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