed *k;
 867         struct bch_extent_ptr *ptr;
 868         struct bset *i;
 869         bool used_mempool, blacklisted;
 870         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 871                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 872         unsigned u64s;
 873         unsigned nonblacklisted_written = 0;
 874         int ret, retry_read = 0, write = READ;
 875
 876         b->version_ondisk = U16_MAX;
 877
 878         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 879         sort_iter_init(iter, b);
 880         iter->size = (btree_blocks(c) + 1) * 2;
 881
 882         if (bch2_meta_read_fault("btree"))
 883                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 884                           "dynamic fault");
 885
 886         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 887                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 888                      "bad magic");
 889
 890         btree_err_on(!b->data->keys.seq,
 891                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 892                      "bad btree header");
 893
 894         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 895                 struct bch_btree_ptr_v2 *bp =
 896                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 897
 898                 btree_err_on(b->data->keys.seq != bp->seq,
 899                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 900                              "got wrong btree node (seq %llx want %llx)",
 901                              b->data->keys.seq, bp->seq);
 902         }
 903
 904         while (b->written < c->opts.btree_node_size) {
 905                 unsigned sectors, whiteout_u64s = 0;
 906                 struct nonce nonce;
 907                 struct bch_csum csum;
 908                 bool first = !b->written;
 909
 910                 if (!b->written) {
 911                         i = &b->data->keys;
 912
 913                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 914                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 915                                      "unknown checksum type %llu",
 916                                      BSET_CSUM_TYPE(i));
 917
 918                         nonce = btree_nonce(i, b->written << 9);
 919                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 920
 921                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 922                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 923                                      "invalid checksum");
 924
 925                         bset_encrypt(c, i, b->written << 9);
 926
 927                         btree_err_on(btree_node_is_extents(b) &&
 928                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 929                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 930                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 931
 932                         sectors = vstruct_sectors(b->data, c->block_bits);
 933                 } else {
 934                         bne = write_block(b);
 935                         i = &bne->keys;
 936
 937                         if (i->seq != b->data->keys.seq)
 938                                 break;
 939
 940                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 941                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 942                                      "unknown checksum type %llu",
 943                                      BSET_CSUM_TYPE(i));
 944
 945                         nonce = btree_nonce(i, b->written << 9);
 946                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 947
 948                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 949                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 950                                      "invalid checksum");
 951
 952                         bset_encrypt(c, i, b->written << 9);
 953
 954                         sectors = vstruct_sectors(bne, c->block_bits);
 955                 }
 956
 957                 b->version_ondisk = min(b->version_ondisk,
 958                                         le16_to_cpu(i->version));
 959
 960                 ret = validate_bset(c, ca, b, i, sectors,
 961                                     READ, have_retry);
 962                 if (ret)
 963                         goto fsck_err;
 964
 965                 if (!b->written)
 966                         btree_node_set_format(b, b->data->format);
 967
 968                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 969                                     READ, have_retry);
 970                 if (ret)
 971                         goto fsck_err;
 972
 973                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 974
 975                 b->written += sectors;
 976
 977                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 978                                         le64_to_cpu(i->journal_seq),
 979                                         true);
 980
 981                 btree_err_on(blacklisted && first,
 982                              BTREE_ERR_FIXABLE, c, ca, b, i,
 983                              "first btree node bset has blacklisted journal seq");
 984                 if (blacklisted && !first)
 985                         continue;
 986
 987                 sort_iter_add(iter, i->start,
 988                               vstruct_idx(i, whiteout_u64s));
 989
 990                 sort_iter_add(iter,
 991                               vstruct_idx(i, whiteout_u64s),
 992                               vstruct_last(i));
 993
 994                 nonblacklisted_written = b->written;
 995         }
 996
 997         for (bne = write_block(b);
 998              bset_byte_offset(b, bne) < btree_bytes(c);
 999              bne = (void *) bne + block_bytes(c))
1000                 btree_err_on(bne->keys.seq == b->data->keys.seq &&
1001                              !bch2_journal_seq_is_blacklisted(c,
1002                                         le64_to_cpu(bne->keys.journal_seq),
1003                                         true),
1004                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1005                              "found bset signature after last bset");
1006
1007         /*
1008          * Blacklisted bsets are those that were written after the most recent
1009          * (flush) journal write. Since there wasn't a flush, they may not have
1010          * made it to all devices - which means we shouldn't write new bsets
1011          * after them, as that could leave a gap and then reads from that device
1012          * wouldn't find all the bsets in that btree node - which means it's
1013          * important that we start writing new bsets after the most recent _non_
1014          * blacklisted bset:
1015          */
1016         b->written = nonblacklisted_written;
1017
1018         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1019         sorted->keys.u64s = 0;
1020
1021         set_btree_bset(b, b->set, &b->data->keys);
1022
1023         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1024
1025         u64s = le16_to_cpu(sorted->keys.u64s);
1026         *sorted = *b->data;
1027         sorted->keys.u64s = cpu_to_le16(u64s);
1028         swap(sorted, b->data);
1029         set_btree_bset(b, b->set, &b->data->keys);
1030         b->nsets = 1;
1031
1032         BUG_ON(b->nr.live_u64s != u64s);
1033
1034         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1035
1036         if (updated_range)
1037                 bch2_btree_node_drop_keys_outside_node(b);
1038
1039         i = &b->data->keys;
1040         for (k = i->start; k != vstruct_last(i);) {
1041                 struct bkey tmp;
1042                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1043                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
1044
1045                 if (invalid ||
1046                     (bch2_inject_invalid_keys &&
1047                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1048                         char buf[160];
1049
1050                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
1051                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
1052                                   "invalid bkey %s: %s", buf, invalid);
1053
1054                         btree_keys_account_key_drop(&b->nr, 0, k);
1055
1056                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1057                         memmove_u64s_down(k, bkey_next(k),
1058                                           (u64 *) vstruct_end(i) - (u64 *) k);
1059                         set_btree_bset_end(b, b->set);
1060                         continue;
1061                 }
1062
1063                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1064                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1065
1066                         bp.v->mem_ptr = 0;
1067                 }
1068
1069                 k = bkey_next(k);
1070         }
1071
1072         bch2_bset_build_aux_tree(b, b->set, false);
1073
1074         set_needs_whiteout(btree_bset_first(b), true);
1075
1076         btree_node_reset_sib_u64s(b);
1077
1078         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1079                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1080
1081                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1082                         set_btree_node_need_rewrite(b);
1083         }
1084 out:
1085         mempool_free(iter, &c->fill_iter);
1086         return retry_read;
1087 fsck_err:
1088         if (ret == BTREE_RETRY_READ) {
1089                 retry_read = 1;
1090         } else {
1091                 bch2_inconsistent_error(c);
1092                 set_btree_node_read_error(b);
1093         }
1094         goto out;
1095 }
1096
1097 static void btree_node_read_work(struct work_struct *work)
1098 {
1099         struct btree_read_bio *rb =
1100                 container_of(work, struct btree_read_bio, work);
1101         struct bch_fs *c        = rb->c;
1102         struct btree *b         = rb->b;
1103         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1104         struct bio *bio         = &rb->bio;
1105         struct bch_io_failures failed = { .nr = 0 };
1106         char buf[200];
1107         struct printbuf out;
1108         bool saw_error = false;
1109         bool can_retry;
1110
1111         goto start;
1112         while (1) {
1113                 bch_info(c, "retrying read");
1114                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1115                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1116                 bio_reset(bio);
1117                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1118                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1119                 bio->bi_iter.bi_size    = btree_bytes(c);
1120
1121                 if (rb->have_ioref) {
1122                         bio_set_dev(bio, ca->disk_sb.bdev);
1123                         submit_bio_wait(bio);
1124                 } else {
1125                         bio->bi_status = BLK_STS_REMOVED;
1126                 }
1127 start:
1128                 out = PBUF(buf);
1129                 btree_pos_to_text(&out, c, b);
1130                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1131                                    bch2_blk_status_to_str(bio->bi_status), buf);
1132                 if (rb->have_ioref)
1133                         percpu_ref_put(&ca->io_ref);
1134                 rb->have_ioref = false;
1135
1136                 bch2_mark_io_failure(&failed, &rb->pick);
1137
1138                 can_retry = bch2_bkey_pick_read_device(c,
1139                                 bkey_i_to_s_c(&b->key),
1140                                 &failed, &rb->pick) > 0;
1141
1142                 if (!bio->bi_status &&
1143                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1144                         break;
1145
1146                 saw_error = true;
1147
1148                 if (!can_retry) {
1149                         set_btree_node_read_error(b);
1150                         break;
1151                 }
1152         }
1153
1154         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1155                                rb->start_time);
1156         bio_put(&rb->bio);
1157
1158         if (saw_error && !btree_node_read_error(b))
1159                 bch2_btree_node_rewrite_async(c, b);
1160
1161         clear_btree_node_read_in_flight(b);
1162         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1163 }
1164
1165 static void btree_node_read_endio(struct bio *bio)
1166 {
1167         struct btree_read_bio *rb =
1168                 container_of(bio, struct btree_read_bio, bio);
1169         struct bch_fs *c        = rb->c;
1170
1171         if (rb->have_ioref) {
1172                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1173                 bch2_latency_acct(ca, rb->start_time, READ);
1174         }
1175
1176         queue_work(c->io_complete_wq, &rb->work);
1177 }
1178
1179 struct btree_node_read_all {
1180         struct closure          cl;
1181         struct bch_fs           *c;
1182         struct btree            *b;
1183         unsigned                nr;
1184         void                    *buf[BCH_REPLICAS_MAX];
1185         struct bio              *bio[BCH_REPLICAS_MAX];
1186         int                     err[BCH_REPLICAS_MAX];
1187 };
1188
1189 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1190 {
1191         struct btree_node *bn = data;
1192         struct btree_node_entry *bne;
1193         unsigned offset = 0;
1194
1195         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1196                 return 0;
1197
1198         while (offset < c->opts.btree_node_size) {
1199                 if (!offset) {
1200                         offset += vstruct_sectors(bn, c->block_bits);
1201                 } else {
1202                         bne = data + (offset << 9);
1203                         if (bne->keys.seq != bn->keys.seq)
1204                                 break;
1205                         offset += vstruct_sectors(bne, c->block_bits);
1206                 }
1207         }
1208
1209         return offset;
1210 }
1211
1212 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1213 {
1214         struct btree_node *bn = data;
1215         struct btree_node_entry *bne;
1216
1217         if (!offset)
1218                 return false;
1219
1220         while (offset < c->opts.btree_node_size) {
1221                 bne = data + (offset << 9);
1222                 if (bne->keys.seq == bn->keys.seq)
1223                         return true;
1224                 offset++;
1225         }
1226
1227         return false;
1228         return offset;
1229 }
1230
1231 static void btree_node_read_all_replicas_done(struct closure *cl)
1232 {
1233         struct btree_node_read_all *ra =
1234                 container_of(cl, struct btree_node_read_all, cl);
1235         struct bch_fs *c = ra->c;
1236         struct btree *b = ra->b;
1237         bool dump_bset_maps = false;
1238         bool have_retry = false;
1239         int ret = 0, best = -1, write = READ;
1240         unsigned i, written, written2;
1241         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1242                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1243
1244         for (i = 0; i < ra->nr; i++) {
1245                 struct btree_node *bn = ra->buf[i];
1246
1247                 if (ra->err[i])
1248                         continue;
1249
1250                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1251                     (seq && seq != bn->keys.seq))
1252                         continue;
1253
1254                 if (best < 0) {
1255                         best = i;
1256                         written = btree_node_sectors_written(c, bn);
1257                         continue;
1258                 }
1259
1260                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1261                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1262                                  "btree node sectors written mismatch: %u != %u",
1263                                  written, written2) ||
1264                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1265                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1266                                  "found bset signature after last bset") ||
1267                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1268                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1269                                  "btree node replicas content mismatch"))
1270                         dump_bset_maps = true;
1271
1272                 if (written2 > written) {
1273                         written = written2;
1274                         best = i;
1275                 }
1276         }
1277 fsck_err:
1278         if (dump_bset_maps) {
1279                 for (i = 0; i < ra->nr; i++) {
1280                         char buf[200];
1281                         struct printbuf out = PBUF(buf);
1282                         struct btree_node *bn = ra->buf[i];
1283                         struct btree_node_entry *bne = NULL;
1284                         unsigned offset = 0, sectors;
1285                         bool gap = false;
1286
1287                         if (ra->err[i])
1288                                 continue;
1289
1290                         while (offset < c->opts.btree_node_size) {
1291                                 if (!offset) {
1292                                         sectors = vstruct_sectors(bn, c->block_bits);
1293                                 } else {
1294                                         bne = ra->buf[i] + (offset << 9);
1295                                         if (bne->keys.seq != bn->keys.seq)
1296                                                 break;
1297                                         sectors = vstruct_sectors(bne, c->block_bits);
1298                                 }
1299
1300                                 pr_buf(&out, " %u-%u", offset, offset + sectors);
1301                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1302                                                         le64_to_cpu(bne->keys.journal_seq), false))
1303                                         pr_buf(&out, "*");
1304                                 offset += sectors;
1305                         }
1306
1307                         while (offset < c->opts.btree_node_size) {
1308                                 bne = ra->buf[i] + (offset << 9);
1309                                 if (bne->keys.seq == bn->keys.seq) {
1310                                         if (!gap)
1311                                                 pr_buf(&out, " GAP");
1312                                         gap = true;
1313
1314                                         sectors = vstruct_sectors(bne, c->block_bits);
1315                                         pr_buf(&out, " %u-%u", offset, offset + sectors);
1316                                         if (bch2_journal_seq_is_blacklisted(c,
1317                                                         le64_to_cpu(bne->keys.journal_seq), false))
1318                                                 pr_buf(&out, "*");
1319                                 }
1320                                 offset++;
1321                         }
1322
1323                         bch_err(c, "replica %u:%s", i, buf);
1324                 }
1325         }
1326
1327         if (best >= 0) {
1328                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1329                 ret = bch2_btree_node_read_done(c, NULL, b, false);
1330         } else {
1331                 ret = -1;
1332         }
1333
1334         if (ret)
1335                 set_btree_node_read_error(b);
1336
1337         for (i = 0; i < ra->nr; i++) {
1338                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1339                 bio_put(ra->bio[i]);
1340         }
1341
1342         closure_debug_destroy(&ra->cl);
1343         kfree(ra);
1344
1345         clear_btree_node_read_in_flight(b);
1346         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1347 }
1348
1349 static void btree_node_read_all_replicas_endio(struct bio *bio)
1350 {
1351         struct btree_read_bio *rb =
1352                 container_of(bio, struct btree_read_bio, bio);
1353         struct bch_fs *c        = rb->c;
1354         struct btree_node_read_all *ra = rb->ra;
1355
1356         if (rb->have_ioref) {
1357                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1358                 bch2_latency_acct(ca, rb->start_time, READ);
1359         }
1360
1361         ra->err[rb->idx] = bio->bi_status;
1362         closure_put(&ra->cl);
1363 }
1364
1365 /*
1366  * XXX This allocates multiple times from the same mempools, and can deadlock
1367  * under sufficient memory pressure (but is only a debug path)
1368  */
1369 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1370 {
1371         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1372         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1373         const union bch_extent_entry *entry;
1374         struct extent_ptr_decoded pick;
1375         struct btree_node_read_all *ra;
1376         unsigned i;
1377
1378         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1379         if (!ra)
1380                 return -ENOMEM;
1381
1382         closure_init(&ra->cl, NULL);
1383         ra->c   = c;
1384         ra->b   = b;
1385         ra->nr  = bch2_bkey_nr_ptrs(k);
1386
1387         for (i = 0; i < ra->nr; i++) {
1388                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1389                 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1390                                                                   btree_bytes(c)),
1391                                               &c->btree_bio);
1392         }
1393
1394         i = 0;
1395         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1396                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1397                 struct btree_read_bio *rb =
1398                         container_of(ra->bio[i], struct btree_read_bio, bio);
1399                 rb->c                   = c;
1400                 rb->b                   = b;
1401                 rb->ra                  = ra;
1402                 rb->start_time          = local_clock();
1403                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1404                 rb->idx                 = i;
1405                 rb->pick                = pick;
1406                 rb->bio.bi_opf          = REQ_OP_READ|REQ_SYNC|REQ_META;
1407                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1408                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1409                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1410
1411                 if (rb->have_ioref) {
1412                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1413                                      bio_sectors(&rb->bio));
1414                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1415
1416                         closure_get(&ra->cl);
1417                         submit_bio(&rb->bio);
1418                 } else {
1419                         ra->err[i] = BLK_STS_REMOVED;
1420                 }
1421
1422                 i++;
1423         }
1424
1425         if (sync) {
1426                 closure_sync(&ra->cl);
1427                 btree_node_read_all_replicas_done(&ra->cl);
1428         } else {
1429                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1430                             c->io_complete_wq);
1431         }
1432
1433         return 0;
1434 }
1435
1436 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1437                           bool sync)
1438 {
1439         struct extent_ptr_decoded pick;
1440         struct btree_read_bio *rb;
1441         struct bch_dev *ca;
1442         struct bio *bio;
1443         char buf[200];
1444         int ret;
1445
1446         btree_pos_to_text(&PBUF(buf), c, b);
1447         trace_btree_read(c, b);
1448
1449         if (bch2_verify_all_btree_replicas &&
1450             !btree_node_read_all_replicas(c, b, sync))
1451                 return;
1452
1453         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1454                                          NULL, &pick);
1455         if (bch2_fs_fatal_err_on(ret <= 0, c,
1456                         "btree node read error: no device to read from\n"
1457                         " at %s", buf)) {
1458                 set_btree_node_read_error(b);
1459                 return;
1460         }
1461
1462         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1463
1464         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1465                                                    btree_bytes(c)),
1466                                &c->btree_bio);
1467         rb = container_of(bio, struct btree_read_bio, bio);
1468         rb->c                   = c;
1469         rb->b                   = b;
1470         rb->ra                  = NULL;
1471         rb->start_time          = local_clock();
1472         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1473         rb->pick                = pick;
1474         INIT_WORK(&rb->work, btree_node_read_work);
1475         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1476         bio->bi_iter.bi_sector  = pick.ptr.offset;
1477         bio->bi_end_io          = btree_node_read_endio;
1478         bch2_bio_map(bio, b->data, btree_bytes(c));
1479
1480         if (rb->have_ioref) {
1481                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1482                              bio_sectors(bio));
1483                 bio_set_dev(bio, ca->disk_sb.bdev);
1484
1485                 if (sync) {
1486                         submit_bio_wait(bio);
1487
1488                         btree_node_read_work(&rb->work);
1489                 } else {
1490                         submit_bio(bio);
1491                 }
1492         } else {
1493                 bio->bi_status = BLK_STS_REMOVED;
1494
1495                 if (sync)
1496                         btree_node_read_work(&rb->work);
1497                 else
1498                         queue_work(c->io_complete_wq, &rb->work);
1499         }
1500 }
1501
1502 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1503                         const struct bkey_i *k, unsigned level)
1504 {
1505         struct closure cl;
1506         struct btree *b;
1507         int ret;
1508
1509         closure_init_stack(&cl);
1510
1511         do {
1512                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1513                 closure_sync(&cl);
1514         } while (ret);
1515
1516         b = bch2_btree_node_mem_alloc(c);
1517         bch2_btree_cache_cannibalize_unlock(c);
1518
1519         BUG_ON(IS_ERR(b));
1520
1521         bkey_copy(&b->key, k);
1522         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1523
1524         set_btree_node_read_in_flight(b);
1525
1526         bch2_btree_node_read(c, b, true);
1527
1528         if (btree_node_read_error(b)) {
1529                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1530
1531                 mutex_lock(&c->btree_cache.lock);
1532                 list_move(&b->list, &c->btree_cache.freeable);
1533                 mutex_unlock(&c->btree_cache.lock);
1534
1535                 ret = -EIO;
1536                 goto err;
1537         }
1538
1539         bch2_btree_set_root_for_read(c, b);
1540 err:
1541         six_unlock_write(&b->c.lock);
1542         six_unlock_intent(&b->c.lock);
1543
1544         return ret;
1545 }
1546
1547 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1548                               struct btree_write *w)
1549 {
1550         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1551
1552         do {
1553                 old = new = v;
1554                 if (!(old & 1))
1555                         break;
1556
1557                 new &= ~1UL;
1558         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1559
1560         if (old & 1)
1561                 closure_put(&((struct btree_update *) new)->cl);
1562
1563         bch2_journal_pin_drop(&c->journal, &w->journal);
1564 }
1565
1566 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1567 {
1568         struct btree_write *w = btree_prev_write(b);
1569         unsigned long old, new, v;
1570
1571         bch2_btree_complete_write(c, b, w);
1572
1573         v = READ_ONCE(b->flags);
1574         do {
1575                 old = new = v;
1576
1577                 if (old & (1U << BTREE_NODE_need_write))
1578                         goto do_write;
1579
1580                 new &= ~(1U << BTREE_NODE_write_in_flight);
1581         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1582
1583         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1584         return;
1585
1586 do_write:
1587         six_lock_read(&b->c.lock, NULL, NULL);
1588         v = READ_ONCE(b->flags);
1589         do {
1590                 old = new = v;
1591
1592                 if ((old & (1U << BTREE_NODE_dirty)) &&
1593                     (old & (1U << BTREE_NODE_need_write)) &&
1594                     !(old & (1U << BTREE_NODE_never_write)) &&
1595                     btree_node_may_write(b)) {
1596                         new &= ~(1U << BTREE_NODE_dirty);
1597                         new &= ~(1U << BTREE_NODE_need_write);
1598                         new |=  (1U << BTREE_NODE_write_in_flight);
1599                         new |=  (1U << BTREE_NODE_just_written);
1600                         new ^=  (1U << BTREE_NODE_write_idx);
1601                 } else {
1602                         new &= ~(1U << BTREE_NODE_write_in_flight);
1603                 }
1604         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1605
1606         if (new & (1U << BTREE_NODE_write_in_flight))
1607                 __bch2_btree_node_write(c, b, true);
1608
1609         six_unlock_read(&b->c.lock);
1610 }
1611
1612 static void bch2_btree_node_write_error(struct bch_fs *c,
1613                                         struct btree_write_bio *wbio)
1614 {
1615         struct btree *b         = wbio->wbio.bio.bi_private;
1616         struct bkey_buf k;
1617         struct bch_extent_ptr *ptr;
1618         struct btree_trans trans;
1619         struct btree_iter *iter;
1620         int ret;
1621
1622         bch2_bkey_buf_init(&k);
1623         bch2_trans_init(&trans, c, 0, 0);
1624
1625         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1626                                         BTREE_MAX_DEPTH, b->c.level, 0);
1627 retry:
1628         ret = bch2_btree_iter_traverse(iter);
1629         if (ret)
1630                 goto err;
1631
1632         /* has node been freed? */
1633         if (iter->l[b->c.level].b != b) {
1634                 /* node has been freed: */
1635                 BUG_ON(!btree_node_dying(b));
1636                 goto out;
1637         }
1638
1639         BUG_ON(!btree_node_hashed(b));
1640
1641         bch2_bkey_buf_copy(&k, c, &b->key);
1642
1643         bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1644                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1645
1646         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1647                 goto err;
1648
1649         ret = bch2_btree_node_update_key(&trans, iter, b, k.k);
1650         if (ret == -EINTR)
1651                 goto retry;
1652         if (ret)
1653                 goto err;
1654 out:
1655         bch2_trans_iter_put(&trans, iter);
1656         bch2_trans_exit(&trans);
1657         bch2_bkey_buf_exit(&k, c);
1658         bio_put(&wbio->wbio.bio);
1659         btree_node_write_done(c, b);
1660         return;
1661 err:
1662         set_btree_node_noevict(b);
1663         bch2_fs_fatal_error(c, "fatal error writing btree node");
1664         goto out;
1665 }
1666
1667 void bch2_btree_write_error_work(struct work_struct *work)
1668 {
1669         struct bch_fs *c = container_of(work, struct bch_fs,
1670                                         btree_write_error_work);
1671         struct bio *bio;
1672
1673         while (1) {
1674                 spin_lock_irq(&c->btree_write_error_lock);
1675                 bio = bio_list_pop(&c->btree_write_error_list);
1676                 spin_unlock_irq(&c->btree_write_error_lock);
1677
1678                 if (!bio)
1679                         break;
1680
1681                 bch2_btree_node_write_error(c,
1682                         container_of(bio, struct btree_write_bio, wbio.bio));
1683         }
1684 }
1685
1686 static void btree_node_write_work(struct work_struct *work)
1687 {
1688         struct btree_write_bio *wbio =
1689                 container_of(work, struct btree_write_bio, work);
1690         struct bch_fs *c        = wbio->wbio.c;
1691         struct btree *b         = wbio->wbio.bio.bi_private;
1692
1693         btree_bounce_free(c,
1694                 wbio->bytes,
1695                 wbio->wbio.used_mempool,
1696                 wbio->data);
1697
1698         if (wbio->wbio.failed.nr) {
1699                 unsigned long flags;
1700
1701                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1702                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1703                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1704
1705                 queue_work(c->btree_error_wq, &c->btree_write_error_work);
1706                 return;
1707         }
1708
1709         bio_put(&wbio->wbio.bio);
1710         btree_node_write_done(c, b);
1711 }
1712
1713 static void btree_node_write_endio(struct bio *bio)
1714 {
1715         struct bch_write_bio *wbio      = to_wbio(bio);
1716         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1717         struct bch_write_bio *orig      = parent ?: wbio;
1718         struct bch_fs *c                = wbio->c;
1719         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1720         unsigned long flags;
1721
1722         if (wbio->have_ioref)
1723                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1724
1725         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1726                                bch2_blk_status_to_str(bio->bi_status)) ||
1727             bch2_meta_write_fault("btree")) {
1728                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1729                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1730                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1731         }
1732
1733         if (wbio->have_ioref)
1734                 percpu_ref_put(&ca->io_ref);
1735
1736         if (parent) {
1737                 bio_put(bio);
1738                 bio_endio(&parent->bio);
1739         } else {
1740                 struct btree_write_bio *wb =
1741                         container_of(orig, struct btree_write_bio, wbio);
1742
1743                 INIT_WORK(&wb->work, btree_node_write_work);
1744                 queue_work(c->io_complete_wq, &wb->work);
1745         }
1746 }
1747
1748 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1749                                    struct bset *i, unsigned sectors)
1750 {
1751         unsigned whiteout_u64s = 0;
1752         int ret;
1753
1754         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1755                 return -1;
1756
1757         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1758                 validate_bset(c, NULL, b, i, sectors, WRITE, false);
1759         if (ret) {
1760                 bch2_inconsistent_error(c);
1761                 dump_stack();
1762         }
1763
1764         return ret;
1765 }
1766
1767 static void btree_write_submit(struct work_struct *work)
1768 {
1769         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1770
1771         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &wbio->key);
1772 }
1773
1774 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, bool already_started)
1775 {
1776         struct btree_write_bio *wbio;
1777         struct bset_tree *t;
1778         struct bset *i;
1779         struct btree_node *bn = NULL;
1780         struct btree_node_entry *bne = NULL;
1781         struct bch_extent_ptr *ptr;
1782         struct sort_iter sort_iter;
1783         struct nonce nonce;
1784         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1785         u64 seq = 0;
1786         bool used_mempool;
1787         unsigned long old, new;
1788         bool validate_before_checksum = false;
1789         void *data;
1790
1791         if (already_started)
1792                 goto do_write;
1793
1794         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1795                 return;
1796
1797         /*
1798          * We may only have a read lock on the btree node - the dirty bit is our
1799          * "lock" against racing with other threads that may be trying to start
1800          * a write, we do a write iff we clear the dirty bit. Since setting the
1801          * dirty bit requires a write lock, we can't race with other threads
1802          * redirtying it:
1803          */
1804         do {
1805                 old = new = READ_ONCE(b->flags);
1806
1807                 if (!(old & (1 << BTREE_NODE_dirty)))
1808                         return;
1809
1810                 if (!btree_node_may_write(b))
1811                         return;
1812
1813                 if (old & (1 << BTREE_NODE_never_write))
1814                         return;
1815
1816                 BUG_ON(old & (1 << BTREE_NODE_write_in_flight));
1817
1818                 new &= ~(1 << BTREE_NODE_dirty);
1819                 new &= ~(1 << BTREE_NODE_need_write);
1820                 new |=  (1 << BTREE_NODE_write_in_flight);
1821                 new |=  (1 << BTREE_NODE_just_written);
1822                 new ^=  (1 << BTREE_NODE_write_idx);
1823         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1824
1825         if (new & (1U << BTREE_NODE_need_write))
1826                 return;
1827 do_write:
1828         atomic_dec(&c->btree_cache.dirty);
1829
1830         BUG_ON(btree_node_fake(b));
1831         BUG_ON((b->will_make_reachable != 0) != !b->written);
1832
1833         BUG_ON(b->written >= c->opts.btree_node_size);
1834         BUG_ON(b->written & (c->opts.block_size - 1));
1835         BUG_ON(bset_written(b, btree_bset_last(b)));
1836         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1837         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1838
1839         bch2_sort_whiteouts(c, b);
1840
1841         sort_iter_init(&sort_iter, b);
1842
1843         bytes = !b->written
1844                 ? sizeof(struct btree_node)
1845                 : sizeof(struct btree_node_entry);
1846
1847         bytes += b->whiteout_u64s * sizeof(u64);
1848
1849         for_each_bset(b, t) {
1850                 i = bset(b, t);
1851
1852                 if (bset_written(b, i))
1853                         continue;
1854
1855                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1856                 sort_iter_add(&sort_iter,
1857                               btree_bkey_first(b, t),
1858                               btree_bkey_last(b, t));
1859                 seq = max(seq, le64_to_cpu(i->journal_seq));
1860         }
1861
1862         BUG_ON(b->written && !seq);
1863
1864         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1865         bytes += 8;
1866
1867         /* buffer must be a multiple of the block size */
1868         bytes = round_up(bytes, block_bytes(c));
1869
1870         data = btree_bounce_alloc(c, bytes, &used_mempool);
1871
1872         if (!b->written) {
1873                 bn = data;
1874                 *bn = *b->data;
1875                 i = &bn->keys;
1876         } else {
1877                 bne = data;
1878                 bne->keys = b->data->keys;
1879                 i = &bne->keys;
1880         }
1881
1882         i->journal_seq  = cpu_to_le64(seq);
1883         i->u64s         = 0;
1884
1885         sort_iter_add(&sort_iter,
1886                       unwritten_whiteouts_start(c, b),
1887                       unwritten_whiteouts_end(c, b));
1888         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1889
1890         b->whiteout_u64s = 0;
1891
1892         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1893         le16_add_cpu(&i->u64s, u64s);
1894
1895         set_needs_whiteout(i, false);
1896
1897         /* do we have data to write? */
1898         if (b->written && !i->u64s)
1899                 goto nowrite;
1900
1901         bytes_to_write = vstruct_end(i) - data;
1902         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1903
1904         memset(data + bytes_to_write, 0,
1905                (sectors_to_write << 9) - bytes_to_write);
1906
1907         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1908         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1909         BUG_ON(i->seq != b->data->keys.seq);
1910
1911         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1912                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1913                 : cpu_to_le16(c->sb.version);
1914         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1915
1916         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1917                 validate_before_checksum = true;
1918
1919         /* validate_bset will be modifying: */
1920         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1921                 validate_before_checksum = true;
1922
1923         /* if we're going to be encrypting, check metadata validity first: */
1924         if (validate_before_checksum &&
1925             validate_bset_for_write(c, b, i, sectors_to_write))
1926                 goto err;
1927
1928         bset_encrypt(c, i, b->written << 9);
1929
1930         nonce = btree_nonce(i, b->written << 9);
1931
1932         if (bn)
1933                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1934         else
1935                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1936
1937         /* if we're not encrypting, check metadata after checksumming: */
1938         if (!validate_before_checksum &&
1939             validate_bset_for_write(c, b, i, sectors_to_write))
1940                 goto err;
1941
1942         /*
1943          * We handle btree write errors by immediately halting the journal -
1944          * after we've done that, we can't issue any subsequent btree writes
1945          * because they might have pointers to new nodes that failed to write.
1946          *
1947          * Furthermore, there's no point in doing any more btree writes because
1948          * with the journal stopped, we're never going to update the journal to
1949          * reflect that those writes were done and the data flushed from the
1950          * journal:
1951          *
1952          * Also on journal error, the pending write may have updates that were
1953          * never journalled (interior nodes, see btree_update_nodes_written()) -
1954          * it's critical that we don't do the write in that case otherwise we
1955          * will have updates visible that weren't in the journal:
1956          *
1957          * Make sure to update b->written so bch2_btree_init_next() doesn't
1958          * break:
1959          */
1960         if (bch2_journal_error(&c->journal) ||
1961             c->opts.nochanges)
1962                 goto err;
1963
1964         trace_btree_write(b, bytes_to_write, sectors_to_write);
1965
1966         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1967                                 buf_pages(data, sectors_to_write << 9),
1968                                 &c->btree_bio),
1969                             struct btree_write_bio, wbio.bio);
1970         wbio_init(&wbio->wbio.bio);
1971         wbio->data                      = data;
1972         wbio->bytes                     = bytes;
1973         wbio->wbio.c                    = c;
1974         wbio->wbio.used_mempool         = used_mempool;
1975         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1976         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1977         wbio->wbio.bio.bi_private       = b;
1978
1979         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1980
1981         /*
1982          * If we're appending to a leaf node, we don't technically need FUA -
1983          * this write just needs to be persisted before the next journal write,
1984          * which will be marked FLUSH|FUA.
1985          *
1986          * Similarly if we're writing a new btree root - the pointer is going to
1987          * be in the next journal entry.
1988          *
1989          * But if we're writing a new btree node (that isn't a root) or
1990          * appending to a non leaf btree node, we need either FUA or a flush
1991          * when we write the parent with the new pointer. FUA is cheaper than a
1992          * flush, and writes appending to leaf nodes aren't blocking anything so
1993          * just make all btree node writes FUA to keep things sane.
1994          */
1995
1996         bkey_copy(&wbio->key, &b->key);
1997
1998         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&wbio->key)), ptr)
1999                 ptr->offset += b->written;
2000
2001         b->written += sectors_to_write;
2002
2003         atomic64_inc(&c->btree_writes_nr);
2004         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
2005
2006         INIT_WORK(&wbio->work, btree_write_submit);
2007         queue_work(c->io_complete_wq, &wbio->work);
2008         return;
2009 err:
2010         set_btree_node_noevict(b);
2011         b->written += sectors_to_write;
2012 nowrite:
2013         btree_bounce_free(c, bytes, used_mempool, data);
2014         btree_node_write_done(c, b);
2015 }
2016
2017 /*
2018  * Work that must be done with write lock held:
2019  */
2020 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2021 {
2022         bool invalidated_iter = false;
2023         struct btree_node_entry *bne;
2024         struct bset_tree *t;
2025
2026         if (!btree_node_just_written(b))
2027                 return false;
2028
2029         BUG_ON(b->whiteout_u64s);
2030
2031         clear_btree_node_just_written(b);
2032
2033         /*
2034          * Note: immediately after write, bset_written() doesn't work - the
2035          * amount of data we had to write after compaction might have been
2036          * smaller than the offset of the last bset.
2037          *
2038          * However, we know that all bsets have been written here, as long as
2039          * we're still holding the write lock:
2040          */
2041
2042         /*
2043          * XXX: decide if we really want to unconditionally sort down to a
2044          * single bset:
2045          */
2046         if (b->nsets > 1) {
2047                 btree_node_sort(c, b, 0, b->nsets, true);
2048                 invalidated_iter = true;
2049         } else {
2050                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2051         }
2052
2053         for_each_bset(b, t)
2054                 set_needs_whiteout(bset(b, t), true);
2055
2056         bch2_btree_verify(c, b);
2057
2058         /*
2059          * If later we don't unconditionally sort down to a single bset, we have
2060          * to ensure this is still true:
2061          */
2062         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2063
2064         bne = want_new_bset(c, b);
2065         if (bne)
2066                 bch2_bset_init_next(c, b, bne);
2067
2068         bch2_btree_build_aux_trees(b);
2069
2070         return invalidated_iter;
2071 }
2072
2073 /*
2074  * Use this one if the node is intent locked:
2075  */
2076 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2077                            enum six_lock_type lock_type_held)
2078 {
2079         if (lock_type_held == SIX_LOCK_intent ||
2080             (lock_type_held == SIX_LOCK_read &&
2081              six_lock_tryupgrade(&b->c.lock))) {
2082                 __bch2_btree_node_write(c, b, false);
2083
2084                 /* don't cycle lock unnecessarily: */
2085                 if (btree_node_just_written(b) &&
2086                     six_trylock_write(&b->c.lock)) {
2087                         bch2_btree_post_write_cleanup(c, b);
2088                         six_unlock_write(&b->c.lock);
2089                 }
2090
2091                 if (lock_type_held == SIX_LOCK_read)
2092                         six_lock_downgrade(&b->c.lock);
2093         } else {
2094                 __bch2_btree_node_write(c, b, false);
2095                 if (lock_type_held == SIX_LOCK_write &&
2096                     btree_node_just_written(b))
2097                         bch2_btree_post_write_cleanup(c, b);
2098         }
2099 }
2100
2101 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2102 {
2103         struct bucket_table *tbl;
2104         struct rhash_head *pos;
2105         struct btree *b;
2106         unsigned i;
2107 restart:
2108         rcu_read_lock();
2109         for_each_cached_btree(b, c, tbl, i, pos)
2110                 if (test_bit(flag, &b->flags)) {
2111                         rcu_read_unlock();
2112                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2113                         goto restart;
2114
2115                 }
2116         rcu_read_unlock();
2117 }
2118
2119 void bch2_btree_flush_all_reads(struct bch_fs *c)
2120 {
2121         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2122 }
2123
2124 void bch2_btree_flush_all_writes(struct bch_fs *c)
2125 {
2126         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2127 }
2128
2129 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
2130 {
2131         struct bucket_table *tbl;
2132         struct rhash_head *pos;
2133         struct btree *b;
2134         unsigned i;
2135
2136         rcu_read_lock();
2137         for_each_cached_btree(b, c, tbl, i, pos) {
2138                 unsigned long flags = READ_ONCE(b->flags);
2139
2140                 if (!(flags & (1 << BTREE_NODE_dirty)))
2141                         continue;
2142
2143                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
2144                        b,
2145                        (flags & (1 << BTREE_NODE_dirty)) != 0,
2146                        (flags & (1 << BTREE_NODE_need_write)) != 0,
2147                        b->c.level,
2148                        b->written,
2149                        !list_empty_careful(&b->write_blocked),
2150                        b->will_make_reachable != 0,
2151                        b->will_make_reachable & 1);
2152         }
2153         rcu_read_unlock();
2154 }