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(bch2_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 = bch2_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                        bch2_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_pos_to_text(struct printbuf *out, struct bch_fs *c,
 601                           struct btree *b)
 602 {
 603         pr_buf(out, "%s level %u/%u\n  ",
 604                bch2_btree_ids[b->c.btree_id],
 605                b->c.level,
 606                c->btree_roots[b->c.btree_id].level);
 607         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 608 }
 609
 610 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 611                           struct btree *b, struct bset *i,
 612                           unsigned offset, int write)
 613 {
 614         pr_buf(out, "error validating btree node %sat btree ",
 615                write ? "before write " : "");
 616         btree_pos_to_text(out, c, b);
 617
 618         pr_buf(out, "\n  node offset %u", b->written);
 619         if (i)
 620                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 621 }
 622
 623 enum btree_err_type {
 624         BTREE_ERR_FIXABLE,
 625         BTREE_ERR_WANT_RETRY,
 626         BTREE_ERR_MUST_RETRY,
 627         BTREE_ERR_FATAL,
 628 };
 629
 630 enum btree_validate_ret {
 631         BTREE_RETRY_READ = 64,
 632 };
 633
 634 #define btree_err(type, c, b, i, msg, ...)                              \
 635 ({                                                                      \
 636         __label__ out;                                                  \
 637         char _buf[300];                                                 \
 638         char *buf2 = _buf;                                              \
 639         struct printbuf out = PBUF(_buf);                               \
 640                                                                         \
 641         buf2 = kmalloc(4096, GFP_ATOMIC);                               \
 642         if (buf2)                                                       \
 643                 out = _PBUF(buf2, 4986);                                \
 644                                                                         \
 645         btree_err_msg(&out, c, b, i, b->written, write);                \
 646         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 647                                                                         \
 648         if (type == BTREE_ERR_FIXABLE &&                                \
 649             write == READ &&                                            \
 650             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 651                 mustfix_fsck_err(c, "%s", buf2);                        \
 652                 goto out;                                               \
 653         }                                                               \
 654                                                                         \
 655         switch (write) {                                                \
 656         case READ:                                                      \
 657                 bch_err(c, "%s", buf2);                                 \
 658                                                                         \
 659                 switch (type) {                                         \
 660                 case BTREE_ERR_FIXABLE:                                 \
 661                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 662                         goto fsck_err;                                  \
 663                 case BTREE_ERR_WANT_RETRY:                              \
 664                         if (have_retry) {                               \
 665                                 ret = BTREE_RETRY_READ;                 \
 666                                 goto fsck_err;                          \
 667                         }                                               \
 668                         break;                                          \
 669                 case BTREE_ERR_MUST_RETRY:                              \
 670                         ret = BTREE_RETRY_READ;                         \
 671                         goto fsck_err;                                  \
 672                 case BTREE_ERR_FATAL:                                   \
 673                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 674                         goto fsck_err;                                  \
 675                 }                                                       \
 676                 break;                                                  \
 677         case WRITE:                                                     \
 678                 bch_err(c, "corrupt metadata before write: %s", buf2);  \
 679                                                                         \
 680                 if (bch2_fs_inconsistent(c)) {                          \
 681                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 682                         goto fsck_err;                                  \
 683                 }                                                       \
 684                 break;                                                  \
 685         }                                                               \
 686 out:                                                                    \
 687         if (buf2 != _buf)                                               \
 688                 kfree(buf2);                                            \
 689         true;                                                           \
 690 })
 691
 692 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 693
 694 static int validate_bset(struct bch_fs *c, struct btree *b,
 695                          struct bset *i, unsigned sectors,
 696                          int write, bool have_retry)
 697 {
 698         unsigned version = le16_to_cpu(i->version);
 699         const char *err;
 700         int ret = 0;
 701
 702         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 703                       version < bcachefs_metadata_version_min) ||
 704                      version >= bcachefs_metadata_version_max,
 705                      BTREE_ERR_FATAL, c, b, i,
 706                      "unsupported bset version");
 707
 708         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 709                          BTREE_ERR_FIXABLE, c, b, i,
 710                          "bset past end of btree node")) {
 711                 i->u64s = 0;
 712                 return 0;
 713         }
 714
 715         btree_err_on(b->written && !i->u64s,
 716                      BTREE_ERR_FIXABLE, c, b, i,
 717                      "empty bset");
 718
 719         if (!b->written) {
 720                 struct btree_node *bn =
 721                         container_of(i, struct btree_node, keys);
 722                 /* These indicate that we read the wrong btree node: */
 723
 724                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 725                         struct bch_btree_ptr_v2 *bp =
 726                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 727
 728                         /* XXX endianness */
 729                         btree_err_on(bp->seq != bn->keys.seq,
 730                                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 731                                      "incorrect sequence number (wrong btree node)");
 732                 }
 733
 734                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 735                              BTREE_ERR_MUST_RETRY, c, b, i,
 736                              "incorrect btree id");
 737
 738                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 739                              BTREE_ERR_MUST_RETRY, c, b, i,
 740                              "incorrect level");
 741
 742                 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
 743                         u64 *p = (u64 *) &bn->ptr;
 744
 745                         *p = swab64(*p);
 746                 }
 747
 748                 if (!write)
 749                         compat_btree_node(b->c.level, b->c.btree_id, version,
 750                                           BSET_BIG_ENDIAN(i), write, bn);
 751
 752                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 753                         struct bch_btree_ptr_v2 *bp =
 754                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 755
 756                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 757                                 b->data->min_key = bp->min_key;
 758                                 b->data->max_key = b->key.k.p;
 759                         }
 760
 761                         btree_err_on(bkey_cmp(b->data->min_key, bp->min_key),
 762                                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 763                                      "incorrect min_key: got %llu:%llu should be %llu:%llu",
 764                                      b->data->min_key.inode,
 765                                      b->data->min_key.offset,
 766                                      bp->min_key.inode,
 767                                      bp->min_key.offset);
 768                 }
 769
 770                 btree_err_on(bkey_cmp(bn->max_key, b->key.k.p),
 771                              BTREE_ERR_MUST_RETRY, c, b, i,
 772                              "incorrect max key %llu:%llu",
 773                              bn->max_key.inode,
 774                              bn->max_key.offset);
 775
 776                 if (write)
 777                         compat_btree_node(b->c.level, b->c.btree_id, version,
 778                                           BSET_BIG_ENDIAN(i), write, bn);
 779
 780                 /* XXX: ideally we would be validating min_key too */
 781 #if 0
 782                 /*
 783                  * not correct anymore, due to btree node write error
 784                  * handling
 785                  *
 786                  * need to add bn->seq to btree keys and verify
 787                  * against that
 788                  */
 789                 btree_err_on(!extent_contains_ptr(bkey_i_to_s_c_extent(&b->key),
 790                                                   bn->ptr),
 791                              BTREE_ERR_FATAL, c, b, i,
 792                              "incorrect backpointer");
 793 #endif
 794                 err = bch2_bkey_format_validate(&bn->format);
 795                 btree_err_on(err,
 796                              BTREE_ERR_FATAL, c, b, i,
 797                              "invalid bkey format: %s", err);
 798
 799                 compat_bformat(b->c.level, b->c.btree_id, version,
 800                                BSET_BIG_ENDIAN(i), write,
 801                                &bn->format);
 802         }
 803 fsck_err:
 804         return ret;
 805 }
 806
 807 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 808                          struct bset *i, unsigned *whiteout_u64s,
 809                          int write, bool have_retry)
 810 {
 811         unsigned version = le16_to_cpu(i->version);
 812         struct bkey_packed *k, *prev = NULL;
 813         bool seen_non_whiteout = false;
 814         int ret = 0;
 815
 816         if (!BSET_SEPARATE_WHITEOUTS(i)) {
 817                 seen_non_whiteout = true;
 818                 *whiteout_u64s = 0;
 819         }
 820
 821         for (k = i->start;
 822              k != vstruct_last(i);) {
 823                 struct bkey_s u;
 824                 struct bkey tmp;
 825                 const char *invalid;
 826
 827                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 828                                  BTREE_ERR_FIXABLE, c, b, i,
 829                                  "key extends past end of bset")) {
 830                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 831                         break;
 832                 }
 833
 834                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 835                                  BTREE_ERR_FIXABLE, c, b, i,
 836                                  "invalid bkey format %u", k->format)) {
 837                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 838                         memmove_u64s_down(k, bkey_next(k),
 839                                           (u64 *) vstruct_end(i) - (u64 *) k);
 840                         continue;
 841                 }
 842
 843                 /* XXX: validate k->u64s */
 844                 if (!write)
 845                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 846                                     BSET_BIG_ENDIAN(i), write,
 847                                     &b->format, k);
 848
 849                 u = __bkey_disassemble(b, k, &tmp);
 850
 851                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 852                         bch2_bkey_in_btree_node(b, u.s_c) ?:
 853                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 854                 if (invalid) {
 855                         char buf[160];
 856
 857                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 858                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
 859                                   "invalid bkey: %s\n%s", invalid, buf);
 860
 861                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 862                         memmove_u64s_down(k, bkey_next(k),
 863                                           (u64 *) vstruct_end(i) - (u64 *) k);
 864                         continue;
 865                 }
 866
 867                 if (write)
 868                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 869                                     BSET_BIG_ENDIAN(i), write,
 870                                     &b->format, k);
 871
 872                 /*
 873                  * with the separate whiteouts thing (used for extents), the
 874                  * second set of keys actually can have whiteouts too, so we
 875                  * can't solely go off bkey_whiteout()...
 876                  */
 877
 878                 if (!seen_non_whiteout &&
 879                     (!bkey_whiteout(k) ||
 880                      (prev && bkey_iter_cmp(b, prev, k) > 0))) {
 881                         *whiteout_u64s = k->_data - i->_data;
 882                         seen_non_whiteout = true;
 883                 } else if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 884                         char buf1[80];
 885                         char buf2[80];
 886                         struct bkey up = bkey_unpack_key(b, prev);
 887
 888                         bch2_bkey_to_text(&PBUF(buf1), &up);
 889                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 890
 891                         bch2_dump_bset(c, b, i, 0);
 892                         btree_err(BTREE_ERR_FATAL, c, b, i,
 893                                   "keys out of order: %s > %s",
 894                                   buf1, buf2);
 895                         /* XXX: repair this */
 896                 }
 897
 898                 prev = k;
 899                 k = bkey_next_skip_noops(k, vstruct_last(i));
 900         }
 901 fsck_err:
 902         return ret;
 903 }
 904
 905 int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry)
 906 {
 907         struct btree_node_entry *bne;
 908         struct sort_iter *iter;
 909         struct btree_node *sorted;
 910         struct bkey_packed *k;
 911         struct bch_extent_ptr *ptr;
 912         struct bset *i;
 913         bool used_mempool, blacklisted;
 914         unsigned u64s;
 915         int ret, retry_read = 0, write = READ;
 916
 917         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 918         sort_iter_init(iter, b);
 919         iter->size = (btree_blocks(c) + 1) * 2;
 920
 921         if (bch2_meta_read_fault("btree"))
 922                 btree_err(BTREE_ERR_MUST_RETRY, c, b, NULL,
 923                           "dynamic fault");
 924
 925         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 926                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 927                      "bad magic");
 928
 929         btree_err_on(!b->data->keys.seq,
 930                      BTREE_ERR_MUST_RETRY, c, b, NULL,
 931                      "bad btree header");
 932
 933         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 934                 struct bch_btree_ptr_v2 *bp =
 935                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 936
 937                 btree_err_on(b->data->keys.seq != bp->seq,
 938                              BTREE_ERR_MUST_RETRY, c, b, NULL,
 939                              "got wrong btree node (seq %llx want %llx)",
 940                              b->data->keys.seq, bp->seq);
 941         }
 942
 943         while (b->written < c->opts.btree_node_size) {
 944                 unsigned sectors, whiteout_u64s = 0;
 945                 struct nonce nonce;
 946                 struct bch_csum csum;
 947                 bool first = !b->written;
 948
 949                 if (!b->written) {
 950                         i = &b->data->keys;
 951
 952                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 953                                      BTREE_ERR_WANT_RETRY, c, b, i,
 954                                      "unknown checksum type %llu",
 955                                      BSET_CSUM_TYPE(i));
 956
 957                         nonce = btree_nonce(i, b->written << 9);
 958                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 959
 960                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 961                                      BTREE_ERR_WANT_RETRY, c, b, i,
 962                                      "invalid checksum");
 963
 964                         bset_encrypt(c, i, b->written << 9);
 965
 966                         if (btree_node_is_extents(b) &&
 967                             !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data)) {
 968                                 set_btree_node_old_extent_overwrite(b);
 969                                 set_btree_node_need_rewrite(b);
 970                         }
 971
 972                         sectors = vstruct_sectors(b->data, c->block_bits);
 973                 } else {
 974                         bne = write_block(b);
 975                         i = &bne->keys;
 976
 977                         if (i->seq != b->data->keys.seq)
 978                                 break;
 979
 980                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 981                                      BTREE_ERR_WANT_RETRY, c, b, i,
 982                                      "unknown checksum type %llu",
 983                                      BSET_CSUM_TYPE(i));
 984
 985                         nonce = btree_nonce(i, b->written << 9);
 986                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 987
 988                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 989                                      BTREE_ERR_WANT_RETRY, c, b, i,
 990                                      "invalid checksum");
 991
 992                         bset_encrypt(c, i, b->written << 9);
 993
 994                         sectors = vstruct_sectors(bne, c->block_bits);
 995                 }
 996
 997                 ret = validate_bset(c, b, i, sectors,
 998                                     READ, have_retry);
 999                 if (ret)
1000                         goto fsck_err;
1001
1002                 if (!b->written)
1003                         btree_node_set_format(b, b->data->format);
1004
1005                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
1006                                     READ, have_retry);
1007                 if (ret)
1008                         goto fsck_err;
1009
1010                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1011
1012                 b->written += sectors;
1013
1014                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1015                                         le64_to_cpu(i->journal_seq),
1016                                         true);
1017
1018                 btree_err_on(blacklisted && first,
1019                              BTREE_ERR_FIXABLE, c, b, i,
1020                              "first btree node bset has blacklisted journal seq");
1021                 if (blacklisted && !first)
1022                         continue;
1023
1024                 sort_iter_add(iter, i->start,
1025                               vstruct_idx(i, whiteout_u64s));
1026
1027                 sort_iter_add(iter,
1028                               vstruct_idx(i, whiteout_u64s),
1029                               vstruct_last(i));
1030         }
1031
1032         for (bne = write_block(b);
1033              bset_byte_offset(b, bne) < btree_bytes(c);
1034              bne = (void *) bne + block_bytes(c))
1035                 btree_err_on(bne->keys.seq == b->data->keys.seq,
1036                              BTREE_ERR_WANT_RETRY, c, b, NULL,
1037                              "found bset signature after last bset");
1038
1039         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1040         sorted->keys.u64s = 0;
1041
1042         set_btree_bset(b, b->set, &b->data->keys);
1043
1044         b->nr = (btree_node_old_extent_overwrite(b)
1045                  ? bch2_extent_sort_fix_overlapping
1046                  : bch2_key_sort_fix_overlapping)(c, &sorted->keys, iter);
1047
1048         u64s = le16_to_cpu(sorted->keys.u64s);
1049         *sorted = *b->data;
1050         sorted->keys.u64s = cpu_to_le16(u64s);
1051         swap(sorted, b->data);
1052         set_btree_bset(b, b->set, &b->data->keys);
1053         b->nsets = 1;
1054
1055         BUG_ON(b->nr.live_u64s != u64s);
1056
1057         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1058
1059         i = &b->data->keys;
1060         for (k = i->start; k != vstruct_last(i);) {
1061                 struct bkey tmp;
1062                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1063                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
1064
1065                 if (invalid ||
1066                     (bch2_inject_invalid_keys &&
1067                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1068                         char buf[160];
1069
1070                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
1071                         btree_err(BTREE_ERR_FIXABLE, c, b, i,
1072                                   "invalid bkey %s: %s", buf, invalid);
1073
1074                         btree_keys_account_key_drop(&b->nr, 0, k);
1075
1076                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1077                         memmove_u64s_down(k, bkey_next(k),
1078                                           (u64 *) vstruct_end(i) - (u64 *) k);
1079                         set_btree_bset_end(b, b->set);
1080                         continue;
1081                 }
1082
1083                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1084                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1085
1086                         bp.v->mem_ptr = 0;
1087                 }
1088
1089                 k = bkey_next_skip_noops(k, vstruct_last(i));
1090         }
1091
1092         bch2_bset_build_aux_tree(b, b->set, false);
1093
1094         set_needs_whiteout(btree_bset_first(b), true);
1095
1096         btree_node_reset_sib_u64s(b);
1097
1098         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1099                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1100
1101                 if (ca->mi.state != BCH_MEMBER_STATE_RW)
1102                         set_btree_node_need_rewrite(b);
1103         }
1104 out:
1105         mempool_free(iter, &c->fill_iter);
1106         return retry_read;
1107 fsck_err:
1108         if (ret == BTREE_RETRY_READ) {
1109                 retry_read = 1;
1110         } else {
1111                 bch2_inconsistent_error(c);
1112                 set_btree_node_read_error(b);
1113         }
1114         goto out;
1115 }
1116
1117 static void btree_node_read_work(struct work_struct *work)
1118 {
1119         struct btree_read_bio *rb =
1120                 container_of(work, struct btree_read_bio, work);
1121         struct bch_fs *c        = rb->c;
1122         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1123         struct btree *b         = rb->bio.bi_private;
1124         struct bio *bio         = &rb->bio;
1125         struct bch_io_failures failed = { .nr = 0 };
1126         char buf[200];
1127         struct printbuf out;
1128         bool can_retry;
1129
1130         goto start;
1131         while (1) {
1132                 bch_info(c, "retrying read");
1133                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1134                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1135                 bio_reset(bio);
1136                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1137                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1138                 bio->bi_iter.bi_size    = btree_bytes(c);
1139
1140                 if (rb->have_ioref) {
1141                         bio_set_dev(bio, ca->disk_sb.bdev);
1142                         submit_bio_wait(bio);
1143                 } else {
1144                         bio->bi_status = BLK_STS_REMOVED;
1145                 }
1146 start:
1147                 out = PBUF(buf);
1148                 btree_pos_to_text(&out, c, b);
1149                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1150                                    bch2_blk_status_to_str(bio->bi_status), buf);
1151                 if (rb->have_ioref)
1152                         percpu_ref_put(&ca->io_ref);
1153                 rb->have_ioref = false;
1154
1155                 bch2_mark_io_failure(&failed, &rb->pick);
1156
1157                 can_retry = bch2_bkey_pick_read_device(c,
1158                                 bkey_i_to_s_c(&b->key),
1159                                 &failed, &rb->pick) > 0;
1160
1161                 if (!bio->bi_status &&
1162                     !bch2_btree_node_read_done(c, b, can_retry))
1163                         break;
1164
1165                 if (!can_retry) {
1166                         set_btree_node_read_error(b);
1167                         break;
1168                 }
1169         }
1170
1171         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1172                                rb->start_time);
1173         bio_put(&rb->bio);
1174         clear_btree_node_read_in_flight(b);
1175         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1176 }
1177
1178 static void btree_node_read_endio(struct bio *bio)
1179 {
1180         struct btree_read_bio *rb =
1181                 container_of(bio, struct btree_read_bio, bio);
1182         struct bch_fs *c        = rb->c;
1183
1184         if (rb->have_ioref) {
1185                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1186                 bch2_latency_acct(ca, rb->start_time, READ);
1187         }
1188
1189         queue_work(system_unbound_wq, &rb->work);
1190 }
1191
1192 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1193                           bool sync)
1194 {
1195         struct extent_ptr_decoded pick;
1196         struct btree_read_bio *rb;
1197         struct bch_dev *ca;
1198         struct bio *bio;
1199         int ret;
1200
1201         trace_btree_read(c, b);
1202
1203         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1204                                          NULL, &pick);
1205         if (bch2_fs_fatal_err_on(ret <= 0, c,
1206                         "btree node read error: no device to read from")) {
1207                 set_btree_node_read_error(b);
1208                 return;
1209         }
1210
1211         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1212
1213         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1214                                                    btree_bytes(c)),
1215                                &c->btree_bio);
1216         rb = container_of(bio, struct btree_read_bio, bio);
1217         rb->c                   = c;
1218         rb->start_time          = local_clock();
1219         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1220         rb->pick                = pick;
1221         INIT_WORK(&rb->work, btree_node_read_work);
1222         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1223         bio->bi_iter.bi_sector  = pick.ptr.offset;
1224         bio->bi_end_io          = btree_node_read_endio;
1225         bio->bi_private         = b;
1226         bch2_bio_map(bio, b->data, btree_bytes(c));
1227
1228         set_btree_node_read_in_flight(b);
1229
1230         if (rb->have_ioref) {
1231                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1232                              bio_sectors(bio));
1233                 bio_set_dev(bio, ca->disk_sb.bdev);
1234
1235                 if (sync) {
1236                         submit_bio_wait(bio);
1237
1238                         bio->bi_private = b;
1239                         btree_node_read_work(&rb->work);
1240                 } else {
1241                         submit_bio(bio);
1242                 }
1243         } else {
1244                 bio->bi_status = BLK_STS_REMOVED;
1245
1246                 if (sync)
1247                         btree_node_read_work(&rb->work);
1248                 else
1249                         queue_work(system_unbound_wq, &rb->work);
1250
1251         }
1252 }
1253
1254 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1255                         const struct bkey_i *k, unsigned level)
1256 {
1257         struct closure cl;
1258         struct btree *b;
1259         int ret;
1260
1261         closure_init_stack(&cl);
1262
1263         do {
1264                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1265                 closure_sync(&cl);
1266         } while (ret);
1267
1268         b = bch2_btree_node_mem_alloc(c);
1269         bch2_btree_cache_cannibalize_unlock(c);
1270
1271         BUG_ON(IS_ERR(b));
1272
1273         bkey_copy(&b->key, k);
1274         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1275
1276         bch2_btree_node_read(c, b, true);
1277
1278         if (btree_node_read_error(b)) {
1279                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1280
1281                 mutex_lock(&c->btree_cache.lock);
1282                 list_move(&b->list, &c->btree_cache.freeable);
1283                 mutex_unlock(&c->btree_cache.lock);
1284
1285                 ret = -EIO;
1286                 goto err;
1287         }
1288
1289         bch2_btree_set_root_for_read(c, b);
1290 err:
1291         six_unlock_write(&b->c.lock);
1292         six_unlock_intent(&b->c.lock);
1293
1294         return ret;
1295 }
1296
1297 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1298                               struct btree_write *w)
1299 {
1300         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1301
1302         do {
1303                 old = new = v;
1304                 if (!(old & 1))
1305                         break;
1306
1307                 new &= ~1UL;
1308         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1309
1310         if (old & 1)
1311                 closure_put(&((struct btree_update *) new)->cl);
1312
1313         bch2_journal_pin_drop(&c->journal, &w->journal);
1314 }
1315
1316 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1317 {
1318         struct btree_write *w = btree_prev_write(b);
1319
1320         bch2_btree_complete_write(c, b, w);
1321         btree_node_io_unlock(b);
1322 }
1323
1324 static void bch2_btree_node_write_error(struct bch_fs *c,
1325                                         struct btree_write_bio *wbio)
1326 {
1327         struct btree *b         = wbio->wbio.bio.bi_private;
1328         struct bkey_buf k;
1329         struct bch_extent_ptr *ptr;
1330         struct btree_trans trans;
1331         struct btree_iter *iter;
1332         int ret;
1333
1334         bch2_bkey_buf_init(&k);
1335         bch2_trans_init(&trans, c, 0, 0);
1336
1337         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1338                                         BTREE_MAX_DEPTH, b->c.level, 0);
1339 retry:
1340         ret = bch2_btree_iter_traverse(iter);
1341         if (ret)
1342                 goto err;
1343
1344         /* has node been freed? */
1345         if (iter->l[b->c.level].b != b) {
1346                 /* node has been freed: */
1347                 BUG_ON(!btree_node_dying(b));
1348                 goto out;
1349         }
1350
1351         BUG_ON(!btree_node_hashed(b));
1352
1353         bch2_bkey_buf_copy(&k, c, &b->key);
1354
1355         bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1356                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1357
1358         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1359                 goto err;
1360
1361         ret = bch2_btree_node_update_key(c, iter, b, k.k);
1362         if (ret == -EINTR)
1363                 goto retry;
1364         if (ret)
1365                 goto err;
1366 out:
1367         bch2_trans_exit(&trans);
1368         bch2_bkey_buf_exit(&k, c);
1369         bio_put(&wbio->wbio.bio);
1370         btree_node_write_done(c, b);
1371         return;
1372 err:
1373         set_btree_node_noevict(b);
1374         bch2_fs_fatal_error(c, "fatal error writing btree node");
1375         goto out;
1376 }
1377
1378 void bch2_btree_write_error_work(struct work_struct *work)
1379 {
1380         struct bch_fs *c = container_of(work, struct bch_fs,
1381                                         btree_write_error_work);
1382         struct bio *bio;
1383
1384         while (1) {
1385                 spin_lock_irq(&c->btree_write_error_lock);
1386                 bio = bio_list_pop(&c->btree_write_error_list);
1387                 spin_unlock_irq(&c->btree_write_error_lock);
1388
1389                 if (!bio)
1390                         break;
1391
1392                 bch2_btree_node_write_error(c,
1393                         container_of(bio, struct btree_write_bio, wbio.bio));
1394         }
1395 }
1396
1397 static void btree_node_write_work(struct work_struct *work)
1398 {
1399         struct btree_write_bio *wbio =
1400                 container_of(work, struct btree_write_bio, work);
1401         struct bch_fs *c        = wbio->wbio.c;
1402         struct btree *b         = wbio->wbio.bio.bi_private;
1403
1404         btree_bounce_free(c,
1405                 wbio->bytes,
1406                 wbio->wbio.used_mempool,
1407                 wbio->data);
1408
1409         if (wbio->wbio.failed.nr) {
1410                 unsigned long flags;
1411
1412                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1413                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1414                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1415
1416                 queue_work(c->wq, &c->btree_write_error_work);
1417                 return;
1418         }
1419
1420         bio_put(&wbio->wbio.bio);
1421         btree_node_write_done(c, b);
1422 }
1423
1424 static void btree_node_write_endio(struct bio *bio)
1425 {
1426         struct bch_write_bio *wbio      = to_wbio(bio);
1427         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1428         struct bch_write_bio *orig      = parent ?: wbio;
1429         struct bch_fs *c                = wbio->c;
1430         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1431         unsigned long flags;
1432
1433         if (wbio->have_ioref)
1434                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1435
1436         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1437                                bch2_blk_status_to_str(bio->bi_status)) ||
1438             bch2_meta_write_fault("btree")) {
1439                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1440                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1441                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1442         }
1443
1444         if (wbio->have_ioref)
1445                 percpu_ref_put(&ca->io_ref);
1446
1447         if (parent) {
1448                 bio_put(bio);
1449                 bio_endio(&parent->bio);
1450         } else {
1451                 struct btree_write_bio *wb =
1452                         container_of(orig, struct btree_write_bio, wbio);
1453
1454                 INIT_WORK(&wb->work, btree_node_write_work);
1455                 queue_work(system_unbound_wq, &wb->work);
1456         }
1457 }
1458
1459 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1460                                    struct bset *i, unsigned sectors)
1461 {
1462         unsigned whiteout_u64s = 0;
1463         int ret;
1464
1465         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_BTREE))
1466                 return -1;
1467
1468         ret = validate_bset(c, b, i, sectors, WRITE, false) ?:
1469                 validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false);
1470         if (ret) {
1471                 bch2_inconsistent_error(c);
1472                 dump_stack();
1473         }
1474
1475         return ret;
1476 }
1477
1478 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1479                             enum six_lock_type lock_type_held)
1480 {
1481         struct btree_write_bio *wbio;
1482         struct bset_tree *t;
1483         struct bset *i;
1484         struct btree_node *bn = NULL;
1485         struct btree_node_entry *bne = NULL;
1486         struct bkey_buf k;
1487         struct bch_extent_ptr *ptr;
1488         struct sort_iter sort_iter;
1489         struct nonce nonce;
1490         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1491         u64 seq = 0;
1492         bool used_mempool;
1493         unsigned long old, new;
1494         bool validate_before_checksum = false;
1495         void *data;
1496
1497         bch2_bkey_buf_init(&k);
1498
1499         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1500                 return;
1501
1502         /*
1503          * We may only have a read lock on the btree node - the dirty bit is our
1504          * "lock" against racing with other threads that may be trying to start
1505          * a write, we do a write iff we clear the dirty bit. Since setting the
1506          * dirty bit requires a write lock, we can't race with other threads
1507          * redirtying it:
1508          */
1509         do {
1510                 old = new = READ_ONCE(b->flags);
1511
1512                 if (!(old & (1 << BTREE_NODE_dirty)))
1513                         return;
1514
1515                 if (!btree_node_may_write(b))
1516                         return;
1517
1518                 if (old & (1 << BTREE_NODE_never_write))
1519                         return;
1520
1521                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1522                         btree_node_wait_on_io(b);
1523                         continue;
1524                 }
1525
1526                 new &= ~(1 << BTREE_NODE_dirty);
1527                 new &= ~(1 << BTREE_NODE_need_write);
1528                 new |=  (1 << BTREE_NODE_write_in_flight);
1529                 new |=  (1 << BTREE_NODE_just_written);
1530                 new ^=  (1 << BTREE_NODE_write_idx);
1531         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1532
1533         atomic_dec(&c->btree_cache.dirty);
1534
1535         BUG_ON(btree_node_fake(b));
1536         BUG_ON((b->will_make_reachable != 0) != !b->written);
1537
1538         BUG_ON(b->written >= c->opts.btree_node_size);
1539         BUG_ON(b->written & (c->opts.block_size - 1));
1540         BUG_ON(bset_written(b, btree_bset_last(b)));
1541         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1542         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1543
1544         bch2_sort_whiteouts(c, b);
1545
1546         sort_iter_init(&sort_iter, b);
1547
1548         bytes = !b->written
1549                 ? sizeof(struct btree_node)
1550                 : sizeof(struct btree_node_entry);
1551
1552         bytes += b->whiteout_u64s * sizeof(u64);
1553
1554         for_each_bset(b, t) {
1555                 i = bset(b, t);
1556
1557                 if (bset_written(b, i))
1558                         continue;
1559
1560                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1561                 sort_iter_add(&sort_iter,
1562                               btree_bkey_first(b, t),
1563                               btree_bkey_last(b, t));
1564                 seq = max(seq, le64_to_cpu(i->journal_seq));
1565         }
1566
1567         BUG_ON(b->written && !seq);
1568
1569         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1570         bytes += 8;
1571
1572         data = btree_bounce_alloc(c, bytes, &used_mempool);
1573
1574         if (!b->written) {
1575                 bn = data;
1576                 *bn = *b->data;
1577                 i = &bn->keys;
1578         } else {
1579                 bne = data;
1580                 bne->keys = b->data->keys;
1581                 i = &bne->keys;
1582         }
1583
1584         i->journal_seq  = cpu_to_le64(seq);
1585         i->u64s         = 0;
1586
1587         if (!btree_node_old_extent_overwrite(b)) {
1588                 sort_iter_add(&sort_iter,
1589                               unwritten_whiteouts_start(c, b),
1590                               unwritten_whiteouts_end(c, b));
1591                 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1592         } else {
1593                 memcpy_u64s(i->start,
1594                             unwritten_whiteouts_start(c, b),
1595                             b->whiteout_u64s);
1596                 i->u64s = cpu_to_le16(b->whiteout_u64s);
1597                 SET_BSET_SEPARATE_WHITEOUTS(i, true);
1598         }
1599
1600         b->whiteout_u64s = 0;
1601
1602         u64s = btree_node_old_extent_overwrite(b)
1603                 ? bch2_sort_extents(vstruct_last(i), &sort_iter, false)
1604                 : bch2_sort_keys(i->start, &sort_iter, false);
1605         le16_add_cpu(&i->u64s, u64s);
1606
1607         set_needs_whiteout(i, false);
1608
1609         /* do we have data to write? */
1610         if (b->written && !i->u64s)
1611                 goto nowrite;
1612
1613         bytes_to_write = vstruct_end(i) - data;
1614         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1615
1616         memset(data + bytes_to_write, 0,
1617                (sectors_to_write << 9) - bytes_to_write);
1618
1619         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1620         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1621         BUG_ON(i->seq != b->data->keys.seq);
1622
1623         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1624                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1625                 : cpu_to_le16(c->sb.version);
1626         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1627
1628         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1629                 validate_before_checksum = true;
1630
1631         /* validate_bset will be modifying: */
1632         if (le16_to_cpu(i->version) <= bcachefs_metadata_version_inode_btree_change)
1633                 validate_before_checksum = true;
1634
1635         /* if we're going to be encrypting, check metadata validity first: */
1636         if (validate_before_checksum &&
1637             validate_bset_for_write(c, b, i, sectors_to_write))
1638                 goto err;
1639
1640         bset_encrypt(c, i, b->written << 9);
1641
1642         nonce = btree_nonce(i, b->written << 9);
1643
1644         if (bn)
1645                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1646         else
1647                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1648
1649         /* if we're not encrypting, check metadata after checksumming: */
1650         if (!validate_before_checksum &&
1651             validate_bset_for_write(c, b, i, sectors_to_write))
1652                 goto err;
1653
1654         /*
1655          * We handle btree write errors by immediately halting the journal -
1656          * after we've done that, we can't issue any subsequent btree writes
1657          * because they might have pointers to new nodes that failed to write.
1658          *
1659          * Furthermore, there's no point in doing any more btree writes because
1660          * with the journal stopped, we're never going to update the journal to
1661          * reflect that those writes were done and the data flushed from the
1662          * journal:
1663          *
1664          * Also on journal error, the pending write may have updates that were
1665          * never journalled (interior nodes, see btree_update_nodes_written()) -
1666          * it's critical that we don't do the write in that case otherwise we
1667          * will have updates visible that weren't in the journal:
1668          *
1669          * Make sure to update b->written so bch2_btree_init_next() doesn't
1670          * break:
1671          */
1672         if (bch2_journal_error(&c->journal) ||
1673             c->opts.nochanges)
1674                 goto err;
1675
1676         trace_btree_write(b, bytes_to_write, sectors_to_write);
1677
1678         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1679                                 buf_pages(data, sectors_to_write << 9),
1680                                 &c->btree_bio),
1681                             struct btree_write_bio, wbio.bio);
1682         wbio_init(&wbio->wbio.bio);
1683         wbio->data                      = data;
1684         wbio->bytes                     = bytes;
1685         wbio->wbio.used_mempool         = used_mempool;
1686         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1687         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1688         wbio->wbio.bio.bi_private       = b;
1689
1690         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1691
1692         /*
1693          * If we're appending to a leaf node, we don't technically need FUA -
1694          * this write just needs to be persisted before the next journal write,
1695          * which will be marked FLUSH|FUA.
1696          *
1697          * Similarly if we're writing a new btree root - the pointer is going to
1698          * be in the next journal entry.
1699          *
1700          * But if we're writing a new btree node (that isn't a root) or
1701          * appending to a non leaf btree node, we need either FUA or a flush
1702          * when we write the parent with the new pointer. FUA is cheaper than a
1703          * flush, and writes appending to leaf nodes aren't blocking anything so
1704          * just make all btree node writes FUA to keep things sane.
1705          */
1706
1707         bch2_bkey_buf_copy(&k, c, &b->key);
1708
1709         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(k.k)), ptr)
1710                 ptr->offset += b->written;
1711
1712         b->written += sectors_to_write;
1713
1714         /* XXX: submitting IO with btree locks held: */
1715         bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_btree, k.k);
1716         bch2_bkey_buf_exit(&k, c);
1717         return;
1718 err:
1719         set_btree_node_noevict(b);
1720         b->written += sectors_to_write;
1721 nowrite:
1722         btree_bounce_free(c, bytes, used_mempool, data);
1723         btree_node_write_done(c, b);
1724 }
1725
1726 /*
1727  * Work that must be done with write lock held:
1728  */
1729 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1730 {
1731         bool invalidated_iter = false;
1732         struct btree_node_entry *bne;
1733         struct bset_tree *t;
1734
1735         if (!btree_node_just_written(b))
1736                 return false;
1737
1738         BUG_ON(b->whiteout_u64s);
1739
1740         clear_btree_node_just_written(b);
1741
1742         /*
1743          * Note: immediately after write, bset_written() doesn't work - the
1744          * amount of data we had to write after compaction might have been
1745          * smaller than the offset of the last bset.
1746          *
1747          * However, we know that all bsets have been written here, as long as
1748          * we're still holding the write lock:
1749          */
1750
1751         /*
1752          * XXX: decide if we really want to unconditionally sort down to a
1753          * single bset:
1754          */
1755         if (b->nsets > 1) {
1756                 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1757                 invalidated_iter = true;
1758         } else {
1759                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1760         }
1761
1762         for_each_bset(b, t)
1763                 set_needs_whiteout(bset(b, t), true);
1764
1765         bch2_btree_verify(c, b);
1766
1767         /*
1768          * If later we don't unconditionally sort down to a single bset, we have
1769          * to ensure this is still true:
1770          */
1771         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1772
1773         bne = want_new_bset(c, b);
1774         if (bne)
1775                 bch2_bset_init_next(c, b, bne);
1776
1777         bch2_btree_build_aux_trees(b);
1778
1779         return invalidated_iter;
1780 }
1781
1782 /*
1783  * Use this one if the node is intent locked:
1784  */
1785 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1786                           enum six_lock_type lock_type_held)
1787 {
1788         BUG_ON(lock_type_held == SIX_LOCK_write);
1789
1790         if (lock_type_held == SIX_LOCK_intent ||
1791             six_lock_tryupgrade(&b->c.lock)) {
1792                 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1793
1794                 /* don't cycle lock unnecessarily: */
1795                 if (btree_node_just_written(b) &&
1796                     six_trylock_write(&b->c.lock)) {
1797                         bch2_btree_post_write_cleanup(c, b);
1798                         six_unlock_write(&b->c.lock);
1799                 }
1800
1801                 if (lock_type_held == SIX_LOCK_read)
1802                         six_lock_downgrade(&b->c.lock);
1803         } else {
1804                 __bch2_btree_node_write(c, b, SIX_LOCK_read);
1805         }
1806 }
1807
1808 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
1809 {
1810         struct bucket_table *tbl;
1811         struct rhash_head *pos;
1812         struct btree *b;
1813         unsigned i;
1814 restart:
1815         rcu_read_lock();
1816         for_each_cached_btree(b, c, tbl, i, pos)
1817                 if (test_bit(flag, &b->flags)) {
1818                         rcu_read_unlock();
1819                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
1820                         goto restart;
1821
1822                 }
1823         rcu_read_unlock();
1824 }
1825
1826 void bch2_btree_flush_all_reads(struct bch_fs *c)
1827 {
1828         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
1829 }
1830
1831 void bch2_btree_flush_all_writes(struct bch_fs *c)
1832 {
1833         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
1834 }
1835
1836 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
1837 {
1838         struct bucket_table *tbl;
1839         struct rhash_head *pos;
1840         struct btree *b;
1841         unsigned i;
1842
1843         rcu_read_lock();
1844         for_each_cached_btree(b, c, tbl, i, pos) {
1845                 unsigned long flags = READ_ONCE(b->flags);
1846
1847                 if (!(flags & (1 << BTREE_NODE_dirty)))
1848                         continue;
1849
1850                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
1851                        b,
1852                        (flags & (1 << BTREE_NODE_dirty)) != 0,
1853                        (flags & (1 << BTREE_NODE_need_write)) != 0,
1854                        b->c.level,
1855                        b->written,
1856                        !list_empty_careful(&b->write_blocked),
1857                        b->will_make_reachable != 0,
1858                        b->will_make_reachable & 1);
1859         }
1860         rcu_read_unlock();
1861 }