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_reclaim.h"
  17 #include "journal_seq_blacklist.h"
  18 #include "super-io.h"
  19
  20 #include <trace/events/bcachefs.h>
  21
  22 /* btree_node_iter_large: */
  23
  24 #define btree_node_iter_cmp_heap(h, _l, _r) btree_node_iter_cmp(b, _l, _r)
  25
  26 void bch2_btree_node_iter_large_push(struct btree_node_iter_large *iter,
  27                                      struct btree *b,
  28                                      const struct bkey_packed *k,
  29                                      const struct bkey_packed *end)
  30 {
  31         if (k != end) {
  32                 struct btree_node_iter_set n =
  33                         ((struct btree_node_iter_set) {
  34                                  __btree_node_key_to_offset(b, k),
  35                                  __btree_node_key_to_offset(b, end)
  36                          });
  37
  38                 __heap_add(iter, n, btree_node_iter_cmp_heap, NULL);
  39         }
  40 }
  41
  42 void bch2_btree_node_iter_large_advance(struct btree_node_iter_large *iter,
  43                                         struct btree *b)
  44 {
  45         iter->data->k += __btree_node_offset_to_key(b, iter->data->k)->u64s;
  46
  47         EBUG_ON(!iter->used);
  48         EBUG_ON(iter->data->k > iter->data->end);
  49
  50         if (iter->data->k == iter->data->end)
  51                 heap_del(iter, 0, btree_node_iter_cmp_heap, NULL);
  52         else
  53                 heap_sift_down(iter, 0, btree_node_iter_cmp_heap, NULL);
  54 }
  55
  56 static void verify_no_dups(struct btree *b,
  57                            struct bkey_packed *start,
  58                            struct bkey_packed *end)
  59 {
  60 #ifdef CONFIG_BCACHEFS_DEBUG
  61         struct bkey_packed *k;
  62
  63         for (k = start; k != end && bkey_next(k) != end; k = bkey_next(k)) {
  64                 struct bkey l = bkey_unpack_key(b, k);
  65                 struct bkey r = bkey_unpack_key(b, bkey_next(k));
  66
  67                 BUG_ON(btree_node_is_extents(b)
  68                        ? bkey_cmp(l.p, bkey_start_pos(&r)) > 0
  69                        : bkey_cmp(l.p, bkey_start_pos(&r)) >= 0);
  70                 //BUG_ON(bkey_cmp_packed(&b->format, k, bkey_next(k)) >= 0);
  71         }
  72 #endif
  73 }
  74
  75 static void clear_needs_whiteout(struct bset *i)
  76 {
  77         struct bkey_packed *k;
  78
  79         for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
  80                 k->needs_whiteout = false;
  81 }
  82
  83 static void set_needs_whiteout(struct bset *i)
  84 {
  85         struct bkey_packed *k;
  86
  87         for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
  88                 k->needs_whiteout = true;
  89 }
  90
  91 static void btree_bounce_free(struct bch_fs *c, unsigned order,
  92                               bool used_mempool, void *p)
  93 {
  94         if (used_mempool)
  95                 mempool_free(p, &c->btree_bounce_pool);
  96         else
  97                 vpfree(p, PAGE_SIZE << order);
  98 }
  99
 100 static void *btree_bounce_alloc(struct bch_fs *c, unsigned order,
 101                                 bool *used_mempool)
 102 {
 103         void *p;
 104
 105         BUG_ON(order > btree_page_order(c));
 106
 107         *used_mempool = false;
 108         p = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT, order);
 109         if (p)
 110                 return p;
 111
 112         *used_mempool = true;
 113         return mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
 114 }
 115
 116 typedef int (*sort_cmp_fn)(struct btree *,
 117                            struct bkey_packed *,
 118                            struct bkey_packed *);
 119
 120 struct sort_iter {
 121         struct btree    *b;
 122         unsigned                used;
 123
 124         struct sort_iter_set {
 125                 struct bkey_packed *k, *end;
 126         } data[MAX_BSETS + 1];
 127 };
 128
 129 static void sort_iter_init(struct sort_iter *iter, struct btree *b)
 130 {
 131         memset(iter, 0, sizeof(*iter));
 132         iter->b = b;
 133 }
 134
 135 static inline void __sort_iter_sift(struct sort_iter *iter,
 136                                     unsigned from,
 137                                     sort_cmp_fn cmp)
 138 {
 139         unsigned i;
 140
 141         for (i = from;
 142              i + 1 < iter->used &&
 143              cmp(iter->b, iter->data[i].k, iter->data[i + 1].k) > 0;
 144              i++)
 145                 swap(iter->data[i], iter->data[i + 1]);
 146 }
 147
 148 static inline void sort_iter_sift(struct sort_iter *iter, sort_cmp_fn cmp)
 149 {
 150
 151         __sort_iter_sift(iter, 0, cmp);
 152 }
 153
 154 static inline void sort_iter_sort(struct sort_iter *iter, sort_cmp_fn cmp)
 155 {
 156         unsigned i = iter->used;
 157
 158         while (i--)
 159                 __sort_iter_sift(iter, i, cmp);
 160 }
 161
 162 static void sort_iter_add(struct sort_iter *iter,
 163                           struct bkey_packed *k,
 164                           struct bkey_packed *end)
 165 {
 166         BUG_ON(iter->used >= ARRAY_SIZE(iter->data));
 167
 168         if (k != end)
 169                 iter->data[iter->used++] = (struct sort_iter_set) { k, end };
 170 }
 171
 172 static inline struct bkey_packed *sort_iter_peek(struct sort_iter *iter)
 173 {
 174         return iter->used ? iter->data->k : NULL;
 175 }
 176
 177 static inline void sort_iter_advance(struct sort_iter *iter, sort_cmp_fn cmp)
 178 {
 179         iter->data->k = bkey_next(iter->data->k);
 180
 181         BUG_ON(iter->data->k > iter->data->end);
 182
 183         if (iter->data->k == iter->data->end)
 184                 array_remove_item(iter->data, iter->used, 0);
 185         else
 186                 sort_iter_sift(iter, cmp);
 187 }
 188
 189 static inline struct bkey_packed *sort_iter_next(struct sort_iter *iter,
 190                                                  sort_cmp_fn cmp)
 191 {
 192         struct bkey_packed *ret = sort_iter_peek(iter);
 193
 194         if (ret)
 195                 sort_iter_advance(iter, cmp);
 196
 197         return ret;
 198 }
 199
 200 static inline int sort_key_whiteouts_cmp(struct btree *b,
 201                                          struct bkey_packed *l,
 202                                          struct bkey_packed *r)
 203 {
 204         return bkey_cmp_packed(b, l, r);
 205 }
 206
 207 static unsigned sort_key_whiteouts(struct bkey_packed *dst,
 208                                    struct sort_iter *iter)
 209 {
 210         struct bkey_packed *in, *out = dst;
 211
 212         sort_iter_sort(iter, sort_key_whiteouts_cmp);
 213
 214         while ((in = sort_iter_next(iter, sort_key_whiteouts_cmp))) {
 215                 bkey_copy(out, in);
 216                 out = bkey_next(out);
 217         }
 218
 219         return (u64 *) out - (u64 *) dst;
 220 }
 221
 222 static inline int sort_extent_whiteouts_cmp(struct btree *b,
 223                                             struct bkey_packed *l,
 224                                             struct bkey_packed *r)
 225 {
 226         struct bkey ul = bkey_unpack_key(b, l);
 227         struct bkey ur = bkey_unpack_key(b, r);
 228
 229         return bkey_cmp(bkey_start_pos(&ul), bkey_start_pos(&ur));
 230 }
 231
 232 static unsigned sort_extent_whiteouts(struct bkey_packed *dst,
 233                                       struct sort_iter *iter)
 234 {
 235         const struct bkey_format *f = &iter->b->format;
 236         struct bkey_packed *in, *out = dst;
 237         struct bkey_i l, r;
 238         bool prev = false, l_packed = false;
 239         u64 max_packed_size     = bkey_field_max(f, BKEY_FIELD_SIZE);
 240         u64 max_packed_offset   = bkey_field_max(f, BKEY_FIELD_OFFSET);
 241         u64 new_size;
 242
 243         max_packed_size = min_t(u64, max_packed_size, KEY_SIZE_MAX);
 244
 245         sort_iter_sort(iter, sort_extent_whiteouts_cmp);
 246
 247         while ((in = sort_iter_next(iter, sort_extent_whiteouts_cmp))) {
 248                 if (bkey_deleted(in))
 249                         continue;
 250
 251                 EBUG_ON(bkeyp_val_u64s(f, in));
 252                 EBUG_ON(in->type != KEY_TYPE_DISCARD);
 253
 254                 r.k = bkey_unpack_key(iter->b, in);
 255
 256                 if (prev &&
 257                     bkey_cmp(l.k.p, bkey_start_pos(&r.k)) >= 0) {
 258                         if (bkey_cmp(l.k.p, r.k.p) >= 0)
 259                                 continue;
 260
 261                         new_size = l_packed
 262                                 ? min(max_packed_size, max_packed_offset -
 263                                       bkey_start_offset(&l.k))
 264                                 : KEY_SIZE_MAX;
 265
 266                         new_size = min(new_size, r.k.p.offset -
 267                                        bkey_start_offset(&l.k));
 268
 269                         BUG_ON(new_size < l.k.size);
 270
 271                         bch2_key_resize(&l.k, new_size);
 272
 273                         if (bkey_cmp(l.k.p, r.k.p) >= 0)
 274                                 continue;
 275
 276                         bch2_cut_front(l.k.p, &r);
 277                 }
 278
 279                 if (prev) {
 280                         if (!bch2_bkey_pack(out, &l, f)) {
 281                                 BUG_ON(l_packed);
 282                                 bkey_copy(out, &l);
 283                         }
 284                         out = bkey_next(out);
 285                 }
 286
 287                 l = r;
 288                 prev = true;
 289                 l_packed = bkey_packed(in);
 290         }
 291
 292         if (prev) {
 293                 if (!bch2_bkey_pack(out, &l, f)) {
 294                         BUG_ON(l_packed);
 295                         bkey_copy(out, &l);
 296                 }
 297                 out = bkey_next(out);
 298         }
 299
 300         return (u64 *) out - (u64 *) dst;
 301 }
 302
 303 static unsigned should_compact_bset(struct btree *b, struct bset_tree *t,
 304                                     bool compacting,
 305                                     enum compact_mode mode)
 306 {
 307         unsigned bset_u64s = le16_to_cpu(bset(b, t)->u64s);
 308         unsigned dead_u64s = bset_u64s - b->nr.bset_u64s[t - b->set];
 309
 310         if (mode == COMPACT_LAZY) {
 311                 if (should_compact_bset_lazy(b, t) ||
 312                     (compacting && !bset_written(b, bset(b, t))))
 313                         return dead_u64s;
 314         } else {
 315                 if (bset_written(b, bset(b, t)))
 316                         return dead_u64s;
 317         }
 318
 319         return 0;
 320 }
 321
 322 bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 323                              enum compact_mode mode)
 324 {
 325         const struct bkey_format *f = &b->format;
 326         struct bset_tree *t;
 327         struct bkey_packed *whiteouts = NULL;
 328         struct bkey_packed *u_start, *u_pos;
 329         struct sort_iter sort_iter;
 330         unsigned order, whiteout_u64s = 0, u64s;
 331         bool used_mempool, compacting = false;
 332
 333         for_each_bset(b, t)
 334                 whiteout_u64s += should_compact_bset(b, t,
 335                                         whiteout_u64s != 0, mode);
 336
 337         if (!whiteout_u64s)
 338                 return false;
 339
 340         sort_iter_init(&sort_iter, b);
 341
 342         whiteout_u64s += b->whiteout_u64s;
 343         order = get_order(whiteout_u64s * sizeof(u64));
 344
 345         whiteouts = btree_bounce_alloc(c, order, &used_mempool);
 346         u_start = u_pos = whiteouts;
 347
 348         memcpy_u64s(u_pos, unwritten_whiteouts_start(c, b),
 349                     b->whiteout_u64s);
 350         u_pos = (void *) u_pos + b->whiteout_u64s * sizeof(u64);
 351
 352         sort_iter_add(&sort_iter, u_start, u_pos);
 353
 354         for_each_bset(b, t) {
 355                 struct bset *i = bset(b, t);
 356                 struct bkey_packed *k, *n, *out, *start, *end;
 357                 struct btree_node_entry *src = NULL, *dst = NULL;
 358
 359                 if (t != b->set && !bset_written(b, i)) {
 360                         src = container_of(i, struct btree_node_entry, keys);
 361                         dst = max(write_block(b),
 362                                   (void *) btree_bkey_last(b, t -1));
 363                 }
 364
 365                 if (!should_compact_bset(b, t, compacting, mode)) {
 366                         if (src != dst) {
 367                                 memmove(dst, src, sizeof(*src) +
 368                                         le16_to_cpu(src->keys.u64s) *
 369                                         sizeof(u64));
 370                                 i = &dst->keys;
 371                                 set_btree_bset(b, t, i);
 372                         }
 373                         continue;
 374                 }
 375
 376                 compacting = true;
 377                 u_start = u_pos;
 378                 start = i->start;
 379                 end = vstruct_last(i);
 380
 381                 if (src != dst) {
 382                         memmove(dst, src, sizeof(*src));
 383                         i = &dst->keys;
 384                         set_btree_bset(b, t, i);
 385                 }
 386
 387                 out = i->start;
 388
 389                 for (k = start; k != end; k = n) {
 390                         n = bkey_next(k);
 391
 392                         if (bkey_deleted(k) && btree_node_is_extents(b))
 393                                 continue;
 394
 395                         if (bkey_whiteout(k) && !k->needs_whiteout)
 396                                 continue;
 397
 398                         if (bkey_whiteout(k)) {
 399                                 unreserve_whiteout(b, k);
 400                                 memcpy_u64s(u_pos, k, bkeyp_key_u64s(f, k));
 401                                 set_bkeyp_val_u64s(f, u_pos, 0);
 402                                 u_pos = bkey_next(u_pos);
 403                         } else if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK) {
 404                                 bkey_copy(out, k);
 405                                 out = bkey_next(out);
 406                         }
 407                 }
 408
 409                 sort_iter_add(&sort_iter, u_start, u_pos);
 410
 411                 if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK) {
 412                         i->u64s = cpu_to_le16((u64 *) out - i->_data);
 413                         set_btree_bset_end(b, t);
 414                         bch2_bset_set_no_aux_tree(b, t);
 415                 }
 416         }
 417
 418         b->whiteout_u64s = (u64 *) u_pos - (u64 *) whiteouts;
 419
 420         BUG_ON((void *) unwritten_whiteouts_start(c, b) <
 421                (void *) btree_bkey_last(b, bset_tree_last(b)));
 422
 423         u64s = btree_node_is_extents(b)
 424                 ? sort_extent_whiteouts(unwritten_whiteouts_start(c, b),
 425                                         &sort_iter)
 426                 : sort_key_whiteouts(unwritten_whiteouts_start(c, b),
 427                                      &sort_iter);
 428
 429         BUG_ON(u64s > b->whiteout_u64s);
 430         BUG_ON(u64s != b->whiteout_u64s && !btree_node_is_extents(b));
 431         BUG_ON(u_pos != whiteouts && !u64s);
 432
 433         if (u64s != b->whiteout_u64s) {
 434                 void *src = unwritten_whiteouts_start(c, b);
 435
 436                 b->whiteout_u64s = u64s;
 437                 memmove_u64s_up(unwritten_whiteouts_start(c, b), src, u64s);
 438         }
 439
 440         verify_no_dups(b,
 441                        unwritten_whiteouts_start(c, b),
 442                        unwritten_whiteouts_end(c, b));
 443
 444         btree_bounce_free(c, order, used_mempool, whiteouts);
 445
 446         if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK)
 447                 bch2_btree_build_aux_trees(b);
 448
 449         bch_btree_keys_u64s_remaining(c, b);
 450         bch2_verify_btree_nr_keys(b);
 451
 452         return true;
 453 }
 454
 455 static bool bch2_drop_whiteouts(struct btree *b)
 456 {
 457         struct bset_tree *t;
 458         bool ret = false;
 459
 460         for_each_bset(b, t) {
 461                 struct bset *i = bset(b, t);
 462                 struct bkey_packed *k, *n, *out, *start, *end;
 463
 464                 if (!should_compact_bset(b, t, true, COMPACT_WRITTEN))
 465                         continue;
 466
 467                 start   = btree_bkey_first(b, t);
 468                 end     = btree_bkey_last(b, t);
 469
 470                 if (!bset_written(b, i) &&
 471                     t != b->set) {
 472                         struct bset *dst =
 473                                max_t(struct bset *, write_block(b),
 474                                      (void *) btree_bkey_last(b, t -1));
 475
 476                         memmove(dst, i, sizeof(struct bset));
 477                         i = dst;
 478                         set_btree_bset(b, t, i);
 479                 }
 480
 481                 out = i->start;
 482
 483                 for (k = start; k != end; k = n) {
 484                         n = bkey_next(k);
 485
 486                         if (!bkey_whiteout(k)) {
 487                                 bkey_copy(out, k);
 488                                 out = bkey_next(out);
 489                         }
 490                 }
 491
 492                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 493                 bch2_bset_set_no_aux_tree(b, t);
 494                 ret = true;
 495         }
 496
 497         bch2_verify_btree_nr_keys(b);
 498
 499         return ret;
 500 }
 501
 502 static inline int sort_keys_cmp(struct btree *b,
 503                                 struct bkey_packed *l,
 504                                 struct bkey_packed *r)
 505 {
 506         return bkey_cmp_packed(b, l, r) ?:
 507                 (int) bkey_whiteout(r) - (int) bkey_whiteout(l) ?:
 508                 (int) l->needs_whiteout - (int) r->needs_whiteout;
 509 }
 510
 511 static unsigned sort_keys(struct bkey_packed *dst,
 512                           struct sort_iter *iter,
 513                           bool filter_whiteouts)
 514 {
 515         const struct bkey_format *f = &iter->b->format;
 516         struct bkey_packed *in, *next, *out = dst;
 517
 518         sort_iter_sort(iter, sort_keys_cmp);
 519
 520         while ((in = sort_iter_next(iter, sort_keys_cmp))) {
 521                 if (bkey_whiteout(in) &&
 522                     (filter_whiteouts || !in->needs_whiteout))
 523                         continue;
 524
 525                 if (bkey_whiteout(in) &&
 526                     (next = sort_iter_peek(iter)) &&
 527                     !bkey_cmp_packed(iter->b, in, next)) {
 528                         BUG_ON(in->needs_whiteout &&
 529                                next->needs_whiteout);
 530                         /*
 531                          * XXX racy, called with read lock from write path
 532                          *
 533                          * leads to spurious BUG_ON() in bkey_unpack_key() in
 534                          * debug mode
 535                          */
 536                         next->needs_whiteout |= in->needs_whiteout;
 537                         continue;
 538                 }
 539
 540                 if (bkey_whiteout(in)) {
 541                         memcpy_u64s(out, in, bkeyp_key_u64s(f, in));
 542                         set_bkeyp_val_u64s(f, out, 0);
 543                 } else {
 544                         bkey_copy(out, in);
 545                 }
 546                 out = bkey_next(out);
 547         }
 548
 549         return (u64 *) out - (u64 *) dst;
 550 }
 551
 552 static inline int sort_extents_cmp(struct btree *b,
 553                                    struct bkey_packed *l,
 554                                    struct bkey_packed *r)
 555 {
 556         return bkey_cmp_packed(b, l, r) ?:
 557                 (int) bkey_deleted(l) - (int) bkey_deleted(r);
 558 }
 559
 560 static unsigned sort_extents(struct bkey_packed *dst,
 561                              struct sort_iter *iter,
 562                              bool filter_whiteouts)
 563 {
 564         struct bkey_packed *in, *out = dst;
 565
 566         sort_iter_sort(iter, sort_extents_cmp);
 567
 568         while ((in = sort_iter_next(iter, sort_extents_cmp))) {
 569                 if (bkey_deleted(in))
 570                         continue;
 571
 572                 if (bkey_whiteout(in) &&
 573                     (filter_whiteouts || !in->needs_whiteout))
 574                         continue;
 575
 576                 bkey_copy(out, in);
 577                 out = bkey_next(out);
 578         }
 579
 580         return (u64 *) out - (u64 *) dst;
 581 }
 582
 583 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 584                             struct btree_iter *iter,
 585                             unsigned start_idx,
 586                             unsigned end_idx,
 587                             bool filter_whiteouts)
 588 {
 589         struct btree_node *out;
 590         struct sort_iter sort_iter;
 591         struct bset_tree *t;
 592         struct bset *start_bset = bset(b, &b->set[start_idx]);
 593         bool used_mempool = false;
 594         u64 start_time, seq = 0;
 595         unsigned i, u64s = 0, order, shift = end_idx - start_idx - 1;
 596         bool sorting_entire_node = start_idx == 0 &&
 597                 end_idx == b->nsets;
 598
 599         sort_iter_init(&sort_iter, b);
 600
 601         for (t = b->set + start_idx;
 602              t < b->set + end_idx;
 603              t++) {
 604                 u64s += le16_to_cpu(bset(b, t)->u64s);
 605                 sort_iter_add(&sort_iter,
 606                               btree_bkey_first(b, t),
 607                               btree_bkey_last(b, t));
 608         }
 609
 610         order = sorting_entire_node
 611                 ? btree_page_order(c)
 612                 : get_order(__vstruct_bytes(struct btree_node, u64s));
 613
 614         out = btree_bounce_alloc(c, order, &used_mempool);
 615
 616         start_time = local_clock();
 617
 618         if (btree_node_is_extents(b))
 619                 filter_whiteouts = bset_written(b, start_bset);
 620
 621         u64s = btree_node_is_extents(b)
 622                 ? sort_extents(out->keys.start, &sort_iter, filter_whiteouts)
 623                 : sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 624
 625         out->keys.u64s = cpu_to_le16(u64s);
 626
 627         BUG_ON(vstruct_end(&out->keys) > (void *) out + (PAGE_SIZE << order));
 628
 629         if (sorting_entire_node)
 630                 bch2_time_stats_update(&c->times[BCH_TIME_btree_sort],
 631                                        start_time);
 632
 633         /* Make sure we preserve bset journal_seq: */
 634         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 635                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 636         start_bset->journal_seq = cpu_to_le64(seq);
 637
 638         if (sorting_entire_node) {
 639                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 640
 641                 BUG_ON(order != btree_page_order(c));
 642
 643                 /*
 644                  * Our temporary buffer is the same size as the btree node's
 645                  * buffer, we can just swap buffers instead of doing a big
 646                  * memcpy()
 647                  */
 648                 *out = *b->data;
 649                 out->keys.u64s = cpu_to_le16(u64s);
 650                 swap(out, b->data);
 651                 set_btree_bset(b, b->set, &b->data->keys);
 652         } else {
 653                 start_bset->u64s = out->keys.u64s;
 654                 memcpy_u64s(start_bset->start,
 655                             out->keys.start,
 656                             le16_to_cpu(out->keys.u64s));
 657         }
 658
 659         for (i = start_idx + 1; i < end_idx; i++)
 660                 b->nr.bset_u64s[start_idx] +=
 661                         b->nr.bset_u64s[i];
 662
 663         b->nsets -= shift;
 664
 665         for (i = start_idx + 1; i < b->nsets; i++) {
 666                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 667                 b->set[i]               = b->set[i + shift];
 668         }
 669
 670         for (i = b->nsets; i < MAX_BSETS; i++)
 671                 b->nr.bset_u64s[i] = 0;
 672
 673         set_btree_bset_end(b, &b->set[start_idx]);
 674         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 675
 676         btree_bounce_free(c, order, used_mempool, out);
 677
 678         bch2_verify_btree_nr_keys(b);
 679 }
 680
 681 /* Sort + repack in a new format: */
 682 static struct btree_nr_keys sort_repack(struct bset *dst,
 683                                         struct btree *src,
 684                                         struct btree_node_iter *src_iter,
 685                                         struct bkey_format *out_f,
 686                                         bool filter_whiteouts)
 687 {
 688         struct bkey_format *in_f = &src->format;
 689         struct bkey_packed *in, *out = vstruct_last(dst);
 690         struct btree_nr_keys nr;
 691
 692         memset(&nr, 0, sizeof(nr));
 693
 694         while ((in = bch2_btree_node_iter_next_all(src_iter, src))) {
 695                 if (filter_whiteouts && bkey_whiteout(in))
 696                         continue;
 697
 698                 if (bch2_bkey_transform(out_f, out, bkey_packed(in)
 699                                        ? in_f : &bch2_bkey_format_current, in))
 700                         out->format = KEY_FORMAT_LOCAL_BTREE;
 701                 else
 702                         bch2_bkey_unpack(src, (void *) out, in);
 703
 704                 btree_keys_account_key_add(&nr, 0, out);
 705                 out = bkey_next(out);
 706         }
 707
 708         dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
 709         return nr;
 710 }
 711
 712 /* Sort, repack, and merge: */
 713 static struct btree_nr_keys sort_repack_merge(struct bch_fs *c,
 714                                               struct bset *dst,
 715                                               struct btree *src,
 716                                               struct btree_node_iter *iter,
 717                                               struct bkey_format *out_f,
 718                                               bool filter_whiteouts,
 719                                               key_filter_fn filter,
 720                                               key_merge_fn merge)
 721 {
 722         struct bkey_packed *k, *prev = NULL, *out;
 723         struct btree_nr_keys nr;
 724         BKEY_PADDED(k) tmp;
 725
 726         memset(&nr, 0, sizeof(nr));
 727
 728         while ((k = bch2_btree_node_iter_next_all(iter, src))) {
 729                 if (filter_whiteouts && bkey_whiteout(k))
 730                         continue;
 731
 732                 /*
 733                  * The filter might modify pointers, so we have to unpack the
 734                  * key and values to &tmp.k:
 735                  */
 736                 bch2_bkey_unpack(src, &tmp.k, k);
 737
 738                 if (filter && filter(c, src, bkey_i_to_s(&tmp.k)))
 739                         continue;
 740
 741                 /* prev is always unpacked, for key merging: */
 742
 743                 if (prev &&
 744                     merge &&
 745                     merge(c, src, (void *) prev, &tmp.k) == BCH_MERGE_MERGE)
 746                         continue;
 747
 748                 /*
 749                  * the current key becomes the new prev: advance prev, then
 750                  * copy the current key - but first pack prev (in place):
 751                  */
 752                 if (prev) {
 753                         bch2_bkey_pack(prev, (void *) prev, out_f);
 754
 755                         btree_keys_account_key_add(&nr, 0, prev);
 756                         prev = bkey_next(prev);
 757                 } else {
 758                         prev = vstruct_last(dst);
 759                 }
 760
 761                 bkey_copy(prev, &tmp.k);
 762         }
 763
 764         if (prev) {
 765                 bch2_bkey_pack(prev, (void *) prev, out_f);
 766                 btree_keys_account_key_add(&nr, 0, prev);
 767                 out = bkey_next(prev);
 768         } else {
 769                 out = vstruct_last(dst);
 770         }
 771
 772         dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
 773         return nr;
 774 }
 775
 776 void bch2_btree_sort_into(struct bch_fs *c,
 777                          struct btree *dst,
 778                          struct btree *src)
 779 {
 780         struct btree_nr_keys nr;
 781         struct btree_node_iter src_iter;
 782         u64 start_time = local_clock();
 783
 784         BUG_ON(dst->nsets != 1);
 785
 786         bch2_bset_set_no_aux_tree(dst, dst->set);
 787
 788         bch2_btree_node_iter_init_from_start(&src_iter, src);
 789
 790         if (btree_node_ops(src)->key_normalize ||
 791             btree_node_ops(src)->key_merge)
 792                 nr = sort_repack_merge(c, btree_bset_first(dst),
 793                                 src, &src_iter,
 794                                 &dst->format,
 795                                 true,
 796                                 btree_node_ops(src)->key_normalize,
 797                                 btree_node_ops(src)->key_merge);
 798         else
 799                 nr = sort_repack(btree_bset_first(dst),
 800                                 src, &src_iter,
 801                                 &dst->format,
 802                                 true);
 803
 804         bch2_time_stats_update(&c->times[BCH_TIME_btree_sort], start_time);
 805
 806         set_btree_bset_end(dst, dst->set);
 807
 808         dst->nr.live_u64s       += nr.live_u64s;
 809         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 810         dst->nr.packed_keys     += nr.packed_keys;
 811         dst->nr.unpacked_keys   += nr.unpacked_keys;
 812
 813         bch2_verify_btree_nr_keys(dst);
 814 }
 815
 816 #define SORT_CRIT       (4096 / sizeof(u64))
 817
 818 /*
 819  * We're about to add another bset to the btree node, so if there's currently
 820  * too many bsets - sort some of them together:
 821  */
 822 static bool btree_node_compact(struct bch_fs *c, struct btree *b,
 823                                struct btree_iter *iter)
 824 {
 825         unsigned unwritten_idx;
 826         bool ret = false;
 827
 828         for (unwritten_idx = 0;
 829              unwritten_idx < b->nsets;
 830              unwritten_idx++)
 831                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 832                         break;
 833
 834         if (b->nsets - unwritten_idx > 1) {
 835                 btree_node_sort(c, b, iter, unwritten_idx,
 836                                 b->nsets, false);
 837                 ret = true;
 838         }
 839
 840         if (unwritten_idx > 1) {
 841                 btree_node_sort(c, b, iter, 0, unwritten_idx, false);
 842                 ret = true;
 843         }
 844
 845         return ret;
 846 }
 847
 848 void bch2_btree_build_aux_trees(struct btree *b)
 849 {
 850         struct bset_tree *t;
 851
 852         for_each_bset(b, t)
 853                 bch2_bset_build_aux_tree(b, t,
 854                                 !bset_written(b, bset(b, t)) &&
 855                                 t == bset_tree_last(b));
 856 }
 857
 858 /*
 859  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 860  * inserted into
 861  *
 862  * Safe to call if there already is an unwritten bset - will only add a new bset
 863  * if @b doesn't already have one.
 864  *
 865  * Returns true if we sorted (i.e. invalidated iterators
 866  */
 867 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
 868                           struct btree_iter *iter)
 869 {
 870         struct btree_node_entry *bne;
 871         bool did_sort;
 872
 873         EBUG_ON(!(b->lock.state.seq & 1));
 874         EBUG_ON(iter && iter->l[b->level].b != b);
 875
 876         did_sort = btree_node_compact(c, b, iter);
 877
 878         bne = want_new_bset(c, b);
 879         if (bne)
 880                 bch2_bset_init_next(c, b, bne);
 881
 882         bch2_btree_build_aux_trees(b);
 883
 884         if (iter && did_sort)
 885                 bch2_btree_iter_reinit_node(iter, b);
 886 }
 887
 888 static struct nonce btree_nonce(struct bset *i, unsigned offset)
 889 {
 890         return (struct nonce) {{
 891                 [0] = cpu_to_le32(offset),
 892                 [1] = ((__le32 *) &i->seq)[0],
 893                 [2] = ((__le32 *) &i->seq)[1],
 894                 [3] = ((__le32 *) &i->journal_seq)[0]^BCH_NONCE_BTREE,
 895         }};
 896 }
 897
 898 static void bset_encrypt(struct bch_fs *c, struct bset *i, unsigned offset)
 899 {
 900         struct nonce nonce = btree_nonce(i, offset);
 901
 902         if (!offset) {
 903                 struct btree_node *bn = container_of(i, struct btree_node, keys);
 904                 unsigned bytes = (void *) &bn->keys - (void *) &bn->flags;
 905
 906                 bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, &bn->flags,
 907                              bytes);
 908
 909                 nonce = nonce_add(nonce, round_up(bytes, CHACHA20_BLOCK_SIZE));
 910         }
 911
 912         bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, i->_data,
 913                      vstruct_end(i) - (void *) i->_data);
 914 }
 915
 916 static int btree_err_msg(struct bch_fs *c, struct btree *b, struct bset *i,
 917                          unsigned offset, int write, char *buf, size_t len)
 918 {
 919         char *out = buf, *end = buf + len;
 920
 921         out += scnprintf(out, end - out,
 922                          "error validating btree node %s"
 923                          "at btree %u level %u/%u\n"
 924                          "pos %llu:%llu node offset %u",
 925                          write ? "before write " : "",
 926                          b->btree_id, b->level,
 927                          c->btree_roots[b->btree_id].level,
 928                          b->key.k.p.inode, b->key.k.p.offset,
 929                          b->written);
 930         if (i)
 931                 out += scnprintf(out, end - out,
 932                                  " bset u64s %u",
 933                                  le16_to_cpu(i->u64s));
 934
 935         return out - buf;
 936 }
 937
 938 enum btree_err_type {
 939         BTREE_ERR_FIXABLE,
 940         BTREE_ERR_WANT_RETRY,
 941         BTREE_ERR_MUST_RETRY,
 942         BTREE_ERR_FATAL,
 943 };
 944
 945 enum btree_validate_ret {
 946         BTREE_RETRY_READ = 64,
 947 };
 948
 949 #define btree_err(type, c, b, i, msg, ...)                              \
 950 ({                                                                      \
 951         __label__ out;                                                  \
 952         char _buf[300], *out = _buf, *end = out + sizeof(_buf);         \
 953                                                                         \
 954         out += btree_err_msg(c, b, i, b->written, write, out, end - out);\
 955         out += scnprintf(out, end - out, ": " msg, ##__VA_ARGS__);      \
 956                                                                         \
 957         if (type == BTREE_ERR_FIXABLE &&                                \
 958             write == READ &&                                            \
 959             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 960                 mustfix_fsck_err(c, "%s", _buf);                        \
 961                 goto out;                                               \
 962         }                                                               \
 963                                                                         \
 964         switch (write) {                                                \
 965         case READ:                                                      \
 966                 bch_err(c, "%s", _buf);                                 \
 967                                                                         \
 968                 switch (type) {                                         \
 969                 case BTREE_ERR_FIXABLE:                                 \
 970                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 971                         goto fsck_err;                                  \
 972                 case BTREE_ERR_WANT_RETRY:                              \
 973                         if (have_retry) {                               \
 974                                 ret = BTREE_RETRY_READ;                 \
 975                                 goto fsck_err;                          \
 976                         }                                               \
 977                         break;                                          \
 978                 case BTREE_ERR_MUST_RETRY:                              \
 979                         ret = BTREE_RETRY_READ;                         \
 980                         goto fsck_err;                                  \
 981                 case BTREE_ERR_FATAL:                                   \
 982                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 983                         goto fsck_err;                                  \
 984                 }                                                       \
 985                 break;                                                  \
 986         case WRITE:                                                     \
 987                 bch_err(c, "corrupt metadata before write: %s", _buf);  \
 988                                                                         \
 989                 if (bch2_fs_inconsistent(c)) {                          \
 990                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 991                         goto fsck_err;                                  \
 992                 }                                                       \
 993                 break;                                                  \
 994         }                                                               \
 995 out:                                                                    \
 996         true;                                                           \
 997 })
 998
 999 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
1000
1001 static int validate_bset(struct bch_fs *c, struct btree *b,
1002                          struct bset *i, unsigned sectors,
1003                          unsigned *whiteout_u64s, int write,
1004                          bool have_retry)
1005 {
1006         struct bkey_packed *k, *prev = NULL;
1007         struct bpos prev_pos = POS_MIN;
1008         enum bkey_type type = btree_node_type(b);
1009         bool seen_non_whiteout = false;
1010         const char *err;
1011         int ret = 0;
1012
1013         if (i == &b->data->keys) {
1014                 /* These indicate that we read the wrong btree node: */
1015                 btree_err_on(BTREE_NODE_ID(b->data) != b->btree_id,
1016                              BTREE_ERR_MUST_RETRY, c, b, i,
1017                              "incorrect btree id");
1018
1019                 btree_err_on(BTREE_NODE_LEVEL(b->data) != b->level,
1020                              BTREE_ERR_MUST_RETRY, c, b, i,
1021                              "incorrect level");
1022
1023                 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
1024                         u64 *p = (u64 *) &b->data->ptr;
1025
1026                         *p = swab64(*p);
1027                         bch2_bpos_swab(&b->data->min_key);
1028                         bch2_bpos_swab(&b->data->max_key);
1029                 }
1030
1031                 btree_err_on(bkey_cmp(b->data->max_key, b->key.k.p),
1032                              BTREE_ERR_MUST_RETRY, c, b, i,
1033                              "incorrect max key");
1034
1035                 /* XXX: ideally we would be validating min_key too */
1036 #if 0
1037                 /*
1038                  * not correct anymore, due to btree node write error
1039                  * handling
1040                  *
1041                  * need to add b->data->seq to btree keys and verify
1042                  * against that
1043                  */
1044                 btree_err_on(!extent_contains_ptr(bkey_i_to_s_c_extent(&b->key),
1045                                                   b->data->ptr),
1046                              BTREE_ERR_FATAL, c, b, i,
1047                              "incorrect backpointer");
1048 #endif
1049                 err = bch2_bkey_format_validate(&b->data->format);
1050                 btree_err_on(err,
1051                              BTREE_ERR_FATAL, c, b, i,
1052                              "invalid bkey format: %s", err);
1053         }
1054
1055         if (btree_err_on(le16_to_cpu(i->version) != BCACHE_BSET_VERSION,
1056                          BTREE_ERR_FIXABLE, c, b, i,
1057                          "unsupported bset version")) {
1058                 i->version = cpu_to_le16(BCACHE_BSET_VERSION);
1059                 i->u64s = 0;
1060                 return 0;
1061         }
1062
1063         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
1064                          BTREE_ERR_FIXABLE, c, b, i,
1065                          "bset past end of btree node")) {
1066                 i->u64s = 0;
1067                 return 0;
1068         }
1069
1070         btree_err_on(b->written && !i->u64s,
1071                      BTREE_ERR_FIXABLE, c, b, i,
1072                      "empty bset");
1073
1074         if (!BSET_SEPARATE_WHITEOUTS(i)) {
1075                 seen_non_whiteout = true;
1076                 *whiteout_u64s = 0;
1077         }
1078
1079         for (k = i->start;
1080              k != vstruct_last(i);) {
1081                 struct bkey_s_c u;
1082                 struct bkey tmp;
1083                 const char *invalid;
1084
1085                 if (btree_err_on(!k->u64s,
1086                                  BTREE_ERR_FIXABLE, c, b, i,
1087                                  "KEY_U64s 0: %zu bytes of metadata lost",
1088                                  vstruct_end(i) - (void *) k)) {
1089                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
1090                         break;
1091                 }
1092
1093                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
1094                                  BTREE_ERR_FIXABLE, c, b, i,
1095                                  "key extends past end of bset")) {
1096                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
1097                         break;
1098                 }
1099
1100                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
1101                                  BTREE_ERR_FIXABLE, c, b, i,
1102                                  "invalid bkey format %u", k->format)) {
1103                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1104                         memmove_u64s_down(k, bkey_next(k),
1105                                           (u64 *) vstruct_end(i) - (u64 *) k);
1106                         continue;
1107                 }
1108
1109                 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN)
1110                         bch2_bkey_swab(type, &b->format, k);
1111
1112                 u = bkey_disassemble(b, k, &tmp);
1113
1114                 invalid = __bch2_bkey_invalid(c, type, u) ?:
1115                         bch2_bkey_in_btree_node(b, u) ?:
1116                         (write ? bch2_bkey_val_invalid(c, type, u) : NULL);
1117                 if (invalid) {
1118                         char buf[160];
1119
1120                         bch2_bkey_val_to_text(c, type, buf, sizeof(buf), u);
1121                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
1122                                   "invalid bkey:\n%s\n%s", invalid, buf);
1123
1124                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1125                         memmove_u64s_down(k, bkey_next(k),
1126                                           (u64 *) vstruct_end(i) - (u64 *) k);
1127                         continue;
1128                 }
1129
1130                 /*
1131                  * with the separate whiteouts thing (used for extents), the
1132                  * second set of keys actually can have whiteouts too, so we
1133                  * can't solely go off bkey_whiteout()...
1134                  */
1135
1136                 if (!seen_non_whiteout &&
1137                     (!bkey_whiteout(k) ||
1138                      (bkey_cmp(prev_pos, bkey_start_pos(u.k)) > 0))) {
1139                         *whiteout_u64s = k->_data - i->_data;
1140                         seen_non_whiteout = true;
1141                 } else if (bkey_cmp(prev_pos, bkey_start_pos(u.k)) > 0) {
1142                         btree_err(BTREE_ERR_FATAL, c, b, i,
1143                                   "keys out of order: %llu:%llu > %llu:%llu",
1144                                   prev_pos.inode,
1145                                   prev_pos.offset,
1146                                   u.k->p.inode,
1147                                   bkey_start_offset(u.k));
1148                         /* XXX: repair this */
1149                 }
1150
1151                 prev_pos = u.k->p;
1152                 prev = k;
1153                 k = bkey_next(k);
1154         }
1155
1156         SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1157 fsck_err:
1158         return ret;
1159 }
1160
1161 int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry)
1162 {
1163         struct btree_node_entry *bne;
1164         struct btree_node_iter_large *iter;
1165         struct btree_node *sorted;
1166         struct bkey_packed *k;
1167         struct bset *i;
1168         bool used_mempool;
1169         unsigned u64s;
1170         int ret, retry_read = 0, write = READ;
1171
1172         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
1173         iter->used = 0;
1174
1175         if (bch2_meta_read_fault("btree"))
1176                 btree_err(BTREE_ERR_MUST_RETRY, c, b, NULL,
1177                           "dynamic fault");
1178
1179         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
1180                      BTREE_ERR_MUST_RETRY, c, b, NULL,
1181                      "bad magic");
1182
1183         btree_err_on(!b->data->keys.seq,
1184                      BTREE_ERR_MUST_RETRY, c, b, NULL,
1185                      "bad btree header");
1186
1187         while (b->written < c->opts.btree_node_size) {
1188                 unsigned sectors, whiteout_u64s = 0;
1189                 struct nonce nonce;
1190                 struct bch_csum csum;
1191                 bool first = !b->written;
1192
1193                 if (!b->written) {
1194                         i = &b->data->keys;
1195
1196                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1197                                      BTREE_ERR_WANT_RETRY, c, b, i,
1198                                      "unknown checksum type");
1199
1200                         nonce = btree_nonce(i, b->written << 9);
1201                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
1202
1203                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
1204                                      BTREE_ERR_WANT_RETRY, c, b, i,
1205                                      "invalid checksum");
1206
1207                         bset_encrypt(c, i, b->written << 9);
1208
1209                         sectors = vstruct_sectors(b->data, c->block_bits);
1210
1211                         btree_node_set_format(b, b->data->format);
1212                 } else {
1213                         bne = write_block(b);
1214                         i = &bne->keys;
1215
1216                         if (i->seq != b->data->keys.seq)
1217                                 break;
1218
1219                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1220                                      BTREE_ERR_WANT_RETRY, c, b, i,
1221                                      "unknown checksum type");
1222
1223                         nonce = btree_nonce(i, b->written << 9);
1224                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1225
1226                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
1227                                      BTREE_ERR_WANT_RETRY, c, b, i,
1228                                      "invalid checksum");
1229
1230                         bset_encrypt(c, i, b->written << 9);
1231
1232                         sectors = vstruct_sectors(bne, c->block_bits);
1233                 }
1234
1235                 ret = validate_bset(c, b, i, sectors, &whiteout_u64s,
1236                                     READ, have_retry);
1237                 if (ret)
1238                         goto fsck_err;
1239
1240                 b->written += sectors;
1241
1242                 ret = bch2_journal_seq_should_ignore(c, le64_to_cpu(i->journal_seq), b);
1243                 if (ret < 0) {
1244                         btree_err(BTREE_ERR_FATAL, c, b, i,
1245                                   "insufficient memory");
1246                         goto err;
1247                 }
1248
1249                 if (ret) {
1250                         btree_err_on(first,
1251                                      BTREE_ERR_FIXABLE, c, b, i,
1252                                      "first btree node bset has blacklisted journal seq");
1253                         if (!first)
1254                                 continue;
1255                 }
1256
1257                 bch2_btree_node_iter_large_push(iter, b,
1258                                            i->start,
1259                                            vstruct_idx(i, whiteout_u64s));
1260
1261                 bch2_btree_node_iter_large_push(iter, b,
1262                                            vstruct_idx(i, whiteout_u64s),
1263                                            vstruct_last(i));
1264         }
1265
1266         for (bne = write_block(b);
1267              bset_byte_offset(b, bne) < btree_bytes(c);
1268              bne = (void *) bne + block_bytes(c))
1269                 btree_err_on(bne->keys.seq == b->data->keys.seq,
1270                              BTREE_ERR_WANT_RETRY, c, b, NULL,
1271                              "found bset signature after last bset");
1272
1273         sorted = btree_bounce_alloc(c, btree_page_order(c), &used_mempool);
1274         sorted->keys.u64s = 0;
1275
1276         set_btree_bset(b, b->set, &b->data->keys);
1277
1278         b->nr = btree_node_is_extents(b)
1279                 ? bch2_extent_sort_fix_overlapping(c, &sorted->keys, b, iter)
1280                 : bch2_key_sort_fix_overlapping(&sorted->keys, b, iter);
1281
1282         u64s = le16_to_cpu(sorted->keys.u64s);
1283         *sorted = *b->data;
1284         sorted->keys.u64s = cpu_to_le16(u64s);
1285         swap(sorted, b->data);
1286         set_btree_bset(b, b->set, &b->data->keys);
1287         b->nsets = 1;
1288
1289         BUG_ON(b->nr.live_u64s != u64s);
1290
1291         btree_bounce_free(c, btree_page_order(c), used_mempool, sorted);
1292
1293         i = &b->data->keys;
1294         for (k = i->start; k != vstruct_last(i);) {
1295                 enum bkey_type type = btree_node_type(b);
1296                 struct bkey tmp;
1297                 struct bkey_s_c u = bkey_disassemble(b, k, &tmp);
1298                 const char *invalid = bch2_bkey_val_invalid(c, type, u);
1299
1300                 if (invalid ||
1301                     (inject_invalid_keys(c) &&
1302                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1303                         char buf[160];
1304
1305                         bch2_bkey_val_to_text(c, type, buf, sizeof(buf), u);
1306                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
1307                                   "invalid bkey %s: %s", buf, invalid);
1308
1309                         btree_keys_account_key_drop(&b->nr, 0, k);
1310
1311                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1312                         memmove_u64s_down(k, bkey_next(k),
1313                                           (u64 *) vstruct_end(i) - (u64 *) k);
1314                         set_btree_bset_end(b, b->set);
1315                         continue;
1316                 }
1317
1318                 k = bkey_next(k);
1319         }
1320
1321         bch2_bset_build_aux_tree(b, b->set, false);
1322
1323         set_needs_whiteout(btree_bset_first(b));
1324
1325         btree_node_reset_sib_u64s(b);
1326 out:
1327         mempool_free(iter, &c->fill_iter);
1328         return retry_read;
1329 err:
1330 fsck_err:
1331         if (ret == BTREE_RETRY_READ) {
1332                 retry_read = 1;
1333         } else {
1334                 bch2_inconsistent_error(c);
1335                 set_btree_node_read_error(b);
1336         }
1337         goto out;
1338 }
1339
1340 static void btree_node_read_work(struct work_struct *work)
1341 {
1342         struct btree_read_bio *rb =
1343                 container_of(work, struct btree_read_bio, work);
1344         struct bch_fs *c        = rb->c;
1345         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1346         struct btree *b         = rb->bio.bi_private;
1347         struct bio *bio         = &rb->bio;
1348         struct bch_io_failures failed = { .nr = 0 };
1349         bool can_retry;
1350
1351         goto start;
1352         while (1) {
1353                 bch_info(c, "retrying read");
1354                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1355                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1356                 bio_reset(bio);
1357                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1358                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1359                 bio->bi_iter.bi_size    = btree_bytes(c);
1360
1361                 if (rb->have_ioref) {
1362                         bio_set_dev(bio, ca->disk_sb.bdev);
1363                         submit_bio_wait(bio);
1364                 } else {
1365                         bio->bi_status = BLK_STS_REMOVED;
1366                 }
1367 start:
1368                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read");
1369                 if (rb->have_ioref)
1370                         percpu_ref_put(&ca->io_ref);
1371                 rb->have_ioref = false;
1372
1373                 bch2_mark_io_failure(&failed, &rb->pick);
1374
1375                 can_retry = bch2_btree_pick_ptr(c, b, &failed, &rb->pick) > 0;
1376
1377                 if (!bio->bi_status &&
1378                     !bch2_btree_node_read_done(c, b, can_retry))
1379                         break;
1380
1381                 if (!can_retry) {
1382                         set_btree_node_read_error(b);
1383                         break;
1384                 }
1385         }
1386
1387         bch2_time_stats_update(&c->times[BCH_TIME_btree_read], rb->start_time);
1388         bio_put(&rb->bio);
1389         clear_btree_node_read_in_flight(b);
1390         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1391 }
1392
1393 static void btree_node_read_endio(struct bio *bio)
1394 {
1395         struct btree_read_bio *rb =
1396                 container_of(bio, struct btree_read_bio, bio);
1397         struct bch_fs *c        = rb->c;
1398
1399         if (rb->have_ioref) {
1400                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1401                 bch2_latency_acct(ca, rb->start_time, READ);
1402         }
1403
1404         queue_work(system_unbound_wq, &rb->work);
1405 }
1406
1407 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1408                           bool sync)
1409 {
1410         struct extent_ptr_decoded pick;
1411         struct btree_read_bio *rb;
1412         struct bch_dev *ca;
1413         struct bio *bio;
1414         int ret;
1415
1416         trace_btree_read(c, b);
1417
1418         ret = bch2_btree_pick_ptr(c, b, NULL, &pick);
1419         if (bch2_fs_fatal_err_on(ret <= 0, c,
1420                         "btree node read error: no device to read from")) {
1421                 set_btree_node_read_error(b);
1422                 return;
1423         }
1424
1425         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1426
1427         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1428                                                    btree_bytes(c)),
1429                                &c->btree_bio);
1430         rb = container_of(bio, struct btree_read_bio, bio);
1431         rb->c                   = c;
1432         rb->start_time          = local_clock();
1433         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1434         rb->pick                = pick;
1435         INIT_WORK(&rb->work, btree_node_read_work);
1436         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1437         bio->bi_iter.bi_sector  = pick.ptr.offset;
1438         bio->bi_iter.bi_size    = btree_bytes(c);
1439         bio->bi_end_io          = btree_node_read_endio;
1440         bio->bi_private         = b;
1441         bch2_bio_map(bio, b->data);
1442
1443         set_btree_node_read_in_flight(b);
1444
1445         if (rb->have_ioref) {
1446                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_BTREE],
1447                              bio_sectors(bio));
1448                 bio_set_dev(bio, ca->disk_sb.bdev);
1449
1450                 if (sync) {
1451                         submit_bio_wait(bio);
1452
1453                         bio->bi_private = b;
1454                         btree_node_read_work(&rb->work);
1455                 } else {
1456                         submit_bio(bio);
1457                 }
1458         } else {
1459                 bio->bi_status = BLK_STS_REMOVED;
1460
1461                 if (sync)
1462                         btree_node_read_work(&rb->work);
1463                 else
1464                         queue_work(system_unbound_wq, &rb->work);
1465
1466         }
1467 }
1468
1469 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1470                         const struct bkey_i *k, unsigned level)
1471 {
1472         struct closure cl;
1473         struct btree *b;
1474         int ret;
1475
1476         closure_init_stack(&cl);
1477
1478         do {
1479                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1480                 closure_sync(&cl);
1481         } while (ret);
1482
1483         b = bch2_btree_node_mem_alloc(c);
1484         bch2_btree_cache_cannibalize_unlock(c);
1485
1486         BUG_ON(IS_ERR(b));
1487
1488         bkey_copy(&b->key, k);
1489         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1490
1491         bch2_btree_node_read(c, b, true);
1492
1493         if (btree_node_read_error(b)) {
1494                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1495
1496                 mutex_lock(&c->btree_cache.lock);
1497                 list_move(&b->list, &c->btree_cache.freeable);
1498                 mutex_unlock(&c->btree_cache.lock);
1499
1500                 ret = -EIO;
1501                 goto err;
1502         }
1503
1504         bch2_btree_set_root_for_read(c, b);
1505 err:
1506         six_unlock_write(&b->lock);
1507         six_unlock_intent(&b->lock);
1508
1509         return ret;
1510 }
1511
1512 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1513                               struct btree_write *w)
1514 {
1515         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1516
1517         do {
1518                 old = new = v;
1519                 if (!(old & 1))
1520                         break;
1521
1522                 new &= ~1UL;
1523         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1524
1525         if (old & 1)
1526                 closure_put(&((struct btree_update *) new)->cl);
1527
1528         bch2_journal_pin_drop(&c->journal, &w->journal);
1529         closure_wake_up(&w->wait);
1530 }
1531
1532 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1533 {
1534         struct btree_write *w = btree_prev_write(b);
1535
1536         bch2_btree_complete_write(c, b, w);
1537         btree_node_io_unlock(b);
1538 }
1539
1540 static void bch2_btree_node_write_error(struct bch_fs *c,
1541                                         struct btree_write_bio *wbio)
1542 {
1543         struct btree *b         = wbio->wbio.bio.bi_private;
1544         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1545         struct bkey_i_extent *new_key;
1546         struct bkey_s_extent e;
1547         struct bch_extent_ptr *ptr;
1548         struct btree_iter iter;
1549         int ret;
1550
1551         __bch2_btree_iter_init(&iter, c, b->btree_id, b->key.k.p,
1552                                BTREE_MAX_DEPTH,
1553                                b->level, BTREE_ITER_NODES);
1554 retry:
1555         ret = bch2_btree_iter_traverse(&iter);
1556         if (ret)
1557                 goto err;
1558
1559         /* has node been freed? */
1560         if (iter.l[b->level].b != b) {
1561                 /* node has been freed: */
1562                 BUG_ON(!btree_node_dying(b));
1563                 goto out;
1564         }
1565
1566         BUG_ON(!btree_node_hashed(b));
1567
1568         bkey_copy(&tmp.k, &b->key);
1569
1570         new_key = bkey_i_to_extent(&tmp.k);
1571         e = extent_i_to_s(new_key);
1572
1573         bch2_extent_drop_ptrs(e, ptr,
1574                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1575
1576         if (!bch2_extent_nr_ptrs(e.c))
1577                 goto err;
1578
1579         ret = bch2_btree_node_update_key(c, &iter, b, new_key);
1580         if (ret == -EINTR)
1581                 goto retry;
1582         if (ret)
1583                 goto err;
1584 out:
1585         bch2_btree_iter_unlock(&iter);
1586         bio_put(&wbio->wbio.bio);
1587         btree_node_write_done(c, b);
1588         return;
1589 err:
1590         set_btree_node_noevict(b);
1591         bch2_fs_fatal_error(c, "fatal error writing btree node");
1592         goto out;
1593 }
1594
1595 void bch2_btree_write_error_work(struct work_struct *work)
1596 {
1597         struct bch_fs *c = container_of(work, struct bch_fs,
1598                                         btree_write_error_work);
1599         struct bio *bio;
1600
1601         while (1) {
1602                 spin_lock_irq(&c->btree_write_error_lock);
1603                 bio = bio_list_pop(&c->btree_write_error_list);
1604                 spin_unlock_irq(&c->btree_write_error_lock);
1605
1606                 if (!bio)
1607                         break;
1608
1609                 bch2_btree_node_write_error(c,
1610                         container_of(bio, struct btree_write_bio, wbio.bio));
1611         }
1612 }
1613
1614 static void btree_node_write_work(struct work_struct *work)
1615 {
1616         struct btree_write_bio *wbio =
1617                 container_of(work, struct btree_write_bio, work);
1618         struct bch_fs *c        = wbio->wbio.c;
1619         struct btree *b         = wbio->wbio.bio.bi_private;
1620
1621         btree_bounce_free(c,
1622                 wbio->wbio.order,
1623                 wbio->wbio.used_mempool,
1624                 wbio->data);
1625
1626         if (wbio->wbio.failed.nr) {
1627                 unsigned long flags;
1628
1629                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1630                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1631                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1632
1633                 queue_work(c->wq, &c->btree_write_error_work);
1634                 return;
1635         }
1636
1637         bio_put(&wbio->wbio.bio);
1638         btree_node_write_done(c, b);
1639 }
1640
1641 static void btree_node_write_endio(struct bio *bio)
1642 {
1643         struct bch_write_bio *wbio      = to_wbio(bio);
1644         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1645         struct bch_write_bio *orig      = parent ?: wbio;
1646         struct bch_fs *c                = wbio->c;
1647         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1648         unsigned long flags;
1649
1650         if (wbio->have_ioref)
1651                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1652
1653         if (bio->bi_status == BLK_STS_REMOVED ||
1654             bch2_dev_io_err_on(bio->bi_status, ca, "btree write") ||
1655             bch2_meta_write_fault("btree")) {
1656                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1657                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1658                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1659         }
1660
1661         if (wbio->have_ioref)
1662                 percpu_ref_put(&ca->io_ref);
1663
1664         if (parent) {
1665                 bio_put(bio);
1666                 bio_endio(&parent->bio);
1667         } else {
1668                 struct btree_write_bio *wb =
1669                         container_of(orig, struct btree_write_bio, wbio);
1670
1671                 INIT_WORK(&wb->work, btree_node_write_work);
1672                 queue_work(system_unbound_wq, &wb->work);
1673         }
1674 }
1675
1676 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1677                                    struct bset *i, unsigned sectors)
1678 {
1679         const struct bch_extent_ptr *ptr;
1680         unsigned whiteout_u64s = 0;
1681         int ret;
1682
1683         extent_for_each_ptr(bkey_i_to_s_c_extent(&b->key), ptr)
1684                 break;
1685
1686         ret = validate_bset(c, b, i, sectors, &whiteout_u64s, WRITE, false);
1687         if (ret)
1688                 bch2_inconsistent_error(c);
1689
1690         return ret;
1691 }
1692
1693 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1694                             enum six_lock_type lock_type_held)
1695 {
1696         struct btree_write_bio *wbio;
1697         struct bset_tree *t;
1698         struct bset *i;
1699         struct btree_node *bn = NULL;
1700         struct btree_node_entry *bne = NULL;
1701         BKEY_PADDED(key) k;
1702         struct bkey_s_extent e;
1703         struct bch_extent_ptr *ptr;
1704         struct sort_iter sort_iter;
1705         struct nonce nonce;
1706         unsigned bytes_to_write, sectors_to_write, order, bytes, u64s;
1707         u64 seq = 0;
1708         bool used_mempool;
1709         unsigned long old, new;
1710         void *data;
1711
1712         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1713                 return;
1714
1715         /*
1716          * We may only have a read lock on the btree node - the dirty bit is our
1717          * "lock" against racing with other threads that may be trying to start
1718          * a write, we do a write iff we clear the dirty bit. Since setting the
1719          * dirty bit requires a write lock, we can't race with other threads
1720          * redirtying it:
1721          */
1722         do {
1723                 old = new = READ_ONCE(b->flags);
1724
1725                 if (!(old & (1 << BTREE_NODE_dirty)))
1726                         return;
1727
1728                 if (b->written &&
1729                     !btree_node_may_write(b))
1730                         return;
1731
1732                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1733                         btree_node_wait_on_io(b);
1734                         continue;
1735                 }
1736
1737                 new &= ~(1 << BTREE_NODE_dirty);
1738                 new &= ~(1 << BTREE_NODE_need_write);
1739                 new |=  (1 << BTREE_NODE_write_in_flight);
1740                 new |=  (1 << BTREE_NODE_just_written);
1741                 new ^=  (1 << BTREE_NODE_write_idx);
1742         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1743
1744         BUG_ON(btree_node_fake(b));
1745         BUG_ON(!list_empty(&b->write_blocked));
1746         BUG_ON((b->will_make_reachable != 0) != !b->written);
1747
1748         BUG_ON(b->written >= c->opts.btree_node_size);
1749         BUG_ON(b->written & (c->opts.block_size - 1));
1750         BUG_ON(bset_written(b, btree_bset_last(b)));
1751         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1752         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1753
1754         /*
1755          * We can't block on six_lock_write() here; another thread might be
1756          * trying to get a journal reservation with read locks held, and getting
1757          * a journal reservation might be blocked on flushing the journal and
1758          * doing btree writes:
1759          */
1760         if (lock_type_held == SIX_LOCK_intent &&
1761             six_trylock_write(&b->lock)) {
1762                 __bch2_compact_whiteouts(c, b, COMPACT_WRITTEN);
1763                 six_unlock_write(&b->lock);
1764         } else {
1765                 __bch2_compact_whiteouts(c, b, COMPACT_WRITTEN_NO_WRITE_LOCK);
1766         }
1767
1768         BUG_ON(b->uncompacted_whiteout_u64s);
1769
1770         sort_iter_init(&sort_iter, b);
1771
1772         bytes = !b->written
1773                 ? sizeof(struct btree_node)
1774                 : sizeof(struct btree_node_entry);
1775
1776         bytes += b->whiteout_u64s * sizeof(u64);
1777
1778         for_each_bset(b, t) {
1779                 i = bset(b, t);
1780
1781                 if (bset_written(b, i))
1782                         continue;
1783
1784                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1785                 sort_iter_add(&sort_iter,
1786                               btree_bkey_first(b, t),
1787                               btree_bkey_last(b, t));
1788                 seq = max(seq, le64_to_cpu(i->journal_seq));
1789         }
1790
1791         order = get_order(bytes);
1792         data = btree_bounce_alloc(c, order, &used_mempool);
1793
1794         if (!b->written) {
1795                 bn = data;
1796                 *bn = *b->data;
1797                 i = &bn->keys;
1798         } else {
1799                 bne = data;
1800                 bne->keys = b->data->keys;
1801                 i = &bne->keys;
1802         }
1803
1804         i->journal_seq  = cpu_to_le64(seq);
1805         i->u64s         = 0;
1806
1807         if (!btree_node_is_extents(b)) {
1808                 sort_iter_add(&sort_iter,
1809                               unwritten_whiteouts_start(c, b),
1810                               unwritten_whiteouts_end(c, b));
1811                 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1812         } else {
1813                 memcpy_u64s(i->start,
1814                             unwritten_whiteouts_start(c, b),
1815                             b->whiteout_u64s);
1816                 i->u64s = cpu_to_le16(b->whiteout_u64s);
1817                 SET_BSET_SEPARATE_WHITEOUTS(i, true);
1818         }
1819
1820         b->whiteout_u64s = 0;
1821
1822         u64s = btree_node_is_extents(b)
1823                 ? sort_extents(vstruct_last(i), &sort_iter, false)
1824                 : sort_keys(i->start, &sort_iter, false);
1825         le16_add_cpu(&i->u64s, u64s);
1826
1827         clear_needs_whiteout(i);
1828
1829         /* do we have data to write? */
1830         if (b->written && !i->u64s)
1831                 goto nowrite;
1832
1833         bytes_to_write = vstruct_end(i) - data;
1834         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1835
1836         memset(data + bytes_to_write, 0,
1837                (sectors_to_write << 9) - bytes_to_write);
1838
1839         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1840         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1841         BUG_ON(i->seq != b->data->keys.seq);
1842
1843         i->version = cpu_to_le16(BCACHE_BSET_VERSION);
1844         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1845
1846         /* if we're going to be encrypting, check metadata validity first: */
1847         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)) &&
1848             validate_bset_for_write(c, b, i, sectors_to_write))
1849                 goto err;
1850
1851         bset_encrypt(c, i, b->written << 9);
1852
1853         nonce = btree_nonce(i, b->written << 9);
1854
1855         if (bn)
1856                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1857         else
1858                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1859
1860         /* if we're not encrypting, check metadata after checksumming: */
1861         if (!bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)) &&
1862             validate_bset_for_write(c, b, i, sectors_to_write))
1863                 goto err;
1864
1865         /*
1866          * We handle btree write errors by immediately halting the journal -
1867          * after we've done that, we can't issue any subsequent btree writes
1868          * because they might have pointers to new nodes that failed to write.
1869          *
1870          * Furthermore, there's no point in doing any more btree writes because
1871          * with the journal stopped, we're never going to update the journal to
1872          * reflect that those writes were done and the data flushed from the
1873          * journal:
1874          *
1875          * Make sure to update b->written so bch2_btree_init_next() doesn't
1876          * break:
1877          */
1878         if (bch2_journal_error(&c->journal) ||
1879             c->opts.nochanges)
1880                 goto err;
1881
1882         trace_btree_write(b, bytes_to_write, sectors_to_write);
1883
1884         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1885                                 buf_pages(data, sectors_to_write << 9),
1886                                 &c->btree_bio),
1887                             struct btree_write_bio, wbio.bio);
1888         wbio_init(&wbio->wbio.bio);
1889         wbio->data                      = data;
1890         wbio->wbio.order                = order;
1891         wbio->wbio.used_mempool         = used_mempool;
1892         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META|REQ_FUA;
1893         wbio->wbio.bio.bi_iter.bi_size  = sectors_to_write << 9;
1894         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1895         wbio->wbio.bio.bi_private       = b;
1896
1897         bch2_bio_map(&wbio->wbio.bio, data);
1898
1899         /*
1900          * If we're appending to a leaf node, we don't technically need FUA -
1901          * this write just needs to be persisted before the next journal write,
1902          * which will be marked FLUSH|FUA.
1903          *
1904          * Similarly if we're writing a new btree root - the pointer is going to
1905          * be in the next journal entry.
1906          *
1907          * But if we're writing a new btree node (that isn't a root) or
1908          * appending to a non leaf btree node, we need either FUA or a flush
1909          * when we write the parent with the new pointer. FUA is cheaper than a
1910          * flush, and writes appending to leaf nodes aren't blocking anything so
1911          * just make all btree node writes FUA to keep things sane.
1912          */
1913
1914         bkey_copy(&k.key, &b->key);
1915         e = bkey_i_to_s_extent(&k.key);
1916
1917         extent_for_each_ptr(e, ptr)
1918                 ptr->offset += b->written;
1919
1920         b->written += sectors_to_write;
1921
1922         bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_BTREE, &k.key);
1923         return;
1924 err:
1925         set_btree_node_noevict(b);
1926         b->written += sectors_to_write;
1927 nowrite:
1928         btree_bounce_free(c, order, used_mempool, data);
1929         btree_node_write_done(c, b);
1930 }
1931
1932 /*
1933  * Work that must be done with write lock held:
1934  */
1935 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1936 {
1937         bool invalidated_iter = false;
1938         struct btree_node_entry *bne;
1939         struct bset_tree *t;
1940
1941         if (!btree_node_just_written(b))
1942                 return false;
1943
1944         BUG_ON(b->whiteout_u64s);
1945         BUG_ON(b->uncompacted_whiteout_u64s);
1946
1947         clear_btree_node_just_written(b);
1948
1949         /*
1950          * Note: immediately after write, bset_written() doesn't work - the
1951          * amount of data we had to write after compaction might have been
1952          * smaller than the offset of the last bset.
1953          *
1954          * However, we know that all bsets have been written here, as long as
1955          * we're still holding the write lock:
1956          */
1957
1958         /*
1959          * XXX: decide if we really want to unconditionally sort down to a
1960          * single bset:
1961          */
1962         if (b->nsets > 1) {
1963                 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1964                 invalidated_iter = true;
1965         } else {
1966                 invalidated_iter = bch2_drop_whiteouts(b);
1967         }
1968
1969         for_each_bset(b, t)
1970                 set_needs_whiteout(bset(b, t));
1971
1972         bch2_btree_verify(c, b);
1973
1974         /*
1975          * If later we don't unconditionally sort down to a single bset, we have
1976          * to ensure this is still true:
1977          */
1978         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1979
1980         bne = want_new_bset(c, b);
1981         if (bne)
1982                 bch2_bset_init_next(c, b, bne);
1983
1984         bch2_btree_build_aux_trees(b);
1985
1986         return invalidated_iter;
1987 }
1988
1989 /*
1990  * Use this one if the node is intent locked:
1991  */
1992 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1993                           enum six_lock_type lock_type_held)
1994 {
1995         BUG_ON(lock_type_held == SIX_LOCK_write);
1996
1997         if (lock_type_held == SIX_LOCK_intent ||
1998             six_lock_tryupgrade(&b->lock)) {
1999                 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
2000
2001                 /* don't cycle lock unnecessarily: */
2002                 if (btree_node_just_written(b) &&
2003                     six_trylock_write(&b->lock)) {
2004                         bch2_btree_post_write_cleanup(c, b);
2005                         six_unlock_write(&b->lock);
2006                 }
2007
2008                 if (lock_type_held == SIX_LOCK_read)
2009                         six_lock_downgrade(&b->lock);
2010         } else {
2011                 __bch2_btree_node_write(c, b, SIX_LOCK_read);
2012         }
2013 }
2014
2015 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2016 {
2017         struct bucket_table *tbl;
2018         struct rhash_head *pos;
2019         struct btree *b;
2020         unsigned i;
2021 restart:
2022         rcu_read_lock();
2023         for_each_cached_btree(b, c, tbl, i, pos)
2024                 if (test_bit(flag, &b->flags)) {
2025                         rcu_read_unlock();
2026                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2027                         goto restart;
2028
2029                 }
2030         rcu_read_unlock();
2031 }
2032
2033 void bch2_btree_flush_all_reads(struct bch_fs *c)
2034 {
2035         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2036 }
2037
2038 void bch2_btree_flush_all_writes(struct bch_fs *c)
2039 {
2040         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2041 }
2042
2043 void bch2_btree_verify_flushed(struct bch_fs *c)
2044 {
2045         struct bucket_table *tbl;
2046         struct rhash_head *pos;
2047         struct btree *b;
2048         unsigned i;
2049
2050         rcu_read_lock();
2051         for_each_cached_btree(b, c, tbl, i, pos) {
2052                 unsigned long flags = READ_ONCE(b->flags);
2053
2054                 BUG_ON((flags & (1 << BTREE_NODE_dirty)) ||
2055                        (flags & (1 << BTREE_NODE_write_in_flight)));
2056         }
2057         rcu_read_unlock();
2058 }
2059
2060 ssize_t bch2_dirty_btree_nodes_print(struct bch_fs *c, char *buf)
2061 {
2062         char *out = buf, *end = buf + PAGE_SIZE;
2063         struct bucket_table *tbl;
2064         struct rhash_head *pos;
2065         struct btree *b;
2066         unsigned i;
2067
2068         rcu_read_lock();
2069         for_each_cached_btree(b, c, tbl, i, pos) {
2070                 unsigned long flags = READ_ONCE(b->flags);
2071                 unsigned idx = (flags & (1 << BTREE_NODE_write_idx)) != 0;
2072
2073                 if (//!(flags & (1 << BTREE_NODE_dirty)) &&
2074                     !b->writes[0].wait.list.first &&
2075                     !b->writes[1].wait.list.first &&
2076                     !(b->will_make_reachable & 1))
2077                         continue;
2078
2079                 out += scnprintf(out, end - out, "%p d %u l %u w %u b %u r %u:%lu c %u p %u\n",
2080                                  b,
2081                                  (flags & (1 << BTREE_NODE_dirty)) != 0,
2082                                  b->level,
2083                                  b->written,
2084                                  !list_empty_careful(&b->write_blocked),
2085                                  b->will_make_reachable != 0,
2086                                  b->will_make_reachable & 1,
2087                                  b->writes[ idx].wait.list.first != NULL,
2088                                  b->writes[!idx].wait.list.first != NULL);
2089         }
2090         rcu_read_unlock();
2091
2092         return out - buf;
2093 }