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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_sort.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "buckets.h"
  13 #include "checksum.h"
  14 #include "debug.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "io.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_seq_blacklist.h"
  20 #include "super-io.h"
  21
  22 #include <linux/sched/mm.h>
  23 #include <trace/events/bcachefs.h>
  24
  25 static void verify_no_dups(struct btree *b,
  26                            struct bkey_packed *start,
  27                            struct bkey_packed *end,
  28                            bool extents)
  29 {
  30 #ifdef CONFIG_BCACHEFS_DEBUG
  31         struct bkey_packed *k, *p;
  32
  33         if (start == end)
  34                 return;
  35
  36         for (p = start, k = bkey_next_skip_noops(start, end);
  37              k != end;
  38              p = k, k = bkey_next_skip_noops(k, end)) {
  39                 struct bkey l = bkey_unpack_key(b, p);
  40                 struct bkey r = bkey_unpack_key(b, k);
  41
  42                 BUG_ON(extents
  43                        ? bkey_cmp(l.p, bkey_start_pos(&r)) > 0
  44                        : bkey_cmp(l.p, bkey_start_pos(&r)) >= 0);
  45                 //BUG_ON(bkey_cmp_packed(&b->format, p, k) >= 0);
  46         }
  47 #endif
  48 }
  49
  50 static void set_needs_whiteout(struct bset *i, int v)
  51 {
  52         struct bkey_packed *k;
  53
  54         for (k = i->start;
  55              k != vstruct_last(i);
  56              k = bkey_next_skip_noops(k, vstruct_last(i)))
  57                 k->needs_whiteout = v;
  58 }
  59
  60 static void btree_bounce_free(struct bch_fs *c, unsigned order,
  61                               bool used_mempool, void *p)
  62 {
  63         if (used_mempool)
  64                 mempool_free(p, &c->btree_bounce_pool);
  65         else
  66                 vpfree(p, PAGE_SIZE << order);
  67 }
  68
  69 static void *btree_bounce_alloc(struct bch_fs *c, unsigned order,
  70                                 bool *used_mempool)
  71 {
  72         unsigned flags = memalloc_nofs_save();
  73         void *p;
  74
  75         BUG_ON(order > btree_page_order(c));
  76
  77         *used_mempool = false;
  78         p = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT, order);
  79         if (!p) {
  80                 *used_mempool = true;
  81                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
  82         }
  83         memalloc_nofs_restore(flags);
  84         return p;
  85 }
  86
  87 static void sort_bkey_ptrs(const struct btree *bt,
  88                            struct bkey_packed **ptrs, unsigned nr)
  89 {
  90         unsigned n = nr, a = nr / 2, b, c, d;
  91
  92         if (!a)
  93                 return;
  94
  95         /* Heap sort: see lib/sort.c: */
  96         while (1) {
  97                 if (a)
  98                         a--;
  99                 else if (--n)
 100                         swap(ptrs[0], ptrs[n]);
 101                 else
 102                         break;
 103
 104                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 105                         b = bkey_cmp_packed(bt,
 106                                             ptrs[c],
 107                                             ptrs[d]) >= 0 ? c : d;
 108                 if (d == n)
 109                         b = c;
 110
 111                 while (b != a &&
 112                        bkey_cmp_packed(bt,
 113                                        ptrs[a],
 114                                        ptrs[b]) >= 0)
 115                         b = (b - 1) / 2;
 116                 c = b;
 117                 while (b != a) {
 118                         b = (b - 1) / 2;
 119                         swap(ptrs[b], ptrs[c]);
 120                 }
 121         }
 122 }
 123
 124 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 125 {
 126         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 127         bool used_mempool = false;
 128         unsigned order;
 129
 130         if (!b->whiteout_u64s)
 131                 return;
 132
 133         order = get_order(b->whiteout_u64s * sizeof(u64));
 134
 135         new_whiteouts = btree_bounce_alloc(c, order, &used_mempool);
 136
 137         ptrs = ptrs_end = ((void *) new_whiteouts + (PAGE_SIZE << order));
 138
 139         for (k = unwritten_whiteouts_start(c, b);
 140              k != unwritten_whiteouts_end(c, b);
 141              k = bkey_next(k))
 142                 *--ptrs = k;
 143
 144         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 145
 146         k = new_whiteouts;
 147
 148         while (ptrs != ptrs_end) {
 149                 bkey_copy(k, *ptrs);
 150                 k = bkey_next(k);
 151                 ptrs++;
 152         }
 153
 154         verify_no_dups(b, new_whiteouts,
 155                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s),
 156                        btree_node_old_extent_overwrite(b));
 157
 158         memcpy_u64s(unwritten_whiteouts_start(c, b),
 159                     new_whiteouts, b->whiteout_u64s);
 160
 161         btree_bounce_free(c, order, used_mempool, new_whiteouts);
 162 }
 163
 164 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 165                                 bool compacting, enum compact_mode mode)
 166 {
 167         if (!bset_dead_u64s(b, t))
 168                 return false;
 169
 170         switch (mode) {
 171         case COMPACT_LAZY:
 172                 return should_compact_bset_lazy(b, t) ||
 173                         (compacting && !bset_written(b, bset(b, t)));
 174         case COMPACT_ALL:
 175                 return true;
 176         default:
 177                 BUG();
 178         }
 179 }
 180
 181 static bool bch2_compact_extent_whiteouts(struct bch_fs *c,
 182                                           struct btree *b,
 183                                           enum compact_mode mode)
 184 {
 185         const struct bkey_format *f = &b->format;
 186         struct bset_tree *t;
 187         struct bkey_packed *whiteouts = NULL;
 188         struct bkey_packed *u_start, *u_pos;
 189         struct sort_iter sort_iter;
 190         unsigned order, whiteout_u64s = 0, u64s;
 191         bool used_mempool, compacting = false;
 192
 193         BUG_ON(!btree_node_is_extents(b));
 194
 195         for_each_bset(b, t)
 196                 if (should_compact_bset(b, t, whiteout_u64s != 0, mode))
 197                         whiteout_u64s += bset_dead_u64s(b, t);
 198
 199         if (!whiteout_u64s)
 200                 return false;
 201
 202         bch2_sort_whiteouts(c, b);
 203
 204         sort_iter_init(&sort_iter, b);
 205
 206         whiteout_u64s += b->whiteout_u64s;
 207         order = get_order(whiteout_u64s * sizeof(u64));
 208
 209         whiteouts = btree_bounce_alloc(c, order, &used_mempool);
 210         u_start = u_pos = whiteouts;
 211
 212         memcpy_u64s(u_pos, unwritten_whiteouts_start(c, b),
 213                     b->whiteout_u64s);
 214         u_pos = (void *) u_pos + b->whiteout_u64s * sizeof(u64);
 215
 216         sort_iter_add(&sort_iter, u_start, u_pos);
 217
 218         for_each_bset(b, t) {
 219                 struct bset *i = bset(b, t);
 220                 struct bkey_packed *k, *n, *out, *start, *end;
 221                 struct btree_node_entry *src = NULL, *dst = NULL;
 222
 223                 if (t != b->set && !bset_written(b, i)) {
 224                         src = container_of(i, struct btree_node_entry, keys);
 225                         dst = max(write_block(b),
 226                                   (void *) btree_bkey_last(b, t - 1));
 227                 }
 228
 229                 if (src != dst)
 230                         compacting = true;
 231
 232                 if (!should_compact_bset(b, t, compacting, mode)) {
 233                         if (src != dst) {
 234                                 memmove(dst, src, sizeof(*src) +
 235                                         le16_to_cpu(src->keys.u64s) *
 236                                         sizeof(u64));
 237                                 i = &dst->keys;
 238                                 set_btree_bset(b, t, i);
 239                         }
 240                         continue;
 241                 }
 242
 243                 compacting = true;
 244                 u_start = u_pos;
 245                 start = i->start;
 246                 end = vstruct_last(i);
 247
 248                 if (src != dst) {
 249                         memmove(dst, src, sizeof(*src));
 250                         i = &dst->keys;
 251                         set_btree_bset(b, t, i);
 252                 }
 253
 254                 out = i->start;
 255
 256                 for (k = start; k != end; k = n) {
 257                         n = bkey_next_skip_noops(k, end);
 258
 259                         if (bkey_deleted(k))
 260                                 continue;
 261
 262                         BUG_ON(bkey_whiteout(k) &&
 263                                k->needs_whiteout &&
 264                                bkey_written(b, k));
 265
 266                         if (bkey_whiteout(k) && !k->needs_whiteout)
 267                                 continue;
 268
 269                         if (bkey_whiteout(k)) {
 270                                 memcpy_u64s(u_pos, k, bkeyp_key_u64s(f, k));
 271                                 set_bkeyp_val_u64s(f, u_pos, 0);
 272                                 u_pos = bkey_next(u_pos);
 273                         } else {
 274                                 bkey_copy(out, k);
 275                                 out = bkey_next(out);
 276                         }
 277                 }
 278
 279                 sort_iter_add(&sort_iter, u_start, u_pos);
 280
 281                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 282                 set_btree_bset_end(b, t);
 283                 bch2_bset_set_no_aux_tree(b, t);
 284         }
 285
 286         b->whiteout_u64s = (u64 *) u_pos - (u64 *) whiteouts;
 287
 288         BUG_ON((void *) unwritten_whiteouts_start(c, b) <
 289                (void *) btree_bkey_last(b, bset_tree_last(b)));
 290
 291         u64s = bch2_sort_extent_whiteouts(unwritten_whiteouts_start(c, b),
 292                                           &sort_iter);
 293
 294         BUG_ON(u64s > b->whiteout_u64s);
 295         BUG_ON(u_pos != whiteouts && !u64s);
 296
 297         if (u64s != b->whiteout_u64s) {
 298                 void *src = unwritten_whiteouts_start(c, b);
 299
 300                 b->whiteout_u64s = u64s;
 301                 memmove_u64s_up(unwritten_whiteouts_start(c, b), src, u64s);
 302         }
 303
 304         verify_no_dups(b,
 305                        unwritten_whiteouts_start(c, b),
 306                        unwritten_whiteouts_end(c, b),
 307                        true);
 308
 309         btree_bounce_free(c, order, used_mempool, whiteouts);
 310
 311         bch2_btree_build_aux_trees(b);
 312
 313         bch_btree_keys_u64s_remaining(c, b);
 314         bch2_verify_btree_nr_keys(b);
 315
 316         return true;
 317 }
 318
 319 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 320 {
 321         struct bset_tree *t;
 322         bool ret = false;
 323
 324         for_each_bset(b, t) {
 325                 struct bset *i = bset(b, t);
 326                 struct bkey_packed *k, *n, *out, *start, *end;
 327                 struct btree_node_entry *src = NULL, *dst = NULL;
 328
 329                 if (t != b->set && !bset_written(b, i)) {
 330                         src = container_of(i, struct btree_node_entry, keys);
 331                         dst = max(write_block(b),
 332                                   (void *) btree_bkey_last(b, t - 1));
 333                 }
 334
 335                 if (src != dst)
 336                         ret = true;
 337
 338                 if (!should_compact_bset(b, t, ret, mode)) {
 339                         if (src != dst) {
 340                                 memmove(dst, src, sizeof(*src) +
 341                                         le16_to_cpu(src->keys.u64s) *
 342                                         sizeof(u64));
 343                                 i = &dst->keys;
 344                                 set_btree_bset(b, t, i);
 345                         }
 346                         continue;
 347                 }
 348
 349                 start   = btree_bkey_first(b, t);
 350                 end     = btree_bkey_last(b, t);
 351
 352                 if (src != dst) {
 353                         memmove(dst, src, sizeof(*src));
 354                         i = &dst->keys;
 355                         set_btree_bset(b, t, i);
 356                 }
 357
 358                 out = i->start;
 359
 360                 for (k = start; k != end; k = n) {
 361                         n = bkey_next_skip_noops(k, end);
 362
 363                         if (!bkey_whiteout(k)) {
 364                                 bkey_copy(out, k);
 365                                 out = bkey_next(out);
 366                         } else {
 367                                 BUG_ON(k->needs_whiteout);
 368                         }
 369                 }
 370
 371                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 372                 set_btree_bset_end(b, t);
 373                 bch2_bset_set_no_aux_tree(b, t);
 374                 ret = true;
 375         }
 376
 377         bch2_verify_btree_nr_keys(b);
 378
 379         bch2_btree_build_aux_trees(b);
 380
 381         return ret;
 382 }
 383
 384 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 385                             enum compact_mode mode)
 386 {
 387         return !btree_node_old_extent_overwrite(b)
 388                 ? bch2_drop_whiteouts(b, mode)
 389                 : bch2_compact_extent_whiteouts(c, b, mode);
 390 }
 391
 392 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 393                             struct btree_iter *iter,
 394                             unsigned start_idx,
 395                             unsigned end_idx,
 396                             bool filter_whiteouts)
 397 {
 398         struct btree_node *out;
 399         struct sort_iter sort_iter;
 400         struct bset_tree *t;
 401         struct bset *start_bset = bset(b, &b->set[start_idx]);
 402         bool used_mempool = false;
 403         u64 start_time, seq = 0;
 404         unsigned i, u64s = 0, order, shift = end_idx - start_idx - 1;
 405         bool sorting_entire_node = start_idx == 0 &&
 406                 end_idx == b->nsets;
 407
 408         sort_iter_init(&sort_iter, b);
 409
 410         for (t = b->set + start_idx;
 411              t < b->set + end_idx;
 412              t++) {
 413                 u64s += le16_to_cpu(bset(b, t)->u64s);
 414                 sort_iter_add(&sort_iter,
 415                               btree_bkey_first(b, t),
 416                               btree_bkey_last(b, t));
 417         }
 418
 419         order = sorting_entire_node
 420                 ? btree_page_order(c)
 421                 : get_order(__vstruct_bytes(struct btree_node, u64s));
 422
 423         out = btree_bounce_alloc(c, order, &used_mempool);
 424
 425         start_time = local_clock();
 426
 427         if (btree_node_old_extent_overwrite(b))
 428                 filter_whiteouts = bset_written(b, start_bset);
 429
 430         u64s = (btree_node_old_extent_overwrite(b)
 431                 ? bch2_sort_extents
 432                 : bch2_sort_keys)(out->keys.start,
 433                                   &sort_iter,
 434                                   filter_whiteouts);
 435
 436         out->keys.u64s = cpu_to_le16(u64s);
 437
 438         BUG_ON(vstruct_end(&out->keys) > (void *) out + (PAGE_SIZE << order));
 439
 440         if (sorting_entire_node)
 441                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 442                                        start_time);
 443
 444         /* Make sure we preserve bset journal_seq: */
 445         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 446                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 447         start_bset->journal_seq = cpu_to_le64(seq);
 448
 449         if (sorting_entire_node) {
 450                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 451
 452                 BUG_ON(order != btree_page_order(c));
 453
 454                 /*
 455                  * Our temporary buffer is the same size as the btree node's
 456                  * buffer, we can just swap buffers instead of doing a big
 457                  * memcpy()
 458                  */
 459                 *out = *b->data;
 460                 out->keys.u64s = cpu_to_le16(u64s);
 461                 swap(out, b->data);
 462                 set_btree_bset(b, b->set, &b->data->keys);
 463         } else {
 464                 start_bset->u64s = out->keys.u64s;
 465                 memcpy_u64s(start_bset->start,
 466                             out->keys.start,
 467                             le16_to_cpu(out->keys.u64s));
 468         }
 469
 470         for (i = start_idx + 1; i < end_idx; i++)
 471                 b->nr.bset_u64s[start_idx] +=
 472                         b->nr.bset_u64s[i];
 473
 474         b->nsets -= shift;
 475
 476         for (i = start_idx + 1; i < b->nsets; i++) {
 477                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 478                 b->set[i]               = b->set[i + shift];
 479         }
 480
 481         for (i = b->nsets; i < MAX_BSETS; i++)
 482                 b->nr.bset_u64s[i] = 0;
 483
 484         set_btree_bset_end(b, &b->set[start_idx]);
 485         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 486
 487         btree_bounce_free(c, order, used_mempool, out);
 488
 489         bch2_verify_btree_nr_keys(b);
 490 }
 491
 492 void bch2_btree_sort_into(struct bch_fs *c,
 493                          struct btree *dst,
 494                          struct btree *src)
 495 {
 496         struct btree_nr_keys nr;
 497         struct btree_node_iter src_iter;
 498         u64 start_time = local_clock();
 499
 500         BUG_ON(dst->nsets != 1);
 501
 502         bch2_bset_set_no_aux_tree(dst, dst->set);
 503
 504         bch2_btree_node_iter_init_from_start(&src_iter, src);
 505
 506         if (btree_node_is_extents(src))
 507                 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
 508                                 src, &src_iter,
 509                                 &dst->format,
 510                                 true);
 511         else
 512                 nr = bch2_sort_repack(btree_bset_first(dst),
 513                                 src, &src_iter,
 514                                 &dst->format,
 515                                 true);
 516
 517         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 518                                start_time);
 519
 520         set_btree_bset_end(dst, dst->set);
 521
 522         dst->nr.live_u64s       += nr.live_u64s;
 523         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 524         dst->nr.packed_keys     += nr.packed_keys;
 525         dst->nr.unpacked_keys   += nr.unpacked_keys;
 526
 527         bch2_verify_btree_nr_keys(dst);
 528 }
 529
 530 #define SORT_CRIT       (4096 / sizeof(u64))
 531
 532 /*
 533  * We're about to add another bset to the btree node, so if there's currently
 534  * too many bsets - sort some of them together:
 535  */
 536 static bool btree_node_compact(struct bch_fs *c, struct btree *b,
 537                                struct btree_iter *iter)
 538 {
 539         unsigned unwritten_idx;
 540         bool ret = false;
 541
 542         for (unwritten_idx = 0;
 543              unwritten_idx < b->nsets;
 544              unwritten_idx++)
 545                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 546                         break;
 547
 548         if (b->nsets - unwritten_idx > 1) {
 549                 btree_node_sort(c, b, iter, unwritten_idx,
 550                                 b->nsets, false);
 551                 ret = true;
 552         }
 553
 554         if (unwritten_idx > 1) {
 555                 btree_node_sort(c, b, iter, 0, unwritten_idx, false);
 556                 ret = true;
 557         }
 558
 559         return ret;
 560 }
 561
 562 void bch2_btree_build_aux_trees(struct btree *b)
 563 {
 564         struct bset_tree *t;
 565
 566         for_each_bset(b, t)
 567                 bch2_bset_build_aux_tree(b, t,
 568                                 !bset_written(b, bset(b, t)) &&
 569                                 t == bset_tree_last(b));
 570 }
 571
 572 /*
 573  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 574  * inserted into
 575  *
 576  * Safe to call if there already is an unwritten bset - will only add a new bset
 577  * if @b doesn't already have one.
 578  *
 579  * Returns true if we sorted (i.e. invalidated iterators
 580  */
 581 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
 582                           struct btree_iter *iter)
 583 {
 584         struct btree_node_entry *bne;
 585         bool did_sort;
 586
 587         EBUG_ON(!(b->c.lock.state.seq & 1));
 588         EBUG_ON(iter && iter->l[b->c.level].b != b);
 589
 590         did_sort = btree_node_compact(c, b, iter);
 591
 592         bne = want_new_bset(c, b);
 593         if (bne)
 594                 bch2_bset_init_next(c, b, bne);
 595
 596         bch2_btree_build_aux_trees(b);
 597
 598         if (iter && did_sort)
 599                 bch2_btree_iter_reinit_node(iter, b);
 600 }
 601
 602 static struct nonce btree_nonce(struct bset *i, unsigned offset)
 603 {
 604         return (struct nonce) {{
 605                 [0] = cpu_to_le32(offset),
 606                 [1] = ((__le32 *) &i->seq)[0],
 607                 [2] = ((__le32 *) &i->seq)[1],
 608                 [3] = ((__le32 *) &i->journal_seq)[0]^BCH_NONCE_BTREE,
 609         }};
 610 }
 611
 612 static void bset_encrypt(struct bch_fs *c, struct bset *i, unsigned offset)
 613 {
 614         struct nonce nonce = btree_nonce(i, offset);
 615
 616         if (!offset) {
 617                 struct btree_node *bn = container_of(i, struct btree_node, keys);
 618                 unsigned bytes = (void *) &bn->keys - (void *) &bn->flags;
 619
 620                 bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, &bn->flags,
 621                              bytes);
 622
 623                 nonce = nonce_add(nonce, round_up(bytes, CHACHA_BLOCK_SIZE));
 624         }
 625
 626         bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, i->_data,
 627                      vstruct_end(i) - (void *) i->_data);
 628 }
 629
 630 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 631                           struct btree *b, struct bset *i,
 632                           unsigned offset, int write)
 633 {
 634         pr_buf(out, "error validating btree node %sat btree %u level %u/%u\n"
 635                "pos ",
 636                write ? "before write " : "",
 637                b->c.btree_id, b->c.level,
 638                c->btree_roots[b->c.btree_id].level);
 639         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 640
 641         pr_buf(out, " node offset %u", b->written);
 642         if (i)
 643                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 644 }
 645
 646 enum btree_err_type {
 647         BTREE_ERR_FIXABLE,
 648         BTREE_ERR_WANT_RETRY,
 649         BTREE_ERR_MUST_RETRY,
 650         BTREE_ERR_FATAL,
 651 };
 652
 653 enum btree_validate_ret {
 654         BTREE_RETRY_READ = 64,
 655 };
 656
 657 #define btree_err(type, c, b, i, msg, ...)                              \
 658 ({                                                                      \
 659         __label__ out;                                                  \
 660         char _buf[300];                                                 \
 661         struct printbuf out = PBUF(_buf);                               \
 662                                                                         \
 663         btree_err_msg(&out, c, b, i, b->written, write);                \
 664         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 665                                                                         \
 666         if (type == BTREE_ERR_FIXABLE &&                                \
 667             write == READ &&                                            \
 668             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 669                 mustfix_fsck_err(c, "%s", _buf);                        \
 670                 goto out;                                               \
 671         }                                                               \
 672                                                                         \
 673         switch (write) {                                                \
 674         case READ:                                                      \
 675                 bch_err(c, "%s", _buf);                                 \
 676                                                                         \
 677                 switch (type) {                                         \
 678                 case BTREE_ERR_FIXABLE:                                 \
 679                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 680                         goto fsck_err;                                  \
 681                 case BTREE_ERR_WANT_RETRY:                              \
 682                         if (have_retry) {                               \
 683                                 ret = BTREE_RETRY_READ;                 \
 684                                 goto fsck_err;                          \
 685                         }                                               \
 686                         break;                                          \
 687                 case BTREE_ERR_MUST_RETRY:                              \
 688                         ret = BTREE_RETRY_READ;                         \
 689                         goto fsck_err;                                  \
 690                 case BTREE_ERR_FATAL:                                   \
 691                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 692                         goto fsck_err;                                  \
 693                 }                                                       \
 694                 break;                                                  \
 695         case WRITE:                                                     \
 696                 bch_err(c, "corrupt metadata before write: %s", _buf);  \
 697                                                                         \
 698                 if (bch2_fs_inconsistent(c)) {                          \
 699                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 700                         goto fsck_err;                                  \
 701                 }                                                       \
 702                 break;                                                  \
 703         }                                                               \
 704 out:                                                                    \
 705         true;                                                           \
 706 })
 707
 708 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 709
 710 static int validate_bset(struct bch_fs *c, struct btree *b,
 711                          struct bset *i, unsigned sectors,
 712                          int write, bool have_retry)
 713 {
 714         unsigned version = le16_to_cpu(i->version);
 715         const char *err;
 716         int ret = 0;
 717
 718         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 719                       version < bcachefs_metadata_version_min) ||
 720                      version >= bcachefs_metadata_version_max,
 721                      BTREE_ERR_FATAL, c, b, i,
 722                      "unsupported bset version");
 723
 724         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 725                          BTREE_ERR_FIXABLE, c, b, i,
 726                          "bset past end of btree node")) {
 727                 i->u64s = 0;
 728                 return 0;
 729         }
 730
 731         btree_err_on(b->written && !i->u64s,
 732                      BTREE_ERR_FIXABLE, c, b, i,
 733                      "empty bset");
 734
 735         if (!b->written) {
 736                 struct btree_node *bn =
 737                         container_of(i, struct btree_node, keys);
 738                 /* These indicate that we read the wrong btree node: */
 739
 740                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 741                         struct bch_btree_ptr_v2 *bp =
 742                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 743
 744                         /* XXX endianness */
 745                         btree_err_on(bp->seq != bn->keys.seq,
 746                                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 747                                      "incorrect sequence number (wrong btree node)");
 748                 }
 749
 750                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 751                              BTREE_ERR_MUST_RETRY, c, b, i,
 752                              "incorrect btree id");
 753
 754                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 755                              BTREE_ERR_MUST_RETRY, c, b, i,
 756                              "incorrect level");
 757
 758                 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
 759                         u64 *p = (u64 *) &bn->ptr;
 760
 761                         *p = swab64(*p);
 762                 }
 763
 764                 if (!write)
 765                         compat_btree_node(b->c.level, b->c.btree_id, version,
 766                                           BSET_BIG_ENDIAN(i), write, bn);
 767
 768                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 769                         struct bch_btree_ptr_v2 *bp =
 770                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 771
 772                         btree_err_on(bkey_cmp(b->data->min_key, bp->min_key),
 773                                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 774                                      "incorrect min_key: got %llu:%llu should be %llu:%llu",
 775                                      b->data->min_key.inode,
 776                                      b->data->min_key.offset,
 777                                      bp->min_key.inode,
 778                                      bp->min_key.offset);
 779                 }
 780
 781                 btree_err_on(bkey_cmp(bn->max_key, b->key.k.p),
 782                              BTREE_ERR_MUST_RETRY, c, b, i,
 783                              "incorrect max key");
 784
 785                 if (write)
 786                         compat_btree_node(b->c.level, b->c.btree_id, version,
 787                                           BSET_BIG_ENDIAN(i), write, bn);
 788
 789                 /* XXX: ideally we would be validating min_key too */
 790 #if 0
 791                 /*
 792                  * not correct anymore, due to btree node write error
 793                  * handling
 794                  *
 795                  * need to add bn->seq to btree keys and verify
 796                  * against that
 797                  */
 798                 btree_err_on(!extent_contains_ptr(bkey_i_to_s_c_extent(&b->key),
 799                                                   bn->ptr),
 800                              BTREE_ERR_FATAL, c, b, i,
 801                              "incorrect backpointer");
 802 #endif
 803                 err = bch2_bkey_format_validate(&bn->format);
 804                 btree_err_on(err,
 805                              BTREE_ERR_FATAL, c, b, i,
 806                              "invalid bkey format: %s", err);
 807
 808                 compat_bformat(b->c.level, b->c.btree_id, version,
 809                                BSET_BIG_ENDIAN(i), write,
 810                                &bn->format);
 811         }
 812 fsck_err:
 813         return ret;
 814 }
 815
 816 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 817                          struct bset *i, unsigned *whiteout_u64s,
 818                          int write, bool have_retry)
 819 {
 820         unsigned version = le16_to_cpu(i->version);
 821         struct bkey_packed *k, *prev = NULL;
 822         bool seen_non_whiteout = false;
 823         int ret = 0;
 824
 825         if (!BSET_SEPARATE_WHITEOUTS(i)) {
 826                 seen_non_whiteout = true;
 827                 *whiteout_u64s = 0;
 828         }
 829
 830         for (k = i->start;
 831              k != vstruct_last(i);) {
 832                 struct bkey_s u;
 833                 struct bkey tmp;
 834                 const char *invalid;
 835
 836                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 837                                  BTREE_ERR_FIXABLE, c, b, i,
 838                                  "key extends past end of bset")) {
 839                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 840                         break;
 841                 }
 842
 843                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 844                                  BTREE_ERR_FIXABLE, c, b, i,
 845                                  "invalid bkey format %u", k->format)) {
 846                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 847                         memmove_u64s_down(k, bkey_next(k),
 848                                           (u64 *) vstruct_end(i) - (u64 *) k);
 849                         continue;
 850                 }
 851
 852                 /* XXX: validate k->u64s */
 853                 if (!write)
 854                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 855                                     BSET_BIG_ENDIAN(i), write,
 856                                     &b->format, k);
 857
 858                 u = __bkey_disassemble(b, k, &tmp);
 859
 860                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 861                         bch2_bkey_in_btree_node(b, u.s_c) ?:
 862                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 863                 if (invalid) {
 864                         char buf[160];
 865
 866                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 867                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
 868                                   "invalid bkey:\n%s\n%s", invalid, buf);
 869
 870                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 871                         memmove_u64s_down(k, bkey_next(k),
 872                                           (u64 *) vstruct_end(i) - (u64 *) k);
 873                         continue;
 874                 }
 875
 876                 if (write)
 877                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 878                                     BSET_BIG_ENDIAN(i), write,
 879                                     &b->format, k);
 880
 881                 /*
 882                  * with the separate whiteouts thing (used for extents), the
 883                  * second set of keys actually can have whiteouts too, so we
 884                  * can't solely go off bkey_whiteout()...
 885                  */
 886
 887                 if (!seen_non_whiteout &&
 888                     (!bkey_whiteout(k) ||
 889                      (prev && bkey_iter_cmp(b, prev, k) > 0))) {
 890                         *whiteout_u64s = k->_data - i->_data;
 891                         seen_non_whiteout = true;
 892                 } else if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 893                         char buf1[80];
 894                         char buf2[80];
 895                         struct bkey up = bkey_unpack_key(b, prev);
 896
 897                         bch2_bkey_to_text(&PBUF(buf1), &up);
 898                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 899
 900                         bch2_dump_bset(b, i, 0);
 901                         btree_err(BTREE_ERR_FATAL, c, b, i,
 902                                   "keys out of order: %s > %s",
 903                                   buf1, buf2);
 904                         /* XXX: repair this */
 905                 }
 906
 907                 prev = k;
 908                 k = bkey_next_skip_noops(k, vstruct_last(i));
 909         }
 910 fsck_err:
 911         return ret;
 912 }
 913
 914 int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry)
 915 {
 916         struct btree_node_entry *bne;
 917         struct sort_iter *iter;
 918         struct btree_node *sorted;
 919         struct bkey_packed *k;
 920         struct bset *i;
 921         bool used_mempool, blacklisted;
 922         unsigned u64s;
 923         int ret, retry_read = 0, write = READ;
 924
 925         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 926         sort_iter_init(iter, b);
 927         iter->size = (btree_blocks(c) + 1) * 2;
 928
 929         if (bch2_meta_read_fault("btree"))
 930                 btree_err(BTREE_ERR_MUST_RETRY, c, b, NULL,
 931                           "dynamic fault");
 932
 933         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 934                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 935                      "bad magic");
 936
 937         btree_err_on(!b->data->keys.seq,
 938                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 939                      "bad btree header");
 940
 941         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 942                 struct bch_btree_ptr_v2 *bp =
 943                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 944
 945                 btree_err_on(b->data->keys.seq != bp->seq,
 946                              BTREE_ERR_MUST_RETRY, c, b, NULL,
 947                              "got wrong btree node (seq %llx want %llx)",
 948                              b->data->keys.seq, bp->seq);
 949         }
 950
 951         while (b->written < c->opts.btree_node_size) {
 952                 unsigned sectors, whiteout_u64s = 0;
 953                 struct nonce nonce;
 954                 struct bch_csum csum;
 955                 bool first = !b->written;
 956
 957                 if (!b->written) {
 958                         i = &b->data->keys;
 959
 960                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 961                                      BTREE_ERR_WANT_RETRY, c, b, i,
 962                                      "unknown checksum type");
 963
 964                         nonce = btree_nonce(i, b->written << 9);
 965                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 966
 967                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 968                                      BTREE_ERR_WANT_RETRY, c, b, i,
 969                                      "invalid checksum");
 970
 971                         bset_encrypt(c, i, b->written << 9);
 972
 973                         if (btree_node_is_extents(b) &&
 974                             !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data))
 975                                 set_btree_node_old_extent_overwrite(b);
 976
 977                         sectors = vstruct_sectors(b->data, c->block_bits);
 978                 } else {
 979                         bne = write_block(b);
 980                         i = &bne->keys;
 981
 982                         if (i->seq != b->data->keys.seq)
 983                                 break;
 984
 985                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 986                                      BTREE_ERR_WANT_RETRY, c, b, i,
 987                                      "unknown checksum type");
 988
 989                         nonce = btree_nonce(i, b->written << 9);
 990                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 991
 992                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 993                                      BTREE_ERR_WANT_RETRY, c, b, i,
 994                                      "invalid checksum");
 995
 996                         bset_encrypt(c, i, b->written << 9);
 997
 998                         sectors = vstruct_sectors(bne, c->block_bits);
 999                 }
1000
1001                 ret = validate_bset(c, b, i, sectors,
1002                                     READ, have_retry);
1003                 if (ret)
1004                         goto fsck_err;
1005
1006                 if (!b->written)
1007                         btree_node_set_format(b, b->data->format);
1008
1009                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
1010                                     READ, have_retry);
1011                 if (ret)
1012                         goto fsck_err;
1013
1014                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1015
1016                 b->written += sectors;
1017
1018                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1019                                         le64_to_cpu(i->journal_seq),
1020                                         true);
1021
1022                 btree_err_on(blacklisted && first,
1023                              BTREE_ERR_FIXABLE, c, b, i,
1024                              "first btree node bset has blacklisted journal seq");
1025                 if (blacklisted && !first)
1026                         continue;
1027
1028                 sort_iter_add(iter, i->start,
1029                               vstruct_idx(i, whiteout_u64s));
1030
1031                 sort_iter_add(iter,
1032                               vstruct_idx(i, whiteout_u64s),
1033                               vstruct_last(i));
1034         }
1035
1036         for (bne = write_block(b);
1037              bset_byte_offset(b, bne) < btree_bytes(c);
1038              bne = (void *) bne + block_bytes(c))
1039                 btree_err_on(bne->keys.seq == b->data->keys.seq,
1040                              BTREE_ERR_WANT_RETRY, c, b, NULL,
1041                              "found bset signature after last bset");
1042
1043         sorted = btree_bounce_alloc(c, btree_page_order(c), &used_mempool);
1044         sorted->keys.u64s = 0;
1045
1046         set_btree_bset(b, b->set, &b->data->keys);
1047
1048         b->nr = (btree_node_old_extent_overwrite(b)
1049                  ? bch2_extent_sort_fix_overlapping
1050                  : bch2_key_sort_fix_overlapping)(c, &sorted->keys, iter);
1051
1052         u64s = le16_to_cpu(sorted->keys.u64s);
1053         *sorted = *b->data;
1054         sorted->keys.u64s = cpu_to_le16(u64s);
1055         swap(sorted, b->data);
1056         set_btree_bset(b, b->set, &b->data->keys);
1057         b->nsets = 1;
1058
1059         BUG_ON(b->nr.live_u64s != u64s);
1060
1061         btree_bounce_free(c, btree_page_order(c), used_mempool, sorted);
1062
1063         i = &b->data->keys;
1064         for (k = i->start; k != vstruct_last(i);) {
1065                 struct bkey tmp;
1066                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1067                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
1068
1069                 if (invalid ||
1070                     (inject_invalid_keys(c) &&
1071                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1072                         char buf[160];
1073
1074                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
1075                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
1076                                   "invalid bkey %s: %s", buf, invalid);
1077
1078                         btree_keys_account_key_drop(&b->nr, 0, k);
1079
1080                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1081                         memmove_u64s_down(k, bkey_next(k),
1082                                           (u64 *) vstruct_end(i) - (u64 *) k);
1083                         set_btree_bset_end(b, b->set);
1084                         continue;
1085                 }
1086
1087                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1088                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1089
1090                         bp.v->mem_ptr = 0;
1091                 }
1092
1093                 k = bkey_next_skip_noops(k, vstruct_last(i));
1094         }
1095
1096         bch2_bset_build_aux_tree(b, b->set, false);
1097
1098         set_needs_whiteout(btree_bset_first(b), true);
1099
1100         btree_node_reset_sib_u64s(b);
1101 out:
1102         mempool_free(iter, &c->fill_iter);
1103         return retry_read;
1104 fsck_err:
1105         if (ret == BTREE_RETRY_READ) {
1106                 retry_read = 1;
1107         } else {
1108                 bch2_inconsistent_error(c);
1109                 set_btree_node_read_error(b);
1110         }
1111         goto out;
1112 }
1113
1114 static void btree_node_read_work(struct work_struct *work)
1115 {
1116         struct btree_read_bio *rb =
1117                 container_of(work, struct btree_read_bio, work);
1118         struct bch_fs *c        = rb->c;
1119         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1120         struct btree *b         = rb->bio.bi_private;
1121         struct bio *bio         = &rb->bio;
1122         struct bch_io_failures failed = { .nr = 0 };
1123         bool can_retry;
1124
1125         goto start;
1126         while (1) {
1127                 bch_info(c, "retrying read");
1128                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1129                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1130                 bio_reset(bio);
1131                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1132                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1133                 bio->bi_iter.bi_size    = btree_bytes(c);
1134
1135                 if (rb->have_ioref) {
1136                         bio_set_dev(bio, ca->disk_sb.bdev);
1137                         submit_bio_wait(bio);
1138                 } else {
1139                         bio->bi_status = BLK_STS_REMOVED;
1140                 }
1141 start:
1142                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read");
1143                 if (rb->have_ioref)
1144                         percpu_ref_put(&ca->io_ref);
1145                 rb->have_ioref = false;
1146
1147                 bch2_mark_io_failure(&failed, &rb->pick);
1148
1149                 can_retry = bch2_bkey_pick_read_device(c,
1150                                 bkey_i_to_s_c(&b->key),
1151                                 &failed, &rb->pick) > 0;
1152
1153                 if (!bio->bi_status &&
1154                     !bch2_btree_node_read_done(c, b, can_retry))
1155                         break;
1156
1157                 if (!can_retry) {
1158                         set_btree_node_read_error(b);
1159                         break;
1160                 }
1161         }
1162
1163         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1164                                rb->start_time);
1165         bio_put(&rb->bio);
1166         clear_btree_node_read_in_flight(b);
1167         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1168 }
1169
1170 static void btree_node_read_endio(struct bio *bio)
1171 {
1172         struct btree_read_bio *rb =
1173                 container_of(bio, struct btree_read_bio, bio);
1174         struct bch_fs *c        = rb->c;
1175
1176         if (rb->have_ioref) {
1177                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1178                 bch2_latency_acct(ca, rb->start_time, READ);
1179         }
1180
1181         queue_work(system_unbound_wq, &rb->work);
1182 }
1183
1184 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1185                           bool sync)
1186 {
1187         struct extent_ptr_decoded pick;
1188         struct btree_read_bio *rb;
1189         struct bch_dev *ca;
1190         struct bio *bio;
1191         int ret;
1192
1193         trace_btree_read(c, b);
1194
1195         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1196                                          NULL, &pick);
1197         if (bch2_fs_fatal_err_on(ret <= 0, c,
1198                         "btree node read error: no device to read from")) {
1199                 set_btree_node_read_error(b);
1200                 return;
1201         }
1202
1203         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1204
1205         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1206                                                    btree_bytes(c)),
1207                                &c->btree_bio);
1208         rb = container_of(bio, struct btree_read_bio, bio);
1209         rb->c                   = c;
1210         rb->start_time          = local_clock();
1211         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1212         rb->pick                = pick;
1213         INIT_WORK(&rb->work, btree_node_read_work);
1214         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1215         bio->bi_iter.bi_sector  = pick.ptr.offset;
1216         bio->bi_end_io          = btree_node_read_endio;
1217         bio->bi_private         = b;
1218         bch2_bio_map(bio, b->data, btree_bytes(c));
1219
1220         set_btree_node_read_in_flight(b);
1221
1222         if (rb->have_ioref) {
1223                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_BTREE],
1224                              bio_sectors(bio));
1225                 bio_set_dev(bio, ca->disk_sb.bdev);
1226
1227                 if (sync) {
1228                         submit_bio_wait(bio);
1229
1230                         bio->bi_private = b;
1231                         btree_node_read_work(&rb->work);
1232                 } else {
1233                         submit_bio(bio);
1234                 }
1235         } else {
1236                 bio->bi_status = BLK_STS_REMOVED;
1237
1238                 if (sync)
1239                         btree_node_read_work(&rb->work);
1240                 else
1241                         queue_work(system_unbound_wq, &rb->work);
1242
1243         }
1244 }
1245
1246 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1247                         const struct bkey_i *k, unsigned level)
1248 {
1249         struct closure cl;
1250         struct btree *b;
1251         int ret;
1252
1253         closure_init_stack(&cl);
1254
1255         do {
1256                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1257                 closure_sync(&cl);
1258         } while (ret);
1259
1260         b = bch2_btree_node_mem_alloc(c);
1261         bch2_btree_cache_cannibalize_unlock(c);
1262
1263         BUG_ON(IS_ERR(b));
1264
1265         bkey_copy(&b->key, k);
1266         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1267
1268         bch2_btree_node_read(c, b, true);
1269
1270         if (btree_node_read_error(b)) {
1271                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1272
1273                 mutex_lock(&c->btree_cache.lock);
1274                 list_move(&b->list, &c->btree_cache.freeable);
1275                 mutex_unlock(&c->btree_cache.lock);
1276
1277                 ret = -EIO;
1278                 goto err;
1279         }
1280
1281         bch2_btree_set_root_for_read(c, b);
1282 err:
1283         six_unlock_write(&b->c.lock);
1284         six_unlock_intent(&b->c.lock);
1285
1286         return ret;
1287 }
1288
1289 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1290                               struct btree_write *w)
1291 {
1292         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1293
1294         do {
1295                 old = new = v;
1296                 if (!(old & 1))
1297                         break;
1298
1299                 new &= ~1UL;
1300         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1301
1302         if (old & 1)
1303                 closure_put(&((struct btree_update *) new)->cl);
1304
1305         bch2_journal_pin_drop(&c->journal, &w->journal);
1306 }
1307
1308 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1309 {
1310         struct btree_write *w = btree_prev_write(b);
1311
1312         bch2_btree_complete_write(c, b, w);
1313         btree_node_io_unlock(b);
1314 }
1315
1316 static void bch2_btree_node_write_error(struct bch_fs *c,
1317                                         struct btree_write_bio *wbio)
1318 {
1319         struct btree *b         = wbio->wbio.bio.bi_private;
1320         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1321         struct bch_extent_ptr *ptr;
1322         struct btree_trans trans;
1323         struct btree_iter *iter;
1324         int ret;
1325
1326         bch2_trans_init(&trans, c, 0, 0);
1327
1328         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1329                                         BTREE_MAX_DEPTH, b->c.level, 0);
1330 retry:
1331         ret = bch2_btree_iter_traverse(iter);
1332         if (ret)
1333                 goto err;
1334
1335         /* has node been freed? */
1336         if (iter->l[b->c.level].b != b) {
1337                 /* node has been freed: */
1338                 BUG_ON(!btree_node_dying(b));
1339                 goto out;
1340         }
1341
1342         BUG_ON(!btree_node_hashed(b));
1343
1344         bkey_copy(&tmp.k, &b->key);
1345
1346         bch2_bkey_drop_ptrs(bkey_i_to_s(&tmp.k), ptr,
1347                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1348
1349         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&tmp.k)))
1350                 goto err;
1351
1352         ret = bch2_btree_node_update_key(c, iter, b, &tmp.k);
1353         if (ret == -EINTR)
1354                 goto retry;
1355         if (ret)
1356                 goto err;
1357 out:
1358         bch2_trans_exit(&trans);
1359         bio_put(&wbio->wbio.bio);
1360         btree_node_write_done(c, b);
1361         return;
1362 err:
1363         set_btree_node_noevict(b);
1364         bch2_fs_fatal_error(c, "fatal error writing btree node");
1365         goto out;
1366 }
1367
1368 void bch2_btree_write_error_work(struct work_struct *work)
1369 {
1370         struct bch_fs *c = container_of(work, struct bch_fs,
1371                                         btree_write_error_work);
1372         struct bio *bio;
1373
1374         while (1) {
1375                 spin_lock_irq(&c->btree_write_error_lock);
1376                 bio = bio_list_pop(&c->btree_write_error_list);
1377                 spin_unlock_irq(&c->btree_write_error_lock);
1378
1379                 if (!bio)
1380                         break;
1381
1382                 bch2_btree_node_write_error(c,
1383                         container_of(bio, struct btree_write_bio, wbio.bio));
1384         }
1385 }
1386
1387 static void btree_node_write_work(struct work_struct *work)
1388 {
1389         struct btree_write_bio *wbio =
1390                 container_of(work, struct btree_write_bio, work);
1391         struct bch_fs *c        = wbio->wbio.c;
1392         struct btree *b         = wbio->wbio.bio.bi_private;
1393
1394         btree_bounce_free(c,
1395                 wbio->wbio.order,
1396                 wbio->wbio.used_mempool,
1397                 wbio->data);
1398
1399         if (wbio->wbio.failed.nr) {
1400                 unsigned long flags;
1401
1402                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1403                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1404                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1405
1406                 queue_work(c->wq, &c->btree_write_error_work);
1407                 return;
1408         }
1409
1410         bio_put(&wbio->wbio.bio);
1411         btree_node_write_done(c, b);
1412 }
1413
1414 static void btree_node_write_endio(struct bio *bio)
1415 {
1416         struct bch_write_bio *wbio      = to_wbio(bio);
1417         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1418         struct bch_write_bio *orig      = parent ?: wbio;
1419         struct bch_fs *c                = wbio->c;
1420         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1421         unsigned long flags;
1422
1423         if (wbio->have_ioref)
1424                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1425
1426         if (bio->bi_status == BLK_STS_REMOVED ||
1427             bch2_dev_io_err_on(bio->bi_status, ca, "btree write") ||
1428             bch2_meta_write_fault("btree")) {
1429                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1430                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1431                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1432         }
1433
1434         if (wbio->have_ioref)
1435                 percpu_ref_put(&ca->io_ref);
1436
1437         if (parent) {
1438                 bio_put(bio);
1439                 bio_endio(&parent->bio);
1440         } else {
1441                 struct btree_write_bio *wb =
1442                         container_of(orig, struct btree_write_bio, wbio);
1443
1444                 INIT_WORK(&wb->work, btree_node_write_work);
1445                 queue_work(system_unbound_wq, &wb->work);
1446         }
1447 }
1448
1449 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1450                                    struct bset *i, unsigned sectors)
1451 {
1452         unsigned whiteout_u64s = 0;
1453         int ret;
1454
1455         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_BTREE))
1456                 return -1;
1457
1458         ret = validate_bset(c, b, i, sectors, WRITE, false) ?:
1459                 validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false);
1460         if (ret)
1461                 bch2_inconsistent_error(c);
1462
1463         return ret;
1464 }
1465
1466 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1467                             enum six_lock_type lock_type_held)
1468 {
1469         struct btree_write_bio *wbio;
1470         struct bset_tree *t;
1471         struct bset *i;
1472         struct btree_node *bn = NULL;
1473         struct btree_node_entry *bne = NULL;
1474         BKEY_PADDED(key) k;
1475         struct bch_extent_ptr *ptr;
1476         struct sort_iter sort_iter;
1477         struct nonce nonce;
1478         unsigned bytes_to_write, sectors_to_write, order, bytes, u64s;
1479         u64 seq = 0;
1480         bool used_mempool;
1481         unsigned long old, new;
1482         bool validate_before_checksum = false;
1483         void *data;
1484
1485         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1486                 return;
1487
1488         /*
1489          * We may only have a read lock on the btree node - the dirty bit is our
1490          * "lock" against racing with other threads that may be trying to start
1491          * a write, we do a write iff we clear the dirty bit. Since setting the
1492          * dirty bit requires a write lock, we can't race with other threads
1493          * redirtying it:
1494          */
1495         do {
1496                 old = new = READ_ONCE(b->flags);
1497
1498                 if (!(old & (1 << BTREE_NODE_dirty)))
1499                         return;
1500
1501                 if (!btree_node_may_write(b))
1502                         return;
1503
1504                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1505                         btree_node_wait_on_io(b);
1506                         continue;
1507                 }
1508
1509                 new &= ~(1 << BTREE_NODE_dirty);
1510                 new &= ~(1 << BTREE_NODE_need_write);
1511                 new |=  (1 << BTREE_NODE_write_in_flight);
1512                 new |=  (1 << BTREE_NODE_just_written);
1513                 new ^=  (1 << BTREE_NODE_write_idx);
1514         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1515
1516         BUG_ON(btree_node_fake(b));
1517         BUG_ON((b->will_make_reachable != 0) != !b->written);
1518
1519         BUG_ON(b->written >= c->opts.btree_node_size);
1520         BUG_ON(b->written & (c->opts.block_size - 1));
1521         BUG_ON(bset_written(b, btree_bset_last(b)));
1522         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1523         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1524
1525         bch2_sort_whiteouts(c, b);
1526
1527         sort_iter_init(&sort_iter, b);
1528
1529         bytes = !b->written
1530                 ? sizeof(struct btree_node)
1531                 : sizeof(struct btree_node_entry);
1532
1533         bytes += b->whiteout_u64s * sizeof(u64);
1534
1535         for_each_bset(b, t) {
1536                 i = bset(b, t);
1537
1538                 if (bset_written(b, i))
1539                         continue;
1540
1541                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1542                 sort_iter_add(&sort_iter,
1543                               btree_bkey_first(b, t),
1544                               btree_bkey_last(b, t));
1545                 seq = max(seq, le64_to_cpu(i->journal_seq));
1546         }
1547
1548         order = get_order(bytes);
1549         data = btree_bounce_alloc(c, order, &used_mempool);
1550
1551         if (!b->written) {
1552                 bn = data;
1553                 *bn = *b->data;
1554                 i = &bn->keys;
1555         } else {
1556                 bne = data;
1557                 bne->keys = b->data->keys;
1558                 i = &bne->keys;
1559         }
1560
1561         i->journal_seq  = cpu_to_le64(seq);
1562         i->u64s         = 0;
1563
1564         if (!btree_node_old_extent_overwrite(b)) {
1565                 sort_iter_add(&sort_iter,
1566                               unwritten_whiteouts_start(c, b),
1567                               unwritten_whiteouts_end(c, b));
1568                 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1569         } else {
1570                 memcpy_u64s(i->start,
1571                             unwritten_whiteouts_start(c, b),
1572                             b->whiteout_u64s);
1573                 i->u64s = cpu_to_le16(b->whiteout_u64s);
1574                 SET_BSET_SEPARATE_WHITEOUTS(i, true);
1575         }
1576
1577         b->whiteout_u64s = 0;
1578
1579         u64s = btree_node_old_extent_overwrite(b)
1580                 ? bch2_sort_extents(vstruct_last(i), &sort_iter, false)
1581                 : bch2_sort_keys(i->start, &sort_iter, false);
1582         le16_add_cpu(&i->u64s, u64s);
1583
1584         set_needs_whiteout(i, false);
1585
1586         /* do we have data to write? */
1587         if (b->written && !i->u64s)
1588                 goto nowrite;
1589
1590         bytes_to_write = vstruct_end(i) - data;
1591         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1592
1593         memset(data + bytes_to_write, 0,
1594                (sectors_to_write << 9) - bytes_to_write);
1595
1596         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1597         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1598         BUG_ON(i->seq != b->data->keys.seq);
1599
1600         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1601                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1602                 : cpu_to_le16(c->sb.version);
1603         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1604
1605         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1606                 validate_before_checksum = true;
1607
1608         /* validate_bset will be modifying: */
1609         if (le16_to_cpu(i->version) < bcachefs_metadata_version_max)
1610                 validate_before_checksum = true;
1611
1612         /* if we're going to be encrypting, check metadata validity first: */
1613         if (validate_before_checksum &&
1614             validate_bset_for_write(c, b, i, sectors_to_write))
1615                 goto err;
1616
1617         bset_encrypt(c, i, b->written << 9);
1618
1619         nonce = btree_nonce(i, b->written << 9);
1620
1621         if (bn)
1622                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1623         else
1624                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1625
1626         /* if we're not encrypting, check metadata after checksumming: */
1627         if (!validate_before_checksum &&
1628             validate_bset_for_write(c, b, i, sectors_to_write))
1629                 goto err;
1630
1631         /*
1632          * We handle btree write errors by immediately halting the journal -
1633          * after we've done that, we can't issue any subsequent btree writes
1634          * because they might have pointers to new nodes that failed to write.
1635          *
1636          * Furthermore, there's no point in doing any more btree writes because
1637          * with the journal stopped, we're never going to update the journal to
1638          * reflect that those writes were done and the data flushed from the
1639          * journal:
1640          *
1641          * Also on journal error, the pending write may have updates that were
1642          * never journalled (interior nodes, see btree_update_nodes_written()) -
1643          * it's critical that we don't do the write in that case otherwise we
1644          * will have updates visible that weren't in the journal:
1645          *
1646          * Make sure to update b->written so bch2_btree_init_next() doesn't
1647          * break:
1648          */
1649         if (bch2_journal_error(&c->journal) ||
1650             c->opts.nochanges)
1651                 goto err;
1652
1653         trace_btree_write(b, bytes_to_write, sectors_to_write);
1654
1655         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1656                                 buf_pages(data, sectors_to_write << 9),
1657                                 &c->btree_bio),
1658                             struct btree_write_bio, wbio.bio);
1659         wbio_init(&wbio->wbio.bio);
1660         wbio->data                      = data;
1661         wbio->wbio.order                = order;
1662         wbio->wbio.used_mempool         = used_mempool;
1663         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1664         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1665         wbio->wbio.bio.bi_private       = b;
1666
1667         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1668
1669         /*
1670          * If we're appending to a leaf node, we don't technically need FUA -
1671          * this write just needs to be persisted before the next journal write,
1672          * which will be marked FLUSH|FUA.
1673          *
1674          * Similarly if we're writing a new btree root - the pointer is going to
1675          * be in the next journal entry.
1676          *
1677          * But if we're writing a new btree node (that isn't a root) or
1678          * appending to a non leaf btree node, we need either FUA or a flush
1679          * when we write the parent with the new pointer. FUA is cheaper than a
1680          * flush, and writes appending to leaf nodes aren't blocking anything so
1681          * just make all btree node writes FUA to keep things sane.
1682          */
1683
1684         bkey_copy(&k.key, &b->key);
1685
1686         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&k.key)), ptr)
1687                 ptr->offset += b->written;
1688
1689         b->written += sectors_to_write;
1690
1691         /* XXX: submitting IO with btree locks held: */
1692         bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_BTREE, &k.key);
1693         return;
1694 err:
1695         set_btree_node_noevict(b);
1696         b->written += sectors_to_write;
1697 nowrite:
1698         btree_bounce_free(c, order, used_mempool, data);
1699         btree_node_write_done(c, b);
1700 }
1701
1702 /*
1703  * Work that must be done with write lock held:
1704  */
1705 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1706 {
1707         bool invalidated_iter = false;
1708         struct btree_node_entry *bne;
1709         struct bset_tree *t;
1710
1711         if (!btree_node_just_written(b))
1712                 return false;
1713
1714         BUG_ON(b->whiteout_u64s);
1715
1716         clear_btree_node_just_written(b);
1717
1718         /*
1719          * Note: immediately after write, bset_written() doesn't work - the
1720          * amount of data we had to write after compaction might have been
1721          * smaller than the offset of the last bset.
1722          *
1723          * However, we know that all bsets have been written here, as long as
1724          * we're still holding the write lock:
1725          */
1726
1727         /*
1728          * XXX: decide if we really want to unconditionally sort down to a
1729          * single bset:
1730          */
1731         if (b->nsets > 1) {
1732                 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1733                 invalidated_iter = true;
1734         } else {
1735                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1736         }
1737
1738         for_each_bset(b, t)
1739                 set_needs_whiteout(bset(b, t), true);
1740
1741         bch2_btree_verify(c, b);
1742
1743         /*
1744          * If later we don't unconditionally sort down to a single bset, we have
1745          * to ensure this is still true:
1746          */
1747         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1748
1749         bne = want_new_bset(c, b);
1750         if (bne)
1751                 bch2_bset_init_next(c, b, bne);
1752
1753         bch2_btree_build_aux_trees(b);
1754
1755         return invalidated_iter;
1756 }
1757
1758 /*
1759  * Use this one if the node is intent locked:
1760  */
1761 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1762                           enum six_lock_type lock_type_held)
1763 {
1764         BUG_ON(lock_type_held == SIX_LOCK_write);
1765
1766         if (lock_type_held == SIX_LOCK_intent ||
1767             six_lock_tryupgrade(&b->c.lock)) {
1768                 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1769
1770                 /* don't cycle lock unnecessarily: */
1771                 if (btree_node_just_written(b) &&
1772                     six_trylock_write(&b->c.lock)) {
1773                         bch2_btree_post_write_cleanup(c, b);
1774                         six_unlock_write(&b->c.lock);
1775                 }
1776
1777                 if (lock_type_held == SIX_LOCK_read)
1778                         six_lock_downgrade(&b->c.lock);
1779         } else {
1780                 __bch2_btree_node_write(c, b, SIX_LOCK_read);
1781         }
1782 }
1783
1784 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
1785 {
1786         struct bucket_table *tbl;
1787         struct rhash_head *pos;
1788         struct btree *b;
1789         unsigned i;
1790 restart:
1791         rcu_read_lock();
1792         for_each_cached_btree(b, c, tbl, i, pos)
1793                 if (test_bit(flag, &b->flags)) {
1794                         rcu_read_unlock();
1795                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
1796                         goto restart;
1797
1798                 }
1799         rcu_read_unlock();
1800 }
1801
1802 void bch2_btree_flush_all_reads(struct bch_fs *c)
1803 {
1804         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
1805 }
1806
1807 void bch2_btree_flush_all_writes(struct bch_fs *c)
1808 {
1809         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
1810 }
1811
1812 void bch2_btree_verify_flushed(struct bch_fs *c)
1813 {
1814         struct bucket_table *tbl;
1815         struct rhash_head *pos;
1816         struct btree *b;
1817         unsigned i;
1818
1819         rcu_read_lock();
1820         for_each_cached_btree(b, c, tbl, i, pos) {
1821                 unsigned long flags = READ_ONCE(b->flags);
1822
1823                 BUG_ON((flags & (1 << BTREE_NODE_dirty)) ||
1824                        (flags & (1 << BTREE_NODE_write_in_flight)));
1825         }
1826         rcu_read_unlock();
1827 }
1828
1829 ssize_t bch2_dirty_btree_nodes_print(struct bch_fs *c, char *buf)
1830 {
1831         struct printbuf out = _PBUF(buf, PAGE_SIZE);
1832         struct bucket_table *tbl;
1833         struct rhash_head *pos;
1834         struct btree *b;
1835         unsigned i;
1836
1837         rcu_read_lock();
1838         for_each_cached_btree(b, c, tbl, i, pos) {
1839                 unsigned long flags = READ_ONCE(b->flags);
1840
1841                 if (!(flags & (1 << BTREE_NODE_dirty)))
1842                         continue;
1843
1844                 pr_buf(&out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
1845                        b,
1846                        (flags & (1 << BTREE_NODE_dirty)) != 0,
1847                        (flags & (1 << BTREE_NODE_need_write)) != 0,
1848                        b->c.level,
1849                        b->written,
1850                        !list_empty_careful(&b->write_blocked),
1851                        b->will_make_reachable != 0,
1852                        b->will_make_reachable & 1);
1853         }
1854         rcu_read_unlock();
1855
1856         return out.pos - buf;
1857 }