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

   1
   2 #include "bcachefs.h"
   3 #include "bkey_methods.h"
   4 #include "btree_cache.h"
   5 #include "btree_io.h"
   6 #include "btree_iter.h"
   7 #include "btree_locking.h"
   8 #include "btree_update.h"
   9 #include "btree_update_interior.h"
  10 #include "buckets.h"
  11 #include "checksum.h"
  12 #include "debug.h"
  13 #include "error.h"
  14 #include "extents.h"
  15 #include "io.h"
  16 #include "journal.h"
  17 #include "super-io.h"
  18
  19 #include <trace/events/bcachefs.h>
  20
  21 static void verify_no_dups(struct btree *b,
  22                            struct bkey_packed *start,
  23                            struct bkey_packed *end)
  24 {
  25 #ifdef CONFIG_BCACHEFS_DEBUG
  26         struct bkey_packed *k;
  27
  28         for (k = start; k != end && bkey_next(k) != end; k = bkey_next(k)) {
  29                 struct bkey l = bkey_unpack_key(b, k);
  30                 struct bkey r = bkey_unpack_key(b, bkey_next(k));
  31
  32                 BUG_ON(btree_node_is_extents(b)
  33                        ? bkey_cmp(l.p, bkey_start_pos(&r)) > 0
  34                        : bkey_cmp(l.p, bkey_start_pos(&r)) >= 0);
  35                 //BUG_ON(bkey_cmp_packed(&b->format, k, bkey_next(k)) >= 0);
  36         }
  37 #endif
  38 }
  39
  40 static void clear_needs_whiteout(struct bset *i)
  41 {
  42         struct bkey_packed *k;
  43
  44         for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
  45                 k->needs_whiteout = false;
  46 }
  47
  48 static void set_needs_whiteout(struct bset *i)
  49 {
  50         struct bkey_packed *k;
  51
  52         for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
  53                 k->needs_whiteout = true;
  54 }
  55
  56 static void btree_bounce_free(struct bch_fs *c, unsigned order,
  57                               bool used_mempool, void *p)
  58 {
  59         if (used_mempool)
  60                 mempool_free(p, &c->btree_bounce_pool);
  61         else
  62                 vpfree(p, PAGE_SIZE << order);
  63 }
  64
  65 static void *btree_bounce_alloc(struct bch_fs *c, unsigned order,
  66                                 bool *used_mempool)
  67 {
  68         void *p;
  69
  70         BUG_ON(order > btree_page_order(c));
  71
  72         *used_mempool = false;
  73         p = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT, order);
  74         if (p)
  75                 return p;
  76
  77         *used_mempool = true;
  78         return mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
  79 }
  80
  81 typedef int (*sort_cmp_fn)(struct btree *,
  82                            struct bkey_packed *,
  83                            struct bkey_packed *);
  84
  85 struct sort_iter {
  86         struct btree    *b;
  87         unsigned                used;
  88
  89         struct sort_iter_set {
  90                 struct bkey_packed *k, *end;
  91         } data[MAX_BSETS + 1];
  92 };
  93
  94 static void sort_iter_init(struct sort_iter *iter, struct btree *b)
  95 {
  96         memset(iter, 0, sizeof(*iter));
  97         iter->b = b;
  98 }
  99
 100 static inline void __sort_iter_sift(struct sort_iter *iter,
 101                                     unsigned from,
 102                                     sort_cmp_fn cmp)
 103 {
 104         unsigned i;
 105
 106         for (i = from;
 107              i + 1 < iter->used &&
 108              cmp(iter->b, iter->data[i].k, iter->data[i + 1].k) > 0;
 109              i++)
 110                 swap(iter->data[i], iter->data[i + 1]);
 111 }
 112
 113 static inline void sort_iter_sift(struct sort_iter *iter, sort_cmp_fn cmp)
 114 {
 115
 116         __sort_iter_sift(iter, 0, cmp);
 117 }
 118
 119 static inline void sort_iter_sort(struct sort_iter *iter, sort_cmp_fn cmp)
 120 {
 121         unsigned i = iter->used;
 122
 123         while (i--)
 124                 __sort_iter_sift(iter, i, cmp);
 125 }
 126
 127 static void sort_iter_add(struct sort_iter *iter,
 128                           struct bkey_packed *k,
 129                           struct bkey_packed *end)
 130 {
 131         BUG_ON(iter->used >= ARRAY_SIZE(iter->data));
 132
 133         if (k != end)
 134                 iter->data[iter->used++] = (struct sort_iter_set) { k, end };
 135 }
 136
 137 static inline struct bkey_packed *sort_iter_peek(struct sort_iter *iter)
 138 {
 139         return iter->used ? iter->data->k : NULL;
 140 }
 141
 142 static inline void sort_iter_advance(struct sort_iter *iter, sort_cmp_fn cmp)
 143 {
 144         iter->data->k = bkey_next(iter->data->k);
 145
 146         BUG_ON(iter->data->k > iter->data->end);
 147
 148         if (iter->data->k == iter->data->end)
 149                 array_remove_item(iter->data, iter->used, 0);
 150         else
 151                 sort_iter_sift(iter, cmp);
 152 }
 153
 154 static inline struct bkey_packed *sort_iter_next(struct sort_iter *iter,
 155                                                  sort_cmp_fn cmp)
 156 {
 157         struct bkey_packed *ret = sort_iter_peek(iter);
 158
 159         if (ret)
 160                 sort_iter_advance(iter, cmp);
 161
 162         return ret;
 163 }
 164
 165 static inline int sort_key_whiteouts_cmp(struct btree *b,
 166                                          struct bkey_packed *l,
 167                                          struct bkey_packed *r)
 168 {
 169         return bkey_cmp_packed(b, l, r);
 170 }
 171
 172 static unsigned sort_key_whiteouts(struct bkey_packed *dst,
 173                                    struct sort_iter *iter)
 174 {
 175         struct bkey_packed *in, *out = dst;
 176
 177         sort_iter_sort(iter, sort_key_whiteouts_cmp);
 178
 179         while ((in = sort_iter_next(iter, sort_key_whiteouts_cmp))) {
 180                 bkey_copy(out, in);
 181                 out = bkey_next(out);
 182         }
 183
 184         return (u64 *) out - (u64 *) dst;
 185 }
 186
 187 static inline int sort_extent_whiteouts_cmp(struct btree *b,
 188                                             struct bkey_packed *l,
 189                                             struct bkey_packed *r)
 190 {
 191         struct bkey ul = bkey_unpack_key(b, l);
 192         struct bkey ur = bkey_unpack_key(b, r);
 193
 194         return bkey_cmp(bkey_start_pos(&ul), bkey_start_pos(&ur));
 195 }
 196
 197 static unsigned sort_extent_whiteouts(struct bkey_packed *dst,
 198                                       struct sort_iter *iter)
 199 {
 200         const struct bkey_format *f = &iter->b->format;
 201         struct bkey_packed *in, *out = dst;
 202         struct bkey_i l, r;
 203         bool prev = false, l_packed = false;
 204         u64 max_packed_size     = bkey_field_max(f, BKEY_FIELD_SIZE);
 205         u64 max_packed_offset   = bkey_field_max(f, BKEY_FIELD_OFFSET);
 206         u64 new_size;
 207
 208         max_packed_size = min_t(u64, max_packed_size, KEY_SIZE_MAX);
 209
 210         sort_iter_sort(iter, sort_extent_whiteouts_cmp);
 211
 212         while ((in = sort_iter_next(iter, sort_extent_whiteouts_cmp))) {
 213                 EBUG_ON(bkeyp_val_u64s(f, in));
 214                 EBUG_ON(in->type != KEY_TYPE_DISCARD);
 215
 216                 r.k = bkey_unpack_key(iter->b, in);
 217
 218                 if (prev &&
 219                     bkey_cmp(l.k.p, bkey_start_pos(&r.k)) >= 0) {
 220                         if (bkey_cmp(l.k.p, r.k.p) >= 0)
 221                                 continue;
 222
 223                         new_size = l_packed
 224                                 ? min(max_packed_size, max_packed_offset -
 225                                       bkey_start_offset(&l.k))
 226                                 : KEY_SIZE_MAX;
 227
 228                         new_size = min(new_size, r.k.p.offset -
 229                                        bkey_start_offset(&l.k));
 230
 231                         BUG_ON(new_size < l.k.size);
 232
 233                         bch2_key_resize(&l.k, new_size);
 234
 235                         if (bkey_cmp(l.k.p, r.k.p) >= 0)
 236                                 continue;
 237
 238                         bch2_cut_front(l.k.p, &r);
 239                 }
 240
 241                 if (prev) {
 242                         if (!bch2_bkey_pack(out, &l, f)) {
 243                                 BUG_ON(l_packed);
 244                                 bkey_copy(out, &l);
 245                         }
 246                         out = bkey_next(out);
 247                 }
 248
 249                 l = r;
 250                 prev = true;
 251                 l_packed = bkey_packed(in);
 252         }
 253
 254         if (prev) {
 255                 if (!bch2_bkey_pack(out, &l, f)) {
 256                         BUG_ON(l_packed);
 257                         bkey_copy(out, &l);
 258                 }
 259                 out = bkey_next(out);
 260         }
 261
 262         return (u64 *) out - (u64 *) dst;
 263 }
 264
 265 static unsigned should_compact_bset(struct btree *b, struct bset_tree *t,
 266                                     bool compacting,
 267                                     enum compact_mode mode)
 268 {
 269         unsigned bset_u64s = le16_to_cpu(bset(b, t)->u64s);
 270         unsigned dead_u64s = bset_u64s - b->nr.bset_u64s[t - b->set];
 271
 272         if (mode == COMPACT_LAZY) {
 273                 if (should_compact_bset_lazy(b, t) ||
 274                     (compacting && bset_unwritten(b, bset(b, t))))
 275                         return dead_u64s;
 276         } else {
 277                 if (bset_written(b, bset(b, t)))
 278                         return dead_u64s;
 279         }
 280
 281         return 0;
 282 }
 283
 284 bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 285                              enum compact_mode mode)
 286 {
 287         const struct bkey_format *f = &b->format;
 288         struct bset_tree *t;
 289         struct bkey_packed *whiteouts = NULL;
 290         struct bkey_packed *u_start, *u_pos;
 291         struct sort_iter sort_iter;
 292         unsigned order, whiteout_u64s = 0, u64s;
 293         bool used_mempool, compacting = false;
 294
 295         for_each_bset(b, t)
 296                 whiteout_u64s += should_compact_bset(b, t,
 297                                         whiteout_u64s != 0, mode);
 298
 299         if (!whiteout_u64s)
 300                 return false;
 301
 302         sort_iter_init(&sort_iter, b);
 303
 304         whiteout_u64s += b->whiteout_u64s;
 305         order = get_order(whiteout_u64s * sizeof(u64));
 306
 307         whiteouts = btree_bounce_alloc(c, order, &used_mempool);
 308         u_start = u_pos = whiteouts;
 309
 310         memcpy_u64s(u_pos, unwritten_whiteouts_start(c, b),
 311                     b->whiteout_u64s);
 312         u_pos = (void *) u_pos + b->whiteout_u64s * sizeof(u64);
 313
 314         sort_iter_add(&sort_iter, u_start, u_pos);
 315
 316         for_each_bset(b, t) {
 317                 struct bset *i = bset(b, t);
 318                 struct bkey_packed *k, *n, *out, *start, *end;
 319                 struct btree_node_entry *src = NULL, *dst = NULL;
 320
 321                 if (t != b->set && bset_unwritten(b, i)) {
 322                         src = container_of(i, struct btree_node_entry, keys);
 323                         dst = max(write_block(b),
 324                                   (void *) btree_bkey_last(b, t -1));
 325                 }
 326
 327                 if (!should_compact_bset(b, t, compacting, mode)) {
 328                         if (src != dst) {
 329                                 memmove(dst, src, sizeof(*src) +
 330                                         le16_to_cpu(src->keys.u64s) *
 331                                         sizeof(u64));
 332                                 i = &dst->keys;
 333                                 set_btree_bset(b, t, i);
 334                         }
 335                         continue;
 336                 }
 337
 338                 compacting = true;
 339                 u_start = u_pos;
 340                 start = i->start;
 341                 end = vstruct_last(i);
 342
 343                 if (src != dst) {
 344                         memmove(dst, src, sizeof(*src));
 345                         i = &dst->keys;
 346                         set_btree_bset(b, t, i);
 347                 }
 348
 349                 out = i->start;
 350
 351                 for (k = start; k != end; k = n) {
 352                         n = bkey_next(k);
 353
 354                         if (bkey_deleted(k) && btree_node_is_extents(b))
 355                                 continue;
 356
 357                         if (bkey_whiteout(k) && !k->needs_whiteout)
 358                                 continue;
 359
 360                         if (bkey_whiteout(k)) {
 361                                 unreserve_whiteout(b, t, k);
 362                                 memcpy_u64s(u_pos, k, bkeyp_key_u64s(f, k));
 363                                 set_bkeyp_val_u64s(f, u_pos, 0);
 364                                 u_pos = bkey_next(u_pos);
 365                         } else if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK) {
 366                                 bkey_copy(out, k);
 367                                 out = bkey_next(out);
 368                         }
 369                 }
 370
 371                 sort_iter_add(&sort_iter, u_start, u_pos);
 372
 373                 if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK) {
 374                         i->u64s = cpu_to_le16((u64 *) out - i->_data);
 375                         set_btree_bset_end(b, t);
 376                         bch2_bset_set_no_aux_tree(b, t);
 377                 }
 378         }
 379
 380         b->whiteout_u64s = (u64 *) u_pos - (u64 *) whiteouts;
 381
 382         BUG_ON((void *) unwritten_whiteouts_start(c, b) <
 383                (void *) btree_bkey_last(b, bset_tree_last(b)));
 384
 385         u64s = btree_node_is_extents(b)
 386                 ? sort_extent_whiteouts(unwritten_whiteouts_start(c, b),
 387                                         &sort_iter)
 388                 : sort_key_whiteouts(unwritten_whiteouts_start(c, b),
 389                                      &sort_iter);
 390
 391         BUG_ON(u64s > b->whiteout_u64s);
 392         BUG_ON(u64s != b->whiteout_u64s && !btree_node_is_extents(b));
 393         BUG_ON(u_pos != whiteouts && !u64s);
 394
 395         if (u64s != b->whiteout_u64s) {
 396                 void *src = unwritten_whiteouts_start(c, b);
 397
 398                 b->whiteout_u64s = u64s;
 399                 memmove_u64s_up(unwritten_whiteouts_start(c, b), src, u64s);
 400         }
 401
 402         verify_no_dups(b,
 403                        unwritten_whiteouts_start(c, b),
 404                        unwritten_whiteouts_end(c, b));
 405
 406         btree_bounce_free(c, order, used_mempool, whiteouts);
 407
 408         if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK)
 409                 bch2_btree_build_aux_trees(b);
 410
 411         bch_btree_keys_u64s_remaining(c, b);
 412         bch2_verify_btree_nr_keys(b);
 413
 414         return true;
 415 }
 416
 417 static bool bch2_drop_whiteouts(struct btree *b)
 418 {
 419         struct bset_tree *t;
 420         bool ret = false;
 421
 422         for_each_bset(b, t) {
 423                 struct bset *i = bset(b, t);
 424                 struct bkey_packed *k, *n, *out, *start, *end;
 425
 426                 if (!should_compact_bset(b, t, true, COMPACT_WRITTEN))
 427                         continue;
 428
 429                 start   = btree_bkey_first(b, t);
 430                 end     = btree_bkey_last(b, t);
 431
 432                 if (bset_unwritten(b, i) &&
 433                     t != b->set) {
 434                         struct bset *dst =
 435                                max_t(struct bset *, write_block(b),
 436                                      (void *) btree_bkey_last(b, t -1));
 437
 438                         memmove(dst, i, sizeof(struct bset));
 439                         i = dst;
 440                         set_btree_bset(b, t, i);
 441                 }
 442
 443                 out = i->start;
 444
 445                 for (k = start; k != end; k = n) {
 446                         n = bkey_next(k);
 447
 448                         if (!bkey_whiteout(k)) {
 449                                 bkey_copy(out, k);
 450                                 out = bkey_next(out);
 451                         }
 452                 }
 453
 454                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 455                 bch2_bset_set_no_aux_tree(b, t);
 456                 ret = true;
 457         }
 458
 459         bch2_verify_btree_nr_keys(b);
 460
 461         return ret;
 462 }
 463
 464 static inline int sort_keys_cmp(struct btree *b,
 465                                 struct bkey_packed *l,
 466                                 struct bkey_packed *r)
 467 {
 468         return bkey_cmp_packed(b, l, r) ?:
 469                 (int) bkey_whiteout(r) - (int) bkey_whiteout(l) ?:
 470                 (int) l->needs_whiteout - (int) r->needs_whiteout;
 471 }
 472
 473 static unsigned sort_keys(struct bkey_packed *dst,
 474                           struct sort_iter *iter,
 475                           bool filter_whiteouts)
 476 {
 477         const struct bkey_format *f = &iter->b->format;
 478         struct bkey_packed *in, *next, *out = dst;
 479
 480         sort_iter_sort(iter, sort_keys_cmp);
 481
 482         while ((in = sort_iter_next(iter, sort_keys_cmp))) {
 483                 if (bkey_whiteout(in) &&
 484                     (filter_whiteouts || !in->needs_whiteout))
 485                         continue;
 486
 487                 if (bkey_whiteout(in) &&
 488                     (next = sort_iter_peek(iter)) &&
 489                     !bkey_cmp_packed(iter->b, in, next)) {
 490                         BUG_ON(in->needs_whiteout &&
 491                                next->needs_whiteout);
 492                         /*
 493                          * XXX racy, called with read lock from write path
 494                          *
 495                          * leads to spurious BUG_ON() in bkey_unpack_key() in
 496                          * debug mode
 497                          */
 498                         next->needs_whiteout |= in->needs_whiteout;
 499                         continue;
 500                 }
 501
 502                 if (bkey_whiteout(in)) {
 503                         memcpy_u64s(out, in, bkeyp_key_u64s(f, in));
 504                         set_bkeyp_val_u64s(f, out, 0);
 505                 } else {
 506                         bkey_copy(out, in);
 507                 }
 508                 out = bkey_next(out);
 509         }
 510
 511         return (u64 *) out - (u64 *) dst;
 512 }
 513
 514 static inline int sort_extents_cmp(struct btree *b,
 515                                    struct bkey_packed *l,
 516                                    struct bkey_packed *r)
 517 {
 518         return bkey_cmp_packed(b, l, r) ?:
 519                 (int) bkey_deleted(l) - (int) bkey_deleted(r);
 520 }
 521
 522 static unsigned sort_extents(struct bkey_packed *dst,
 523                              struct sort_iter *iter,
 524                              bool filter_whiteouts)
 525 {
 526         struct bkey_packed *in, *out = dst;
 527
 528         sort_iter_sort(iter, sort_extents_cmp);
 529
 530         while ((in = sort_iter_next(iter, sort_extents_cmp))) {
 531                 if (bkey_deleted(in))
 532                         continue;
 533
 534                 if (bkey_whiteout(in) &&
 535                     (filter_whiteouts || !in->needs_whiteout))
 536                         continue;
 537
 538                 bkey_copy(out, in);
 539                 out = bkey_next(out);
 540         }
 541
 542         return (u64 *) out - (u64 *) dst;
 543 }
 544
 545 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 546                             struct btree_iter *iter,
 547                             unsigned start_idx,
 548                             unsigned end_idx,
 549                             bool filter_whiteouts)
 550 {
 551         struct btree_node *out;
 552         struct sort_iter sort_iter;
 553         struct bset_tree *t;
 554         struct bset *start_bset = bset(b, &b->set[start_idx]);
 555         bool used_mempool = false;
 556         u64 start_time, seq = 0;
 557         unsigned i, u64s = 0, order, shift = end_idx - start_idx - 1;
 558         bool sorting_entire_node = start_idx == 0 &&
 559                 end_idx == b->nsets;
 560
 561         sort_iter_init(&sort_iter, b);
 562
 563         for (t = b->set + start_idx;
 564              t < b->set + end_idx;
 565              t++) {
 566                 u64s += le16_to_cpu(bset(b, t)->u64s);
 567                 sort_iter_add(&sort_iter,
 568                               btree_bkey_first(b, t),
 569                               btree_bkey_last(b, t));
 570         }
 571
 572         order = sorting_entire_node
 573                 ? btree_page_order(c)
 574                 : get_order(__vstruct_bytes(struct btree_node, u64s));
 575
 576         out = btree_bounce_alloc(c, order, &used_mempool);
 577
 578         start_time = local_clock();
 579
 580         if (btree_node_is_extents(b))
 581                 filter_whiteouts = bset_written(b, start_bset);
 582
 583         u64s = btree_node_is_extents(b)
 584                 ? sort_extents(out->keys.start, &sort_iter, filter_whiteouts)
 585                 : sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 586
 587         out->keys.u64s = cpu_to_le16(u64s);
 588
 589         BUG_ON(vstruct_end(&out->keys) > (void *) out + (PAGE_SIZE << order));
 590
 591         if (sorting_entire_node)
 592                 bch2_time_stats_update(&c->btree_sort_time, start_time);
 593
 594         /* Make sure we preserve bset journal_seq: */
 595         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 596                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 597         start_bset->journal_seq = cpu_to_le64(seq);
 598
 599         if (sorting_entire_node) {
 600                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 601
 602                 BUG_ON(order != btree_page_order(c));
 603
 604                 /*
 605                  * Our temporary buffer is the same size as the btree node's
 606                  * buffer, we can just swap buffers instead of doing a big
 607                  * memcpy()
 608                  */
 609                 *out = *b->data;
 610                 out->keys.u64s = cpu_to_le16(u64s);
 611                 swap(out, b->data);
 612                 set_btree_bset(b, b->set, &b->data->keys);
 613         } else {
 614                 start_bset->u64s = out->keys.u64s;
 615                 memcpy_u64s(start_bset->start,
 616                             out->keys.start,
 617                             le16_to_cpu(out->keys.u64s));
 618         }
 619
 620         for (i = start_idx + 1; i < end_idx; i++)
 621                 b->nr.bset_u64s[start_idx] +=
 622                         b->nr.bset_u64s[i];
 623
 624         b->nsets -= shift;
 625
 626         for (i = start_idx + 1; i < b->nsets; i++) {
 627                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 628                 b->set[i]               = b->set[i + shift];
 629         }
 630
 631         for (i = b->nsets; i < MAX_BSETS; i++)
 632                 b->nr.bset_u64s[i] = 0;
 633
 634         set_btree_bset_end(b, &b->set[start_idx]);
 635         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 636
 637         btree_bounce_free(c, order, used_mempool, out);
 638
 639         bch2_verify_btree_nr_keys(b);
 640 }
 641
 642 /* Sort + repack in a new format: */
 643 static struct btree_nr_keys sort_repack(struct bset *dst,
 644                                         struct btree *src,
 645                                         struct btree_node_iter *src_iter,
 646                                         struct bkey_format *out_f,
 647                                         bool filter_whiteouts)
 648 {
 649         struct bkey_format *in_f = &src->format;
 650         struct bkey_packed *in, *out = vstruct_last(dst);
 651         struct btree_nr_keys nr;
 652
 653         memset(&nr, 0, sizeof(nr));
 654
 655         while ((in = bch2_btree_node_iter_next_all(src_iter, src))) {
 656                 if (filter_whiteouts && bkey_whiteout(in))
 657                         continue;
 658
 659                 if (bch2_bkey_transform(out_f, out, bkey_packed(in)
 660                                        ? in_f : &bch2_bkey_format_current, in))
 661                         out->format = KEY_FORMAT_LOCAL_BTREE;
 662                 else
 663                         bch2_bkey_unpack(src, (void *) out, in);
 664
 665                 btree_keys_account_key_add(&nr, 0, out);
 666                 out = bkey_next(out);
 667         }
 668
 669         dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
 670         return nr;
 671 }
 672
 673 /* Sort, repack, and merge: */
 674 static struct btree_nr_keys sort_repack_merge(struct bch_fs *c,
 675                                               struct bset *dst,
 676                                               struct btree *src,
 677                                               struct btree_node_iter *iter,
 678                                               struct bkey_format *out_f,
 679                                               bool filter_whiteouts,
 680                                               key_filter_fn filter,
 681                                               key_merge_fn merge)
 682 {
 683         struct bkey_packed *k, *prev = NULL, *out;
 684         struct btree_nr_keys nr;
 685         BKEY_PADDED(k) tmp;
 686
 687         memset(&nr, 0, sizeof(nr));
 688
 689         while ((k = bch2_btree_node_iter_next_all(iter, src))) {
 690                 if (filter_whiteouts && bkey_whiteout(k))
 691                         continue;
 692
 693                 /*
 694                  * The filter might modify pointers, so we have to unpack the
 695                  * key and values to &tmp.k:
 696                  */
 697                 bch2_bkey_unpack(src, &tmp.k, k);
 698
 699                 if (filter && filter(c, src, bkey_i_to_s(&tmp.k)))
 700                         continue;
 701
 702                 /* prev is always unpacked, for key merging: */
 703
 704                 if (prev &&
 705                     merge &&
 706                     merge(c, src, (void *) prev, &tmp.k) == BCH_MERGE_MERGE)
 707                         continue;
 708
 709                 /*
 710                  * the current key becomes the new prev: advance prev, then
 711                  * copy the current key - but first pack prev (in place):
 712                  */
 713                 if (prev) {
 714                         bch2_bkey_pack(prev, (void *) prev, out_f);
 715
 716                         btree_keys_account_key_add(&nr, 0, prev);
 717                         prev = bkey_next(prev);
 718                 } else {
 719                         prev = vstruct_last(dst);
 720                 }
 721
 722                 bkey_copy(prev, &tmp.k);
 723         }
 724
 725         if (prev) {
 726                 bch2_bkey_pack(prev, (void *) prev, out_f);
 727                 btree_keys_account_key_add(&nr, 0, prev);
 728                 out = bkey_next(prev);
 729         } else {
 730                 out = vstruct_last(dst);
 731         }
 732
 733         dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
 734         return nr;
 735 }
 736
 737 void bch2_btree_sort_into(struct bch_fs *c,
 738                          struct btree *dst,
 739                          struct btree *src)
 740 {
 741         struct btree_nr_keys nr;
 742         struct btree_node_iter src_iter;
 743         u64 start_time = local_clock();
 744
 745         BUG_ON(dst->nsets != 1);
 746
 747         bch2_bset_set_no_aux_tree(dst, dst->set);
 748
 749         bch2_btree_node_iter_init_from_start(&src_iter, src,
 750                                             btree_node_is_extents(src));
 751
 752         if (btree_node_ops(src)->key_normalize ||
 753             btree_node_ops(src)->key_merge)
 754                 nr = sort_repack_merge(c, btree_bset_first(dst),
 755                                 src, &src_iter,
 756                                 &dst->format,
 757                                 true,
 758                                 btree_node_ops(src)->key_normalize,
 759                                 btree_node_ops(src)->key_merge);
 760         else
 761                 nr = sort_repack(btree_bset_first(dst),
 762                                 src, &src_iter,
 763                                 &dst->format,
 764                                 true);
 765
 766         bch2_time_stats_update(&c->btree_sort_time, start_time);
 767
 768         set_btree_bset_end(dst, dst->set);
 769
 770         dst->nr.live_u64s       += nr.live_u64s;
 771         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 772         dst->nr.packed_keys     += nr.packed_keys;
 773         dst->nr.unpacked_keys   += nr.unpacked_keys;
 774
 775         bch2_verify_btree_nr_keys(dst);
 776 }
 777
 778 #define SORT_CRIT       (4096 / sizeof(u64))
 779
 780 /*
 781  * We're about to add another bset to the btree node, so if there's currently
 782  * too many bsets - sort some of them together:
 783  */
 784 static bool btree_node_compact(struct bch_fs *c, struct btree *b,
 785                                struct btree_iter *iter)
 786 {
 787         unsigned unwritten_idx;
 788         bool ret = false;
 789
 790         for (unwritten_idx = 0;
 791              unwritten_idx < b->nsets;
 792              unwritten_idx++)
 793                 if (bset_unwritten(b, bset(b, &b->set[unwritten_idx])))
 794                         break;
 795
 796         if (b->nsets - unwritten_idx > 1) {
 797                 btree_node_sort(c, b, iter, unwritten_idx,
 798                                 b->nsets, false);
 799                 ret = true;
 800         }
 801
 802         if (unwritten_idx > 1) {
 803                 btree_node_sort(c, b, iter, 0, unwritten_idx, false);
 804                 ret = true;
 805         }
 806
 807         return ret;
 808 }
 809
 810 void bch2_btree_build_aux_trees(struct btree *b)
 811 {
 812         struct bset_tree *t;
 813
 814         for_each_bset(b, t)
 815                 bch2_bset_build_aux_tree(b, t,
 816                                 bset_unwritten(b, bset(b, t)) &&
 817                                 t == bset_tree_last(b));
 818 }
 819
 820 /*
 821  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 822  * inserted into
 823  *
 824  * Safe to call if there already is an unwritten bset - will only add a new bset
 825  * if @b doesn't already have one.
 826  *
 827  * Returns true if we sorted (i.e. invalidated iterators
 828  */
 829 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
 830                           struct btree_iter *iter)
 831 {
 832         struct btree_node_entry *bne;
 833         bool did_sort;
 834
 835         EBUG_ON(!(b->lock.state.seq & 1));
 836         EBUG_ON(iter && iter->l[b->level].b != b);
 837
 838         did_sort = btree_node_compact(c, b, iter);
 839
 840         bne = want_new_bset(c, b);
 841         if (bne)
 842                 bch2_bset_init_next(b, &bne->keys);
 843
 844         bch2_btree_build_aux_trees(b);
 845
 846         if (iter && did_sort)
 847                 bch2_btree_iter_reinit_node(iter, b);
 848 }
 849
 850 static struct nonce btree_nonce(struct bset *i, unsigned offset)
 851 {
 852         return (struct nonce) {{
 853                 [0] = cpu_to_le32(offset),
 854                 [1] = ((__le32 *) &i->seq)[0],
 855                 [2] = ((__le32 *) &i->seq)[1],
 856                 [3] = ((__le32 *) &i->journal_seq)[0]^BCH_NONCE_BTREE,
 857         }};
 858 }
 859
 860 static void bset_encrypt(struct bch_fs *c, struct bset *i, unsigned offset)
 861 {
 862         struct nonce nonce = btree_nonce(i, offset);
 863
 864         if (!offset) {
 865                 struct btree_node *bn = container_of(i, struct btree_node, keys);
 866                 unsigned bytes = (void *) &bn->keys - (void *) &bn->flags;
 867
 868                 bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, &bn->flags,
 869                              bytes);
 870
 871                 nonce = nonce_add(nonce, round_up(bytes, CHACHA20_BLOCK_SIZE));
 872         }
 873
 874         bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, i->_data,
 875                      vstruct_end(i) - (void *) i->_data);
 876 }
 877
 878 static int btree_err_msg(struct bch_fs *c, struct btree *b, struct bset *i,
 879                          unsigned offset, int write, char *buf, size_t len)
 880 {
 881         char *out = buf, *end = buf + len;
 882
 883         out += scnprintf(out, end - out,
 884                          "error validating btree node %s "
 885                          "at btree %u level %u/%u\n"
 886                          "pos %llu:%llu node offset %u",
 887                          write ? "before write " : "",
 888                          b->btree_id, b->level,
 889                          c->btree_roots[b->btree_id].level,
 890                          b->key.k.p.inode, b->key.k.p.offset,
 891                          b->written);
 892         if (i)
 893                 out += scnprintf(out, end - out,
 894                                  " bset u64s %u",
 895                                  le16_to_cpu(i->u64s));
 896
 897         return out - buf;
 898 }
 899
 900 enum btree_err_type {
 901         BTREE_ERR_FIXABLE,
 902         BTREE_ERR_WANT_RETRY,
 903         BTREE_ERR_MUST_RETRY,
 904         BTREE_ERR_FATAL,
 905 };
 906
 907 enum btree_validate_ret {
 908         BTREE_RETRY_READ = 64,
 909 };
 910
 911 #define btree_err(type, c, b, i, msg, ...)                              \
 912 ({                                                                      \
 913         char buf[200], *out = buf, *end = out + sizeof(buf);            \
 914                                                                         \
 915         out += btree_err_msg(c, b, i, b->written, write, out, end - out);\
 916         out += scnprintf(out, end - out, ": " msg, ##__VA_ARGS__);      \
 917                                                                         \
 918         if (type == BTREE_ERR_FIXABLE &&                                \
 919             write == READ &&                                            \
 920             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 921                 mustfix_fsck_err(c, "%s", buf);                         \
 922         } else {                                                        \
 923                 bch_err(c, "%s", buf);                                  \
 924                                                                         \
 925                 switch (type) {                                         \
 926                 case BTREE_ERR_FIXABLE:                                 \
 927                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 928                         goto fsck_err;                                  \
 929                 case BTREE_ERR_WANT_RETRY:                              \
 930                         if (have_retry) {                               \
 931                                 ret = BTREE_RETRY_READ;                 \
 932                                 goto fsck_err;                          \
 933                         }                                               \
 934                         break;                                          \
 935                 case BTREE_ERR_MUST_RETRY:                              \
 936                         ret = BTREE_RETRY_READ;                         \
 937                         goto fsck_err;                                  \
 938                 case BTREE_ERR_FATAL:                                   \
 939                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 940                         goto fsck_err;                                  \
 941                 }                                                       \
 942         }                                                               \
 943         true;                                                           \
 944 })
 945
 946 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 947
 948 static int validate_bset(struct bch_fs *c, struct btree *b,
 949                          struct bset *i, unsigned sectors,
 950                          unsigned *whiteout_u64s, int write,
 951                          bool have_retry)
 952 {
 953         struct bkey_packed *k, *prev = NULL;
 954         struct bpos prev_pos = POS_MIN;
 955         enum bkey_type type = btree_node_type(b);
 956         bool seen_non_whiteout = false;
 957         const char *err;
 958         int ret = 0;
 959
 960         if (i == &b->data->keys) {
 961                 /* These indicate that we read the wrong btree node: */
 962                 btree_err_on(BTREE_NODE_ID(b->data) != b->btree_id,
 963                              BTREE_ERR_MUST_RETRY, c, b, i,
 964                              "incorrect btree id");
 965
 966                 btree_err_on(BTREE_NODE_LEVEL(b->data) != b->level,
 967                              BTREE_ERR_MUST_RETRY, c, b, i,
 968                              "incorrect level");
 969
 970                 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
 971                         u64 *p = (u64 *) &b->data->ptr;
 972
 973                         *p = swab64(*p);
 974                         bch2_bpos_swab(&b->data->min_key);
 975                         bch2_bpos_swab(&b->data->max_key);
 976                 }
 977
 978                 btree_err_on(bkey_cmp(b->data->max_key, b->key.k.p),
 979                              BTREE_ERR_MUST_RETRY, c, b, i,
 980                              "incorrect max key");
 981
 982                 /* XXX: ideally we would be validating min_key too */
 983 #if 0
 984                 /*
 985                  * not correct anymore, due to btree node write error
 986                  * handling
 987                  *
 988                  * need to add b->data->seq to btree keys and verify
 989                  * against that
 990                  */
 991                 btree_err_on(!extent_contains_ptr(bkey_i_to_s_c_extent(&b->key),
 992                                                   b->data->ptr),
 993                              BTREE_ERR_FATAL, c, b, i,
 994                              "incorrect backpointer");
 995 #endif
 996                 err = bch2_bkey_format_validate(&b->data->format);
 997                 btree_err_on(err,
 998                              BTREE_ERR_FATAL, c, b, i,
 999                              "invalid bkey format: %s", err);
1000         }
1001
1002         if (btree_err_on(le16_to_cpu(i->version) != BCACHE_BSET_VERSION,
1003                          BTREE_ERR_FIXABLE, c, b, i,
1004                          "unsupported bset version")) {
1005                 i->version = cpu_to_le16(BCACHE_BSET_VERSION);
1006                 i->u64s = 0;
1007                 return 0;
1008         }
1009
1010         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
1011                          BTREE_ERR_FIXABLE, c, b, i,
1012                          "bset past end of btree node")) {
1013                 i->u64s = 0;
1014                 return 0;
1015         }
1016
1017         btree_err_on(b->written && !i->u64s,
1018                      BTREE_ERR_FIXABLE, c, b, i,
1019                      "empty bset");
1020
1021         if (!BSET_SEPARATE_WHITEOUTS(i)) {
1022                 seen_non_whiteout = true;
1023                 *whiteout_u64s = 0;
1024         }
1025
1026         for (k = i->start;
1027              k != vstruct_last(i);) {
1028                 struct bkey_s_c u;
1029                 struct bkey tmp;
1030                 const char *invalid;
1031
1032                 if (btree_err_on(!k->u64s,
1033                                  BTREE_ERR_FIXABLE, c, b, i,
1034                                  "KEY_U64s 0: %zu bytes of metadata lost",
1035                                  vstruct_end(i) - (void *) k)) {
1036                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
1037                         break;
1038                 }
1039
1040                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
1041                                  BTREE_ERR_FIXABLE, c, b, i,
1042                                  "key extends past end of bset")) {
1043                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
1044                         break;
1045                 }
1046
1047                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
1048                                  BTREE_ERR_FIXABLE, c, b, i,
1049                                  "invalid bkey format %u", k->format)) {
1050                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1051                         memmove_u64s_down(k, bkey_next(k),
1052                                           (u64 *) vstruct_end(i) - (u64 *) k);
1053                         continue;
1054                 }
1055
1056                 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN)
1057                         bch2_bkey_swab(type, &b->format, k);
1058
1059                 u = bkey_disassemble(b, k, &tmp);
1060
1061                 invalid = __bch2_bkey_invalid(c, type, u) ?:
1062                         bch2_bkey_in_btree_node(b, u) ?:
1063                         (write ? bch2_bkey_val_invalid(c, type, u) : NULL);
1064                 if (invalid) {
1065                         char buf[160];
1066
1067                         bch2_bkey_val_to_text(c, type, buf, sizeof(buf), u);
1068                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
1069                                   "invalid bkey %s: %s", buf, invalid);
1070
1071                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1072                         memmove_u64s_down(k, bkey_next(k),
1073                                           (u64 *) vstruct_end(i) - (u64 *) k);
1074                         continue;
1075                 }
1076
1077                 /*
1078                  * with the separate whiteouts thing (used for extents), the
1079                  * second set of keys actually can have whiteouts too, so we
1080                  * can't solely go off bkey_whiteout()...
1081                  */
1082
1083                 if (!seen_non_whiteout &&
1084                     (!bkey_whiteout(k) ||
1085                      (bkey_cmp(prev_pos, bkey_start_pos(u.k)) > 0))) {
1086                         *whiteout_u64s = k->_data - i->_data;
1087                         seen_non_whiteout = true;
1088                 } else if (bkey_cmp(prev_pos, bkey_start_pos(u.k)) > 0) {
1089                         btree_err(BTREE_ERR_FATAL, c, b, i,
1090                                   "keys out of order: %llu:%llu > %llu:%llu",
1091                                   prev_pos.inode,
1092                                   prev_pos.offset,
1093                                   u.k->p.inode,
1094                                   bkey_start_offset(u.k));
1095                         /* XXX: repair this */
1096                 }
1097
1098                 prev_pos = u.k->p;
1099                 prev = k;
1100                 k = bkey_next(k);
1101         }
1102
1103         SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1104 fsck_err:
1105         return ret;
1106 }
1107
1108 int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry)
1109 {
1110         struct btree_node_entry *bne;
1111         struct btree_node_iter *iter;
1112         struct btree_node *sorted;
1113         struct bkey_packed *k;
1114         struct bset *i;
1115         bool used_mempool;
1116         unsigned u64s;
1117         int ret, retry_read = 0, write = READ;
1118
1119         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
1120         __bch2_btree_node_iter_init(iter, btree_node_is_extents(b));
1121
1122         if (bch2_meta_read_fault("btree"))
1123                 btree_err(BTREE_ERR_MUST_RETRY, c, b, NULL,
1124                           "dynamic fault");
1125
1126         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
1127                      BTREE_ERR_MUST_RETRY, c, b, NULL,
1128                      "bad magic");
1129
1130         btree_err_on(!b->data->keys.seq,
1131                      BTREE_ERR_MUST_RETRY, c, b, NULL,
1132                      "bad btree header");
1133
1134         while (b->written < c->opts.btree_node_size) {
1135                 unsigned sectors, whiteout_u64s = 0;
1136                 struct nonce nonce;
1137                 struct bch_csum csum;
1138
1139                 if (!b->written) {
1140                         i = &b->data->keys;
1141
1142                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1143                                      BTREE_ERR_WANT_RETRY, c, b, i,
1144                                      "unknown checksum type");
1145
1146                         nonce = btree_nonce(i, b->written << 9);
1147                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
1148
1149                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
1150                                      BTREE_ERR_WANT_RETRY, c, b, i,
1151                                      "invalid checksum");
1152
1153                         bset_encrypt(c, i, b->written << 9);
1154
1155                         sectors = vstruct_sectors(b->data, c->block_bits);
1156
1157                         set_btree_bset(b, b->set, &b->data->keys);
1158                         btree_node_set_format(b, b->data->format);
1159                 } else {
1160                         bne = write_block(b);
1161                         i = &bne->keys;
1162
1163                         if (i->seq != b->data->keys.seq)
1164                                 break;
1165
1166                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1167                                      BTREE_ERR_WANT_RETRY, c, b, i,
1168                                      "unknown checksum type");
1169
1170                         nonce = btree_nonce(i, b->written << 9);
1171                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1172
1173                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
1174                                      BTREE_ERR_WANT_RETRY, c, b, i,
1175                                      "invalid checksum");
1176
1177                         bset_encrypt(c, i, b->written << 9);
1178
1179                         sectors = vstruct_sectors(bne, c->block_bits);
1180                 }
1181
1182                 ret = validate_bset(c, b, i, sectors, &whiteout_u64s,
1183                                     READ, have_retry);
1184                 if (ret)
1185                         goto fsck_err;
1186
1187                 b->written += sectors;
1188
1189                 ret = bch2_journal_seq_should_ignore(c, le64_to_cpu(i->journal_seq), b);
1190                 if (ret < 0) {
1191                         btree_err(BTREE_ERR_FATAL, c, b, i,
1192                                   "insufficient memory");
1193                         goto err;
1194                 }
1195
1196                 if (ret) {
1197                         btree_err_on(!b->written,
1198                                      BTREE_ERR_FIXABLE, c, b, i,
1199                                      "first btree node bset has blacklisted journal seq");
1200                         if (b->written)
1201                                 continue;
1202                 }
1203
1204                 __bch2_btree_node_iter_push(iter, b,
1205                                            i->start,
1206                                            vstruct_idx(i, whiteout_u64s));
1207
1208                 __bch2_btree_node_iter_push(iter, b,
1209                                            vstruct_idx(i, whiteout_u64s),
1210                                            vstruct_last(i));
1211         }
1212
1213         for (bne = write_block(b);
1214              bset_byte_offset(b, bne) < btree_bytes(c);
1215              bne = (void *) bne + block_bytes(c))
1216                 btree_err_on(bne->keys.seq == b->data->keys.seq,
1217                              BTREE_ERR_WANT_RETRY, c, b, NULL,
1218                              "found bset signature after last bset");
1219
1220         sorted = btree_bounce_alloc(c, btree_page_order(c), &used_mempool);
1221         sorted->keys.u64s = 0;
1222
1223         b->nr = btree_node_is_extents(b)
1224                 ? bch2_extent_sort_fix_overlapping(c, &sorted->keys, b, iter)
1225                 : bch2_key_sort_fix_overlapping(&sorted->keys, b, iter);
1226
1227         u64s = le16_to_cpu(sorted->keys.u64s);
1228         *sorted = *b->data;
1229         sorted->keys.u64s = cpu_to_le16(u64s);
1230         swap(sorted, b->data);
1231         set_btree_bset(b, b->set, &b->data->keys);
1232         b->nsets = 1;
1233
1234         BUG_ON(b->nr.live_u64s != u64s);
1235
1236         btree_bounce_free(c, btree_page_order(c), used_mempool, sorted);
1237
1238         i = &b->data->keys;
1239         for (k = i->start; k != vstruct_last(i);) {
1240                 enum bkey_type type = btree_node_type(b);
1241                 struct bkey tmp;
1242                 struct bkey_s_c u = bkey_disassemble(b, k, &tmp);
1243                 const char *invalid = bch2_bkey_val_invalid(c, type, u);
1244
1245                 if (invalid) {
1246                         char buf[160];
1247
1248                         bch2_bkey_val_to_text(c, type, buf, sizeof(buf), u);
1249                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
1250                                   "invalid bkey %s: %s", buf, invalid);
1251
1252                         btree_keys_account_key_drop(&b->nr, 0, k);
1253
1254                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1255                         memmove_u64s_down(k, bkey_next(k),
1256                                           (u64 *) vstruct_end(i) - (u64 *) k);
1257                         continue;
1258                 }
1259
1260                 k = bkey_next(k);
1261         }
1262
1263         bch2_bset_build_aux_tree(b, b->set, false);
1264
1265         set_needs_whiteout(btree_bset_first(b));
1266
1267         btree_node_reset_sib_u64s(b);
1268 out:
1269         mempool_free(iter, &c->fill_iter);
1270         return retry_read;
1271 err:
1272 fsck_err:
1273         if (ret == BTREE_RETRY_READ) {
1274                 retry_read = 1;
1275         } else {
1276                 bch2_inconsistent_error(c);
1277                 set_btree_node_read_error(b);
1278         }
1279         goto out;
1280 }
1281
1282 static void btree_node_read_work(struct work_struct *work)
1283 {
1284         struct btree_read_bio *rb =
1285                 container_of(work, struct btree_read_bio, work);
1286         struct bch_fs *c        = rb->c;
1287         struct btree *b         = rb->bio.bi_private;
1288         struct bio *bio         = &rb->bio;
1289         struct bch_devs_mask avoid;
1290
1291         memset(&avoid, 0, sizeof(avoid));
1292
1293         goto start;
1294         do {
1295                 bch_info(c, "retrying read");
1296                 bio_reset(bio);
1297                 bio_set_dev(bio, rb->pick.ca->disk_sb.bdev);
1298                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1299                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1300                 bio->bi_iter.bi_size    = btree_bytes(c);
1301                 submit_bio_wait(bio);
1302 start:
1303                 bch2_dev_io_err_on(bio->bi_status, rb->pick.ca, "btree read");
1304                 percpu_ref_put(&rb->pick.ca->io_ref);
1305
1306                 __set_bit(rb->pick.ca->dev_idx, avoid.d);
1307                 rb->pick = bch2_btree_pick_ptr(c, b, &avoid);
1308
1309                 if (!bio->bi_status &&
1310                     !bch2_btree_node_read_done(c, b, !IS_ERR_OR_NULL(rb->pick.ca)))
1311                         goto out;
1312         } while (!IS_ERR_OR_NULL(rb->pick.ca));
1313
1314         set_btree_node_read_error(b);
1315 out:
1316         if (!IS_ERR_OR_NULL(rb->pick.ca))
1317                 percpu_ref_put(&rb->pick.ca->io_ref);
1318
1319         bch2_time_stats_update(&c->btree_read_time, rb->start_time);
1320         bio_put(&rb->bio);
1321         clear_btree_node_read_in_flight(b);
1322         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1323 }
1324
1325 static void btree_node_read_endio(struct bio *bio)
1326 {
1327         struct btree_read_bio *rb =
1328                 container_of(bio, struct btree_read_bio, bio);
1329
1330         bch2_latency_acct(rb->pick.ca, rb->start_time >> 10, READ);
1331
1332         INIT_WORK(&rb->work, btree_node_read_work);
1333         queue_work(system_unbound_wq, &rb->work);
1334 }
1335
1336 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1337                           bool sync)
1338 {
1339         struct extent_pick_ptr pick;
1340         struct btree_read_bio *rb;
1341         struct bio *bio;
1342
1343         trace_btree_read(c, b);
1344
1345         pick = bch2_btree_pick_ptr(c, b, NULL);
1346         if (bch2_fs_fatal_err_on(!pick.ca, c,
1347                         "btree node read error: no device to read from")) {
1348                 set_btree_node_read_error(b);
1349                 return;
1350         }
1351
1352         bio = bio_alloc_bioset(GFP_NOIO, btree_pages(c), &c->btree_bio);
1353         rb = container_of(bio, struct btree_read_bio, bio);
1354         rb->c                   = c;
1355         rb->start_time          = local_clock();
1356         rb->pick                = pick;
1357         bio_set_dev(bio, pick.ca->disk_sb.bdev);
1358         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1359         bio->bi_iter.bi_sector  = pick.ptr.offset;
1360         bio->bi_iter.bi_size    = btree_bytes(c);
1361         bch2_bio_map(bio, b->data);
1362
1363         this_cpu_add(pick.ca->io_done->sectors[READ][BCH_DATA_BTREE],
1364                      bio_sectors(bio));
1365
1366         set_btree_node_read_in_flight(b);
1367
1368         if (sync) {
1369                 submit_bio_wait(bio);
1370                 bio->bi_private = b;
1371                 btree_node_read_work(&rb->work);
1372         } else {
1373                 bio->bi_end_io  = btree_node_read_endio;
1374                 bio->bi_private = b;
1375                 submit_bio(bio);
1376         }
1377 }
1378
1379 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1380                         const struct bkey_i *k, unsigned level)
1381 {
1382         struct closure cl;
1383         struct btree *b;
1384         int ret;
1385
1386         closure_init_stack(&cl);
1387
1388         do {
1389                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1390                 closure_sync(&cl);
1391         } while (ret);
1392
1393         b = bch2_btree_node_mem_alloc(c);
1394         bch2_btree_cache_cannibalize_unlock(c);
1395
1396         BUG_ON(IS_ERR(b));
1397
1398         bkey_copy(&b->key, k);
1399         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1400
1401         bch2_btree_node_read(c, b, true);
1402
1403         if (btree_node_read_error(b)) {
1404                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1405
1406                 mutex_lock(&c->btree_cache.lock);
1407                 list_move(&b->list, &c->btree_cache.freeable);
1408                 mutex_unlock(&c->btree_cache.lock);
1409
1410                 ret = -EIO;
1411                 goto err;
1412         }
1413
1414         bch2_btree_set_root_for_read(c, b);
1415 err:
1416         six_unlock_write(&b->lock);
1417         six_unlock_intent(&b->lock);
1418
1419         return ret;
1420 }
1421
1422 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1423                               struct btree_write *w)
1424 {
1425         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1426
1427         do {
1428                 old = new = v;
1429                 if (!(old & 1))
1430                         break;
1431
1432                 new &= ~1UL;
1433         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1434
1435         if (old & 1)
1436                 closure_put(&((struct btree_update *) new)->cl);
1437
1438         bch2_journal_pin_drop(&c->journal, &w->journal);
1439         closure_wake_up(&w->wait);
1440 }
1441
1442 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1443 {
1444         struct btree_write *w = btree_prev_write(b);
1445
1446         bch2_btree_complete_write(c, b, w);
1447         btree_node_io_unlock(b);
1448 }
1449
1450 static void bch2_btree_node_write_error(struct bch_fs *c,
1451                                         struct btree_write_bio *wbio)
1452 {
1453         struct btree *b         = wbio->wbio.bio.bi_private;
1454         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1455         struct bkey_i_extent *new_key;
1456         struct bkey_s_extent e;
1457         struct bch_extent_ptr *ptr;
1458         struct btree_iter iter;
1459         int ret;
1460
1461         __bch2_btree_iter_init(&iter, c, b->btree_id, b->key.k.p,
1462                                BTREE_MAX_DEPTH,
1463                                b->level, 0);
1464 retry:
1465         ret = bch2_btree_iter_traverse(&iter);
1466         if (ret)
1467                 goto err;
1468
1469         /* has node been freed? */
1470         if (iter.l[b->level].b != b) {
1471                 /* node has been freed: */
1472                 BUG_ON(!btree_node_dying(b));
1473                 goto out;
1474         }
1475
1476         BUG_ON(!btree_node_hashed(b));
1477
1478         bkey_copy(&tmp.k, &b->key);
1479
1480         new_key = bkey_i_to_extent(&tmp.k);
1481         e = extent_i_to_s(new_key);
1482         extent_for_each_ptr_backwards(e, ptr)
1483                 if (bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev))
1484                         bch2_extent_drop_ptr(e, ptr);
1485
1486         if (!bch2_extent_nr_ptrs(e.c))
1487                 goto err;
1488
1489         ret = bch2_btree_node_update_key(c, &iter, b, new_key);
1490         if (ret == -EINTR)
1491                 goto retry;
1492         if (ret)
1493                 goto err;
1494 out:
1495         bch2_btree_iter_unlock(&iter);
1496         bio_put(&wbio->wbio.bio);
1497         btree_node_write_done(c, b);
1498         return;
1499 err:
1500         set_btree_node_noevict(b);
1501         bch2_fs_fatal_error(c, "fatal error writing btree node");
1502         goto out;
1503 }
1504
1505 void bch2_btree_write_error_work(struct work_struct *work)
1506 {
1507         struct bch_fs *c = container_of(work, struct bch_fs,
1508                                         btree_write_error_work);
1509         struct bio *bio;
1510
1511         while (1) {
1512                 spin_lock_irq(&c->btree_write_error_lock);
1513                 bio = bio_list_pop(&c->btree_write_error_list);
1514                 spin_unlock_irq(&c->btree_write_error_lock);
1515
1516                 if (!bio)
1517                         break;
1518
1519                 bch2_btree_node_write_error(c,
1520                         container_of(bio, struct btree_write_bio, wbio.bio));
1521         }
1522 }
1523
1524 static void btree_node_write_work(struct work_struct *work)
1525 {
1526         struct btree_write_bio *wbio =
1527                 container_of(work, struct btree_write_bio, work);
1528         struct bch_fs *c        = wbio->wbio.c;
1529         struct btree *b         = wbio->wbio.bio.bi_private;
1530
1531         btree_bounce_free(c,
1532                 wbio->wbio.order,
1533                 wbio->wbio.used_mempool,
1534                 wbio->data);
1535
1536         if (wbio->wbio.failed.nr) {
1537                 unsigned long flags;
1538
1539                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1540                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1541                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1542
1543                 queue_work(c->wq, &c->btree_write_error_work);
1544                 return;
1545         }
1546
1547         bio_put(&wbio->wbio.bio);
1548         btree_node_write_done(c, b);
1549 }
1550
1551 static void btree_node_write_endio(struct bio *bio)
1552 {
1553         struct bch_write_bio *wbio      = to_wbio(bio);
1554         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1555         struct bch_write_bio *orig      = parent ?: wbio;
1556         struct bch_fs *c                = wbio->c;
1557         struct bch_dev *ca              = wbio->ca;
1558         unsigned long flags;
1559
1560         bch2_latency_acct(ca, wbio->submit_time_us, WRITE);
1561
1562         if (bio->bi_status == BLK_STS_REMOVED ||
1563             bch2_dev_io_err_on(bio->bi_status, ca, "btree write") ||
1564             bch2_meta_write_fault("btree")) {
1565                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1566                 bch2_dev_list_add_dev(&orig->failed, ca->dev_idx);
1567                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1568         }
1569
1570         if (wbio->have_io_ref)
1571                 percpu_ref_put(&ca->io_ref);
1572
1573         if (parent) {
1574                 bio_put(bio);
1575                 bio_endio(&parent->bio);
1576         } else {
1577                 struct btree_write_bio *wb =
1578                         container_of(orig, struct btree_write_bio, wbio);
1579
1580                 INIT_WORK(&wb->work, btree_node_write_work);
1581                 queue_work(system_unbound_wq, &wb->work);
1582         }
1583 }
1584
1585 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1586                                    struct bset *i, unsigned sectors)
1587 {
1588         const struct bch_extent_ptr *ptr;
1589         unsigned whiteout_u64s = 0;
1590         int ret;
1591
1592         extent_for_each_ptr(bkey_i_to_s_c_extent(&b->key), ptr)
1593                 break;
1594
1595         ret = validate_bset(c, b, i, sectors, &whiteout_u64s, WRITE, false);
1596         if (ret)
1597                 bch2_inconsistent_error(c);
1598
1599         return ret;
1600 }
1601
1602 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1603                             enum six_lock_type lock_type_held)
1604 {
1605         struct btree_write_bio *wbio;
1606         struct bset_tree *t;
1607         struct bset *i;
1608         struct btree_node *bn = NULL;
1609         struct btree_node_entry *bne = NULL;
1610         BKEY_PADDED(key) k;
1611         struct bkey_s_extent e;
1612         struct bch_extent_ptr *ptr;
1613         struct sort_iter sort_iter;
1614         struct nonce nonce;
1615         unsigned bytes_to_write, sectors_to_write, order, bytes, u64s;
1616         u64 seq = 0;
1617         bool used_mempool;
1618         unsigned long old, new;
1619         void *data;
1620
1621         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1622                 return;
1623
1624         /*
1625          * We may only have a read lock on the btree node - the dirty bit is our
1626          * "lock" against racing with other threads that may be trying to start
1627          * a write, we do a write iff we clear the dirty bit. Since setting the
1628          * dirty bit requires a write lock, we can't race with other threads
1629          * redirtying it:
1630          */
1631         do {
1632                 old = new = READ_ONCE(b->flags);
1633
1634                 if (!(old & (1 << BTREE_NODE_dirty)))
1635                         return;
1636
1637                 if (b->written &&
1638                     !btree_node_may_write(b))
1639                         return;
1640
1641                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1642                         btree_node_wait_on_io(b);
1643                         continue;
1644                 }
1645
1646                 new &= ~(1 << BTREE_NODE_dirty);
1647                 new &= ~(1 << BTREE_NODE_need_write);
1648                 new |=  (1 << BTREE_NODE_write_in_flight);
1649                 new |=  (1 << BTREE_NODE_just_written);
1650                 new ^=  (1 << BTREE_NODE_write_idx);
1651         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1652
1653         BUG_ON(btree_node_fake(b));
1654         BUG_ON(!list_empty(&b->write_blocked));
1655         BUG_ON((b->will_make_reachable != 0) != !b->written);
1656
1657         BUG_ON(b->written >= c->opts.btree_node_size);
1658         BUG_ON(bset_written(b, btree_bset_last(b)));
1659         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1660         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1661
1662         /*
1663          * We can't block on six_lock_write() here; another thread might be
1664          * trying to get a journal reservation with read locks held, and getting
1665          * a journal reservation might be blocked on flushing the journal and
1666          * doing btree writes:
1667          */
1668         if (lock_type_held == SIX_LOCK_intent &&
1669             six_trylock_write(&b->lock)) {
1670                 __bch2_compact_whiteouts(c, b, COMPACT_WRITTEN);
1671                 six_unlock_write(&b->lock);
1672         } else {
1673                 __bch2_compact_whiteouts(c, b, COMPACT_WRITTEN_NO_WRITE_LOCK);
1674         }
1675
1676         BUG_ON(b->uncompacted_whiteout_u64s);
1677
1678         sort_iter_init(&sort_iter, b);
1679
1680         bytes = !b->written
1681                 ? sizeof(struct btree_node)
1682                 : sizeof(struct btree_node_entry);
1683
1684         bytes += b->whiteout_u64s * sizeof(u64);
1685
1686         for_each_bset(b, t) {
1687                 i = bset(b, t);
1688
1689                 if (bset_written(b, i))
1690                         continue;
1691
1692                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1693                 sort_iter_add(&sort_iter,
1694                               btree_bkey_first(b, t),
1695                               btree_bkey_last(b, t));
1696                 seq = max(seq, le64_to_cpu(i->journal_seq));
1697         }
1698
1699         order = get_order(bytes);
1700         data = btree_bounce_alloc(c, order, &used_mempool);
1701
1702         if (!b->written) {
1703                 bn = data;
1704                 *bn = *b->data;
1705                 i = &bn->keys;
1706         } else {
1707                 bne = data;
1708                 bne->keys = b->data->keys;
1709                 i = &bne->keys;
1710         }
1711
1712         i->journal_seq  = cpu_to_le64(seq);
1713         i->u64s         = 0;
1714
1715         if (!btree_node_is_extents(b)) {
1716                 sort_iter_add(&sort_iter,
1717                               unwritten_whiteouts_start(c, b),
1718                               unwritten_whiteouts_end(c, b));
1719                 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1720         } else {
1721                 memcpy_u64s(i->start,
1722                             unwritten_whiteouts_start(c, b),
1723                             b->whiteout_u64s);
1724                 i->u64s = cpu_to_le16(b->whiteout_u64s);
1725                 SET_BSET_SEPARATE_WHITEOUTS(i, true);
1726         }
1727
1728         b->whiteout_u64s = 0;
1729
1730         u64s = btree_node_is_extents(b)
1731                 ? sort_extents(vstruct_last(i), &sort_iter, false)
1732                 : sort_keys(i->start, &sort_iter, false);
1733         le16_add_cpu(&i->u64s, u64s);
1734
1735         clear_needs_whiteout(i);
1736
1737         /* do we have data to write? */
1738         if (b->written && !i->u64s)
1739                 goto nowrite;
1740
1741         bytes_to_write = vstruct_end(i) - data;
1742         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1743
1744         memset(data + bytes_to_write, 0,
1745                (sectors_to_write << 9) - bytes_to_write);
1746
1747         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1748         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1749         BUG_ON(i->seq != b->data->keys.seq);
1750
1751         i->version = cpu_to_le16(BCACHE_BSET_VERSION);
1752         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1753
1754         /* if we're going to be encrypting, check metadata validity first: */
1755         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)) &&
1756             validate_bset_for_write(c, b, i, sectors_to_write))
1757                 goto err;
1758
1759         bset_encrypt(c, i, b->written << 9);
1760
1761         nonce = btree_nonce(i, b->written << 9);
1762
1763         if (bn)
1764                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1765         else
1766                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1767
1768         /* if we're not encrypting, check metadata after checksumming: */
1769         if (!bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)) &&
1770             validate_bset_for_write(c, b, i, sectors_to_write))
1771                 goto err;
1772
1773         /*
1774          * We handle btree write errors by immediately halting the journal -
1775          * after we've done that, we can't issue any subsequent btree writes
1776          * because they might have pointers to new nodes that failed to write.
1777          *
1778          * Furthermore, there's no point in doing any more btree writes because
1779          * with the journal stopped, we're never going to update the journal to
1780          * reflect that those writes were done and the data flushed from the
1781          * journal:
1782          *
1783          * Make sure to update b->written so bch2_btree_init_next() doesn't
1784          * break:
1785          */
1786         if (bch2_journal_error(&c->journal) ||
1787             c->opts.nochanges)
1788                 goto err;
1789
1790         trace_btree_write(b, bytes_to_write, sectors_to_write);
1791
1792         wbio = container_of(bio_alloc_bioset(GFP_NOIO, 1 << order, &c->btree_bio),
1793                             struct btree_write_bio, wbio.bio);
1794         wbio_init(&wbio->wbio.bio);
1795         wbio->data                      = data;
1796         wbio->wbio.order                = order;
1797         wbio->wbio.used_mempool         = used_mempool;
1798         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META|REQ_FUA;
1799         wbio->wbio.bio.bi_iter.bi_size  = sectors_to_write << 9;
1800         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1801         wbio->wbio.bio.bi_private       = b;
1802
1803         bch2_bio_map(&wbio->wbio.bio, data);
1804
1805         /*
1806          * If we're appending to a leaf node, we don't technically need FUA -
1807          * this write just needs to be persisted before the next journal write,
1808          * which will be marked FLUSH|FUA.
1809          *
1810          * Similarly if we're writing a new btree root - the pointer is going to
1811          * be in the next journal entry.
1812          *
1813          * But if we're writing a new btree node (that isn't a root) or
1814          * appending to a non leaf btree node, we need either FUA or a flush
1815          * when we write the parent with the new pointer. FUA is cheaper than a
1816          * flush, and writes appending to leaf nodes aren't blocking anything so
1817          * just make all btree node writes FUA to keep things sane.
1818          */
1819
1820         bkey_copy(&k.key, &b->key);
1821         e = bkey_i_to_s_extent(&k.key);
1822
1823         extent_for_each_ptr(e, ptr)
1824                 ptr->offset += b->written;
1825
1826         b->written += sectors_to_write;
1827
1828         bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_BTREE, &k.key);
1829         return;
1830 err:
1831         set_btree_node_noevict(b);
1832         b->written += sectors_to_write;
1833 nowrite:
1834         btree_bounce_free(c, order, used_mempool, data);
1835         btree_node_write_done(c, b);
1836 }
1837
1838 /*
1839  * Work that must be done with write lock held:
1840  */
1841 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1842 {
1843         bool invalidated_iter = false;
1844         struct btree_node_entry *bne;
1845         struct bset_tree *t;
1846
1847         if (!btree_node_just_written(b))
1848                 return false;
1849
1850         BUG_ON(b->whiteout_u64s);
1851         BUG_ON(b->uncompacted_whiteout_u64s);
1852
1853         clear_btree_node_just_written(b);
1854
1855         /*
1856          * Note: immediately after write, bset_unwritten()/bset_written() don't
1857          * work - the amount of data we had to write after compaction might have
1858          * been smaller than the offset of the last bset.
1859          *
1860          * However, we know that all bsets have been written here, as long as
1861          * we're still holding the write lock:
1862          */
1863
1864         /*
1865          * XXX: decide if we really want to unconditionally sort down to a
1866          * single bset:
1867          */
1868         if (b->nsets > 1) {
1869                 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1870                 invalidated_iter = true;
1871         } else {
1872                 invalidated_iter = bch2_drop_whiteouts(b);
1873         }
1874
1875         for_each_bset(b, t)
1876                 set_needs_whiteout(bset(b, t));
1877
1878         bch2_btree_verify(c, b);
1879
1880         /*
1881          * If later we don't unconditionally sort down to a single bset, we have
1882          * to ensure this is still true:
1883          */
1884         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1885
1886         bne = want_new_bset(c, b);
1887         if (bne)
1888                 bch2_bset_init_next(b, &bne->keys);
1889
1890         bch2_btree_build_aux_trees(b);
1891
1892         return invalidated_iter;
1893 }
1894
1895 /*
1896  * Use this one if the node is intent locked:
1897  */
1898 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1899                           enum six_lock_type lock_type_held)
1900 {
1901         BUG_ON(lock_type_held == SIX_LOCK_write);
1902
1903         if (lock_type_held == SIX_LOCK_intent ||
1904             six_lock_tryupgrade(&b->lock)) {
1905                 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1906
1907                 /* don't cycle lock unnecessarily: */
1908                 if (btree_node_just_written(b) &&
1909                     six_trylock_write(&b->lock)) {
1910                         bch2_btree_post_write_cleanup(c, b);
1911                         six_unlock_write(&b->lock);
1912                 }
1913
1914                 if (lock_type_held == SIX_LOCK_read)
1915                         six_lock_downgrade(&b->lock);
1916         } else {
1917                 __bch2_btree_node_write(c, b, SIX_LOCK_read);
1918         }
1919 }
1920
1921 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
1922 {
1923         struct bucket_table *tbl;
1924         struct rhash_head *pos;
1925         struct btree *b;
1926         unsigned i;
1927 restart:
1928         rcu_read_lock();
1929         for_each_cached_btree(b, c, tbl, i, pos)
1930                 if (test_bit(flag, &b->flags)) {
1931                         rcu_read_unlock();
1932                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
1933                         goto restart;
1934
1935                 }
1936         rcu_read_unlock();
1937 }
1938
1939 void bch2_btree_flush_all_reads(struct bch_fs *c)
1940 {
1941         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
1942 }
1943
1944 void bch2_btree_flush_all_writes(struct bch_fs *c)
1945 {
1946         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
1947 }
1948
1949 void bch2_btree_verify_flushed(struct bch_fs *c)
1950 {
1951         struct bucket_table *tbl;
1952         struct rhash_head *pos;
1953         struct btree *b;
1954         unsigned i;
1955
1956         rcu_read_lock();
1957         for_each_cached_btree(b, c, tbl, i, pos) {
1958                 unsigned long flags = READ_ONCE(b->flags);
1959
1960                 BUG_ON((flags & (1 << BTREE_NODE_dirty)) ||
1961                        (flags & (1 << BTREE_NODE_write_in_flight)));
1962         }
1963         rcu_read_unlock();
1964 }
1965
1966 ssize_t bch2_dirty_btree_nodes_print(struct bch_fs *c, char *buf)
1967 {
1968         char *out = buf, *end = buf + PAGE_SIZE;
1969         struct bucket_table *tbl;
1970         struct rhash_head *pos;
1971         struct btree *b;
1972         unsigned i;
1973
1974         rcu_read_lock();
1975         for_each_cached_btree(b, c, tbl, i, pos) {
1976                 unsigned long flags = READ_ONCE(b->flags);
1977                 unsigned idx = (flags & (1 << BTREE_NODE_write_idx)) != 0;
1978
1979                 if (//!(flags & (1 << BTREE_NODE_dirty)) &&
1980                     !b->writes[0].wait.list.first &&
1981                     !b->writes[1].wait.list.first &&
1982                     !(b->will_make_reachable & 1))
1983                         continue;
1984
1985                 out += scnprintf(out, end - out, "%p d %u l %u w %u b %u r %u:%lu c %u p %u\n",
1986                                  b,
1987                                  (flags & (1 << BTREE_NODE_dirty)) != 0,
1988                                  b->level,
1989                                  b->written,
1990                                  !list_empty_careful(&b->write_blocked),
1991                                  b->will_make_reachable != 0,
1992                                  b->will_make_reachable & 1,
1993                                  b->writes[ idx].wait.list.first != NULL,
1994                                  b->writes[!idx].wait.list.first != NULL);
1995         }
1996         rcu_read_unlock();
1997
1998         return out - buf;
1999 }