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, size_t size,
  61                               bool used_mempool, void *p)
  62 {
  63         if (used_mempool)
  64                 mempool_free(p, &c->btree_bounce_pool);
  65         else
  66                 vpfree(p, size);
  67 }
  68
  69 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
  70                                 bool *used_mempool)
  71 {
  72         unsigned flags = memalloc_nofs_save();
  73         void *p;
  74
  75         BUG_ON(size > btree_bytes(c));
  76
  77         *used_mempool = false;
  78         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
  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         size_t bytes = b->whiteout_u64s * sizeof(u64);
 129
 130         if (!b->whiteout_u64s)
 131                 return;
 132
 133         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 134
 135         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 136
 137         for (k = unwritten_whiteouts_start(c, b);
 138              k != unwritten_whiteouts_end(c, b);
 139              k = bkey_next(k))
 140                 *--ptrs = k;
 141
 142         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 143
 144         k = new_whiteouts;
 145
 146         while (ptrs != ptrs_end) {
 147                 bkey_copy(k, *ptrs);
 148                 k = bkey_next(k);
 149                 ptrs++;
 150         }
 151
 152         verify_no_dups(b, new_whiteouts,
 153                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s),
 154                        btree_node_old_extent_overwrite(b));
 155
 156         memcpy_u64s(unwritten_whiteouts_start(c, b),
 157                     new_whiteouts, b->whiteout_u64s);
 158
 159         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 160 }
 161
 162 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 163                                 bool compacting, enum compact_mode mode)
 164 {
 165         if (!bset_dead_u64s(b, t))
 166                 return false;
 167
 168         switch (mode) {
 169         case COMPACT_LAZY:
 170                 return should_compact_bset_lazy(b, t) ||
 171                         (compacting && !bset_written(b, bset(b, t)));
 172         case COMPACT_ALL:
 173                 return true;
 174         default:
 175                 BUG();
 176         }
 177 }
 178
 179 static bool bch2_compact_extent_whiteouts(struct bch_fs *c,
 180                                           struct btree *b,
 181                                           enum compact_mode mode)
 182 {
 183         const struct bkey_format *f = &b->format;
 184         struct bset_tree *t;
 185         struct bkey_packed *whiteouts = NULL;
 186         struct bkey_packed *u_start, *u_pos;
 187         struct sort_iter sort_iter;
 188         unsigned bytes, whiteout_u64s = 0, u64s;
 189         bool used_mempool, compacting = false;
 190
 191         BUG_ON(!btree_node_is_extents(b));
 192
 193         for_each_bset(b, t)
 194                 if (should_compact_bset(b, t, whiteout_u64s != 0, mode))
 195                         whiteout_u64s += bset_dead_u64s(b, t);
 196
 197         if (!whiteout_u64s)
 198                 return false;
 199
 200         bch2_sort_whiteouts(c, b);
 201
 202         sort_iter_init(&sort_iter, b);
 203
 204         whiteout_u64s += b->whiteout_u64s;
 205         bytes = whiteout_u64s * sizeof(u64);
 206
 207         whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 208         u_start = u_pos = whiteouts;
 209
 210         memcpy_u64s(u_pos, unwritten_whiteouts_start(c, b),
 211                     b->whiteout_u64s);
 212         u_pos = (void *) u_pos + b->whiteout_u64s * sizeof(u64);
 213
 214         sort_iter_add(&sort_iter, u_start, u_pos);
 215
 216         for_each_bset(b, t) {
 217                 struct bset *i = bset(b, t);
 218                 struct bkey_packed *k, *n, *out, *start, *end;
 219                 struct btree_node_entry *src = NULL, *dst = NULL;
 220
 221                 if (t != b->set && !bset_written(b, i)) {
 222                         src = container_of(i, struct btree_node_entry, keys);
 223                         dst = max(write_block(b),
 224                                   (void *) btree_bkey_last(b, t - 1));
 225                 }
 226
 227                 if (src != dst)
 228                         compacting = true;
 229
 230                 if (!should_compact_bset(b, t, compacting, mode)) {
 231                         if (src != dst) {
 232                                 memmove(dst, src, sizeof(*src) +
 233                                         le16_to_cpu(src->keys.u64s) *
 234                                         sizeof(u64));
 235                                 i = &dst->keys;
 236                                 set_btree_bset(b, t, i);
 237                         }
 238                         continue;
 239                 }
 240
 241                 compacting = true;
 242                 u_start = u_pos;
 243                 start = i->start;
 244                 end = vstruct_last(i);
 245
 246                 if (src != dst) {
 247                         memmove(dst, src, sizeof(*src));
 248                         i = &dst->keys;
 249                         set_btree_bset(b, t, i);
 250                 }
 251
 252                 out = i->start;
 253
 254                 for (k = start; k != end; k = n) {
 255                         n = bkey_next_skip_noops(k, end);
 256
 257                         if (bkey_deleted(k))
 258                                 continue;
 259
 260                         BUG_ON(bkey_whiteout(k) &&
 261                                k->needs_whiteout &&
 262                                bkey_written(b, k));
 263
 264                         if (bkey_whiteout(k) && !k->needs_whiteout)
 265                                 continue;
 266
 267                         if (bkey_whiteout(k)) {
 268                                 memcpy_u64s(u_pos, k, bkeyp_key_u64s(f, k));
 269                                 set_bkeyp_val_u64s(f, u_pos, 0);
 270                                 u_pos = bkey_next(u_pos);
 271                         } else {
 272                                 bkey_copy(out, k);
 273                                 out = bkey_next(out);
 274                         }
 275                 }
 276
 277                 sort_iter_add(&sort_iter, u_start, u_pos);
 278
 279                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 280                 set_btree_bset_end(b, t);
 281                 bch2_bset_set_no_aux_tree(b, t);
 282         }
 283
 284         b->whiteout_u64s = (u64 *) u_pos - (u64 *) whiteouts;
 285
 286         BUG_ON((void *) unwritten_whiteouts_start(c, b) <
 287                (void *) btree_bkey_last(b, bset_tree_last(b)));
 288
 289         u64s = bch2_sort_extent_whiteouts(unwritten_whiteouts_start(c, b),
 290                                           &sort_iter);
 291
 292         BUG_ON(u64s > b->whiteout_u64s);
 293         BUG_ON(u_pos != whiteouts && !u64s);
 294
 295         if (u64s != b->whiteout_u64s) {
 296                 void *src = unwritten_whiteouts_start(c, b);
 297
 298                 b->whiteout_u64s = u64s;
 299                 memmove_u64s_up(unwritten_whiteouts_start(c, b), src, u64s);
 300         }
 301
 302         verify_no_dups(b,
 303                        unwritten_whiteouts_start(c, b),
 304                        unwritten_whiteouts_end(c, b),
 305                        true);
 306
 307         btree_bounce_free(c, bytes, used_mempool, whiteouts);
 308
 309         bch2_btree_build_aux_trees(b);
 310
 311         bch_btree_keys_u64s_remaining(c, b);
 312         bch2_verify_btree_nr_keys(b);
 313
 314         return true;
 315 }
 316
 317 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 318 {
 319         struct bset_tree *t;
 320         bool ret = false;
 321
 322         for_each_bset(b, t) {
 323                 struct bset *i = bset(b, t);
 324                 struct bkey_packed *k, *n, *out, *start, *end;
 325                 struct btree_node_entry *src = NULL, *dst = NULL;
 326
 327                 if (t != b->set && !bset_written(b, i)) {
 328                         src = container_of(i, struct btree_node_entry, keys);
 329                         dst = max(write_block(b),
 330                                   (void *) btree_bkey_last(b, t - 1));
 331                 }
 332
 333                 if (src != dst)
 334                         ret = true;
 335
 336                 if (!should_compact_bset(b, t, ret, mode)) {
 337                         if (src != dst) {
 338                                 memmove(dst, src, sizeof(*src) +
 339                                         le16_to_cpu(src->keys.u64s) *
 340                                         sizeof(u64));
 341                                 i = &dst->keys;
 342                                 set_btree_bset(b, t, i);
 343                         }
 344                         continue;
 345                 }
 346
 347                 start   = btree_bkey_first(b, t);
 348                 end     = btree_bkey_last(b, t);
 349
 350                 if (src != dst) {
 351                         memmove(dst, src, sizeof(*src));
 352                         i = &dst->keys;
 353                         set_btree_bset(b, t, i);
 354                 }
 355
 356                 out = i->start;
 357
 358                 for (k = start; k != end; k = n) {
 359                         n = bkey_next_skip_noops(k, end);
 360
 361                         if (!bkey_whiteout(k)) {
 362                                 bkey_copy(out, k);
 363                                 out = bkey_next(out);
 364                         } else {
 365                                 BUG_ON(k->needs_whiteout);
 366                         }
 367                 }
 368
 369                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 370                 set_btree_bset_end(b, t);
 371                 bch2_bset_set_no_aux_tree(b, t);
 372                 ret = true;
 373         }
 374
 375         bch2_verify_btree_nr_keys(b);
 376
 377         bch2_btree_build_aux_trees(b);
 378
 379         return ret;
 380 }
 381
 382 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 383                             enum compact_mode mode)
 384 {
 385         return !btree_node_old_extent_overwrite(b)
 386                 ? bch2_drop_whiteouts(b, mode)
 387                 : bch2_compact_extent_whiteouts(c, b, mode);
 388 }
 389
 390 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 391                             struct btree_iter *iter,
 392                             unsigned start_idx,
 393                             unsigned end_idx,
 394                             bool filter_whiteouts)
 395 {
 396         struct btree_node *out;
 397         struct sort_iter sort_iter;
 398         struct bset_tree *t;
 399         struct bset *start_bset = bset(b, &b->set[start_idx]);
 400         bool used_mempool = false;
 401         u64 start_time, seq = 0;
 402         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 403         bool sorting_entire_node = start_idx == 0 &&
 404                 end_idx == b->nsets;
 405
 406         sort_iter_init(&sort_iter, b);
 407
 408         for (t = b->set + start_idx;
 409              t < b->set + end_idx;
 410              t++) {
 411                 u64s += le16_to_cpu(bset(b, t)->u64s);
 412                 sort_iter_add(&sort_iter,
 413                               btree_bkey_first(b, t),
 414                               btree_bkey_last(b, t));
 415         }
 416
 417         bytes = sorting_entire_node
 418                 ? btree_bytes(c)
 419                 : __vstruct_bytes(struct btree_node, u64s);
 420
 421         out = btree_bounce_alloc(c, bytes, &used_mempool);
 422
 423         start_time = local_clock();
 424
 425         if (btree_node_old_extent_overwrite(b))
 426                 filter_whiteouts = bset_written(b, start_bset);
 427
 428         u64s = (btree_node_old_extent_overwrite(b)
 429                 ? bch2_sort_extents
 430                 : bch2_sort_keys)(out->keys.start,
 431                                   &sort_iter,
 432                                   filter_whiteouts);
 433
 434         out->keys.u64s = cpu_to_le16(u64s);
 435
 436         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 437
 438         if (sorting_entire_node)
 439                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 440                                        start_time);
 441
 442         /* Make sure we preserve bset journal_seq: */
 443         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 444                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 445         start_bset->journal_seq = cpu_to_le64(seq);
 446
 447         if (sorting_entire_node) {
 448                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 449
 450                 BUG_ON(bytes != btree_bytes(c));
 451
 452                 /*
 453                  * Our temporary buffer is the same size as the btree node's
 454                  * buffer, we can just swap buffers instead of doing a big
 455                  * memcpy()
 456                  */
 457                 *out = *b->data;
 458                 out->keys.u64s = cpu_to_le16(u64s);
 459                 swap(out, b->data);
 460                 set_btree_bset(b, b->set, &b->data->keys);
 461         } else {
 462                 start_bset->u64s = out->keys.u64s;
 463                 memcpy_u64s(start_bset->start,
 464                             out->keys.start,
 465                             le16_to_cpu(out->keys.u64s));
 466         }
 467
 468         for (i = start_idx + 1; i < end_idx; i++)
 469                 b->nr.bset_u64s[start_idx] +=
 470                         b->nr.bset_u64s[i];
 471
 472         b->nsets -= shift;
 473
 474         for (i = start_idx + 1; i < b->nsets; i++) {
 475                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 476                 b->set[i]               = b->set[i + shift];
 477         }
 478
 479         for (i = b->nsets; i < MAX_BSETS; i++)
 480                 b->nr.bset_u64s[i] = 0;
 481
 482         set_btree_bset_end(b, &b->set[start_idx]);
 483         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 484
 485         btree_bounce_free(c, bytes, used_mempool, out);
 486
 487         bch2_verify_btree_nr_keys(b);
 488 }
 489
 490 void bch2_btree_sort_into(struct bch_fs *c,
 491                          struct btree *dst,
 492                          struct btree *src)
 493 {
 494         struct btree_nr_keys nr;
 495         struct btree_node_iter src_iter;
 496         u64 start_time = local_clock();
 497
 498         BUG_ON(dst->nsets != 1);
 499
 500         bch2_bset_set_no_aux_tree(dst, dst->set);
 501
 502         bch2_btree_node_iter_init_from_start(&src_iter, src);
 503
 504         if (btree_node_is_extents(src))
 505                 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
 506                                 src, &src_iter,
 507                                 &dst->format,
 508                                 true);
 509         else
 510                 nr = bch2_sort_repack(btree_bset_first(dst),
 511                                 src, &src_iter,
 512                                 &dst->format,
 513                                 true);
 514
 515         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 516                                start_time);
 517
 518         set_btree_bset_end(dst, dst->set);
 519
 520         dst->nr.live_u64s       += nr.live_u64s;
 521         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 522         dst->nr.packed_keys     += nr.packed_keys;
 523         dst->nr.unpacked_keys   += nr.unpacked_keys;
 524
 525         bch2_verify_btree_nr_keys(dst);
 526 }
 527
 528 #define SORT_CRIT       (4096 / sizeof(u64))
 529
 530 /*
 531  * We're about to add another bset to the btree node, so if there's currently
 532  * too many bsets - sort some of them together:
 533  */
 534 static bool btree_node_compact(struct bch_fs *c, struct btree *b,
 535                                struct btree_iter *iter)
 536 {
 537         unsigned unwritten_idx;
 538         bool ret = false;
 539
 540         for (unwritten_idx = 0;
 541              unwritten_idx < b->nsets;
 542              unwritten_idx++)
 543                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 544                         break;
 545
 546         if (b->nsets - unwritten_idx > 1) {
 547                 btree_node_sort(c, b, iter, unwritten_idx,
 548                                 b->nsets, false);
 549                 ret = true;
 550         }
 551
 552         if (unwritten_idx > 1) {
 553                 btree_node_sort(c, b, iter, 0, unwritten_idx, false);
 554                 ret = true;
 555         }
 556
 557         return ret;
 558 }
 559
 560 void bch2_btree_build_aux_trees(struct btree *b)
 561 {
 562         struct bset_tree *t;
 563
 564         for_each_bset(b, t)
 565                 bch2_bset_build_aux_tree(b, t,
 566                                 !bset_written(b, bset(b, t)) &&
 567                                 t == bset_tree_last(b));
 568 }
 569
 570 /*
 571  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 572  * inserted into
 573  *
 574  * Safe to call if there already is an unwritten bset - will only add a new bset
 575  * if @b doesn't already have one.
 576  *
 577  * Returns true if we sorted (i.e. invalidated iterators
 578  */
 579 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
 580                           struct btree_iter *iter)
 581 {
 582         struct btree_node_entry *bne;
 583         bool did_sort;
 584
 585         EBUG_ON(!(b->c.lock.state.seq & 1));
 586         EBUG_ON(iter && iter->l[b->c.level].b != b);
 587
 588         did_sort = btree_node_compact(c, b, iter);
 589
 590         bne = want_new_bset(c, b);
 591         if (bne)
 592                 bch2_bset_init_next(c, b, bne);
 593
 594         bch2_btree_build_aux_trees(b);
 595
 596         if (iter && did_sort)
 597                 bch2_btree_iter_reinit_node(iter, b);
 598 }
 599
 600 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 601                           struct btree *b, struct bset *i,
 602                           unsigned offset, int write)
 603 {
 604         pr_buf(out, "error validating btree node %sat btree %u level %u/%u\n"
 605                "pos ",
 606                write ? "before write " : "",
 607                b->c.btree_id, b->c.level,
 608                c->btree_roots[b->c.btree_id].level);
 609         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 610
 611         pr_buf(out, " node offset %u", b->written);
 612         if (i)
 613                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 614 }
 615
 616 enum btree_err_type {
 617         BTREE_ERR_FIXABLE,
 618         BTREE_ERR_WANT_RETRY,
 619         BTREE_ERR_MUST_RETRY,
 620         BTREE_ERR_FATAL,
 621 };
 622
 623 enum btree_validate_ret {
 624         BTREE_RETRY_READ = 64,
 625 };
 626
 627 #define btree_err(type, c, b, i, msg, ...)                              \
 628 ({                                                                      \
 629         __label__ out;                                                  \
 630         char _buf[300];                                                 \
 631         struct printbuf out = PBUF(_buf);                               \
 632                                                                         \
 633         btree_err_msg(&out, c, b, i, b->written, write);                \
 634         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 635                                                                         \
 636         if (type == BTREE_ERR_FIXABLE &&                                \
 637             write == READ &&                                            \
 638             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 639                 mustfix_fsck_err(c, "%s", _buf);                        \
 640                 goto out;                                               \
 641         }                                                               \
 642                                                                         \
 643         switch (write) {                                                \
 644         case READ:                                                      \
 645                 bch_err(c, "%s", _buf);                                 \
 646                                                                         \
 647                 switch (type) {                                         \
 648                 case BTREE_ERR_FIXABLE:                                 \
 649                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 650                         goto fsck_err;                                  \
 651                 case BTREE_ERR_WANT_RETRY:                              \
 652                         if (have_retry) {                               \
 653                                 ret = BTREE_RETRY_READ;                 \
 654                                 goto fsck_err;                          \
 655                         }                                               \
 656                         break;                                          \
 657                 case BTREE_ERR_MUST_RETRY:                              \
 658                         ret = BTREE_RETRY_READ;                         \
 659                         goto fsck_err;                                  \
 660                 case BTREE_ERR_FATAL:                                   \
 661                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 662                         goto fsck_err;                                  \
 663                 }                                                       \
 664                 break;                                                  \
 665         case WRITE:                                                     \
 666                 bch_err(c, "corrupt metadata before write: %s", _buf);  \
 667                                                                         \
 668                 if (bch2_fs_inconsistent(c)) {                          \
 669                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 670                         goto fsck_err;                                  \
 671                 }                                                       \
 672                 break;                                                  \
 673         }                                                               \
 674 out:                                                                    \
 675         true;                                                           \
 676 })
 677
 678 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 679
 680 static int validate_bset(struct bch_fs *c, struct btree *b,
 681                          struct bset *i, unsigned sectors,
 682                          int write, bool have_retry)
 683 {
 684         unsigned version = le16_to_cpu(i->version);
 685         const char *err;
 686         int ret = 0;
 687
 688         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 689                       version < bcachefs_metadata_version_min) ||
 690                      version >= bcachefs_metadata_version_max,
 691                      BTREE_ERR_FATAL, c, b, i,
 692                      "unsupported bset version");
 693
 694         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 695                          BTREE_ERR_FIXABLE, c, b, i,
 696                          "bset past end of btree node")) {
 697                 i->u64s = 0;
 698                 return 0;
 699         }
 700
 701         btree_err_on(b->written && !i->u64s,
 702                      BTREE_ERR_FIXABLE, c, b, i,
 703                      "empty bset");
 704
 705         if (!b->written) {
 706                 struct btree_node *bn =
 707                         container_of(i, struct btree_node, keys);
 708                 /* These indicate that we read the wrong btree node: */
 709
 710                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 711                         struct bch_btree_ptr_v2 *bp =
 712                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 713
 714                         /* XXX endianness */
 715                         btree_err_on(bp->seq != bn->keys.seq,
 716                                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 717                                      "incorrect sequence number (wrong btree node)");
 718                 }
 719
 720                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 721                              BTREE_ERR_MUST_RETRY, c, b, i,
 722                              "incorrect btree id");
 723
 724                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 725                              BTREE_ERR_MUST_RETRY, c, b, i,
 726                              "incorrect level");
 727
 728                 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
 729                         u64 *p = (u64 *) &bn->ptr;
 730
 731                         *p = swab64(*p);
 732                 }
 733
 734                 if (!write)
 735                         compat_btree_node(b->c.level, b->c.btree_id, version,
 736                                           BSET_BIG_ENDIAN(i), write, bn);
 737
 738                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 739                         struct bch_btree_ptr_v2 *bp =
 740                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 741
 742                         btree_err_on(bkey_cmp(b->data->min_key, bp->min_key),
 743                                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 744                                      "incorrect min_key: got %llu:%llu should be %llu:%llu",
 745                                      b->data->min_key.inode,
 746                                      b->data->min_key.offset,
 747                                      bp->min_key.inode,
 748                                      bp->min_key.offset);
 749                 }
 750
 751                 btree_err_on(bkey_cmp(bn->max_key, b->key.k.p),
 752                              BTREE_ERR_MUST_RETRY, c, b, i,
 753                              "incorrect max key");
 754
 755                 if (write)
 756                         compat_btree_node(b->c.level, b->c.btree_id, version,
 757                                           BSET_BIG_ENDIAN(i), write, bn);
 758
 759                 /* XXX: ideally we would be validating min_key too */
 760 #if 0
 761                 /*
 762                  * not correct anymore, due to btree node write error
 763                  * handling
 764                  *
 765                  * need to add bn->seq to btree keys and verify
 766                  * against that
 767                  */
 768                 btree_err_on(!extent_contains_ptr(bkey_i_to_s_c_extent(&b->key),
 769                                                   bn->ptr),
 770                              BTREE_ERR_FATAL, c, b, i,
 771                              "incorrect backpointer");
 772 #endif
 773                 err = bch2_bkey_format_validate(&bn->format);
 774                 btree_err_on(err,
 775                              BTREE_ERR_FATAL, c, b, i,
 776                              "invalid bkey format: %s", err);
 777
 778                 compat_bformat(b->c.level, b->c.btree_id, version,
 779                                BSET_BIG_ENDIAN(i), write,
 780                                &bn->format);
 781         }
 782 fsck_err:
 783         return ret;
 784 }
 785
 786 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 787                          struct bset *i, unsigned *whiteout_u64s,
 788                          int write, bool have_retry)
 789 {
 790         unsigned version = le16_to_cpu(i->version);
 791         struct bkey_packed *k, *prev = NULL;
 792         bool seen_non_whiteout = false;
 793         int ret = 0;
 794
 795         if (!BSET_SEPARATE_WHITEOUTS(i)) {
 796                 seen_non_whiteout = true;
 797                 *whiteout_u64s = 0;
 798         }
 799
 800         for (k = i->start;
 801              k != vstruct_last(i);) {
 802                 struct bkey_s u;
 803                 struct bkey tmp;
 804                 const char *invalid;
 805
 806                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 807                                  BTREE_ERR_FIXABLE, c, b, i,
 808                                  "key extends past end of bset")) {
 809                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 810                         break;
 811                 }
 812
 813                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 814                                  BTREE_ERR_FIXABLE, c, b, i,
 815                                  "invalid bkey format %u", k->format)) {
 816                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 817                         memmove_u64s_down(k, bkey_next(k),
 818                                           (u64 *) vstruct_end(i) - (u64 *) k);
 819                         continue;
 820                 }
 821
 822                 /* XXX: validate k->u64s */
 823                 if (!write)
 824                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 825                                     BSET_BIG_ENDIAN(i), write,
 826                                     &b->format, k);
 827
 828                 u = __bkey_disassemble(b, k, &tmp);
 829
 830                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 831                         bch2_bkey_in_btree_node(b, u.s_c) ?:
 832                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 833                 if (invalid) {
 834                         char buf[160];
 835
 836                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 837                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
 838                                   "invalid bkey:\n%s\n%s", invalid, buf);
 839
 840                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 841                         memmove_u64s_down(k, bkey_next(k),
 842                                           (u64 *) vstruct_end(i) - (u64 *) k);
 843                         continue;
 844                 }
 845
 846                 if (write)
 847                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 848                                     BSET_BIG_ENDIAN(i), write,
 849                                     &b->format, k);
 850
 851                 /*
 852                  * with the separate whiteouts thing (used for extents), the
 853                  * second set of keys actually can have whiteouts too, so we
 854                  * can't solely go off bkey_whiteout()...
 855                  */
 856
 857                 if (!seen_non_whiteout &&
 858                     (!bkey_whiteout(k) ||
 859                      (prev && bkey_iter_cmp(b, prev, k) > 0))) {
 860                         *whiteout_u64s = k->_data - i->_data;
 861                         seen_non_whiteout = true;
 862                 } else if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 863                         char buf1[80];
 864                         char buf2[80];
 865                         struct bkey up = bkey_unpack_key(b, prev);
 866
 867                         bch2_bkey_to_text(&PBUF(buf1), &up);
 868                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 869
 870                         bch2_dump_bset(c, b, i, 0);
 871                         btree_err(BTREE_ERR_FATAL, c, b, i,
 872                                   "keys out of order: %s > %s",
 873                                   buf1, buf2);
 874                         /* XXX: repair this */
 875                 }
 876
 877                 prev = k;
 878                 k = bkey_next_skip_noops(k, vstruct_last(i));
 879         }
 880 fsck_err:
 881         return ret;
 882 }
 883
 884 int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry)
 885 {
 886         struct btree_node_entry *bne;
 887         struct sort_iter *iter;
 888         struct btree_node *sorted;
 889         struct bkey_packed *k;
 890         struct bch_extent_ptr *ptr;
 891         struct bset *i;
 892         bool used_mempool, blacklisted;
 893         unsigned u64s;
 894         int ret, retry_read = 0, write = READ;
 895
 896         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 897         sort_iter_init(iter, b);
 898         iter->size = (btree_blocks(c) + 1) * 2;
 899
 900         if (bch2_meta_read_fault("btree"))
 901                 btree_err(BTREE_ERR_MUST_RETRY, c, b, NULL,
 902                           "dynamic fault");
 903
 904         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 905                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 906                      "bad magic");
 907
 908         btree_err_on(!b->data->keys.seq,
 909                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 910                      "bad btree header");
 911
 912         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 913                 struct bch_btree_ptr_v2 *bp =
 914                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 915
 916                 btree_err_on(b->data->keys.seq != bp->seq,
 917                              BTREE_ERR_MUST_RETRY, c, b, NULL,
 918                              "got wrong btree node (seq %llx want %llx)",
 919                              b->data->keys.seq, bp->seq);
 920         }
 921
 922         while (b->written < c->opts.btree_node_size) {
 923                 unsigned sectors, whiteout_u64s = 0;
 924                 struct nonce nonce;
 925                 struct bch_csum csum;
 926                 bool first = !b->written;
 927
 928                 if (!b->written) {
 929                         i = &b->data->keys;
 930
 931                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 932                                      BTREE_ERR_WANT_RETRY, c, b, i,
 933                                      "unknown checksum type");
 934
 935                         nonce = btree_nonce(i, b->written << 9);
 936                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 937
 938                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 939                                      BTREE_ERR_WANT_RETRY, c, b, i,
 940                                      "invalid checksum");
 941
 942                         bset_encrypt(c, i, b->written << 9);
 943
 944                         if (btree_node_is_extents(b) &&
 945                             !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data)) {
 946                                 set_btree_node_old_extent_overwrite(b);
 947                                 set_btree_node_need_rewrite(b);
 948                         }
 949
 950                         sectors = vstruct_sectors(b->data, c->block_bits);
 951                 } else {
 952                         bne = write_block(b);
 953                         i = &bne->keys;
 954
 955                         if (i->seq != b->data->keys.seq)
 956                                 break;
 957
 958                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 959                                      BTREE_ERR_WANT_RETRY, c, b, i,
 960                                      "unknown checksum type");
 961
 962                         nonce = btree_nonce(i, b->written << 9);
 963                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 964
 965                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 966                                      BTREE_ERR_WANT_RETRY, c, b, i,
 967                                      "invalid checksum");
 968
 969                         bset_encrypt(c, i, b->written << 9);
 970
 971                         sectors = vstruct_sectors(bne, c->block_bits);
 972                 }
 973
 974                 ret = validate_bset(c, b, i, sectors,
 975                                     READ, have_retry);
 976                 if (ret)
 977                         goto fsck_err;
 978
 979                 if (!b->written)
 980                         btree_node_set_format(b, b->data->format);
 981
 982                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 983                                     READ, have_retry);
 984                 if (ret)
 985                         goto fsck_err;
 986
 987                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 988
 989                 b->written += sectors;
 990
 991                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 992                                         le64_to_cpu(i->journal_seq),
 993                                         true);
 994
 995                 btree_err_on(blacklisted && first,
 996                              BTREE_ERR_FIXABLE, c, b, i,
 997                              "first btree node bset has blacklisted journal seq");
 998                 if (blacklisted && !first)
 999                         continue;
1000
1001                 sort_iter_add(iter, i->start,
1002                               vstruct_idx(i, whiteout_u64s));
1003
1004                 sort_iter_add(iter,
1005                               vstruct_idx(i, whiteout_u64s),
1006                               vstruct_last(i));
1007         }
1008
1009         for (bne = write_block(b);
1010              bset_byte_offset(b, bne) < btree_bytes(c);
1011              bne = (void *) bne + block_bytes(c))
1012                 btree_err_on(bne->keys.seq == b->data->keys.seq,
1013                              BTREE_ERR_WANT_RETRY, c, b, NULL,
1014                              "found bset signature after last bset");
1015
1016         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1017         sorted->keys.u64s = 0;
1018
1019         set_btree_bset(b, b->set, &b->data->keys);
1020
1021         b->nr = (btree_node_old_extent_overwrite(b)
1022                  ? bch2_extent_sort_fix_overlapping
1023                  : bch2_key_sort_fix_overlapping)(c, &sorted->keys, iter);
1024
1025         u64s = le16_to_cpu(sorted->keys.u64s);
1026         *sorted = *b->data;
1027         sorted->keys.u64s = cpu_to_le16(u64s);
1028         swap(sorted, b->data);
1029         set_btree_bset(b, b->set, &b->data->keys);
1030         b->nsets = 1;
1031
1032         BUG_ON(b->nr.live_u64s != u64s);
1033
1034         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1035
1036         i = &b->data->keys;
1037         for (k = i->start; k != vstruct_last(i);) {
1038                 struct bkey tmp;
1039                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1040                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
1041
1042                 if (invalid ||
1043                     (inject_invalid_keys(c) &&
1044                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1045                         char buf[160];
1046
1047                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
1048                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
1049                                   "invalid bkey %s: %s", buf, invalid);
1050
1051                         btree_keys_account_key_drop(&b->nr, 0, k);
1052
1053                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1054                         memmove_u64s_down(k, bkey_next(k),
1055                                           (u64 *) vstruct_end(i) - (u64 *) k);
1056                         set_btree_bset_end(b, b->set);
1057                         continue;
1058                 }
1059
1060                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1061                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1062
1063                         bp.v->mem_ptr = 0;
1064                 }
1065
1066                 k = bkey_next_skip_noops(k, vstruct_last(i));
1067         }
1068
1069         bch2_bset_build_aux_tree(b, b->set, false);
1070
1071         set_needs_whiteout(btree_bset_first(b), true);
1072
1073         btree_node_reset_sib_u64s(b);
1074
1075         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1076                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1077
1078                 if (ca->mi.state != BCH_MEMBER_STATE_RW)
1079                         set_btree_node_need_rewrite(b);
1080         }
1081 out:
1082         mempool_free(iter, &c->fill_iter);
1083         return retry_read;
1084 fsck_err:
1085         if (ret == BTREE_RETRY_READ) {
1086                 retry_read = 1;
1087         } else {
1088                 bch2_inconsistent_error(c);
1089                 set_btree_node_read_error(b);
1090         }
1091         goto out;
1092 }
1093
1094 static void btree_node_read_work(struct work_struct *work)
1095 {
1096         struct btree_read_bio *rb =
1097                 container_of(work, struct btree_read_bio, work);
1098         struct bch_fs *c        = rb->c;
1099         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1100         struct btree *b         = rb->bio.bi_private;
1101         struct bio *bio         = &rb->bio;
1102         struct bch_io_failures failed = { .nr = 0 };
1103         bool can_retry;
1104
1105         goto start;
1106         while (1) {
1107                 bch_info(c, "retrying read");
1108                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1109                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1110                 bio_reset(bio);
1111                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1112                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1113                 bio->bi_iter.bi_size    = btree_bytes(c);
1114
1115                 if (rb->have_ioref) {
1116                         bio_set_dev(bio, ca->disk_sb.bdev);
1117                         submit_bio_wait(bio);
1118                 } else {
1119                         bio->bi_status = BLK_STS_REMOVED;
1120                 }
1121 start:
1122                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read: %s",
1123                                    bch2_blk_status_to_str(bio->bi_status));
1124                 if (rb->have_ioref)
1125                         percpu_ref_put(&ca->io_ref);
1126                 rb->have_ioref = false;
1127
1128                 bch2_mark_io_failure(&failed, &rb->pick);
1129
1130                 can_retry = bch2_bkey_pick_read_device(c,
1131                                 bkey_i_to_s_c(&b->key),
1132                                 &failed, &rb->pick) > 0;
1133
1134                 if (!bio->bi_status &&
1135                     !bch2_btree_node_read_done(c, b, can_retry))
1136                         break;
1137
1138                 if (!can_retry) {
1139                         set_btree_node_read_error(b);
1140                         break;
1141                 }
1142         }
1143
1144         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1145                                rb->start_time);
1146         bio_put(&rb->bio);
1147         clear_btree_node_read_in_flight(b);
1148         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1149 }
1150
1151 static void btree_node_read_endio(struct bio *bio)
1152 {
1153         struct btree_read_bio *rb =
1154                 container_of(bio, struct btree_read_bio, bio);
1155         struct bch_fs *c        = rb->c;
1156
1157         if (rb->have_ioref) {
1158                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1159                 bch2_latency_acct(ca, rb->start_time, READ);
1160         }
1161
1162         queue_work(system_unbound_wq, &rb->work);
1163 }
1164
1165 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1166                           bool sync)
1167 {
1168         struct extent_ptr_decoded pick;
1169         struct btree_read_bio *rb;
1170         struct bch_dev *ca;
1171         struct bio *bio;
1172         int ret;
1173
1174         trace_btree_read(c, b);
1175
1176         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1177                                          NULL, &pick);
1178         if (bch2_fs_fatal_err_on(ret <= 0, c,
1179                         "btree node read error: no device to read from")) {
1180                 set_btree_node_read_error(b);
1181                 return;
1182         }
1183
1184         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1185
1186         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1187                                                    btree_bytes(c)),
1188                                &c->btree_bio);
1189         rb = container_of(bio, struct btree_read_bio, bio);
1190         rb->c                   = c;
1191         rb->start_time          = local_clock();
1192         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1193         rb->pick                = pick;
1194         INIT_WORK(&rb->work, btree_node_read_work);
1195         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1196         bio->bi_iter.bi_sector  = pick.ptr.offset;
1197         bio->bi_end_io          = btree_node_read_endio;
1198         bio->bi_private         = b;
1199         bch2_bio_map(bio, b->data, btree_bytes(c));
1200
1201         set_btree_node_read_in_flight(b);
1202
1203         if (rb->have_ioref) {
1204                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1205                              bio_sectors(bio));
1206                 bio_set_dev(bio, ca->disk_sb.bdev);
1207
1208                 if (sync) {
1209                         submit_bio_wait(bio);
1210
1211                         bio->bi_private = b;
1212                         btree_node_read_work(&rb->work);
1213                 } else {
1214                         submit_bio(bio);
1215                 }
1216         } else {
1217                 bio->bi_status = BLK_STS_REMOVED;
1218
1219                 if (sync)
1220                         btree_node_read_work(&rb->work);
1221                 else
1222                         queue_work(system_unbound_wq, &rb->work);
1223
1224         }
1225 }
1226
1227 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1228                         const struct bkey_i *k, unsigned level)
1229 {
1230         struct closure cl;
1231         struct btree *b;
1232         int ret;
1233
1234         closure_init_stack(&cl);
1235
1236         do {
1237                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1238                 closure_sync(&cl);
1239         } while (ret);
1240
1241         b = bch2_btree_node_mem_alloc(c);
1242         bch2_btree_cache_cannibalize_unlock(c);
1243
1244         BUG_ON(IS_ERR(b));
1245
1246         bkey_copy(&b->key, k);
1247         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1248
1249         bch2_btree_node_read(c, b, true);
1250
1251         if (btree_node_read_error(b)) {
1252                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1253
1254                 mutex_lock(&c->btree_cache.lock);
1255                 list_move(&b->list, &c->btree_cache.freeable);
1256                 mutex_unlock(&c->btree_cache.lock);
1257
1258                 ret = -EIO;
1259                 goto err;
1260         }
1261
1262         bch2_btree_set_root_for_read(c, b);
1263 err:
1264         six_unlock_write(&b->c.lock);
1265         six_unlock_intent(&b->c.lock);
1266
1267         return ret;
1268 }
1269
1270 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1271                               struct btree_write *w)
1272 {
1273         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1274
1275         do {
1276                 old = new = v;
1277                 if (!(old & 1))
1278                         break;
1279
1280                 new &= ~1UL;
1281         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1282
1283         if (old & 1)
1284                 closure_put(&((struct btree_update *) new)->cl);
1285
1286         bch2_journal_pin_drop(&c->journal, &w->journal);
1287 }
1288
1289 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1290 {
1291         struct btree_write *w = btree_prev_write(b);
1292
1293         bch2_btree_complete_write(c, b, w);
1294         btree_node_io_unlock(b);
1295 }
1296
1297 static void bch2_btree_node_write_error(struct bch_fs *c,
1298                                         struct btree_write_bio *wbio)
1299 {
1300         struct btree *b         = wbio->wbio.bio.bi_private;
1301         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1302         struct bch_extent_ptr *ptr;
1303         struct btree_trans trans;
1304         struct btree_iter *iter;
1305         int ret;
1306
1307         bch2_trans_init(&trans, c, 0, 0);
1308
1309         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1310                                         BTREE_MAX_DEPTH, b->c.level, 0);
1311 retry:
1312         ret = bch2_btree_iter_traverse(iter);
1313         if (ret)
1314                 goto err;
1315
1316         /* has node been freed? */
1317         if (iter->l[b->c.level].b != b) {
1318                 /* node has been freed: */
1319                 BUG_ON(!btree_node_dying(b));
1320                 goto out;
1321         }
1322
1323         BUG_ON(!btree_node_hashed(b));
1324
1325         bkey_copy(&tmp.k, &b->key);
1326
1327         bch2_bkey_drop_ptrs(bkey_i_to_s(&tmp.k), ptr,
1328                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1329
1330         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&tmp.k)))
1331                 goto err;
1332
1333         ret = bch2_btree_node_update_key(c, iter, b, &tmp.k);
1334         if (ret == -EINTR)
1335                 goto retry;
1336         if (ret)
1337                 goto err;
1338 out:
1339         bch2_trans_exit(&trans);
1340         bio_put(&wbio->wbio.bio);
1341         btree_node_write_done(c, b);
1342         return;
1343 err:
1344         set_btree_node_noevict(b);
1345         bch2_fs_fatal_error(c, "fatal error writing btree node");
1346         goto out;
1347 }
1348
1349 void bch2_btree_write_error_work(struct work_struct *work)
1350 {
1351         struct bch_fs *c = container_of(work, struct bch_fs,
1352                                         btree_write_error_work);
1353         struct bio *bio;
1354
1355         while (1) {
1356                 spin_lock_irq(&c->btree_write_error_lock);
1357                 bio = bio_list_pop(&c->btree_write_error_list);
1358                 spin_unlock_irq(&c->btree_write_error_lock);
1359
1360                 if (!bio)
1361                         break;
1362
1363                 bch2_btree_node_write_error(c,
1364                         container_of(bio, struct btree_write_bio, wbio.bio));
1365         }
1366 }
1367
1368 static void btree_node_write_work(struct work_struct *work)
1369 {
1370         struct btree_write_bio *wbio =
1371                 container_of(work, struct btree_write_bio, work);
1372         struct bch_fs *c        = wbio->wbio.c;
1373         struct btree *b         = wbio->wbio.bio.bi_private;
1374
1375         btree_bounce_free(c,
1376                 wbio->bytes,
1377                 wbio->wbio.used_mempool,
1378                 wbio->data);
1379
1380         if (wbio->wbio.failed.nr) {
1381                 unsigned long flags;
1382
1383                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1384                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1385                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1386
1387                 queue_work(c->wq, &c->btree_write_error_work);
1388                 return;
1389         }
1390
1391         bio_put(&wbio->wbio.bio);
1392         btree_node_write_done(c, b);
1393 }
1394
1395 static void btree_node_write_endio(struct bio *bio)
1396 {
1397         struct bch_write_bio *wbio      = to_wbio(bio);
1398         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1399         struct bch_write_bio *orig      = parent ?: wbio;
1400         struct bch_fs *c                = wbio->c;
1401         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1402         unsigned long flags;
1403
1404         if (wbio->have_ioref)
1405                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1406
1407         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write: %s",
1408                                bch2_blk_status_to_str(bio->bi_status)) ||
1409             bch2_meta_write_fault("btree")) {
1410                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1411                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1412                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1413         }
1414
1415         if (wbio->have_ioref)
1416                 percpu_ref_put(&ca->io_ref);
1417
1418         if (parent) {
1419                 bio_put(bio);
1420                 bio_endio(&parent->bio);
1421         } else {
1422                 struct btree_write_bio *wb =
1423                         container_of(orig, struct btree_write_bio, wbio);
1424
1425                 INIT_WORK(&wb->work, btree_node_write_work);
1426                 queue_work(system_unbound_wq, &wb->work);
1427         }
1428 }
1429
1430 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1431                                    struct bset *i, unsigned sectors)
1432 {
1433         unsigned whiteout_u64s = 0;
1434         int ret;
1435
1436         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_BTREE))
1437                 return -1;
1438
1439         ret = validate_bset(c, b, i, sectors, WRITE, false) ?:
1440                 validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false);
1441         if (ret)
1442                 bch2_inconsistent_error(c);
1443
1444         return ret;
1445 }
1446
1447 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1448                             enum six_lock_type lock_type_held)
1449 {
1450         struct btree_write_bio *wbio;
1451         struct bset_tree *t;
1452         struct bset *i;
1453         struct btree_node *bn = NULL;
1454         struct btree_node_entry *bne = NULL;
1455         BKEY_PADDED(key) k;
1456         struct bch_extent_ptr *ptr;
1457         struct sort_iter sort_iter;
1458         struct nonce nonce;
1459         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1460         u64 seq = 0;
1461         bool used_mempool;
1462         unsigned long old, new;
1463         bool validate_before_checksum = false;
1464         void *data;
1465
1466         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1467                 return;
1468
1469         /*
1470          * We may only have a read lock on the btree node - the dirty bit is our
1471          * "lock" against racing with other threads that may be trying to start
1472          * a write, we do a write iff we clear the dirty bit. Since setting the
1473          * dirty bit requires a write lock, we can't race with other threads
1474          * redirtying it:
1475          */
1476         do {
1477                 old = new = READ_ONCE(b->flags);
1478
1479                 if (!(old & (1 << BTREE_NODE_dirty)))
1480                         return;
1481
1482                 if (!btree_node_may_write(b))
1483                         return;
1484
1485                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1486                         btree_node_wait_on_io(b);
1487                         continue;
1488                 }
1489
1490                 new &= ~(1 << BTREE_NODE_dirty);
1491                 new &= ~(1 << BTREE_NODE_need_write);
1492                 new |=  (1 << BTREE_NODE_write_in_flight);
1493                 new |=  (1 << BTREE_NODE_just_written);
1494                 new ^=  (1 << BTREE_NODE_write_idx);
1495         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1496
1497         BUG_ON(btree_node_fake(b));
1498         BUG_ON((b->will_make_reachable != 0) != !b->written);
1499
1500         BUG_ON(b->written >= c->opts.btree_node_size);
1501         BUG_ON(b->written & (c->opts.block_size - 1));
1502         BUG_ON(bset_written(b, btree_bset_last(b)));
1503         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1504         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1505
1506         bch2_sort_whiteouts(c, b);
1507
1508         sort_iter_init(&sort_iter, b);
1509
1510         bytes = !b->written
1511                 ? sizeof(struct btree_node)
1512                 : sizeof(struct btree_node_entry);
1513
1514         bytes += b->whiteout_u64s * sizeof(u64);
1515
1516         for_each_bset(b, t) {
1517                 i = bset(b, t);
1518
1519                 if (bset_written(b, i))
1520                         continue;
1521
1522                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1523                 sort_iter_add(&sort_iter,
1524                               btree_bkey_first(b, t),
1525                               btree_bkey_last(b, t));
1526                 seq = max(seq, le64_to_cpu(i->journal_seq));
1527         }
1528
1529         data = btree_bounce_alloc(c, bytes, &used_mempool);
1530
1531         if (!b->written) {
1532                 bn = data;
1533                 *bn = *b->data;
1534                 i = &bn->keys;
1535         } else {
1536                 bne = data;
1537                 bne->keys = b->data->keys;
1538                 i = &bne->keys;
1539         }
1540
1541         i->journal_seq  = cpu_to_le64(seq);
1542         i->u64s         = 0;
1543
1544         if (!btree_node_old_extent_overwrite(b)) {
1545                 sort_iter_add(&sort_iter,
1546                               unwritten_whiteouts_start(c, b),
1547                               unwritten_whiteouts_end(c, b));
1548                 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1549         } else {
1550                 memcpy_u64s(i->start,
1551                             unwritten_whiteouts_start(c, b),
1552                             b->whiteout_u64s);
1553                 i->u64s = cpu_to_le16(b->whiteout_u64s);
1554                 SET_BSET_SEPARATE_WHITEOUTS(i, true);
1555         }
1556
1557         b->whiteout_u64s = 0;
1558
1559         u64s = btree_node_old_extent_overwrite(b)
1560                 ? bch2_sort_extents(vstruct_last(i), &sort_iter, false)
1561                 : bch2_sort_keys(i->start, &sort_iter, false);
1562         le16_add_cpu(&i->u64s, u64s);
1563
1564         set_needs_whiteout(i, false);
1565
1566         /* do we have data to write? */
1567         if (b->written && !i->u64s)
1568                 goto nowrite;
1569
1570         bytes_to_write = vstruct_end(i) - data;
1571         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1572
1573         memset(data + bytes_to_write, 0,
1574                (sectors_to_write << 9) - bytes_to_write);
1575
1576         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1577         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1578         BUG_ON(i->seq != b->data->keys.seq);
1579
1580         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1581                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1582                 : cpu_to_le16(c->sb.version);
1583         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1584
1585         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1586                 validate_before_checksum = true;
1587
1588         /* validate_bset will be modifying: */
1589         if (le16_to_cpu(i->version) < bcachefs_metadata_version_max)
1590                 validate_before_checksum = true;
1591
1592         /* if we're going to be encrypting, check metadata validity first: */
1593         if (validate_before_checksum &&
1594             validate_bset_for_write(c, b, i, sectors_to_write))
1595                 goto err;
1596
1597         bset_encrypt(c, i, b->written << 9);
1598
1599         nonce = btree_nonce(i, b->written << 9);
1600
1601         if (bn)
1602                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1603         else
1604                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1605
1606         /* if we're not encrypting, check metadata after checksumming: */
1607         if (!validate_before_checksum &&
1608             validate_bset_for_write(c, b, i, sectors_to_write))
1609                 goto err;
1610
1611         /*
1612          * We handle btree write errors by immediately halting the journal -
1613          * after we've done that, we can't issue any subsequent btree writes
1614          * because they might have pointers to new nodes that failed to write.
1615          *
1616          * Furthermore, there's no point in doing any more btree writes because
1617          * with the journal stopped, we're never going to update the journal to
1618          * reflect that those writes were done and the data flushed from the
1619          * journal:
1620          *
1621          * Also on journal error, the pending write may have updates that were
1622          * never journalled (interior nodes, see btree_update_nodes_written()) -
1623          * it's critical that we don't do the write in that case otherwise we
1624          * will have updates visible that weren't in the journal:
1625          *
1626          * Make sure to update b->written so bch2_btree_init_next() doesn't
1627          * break:
1628          */
1629         if (bch2_journal_error(&c->journal) ||
1630             c->opts.nochanges)
1631                 goto err;
1632
1633         trace_btree_write(b, bytes_to_write, sectors_to_write);
1634
1635         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1636                                 buf_pages(data, sectors_to_write << 9),
1637                                 &c->btree_bio),
1638                             struct btree_write_bio, wbio.bio);
1639         wbio_init(&wbio->wbio.bio);
1640         wbio->data                      = data;
1641         wbio->bytes                     = bytes;
1642         wbio->wbio.used_mempool         = used_mempool;
1643         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1644         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1645         wbio->wbio.bio.bi_private       = b;
1646
1647         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1648
1649         /*
1650          * If we're appending to a leaf node, we don't technically need FUA -
1651          * this write just needs to be persisted before the next journal write,
1652          * which will be marked FLUSH|FUA.
1653          *
1654          * Similarly if we're writing a new btree root - the pointer is going to
1655          * be in the next journal entry.
1656          *
1657          * But if we're writing a new btree node (that isn't a root) or
1658          * appending to a non leaf btree node, we need either FUA or a flush
1659          * when we write the parent with the new pointer. FUA is cheaper than a
1660          * flush, and writes appending to leaf nodes aren't blocking anything so
1661          * just make all btree node writes FUA to keep things sane.
1662          */
1663
1664         bkey_copy(&k.key, &b->key);
1665
1666         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&k.key)), ptr)
1667                 ptr->offset += b->written;
1668
1669         b->written += sectors_to_write;
1670
1671         /* XXX: submitting IO with btree locks held: */
1672         bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_btree, &k.key);
1673         return;
1674 err:
1675         set_btree_node_noevict(b);
1676         b->written += sectors_to_write;
1677 nowrite:
1678         btree_bounce_free(c, bytes, used_mempool, data);
1679         btree_node_write_done(c, b);
1680 }
1681
1682 /*
1683  * Work that must be done with write lock held:
1684  */
1685 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1686 {
1687         bool invalidated_iter = false;
1688         struct btree_node_entry *bne;
1689         struct bset_tree *t;
1690
1691         if (!btree_node_just_written(b))
1692                 return false;
1693
1694         BUG_ON(b->whiteout_u64s);
1695
1696         clear_btree_node_just_written(b);
1697
1698         /*
1699          * Note: immediately after write, bset_written() doesn't work - the
1700          * amount of data we had to write after compaction might have been
1701          * smaller than the offset of the last bset.
1702          *
1703          * However, we know that all bsets have been written here, as long as
1704          * we're still holding the write lock:
1705          */
1706
1707         /*
1708          * XXX: decide if we really want to unconditionally sort down to a
1709          * single bset:
1710          */
1711         if (b->nsets > 1) {
1712                 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1713                 invalidated_iter = true;
1714         } else {
1715                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1716         }
1717
1718         for_each_bset(b, t)
1719                 set_needs_whiteout(bset(b, t), true);
1720
1721         bch2_btree_verify(c, b);
1722
1723         /*
1724          * If later we don't unconditionally sort down to a single bset, we have
1725          * to ensure this is still true:
1726          */
1727         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1728
1729         bne = want_new_bset(c, b);
1730         if (bne)
1731                 bch2_bset_init_next(c, b, bne);
1732
1733         bch2_btree_build_aux_trees(b);
1734
1735         return invalidated_iter;
1736 }
1737
1738 /*
1739  * Use this one if the node is intent locked:
1740  */
1741 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1742                           enum six_lock_type lock_type_held)
1743 {
1744         BUG_ON(lock_type_held == SIX_LOCK_write);
1745
1746         if (lock_type_held == SIX_LOCK_intent ||
1747             six_lock_tryupgrade(&b->c.lock)) {
1748                 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1749
1750                 /* don't cycle lock unnecessarily: */
1751                 if (btree_node_just_written(b) &&
1752                     six_trylock_write(&b->c.lock)) {
1753                         bch2_btree_post_write_cleanup(c, b);
1754                         six_unlock_write(&b->c.lock);
1755                 }
1756
1757                 if (lock_type_held == SIX_LOCK_read)
1758                         six_lock_downgrade(&b->c.lock);
1759         } else {
1760                 __bch2_btree_node_write(c, b, SIX_LOCK_read);
1761         }
1762 }
1763
1764 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
1765 {
1766         struct bucket_table *tbl;
1767         struct rhash_head *pos;
1768         struct btree *b;
1769         unsigned i;
1770 restart:
1771         rcu_read_lock();
1772         for_each_cached_btree(b, c, tbl, i, pos)
1773                 if (test_bit(flag, &b->flags)) {
1774                         rcu_read_unlock();
1775                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
1776                         goto restart;
1777
1778                 }
1779         rcu_read_unlock();
1780 }
1781
1782 void bch2_btree_flush_all_reads(struct bch_fs *c)
1783 {
1784         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
1785 }
1786
1787 void bch2_btree_flush_all_writes(struct bch_fs *c)
1788 {
1789         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
1790 }
1791
1792 void bch2_btree_verify_flushed(struct bch_fs *c)
1793 {
1794         struct bucket_table *tbl;
1795         struct rhash_head *pos;
1796         struct btree *b;
1797         unsigned i;
1798
1799         rcu_read_lock();
1800         for_each_cached_btree(b, c, tbl, i, pos) {
1801                 unsigned long flags = READ_ONCE(b->flags);
1802
1803                 BUG_ON((flags & (1 << BTREE_NODE_dirty)) ||
1804                        (flags & (1 << BTREE_NODE_write_in_flight)));
1805         }
1806         rcu_read_unlock();
1807 }
1808
1809 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
1810 {
1811         struct bucket_table *tbl;
1812         struct rhash_head *pos;
1813         struct btree *b;
1814         unsigned i;
1815
1816         rcu_read_lock();
1817         for_each_cached_btree(b, c, tbl, i, pos) {
1818                 unsigned long flags = READ_ONCE(b->flags);
1819
1820                 if (!(flags & (1 << BTREE_NODE_dirty)))
1821                         continue;
1822
1823                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
1824                        b,
1825                        (flags & (1 << BTREE_NODE_dirty)) != 0,
1826                        (flags & (1 << BTREE_NODE_need_write)) != 0,
1827                        b->c.level,
1828                        b->written,
1829                        !list_empty_careful(&b->write_blocked),
1830                        b->will_make_reachable != 0,
1831                        b->will_make_reachable & 1);
1832         }
1833         rcu_read_unlock();
1834 }