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_write.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_seq_blacklist.h"
  20 #include "recovery.h"
  21 #include "super-io.h"
  22 #include "trace.h"
  23
  24 #include <linux/sched/mm.h>
  25
  26 void bch2_btree_node_io_unlock(struct btree *b)
  27 {
  28         EBUG_ON(!btree_node_write_in_flight(b));
  29
  30         clear_btree_node_write_in_flight_inner(b);
  31         clear_btree_node_write_in_flight(b);
  32         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
  33 }
  34
  35 void bch2_btree_node_io_lock(struct btree *b)
  36 {
  37         bch2_assert_btree_nodes_not_locked();
  38
  39         wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
  40                             TASK_UNINTERRUPTIBLE);
  41 }
  42
  43 void __bch2_btree_node_wait_on_read(struct btree *b)
  44 {
  45         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  46                        TASK_UNINTERRUPTIBLE);
  47 }
  48
  49 void __bch2_btree_node_wait_on_write(struct btree *b)
  50 {
  51         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  52                        TASK_UNINTERRUPTIBLE);
  53 }
  54
  55 void bch2_btree_node_wait_on_read(struct btree *b)
  56 {
  57         bch2_assert_btree_nodes_not_locked();
  58
  59         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  60                        TASK_UNINTERRUPTIBLE);
  61 }
  62
  63 void bch2_btree_node_wait_on_write(struct btree *b)
  64 {
  65         bch2_assert_btree_nodes_not_locked();
  66
  67         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  68                        TASK_UNINTERRUPTIBLE);
  69 }
  70
  71 static void verify_no_dups(struct btree *b,
  72                            struct bkey_packed *start,
  73                            struct bkey_packed *end)
  74 {
  75 #ifdef CONFIG_BCACHEFS_DEBUG
  76         struct bkey_packed *k, *p;
  77
  78         if (start == end)
  79                 return;
  80
  81         for (p = start, k = bkey_p_next(start);
  82              k != end;
  83              p = k, k = bkey_p_next(k)) {
  84                 struct bkey l = bkey_unpack_key(b, p);
  85                 struct bkey r = bkey_unpack_key(b, k);
  86
  87                 BUG_ON(bpos_ge(l.p, bkey_start_pos(&r)));
  88         }
  89 #endif
  90 }
  91
  92 static void set_needs_whiteout(struct bset *i, int v)
  93 {
  94         struct bkey_packed *k;
  95
  96         for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
  97                 k->needs_whiteout = v;
  98 }
  99
 100 static void btree_bounce_free(struct bch_fs *c, size_t size,
 101                               bool used_mempool, void *p)
 102 {
 103         if (used_mempool)
 104                 mempool_free(p, &c->btree_bounce_pool);
 105         else
 106                 vpfree(p, size);
 107 }
 108
 109 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
 110                                 bool *used_mempool)
 111 {
 112         unsigned flags = memalloc_nofs_save();
 113         void *p;
 114
 115         BUG_ON(size > btree_bytes(c));
 116
 117         *used_mempool = false;
 118         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
 119         if (!p) {
 120                 *used_mempool = true;
 121                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
 122         }
 123         memalloc_nofs_restore(flags);
 124         return p;
 125 }
 126
 127 static void sort_bkey_ptrs(const struct btree *bt,
 128                            struct bkey_packed **ptrs, unsigned nr)
 129 {
 130         unsigned n = nr, a = nr / 2, b, c, d;
 131
 132         if (!a)
 133                 return;
 134
 135         /* Heap sort: see lib/sort.c: */
 136         while (1) {
 137                 if (a)
 138                         a--;
 139                 else if (--n)
 140                         swap(ptrs[0], ptrs[n]);
 141                 else
 142                         break;
 143
 144                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 145                         b = bch2_bkey_cmp_packed(bt,
 146                                             ptrs[c],
 147                                             ptrs[d]) >= 0 ? c : d;
 148                 if (d == n)
 149                         b = c;
 150
 151                 while (b != a &&
 152                        bch2_bkey_cmp_packed(bt,
 153                                        ptrs[a],
 154                                        ptrs[b]) >= 0)
 155                         b = (b - 1) / 2;
 156                 c = b;
 157                 while (b != a) {
 158                         b = (b - 1) / 2;
 159                         swap(ptrs[b], ptrs[c]);
 160                 }
 161         }
 162 }
 163
 164 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 165 {
 166         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 167         bool used_mempool = false;
 168         size_t bytes = b->whiteout_u64s * sizeof(u64);
 169
 170         if (!b->whiteout_u64s)
 171                 return;
 172
 173         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 174
 175         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 176
 177         for (k = unwritten_whiteouts_start(c, b);
 178              k != unwritten_whiteouts_end(c, b);
 179              k = bkey_p_next(k))
 180                 *--ptrs = k;
 181
 182         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 183
 184         k = new_whiteouts;
 185
 186         while (ptrs != ptrs_end) {
 187                 bkey_p_copy(k, *ptrs);
 188                 k = bkey_p_next(k);
 189                 ptrs++;
 190         }
 191
 192         verify_no_dups(b, new_whiteouts,
 193                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 194
 195         memcpy_u64s(unwritten_whiteouts_start(c, b),
 196                     new_whiteouts, b->whiteout_u64s);
 197
 198         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 199 }
 200
 201 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 202                                 bool compacting, enum compact_mode mode)
 203 {
 204         if (!bset_dead_u64s(b, t))
 205                 return false;
 206
 207         switch (mode) {
 208         case COMPACT_LAZY:
 209                 return should_compact_bset_lazy(b, t) ||
 210                         (compacting && !bset_written(b, bset(b, t)));
 211         case COMPACT_ALL:
 212                 return true;
 213         default:
 214                 BUG();
 215         }
 216 }
 217
 218 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 219 {
 220         struct bset_tree *t;
 221         bool ret = false;
 222
 223         for_each_bset(b, t) {
 224                 struct bset *i = bset(b, t);
 225                 struct bkey_packed *k, *n, *out, *start, *end;
 226                 struct btree_node_entry *src = NULL, *dst = NULL;
 227
 228                 if (t != b->set && !bset_written(b, i)) {
 229                         src = container_of(i, struct btree_node_entry, keys);
 230                         dst = max(write_block(b),
 231                                   (void *) btree_bkey_last(b, t - 1));
 232                 }
 233
 234                 if (src != dst)
 235                         ret = true;
 236
 237                 if (!should_compact_bset(b, t, ret, mode)) {
 238                         if (src != dst) {
 239                                 memmove(dst, src, sizeof(*src) +
 240                                         le16_to_cpu(src->keys.u64s) *
 241                                         sizeof(u64));
 242                                 i = &dst->keys;
 243                                 set_btree_bset(b, t, i);
 244                         }
 245                         continue;
 246                 }
 247
 248                 start   = btree_bkey_first(b, t);
 249                 end     = btree_bkey_last(b, t);
 250
 251                 if (src != dst) {
 252                         memmove(dst, src, sizeof(*src));
 253                         i = &dst->keys;
 254                         set_btree_bset(b, t, i);
 255                 }
 256
 257                 out = i->start;
 258
 259                 for (k = start; k != end; k = n) {
 260                         n = bkey_p_next(k);
 261
 262                         if (!bkey_deleted(k)) {
 263                                 bkey_p_copy(out, k);
 264                                 out = bkey_p_next(out);
 265                         } else {
 266                                 BUG_ON(k->needs_whiteout);
 267                         }
 268                 }
 269
 270                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 271                 set_btree_bset_end(b, t);
 272                 bch2_bset_set_no_aux_tree(b, t);
 273                 ret = true;
 274         }
 275
 276         bch2_verify_btree_nr_keys(b);
 277
 278         bch2_btree_build_aux_trees(b);
 279
 280         return ret;
 281 }
 282
 283 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 284                             enum compact_mode mode)
 285 {
 286         return bch2_drop_whiteouts(b, mode);
 287 }
 288
 289 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 290                             unsigned start_idx,
 291                             unsigned end_idx,
 292                             bool filter_whiteouts)
 293 {
 294         struct btree_node *out;
 295         struct sort_iter_stack sort_iter;
 296         struct bset_tree *t;
 297         struct bset *start_bset = bset(b, &b->set[start_idx]);
 298         bool used_mempool = false;
 299         u64 start_time, seq = 0;
 300         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 301         bool sorting_entire_node = start_idx == 0 &&
 302                 end_idx == b->nsets;
 303
 304         sort_iter_stack_init(&sort_iter, b);
 305
 306         for (t = b->set + start_idx;
 307              t < b->set + end_idx;
 308              t++) {
 309                 u64s += le16_to_cpu(bset(b, t)->u64s);
 310                 sort_iter_add(&sort_iter.iter,
 311                               btree_bkey_first(b, t),
 312                               btree_bkey_last(b, t));
 313         }
 314
 315         bytes = sorting_entire_node
 316                 ? btree_bytes(c)
 317                 : __vstruct_bytes(struct btree_node, u64s);
 318
 319         out = btree_bounce_alloc(c, bytes, &used_mempool);
 320
 321         start_time = local_clock();
 322
 323         u64s = bch2_sort_keys(out->keys.start, &sort_iter.iter, filter_whiteouts);
 324
 325         out->keys.u64s = cpu_to_le16(u64s);
 326
 327         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 328
 329         if (sorting_entire_node)
 330                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 331                                        start_time);
 332
 333         /* Make sure we preserve bset journal_seq: */
 334         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 335                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 336         start_bset->journal_seq = cpu_to_le64(seq);
 337
 338         if (sorting_entire_node) {
 339                 u64s = le16_to_cpu(out->keys.u64s);
 340
 341                 BUG_ON(bytes != btree_bytes(c));
 342
 343                 /*
 344                  * Our temporary buffer is the same size as the btree node's
 345                  * buffer, we can just swap buffers instead of doing a big
 346                  * memcpy()
 347                  */
 348                 *out = *b->data;
 349                 out->keys.u64s = cpu_to_le16(u64s);
 350                 swap(out, b->data);
 351                 set_btree_bset(b, b->set, &b->data->keys);
 352         } else {
 353                 start_bset->u64s = out->keys.u64s;
 354                 memcpy_u64s(start_bset->start,
 355                             out->keys.start,
 356                             le16_to_cpu(out->keys.u64s));
 357         }
 358
 359         for (i = start_idx + 1; i < end_idx; i++)
 360                 b->nr.bset_u64s[start_idx] +=
 361                         b->nr.bset_u64s[i];
 362
 363         b->nsets -= shift;
 364
 365         for (i = start_idx + 1; i < b->nsets; i++) {
 366                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 367                 b->set[i]               = b->set[i + shift];
 368         }
 369
 370         for (i = b->nsets; i < MAX_BSETS; i++)
 371                 b->nr.bset_u64s[i] = 0;
 372
 373         set_btree_bset_end(b, &b->set[start_idx]);
 374         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 375
 376         btree_bounce_free(c, bytes, used_mempool, out);
 377
 378         bch2_verify_btree_nr_keys(b);
 379 }
 380
 381 void bch2_btree_sort_into(struct bch_fs *c,
 382                          struct btree *dst,
 383                          struct btree *src)
 384 {
 385         struct btree_nr_keys nr;
 386         struct btree_node_iter src_iter;
 387         u64 start_time = local_clock();
 388
 389         BUG_ON(dst->nsets != 1);
 390
 391         bch2_bset_set_no_aux_tree(dst, dst->set);
 392
 393         bch2_btree_node_iter_init_from_start(&src_iter, src);
 394
 395         nr = bch2_sort_repack(btree_bset_first(dst),
 396                         src, &src_iter,
 397                         &dst->format,
 398                         true);
 399
 400         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 401                                start_time);
 402
 403         set_btree_bset_end(dst, dst->set);
 404
 405         dst->nr.live_u64s       += nr.live_u64s;
 406         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 407         dst->nr.packed_keys     += nr.packed_keys;
 408         dst->nr.unpacked_keys   += nr.unpacked_keys;
 409
 410         bch2_verify_btree_nr_keys(dst);
 411 }
 412
 413 /*
 414  * We're about to add another bset to the btree node, so if there's currently
 415  * too many bsets - sort some of them together:
 416  */
 417 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 418 {
 419         unsigned unwritten_idx;
 420         bool ret = false;
 421
 422         for (unwritten_idx = 0;
 423              unwritten_idx < b->nsets;
 424              unwritten_idx++)
 425                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 426                         break;
 427
 428         if (b->nsets - unwritten_idx > 1) {
 429                 btree_node_sort(c, b, unwritten_idx,
 430                                 b->nsets, false);
 431                 ret = true;
 432         }
 433
 434         if (unwritten_idx > 1) {
 435                 btree_node_sort(c, b, 0, unwritten_idx, false);
 436                 ret = true;
 437         }
 438
 439         return ret;
 440 }
 441
 442 void bch2_btree_build_aux_trees(struct btree *b)
 443 {
 444         struct bset_tree *t;
 445
 446         for_each_bset(b, t)
 447                 bch2_bset_build_aux_tree(b, t,
 448                                 !bset_written(b, bset(b, t)) &&
 449                                 t == bset_tree_last(b));
 450 }
 451
 452 /*
 453  * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
 454  *
 455  * The first bset is going to be of similar order to the size of the node, the
 456  * last bset is bounded by btree_write_set_buffer(), which is set to keep the
 457  * memmove on insert from being too expensive: the middle bset should, ideally,
 458  * be the geometric mean of the first and the last.
 459  *
 460  * Returns true if the middle bset is greater than that geometric mean:
 461  */
 462 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
 463 {
 464         unsigned mid_u64s_bits =
 465                 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
 466
 467         return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
 468 }
 469
 470 /*
 471  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 472  * inserted into
 473  *
 474  * Safe to call if there already is an unwritten bset - will only add a new bset
 475  * if @b doesn't already have one.
 476  *
 477  * Returns true if we sorted (i.e. invalidated iterators
 478  */
 479 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 480 {
 481         struct bch_fs *c = trans->c;
 482         struct btree_node_entry *bne;
 483         bool reinit_iter = false;
 484
 485         EBUG_ON(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]);
 486         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 487         BUG_ON(btree_node_just_written(b));
 488
 489         if (b->nsets == MAX_BSETS &&
 490             !btree_node_write_in_flight(b) &&
 491             should_compact_all(c, b)) {
 492                 bch2_btree_node_write(c, b, SIX_LOCK_write,
 493                                       BTREE_WRITE_init_next_bset);
 494                 reinit_iter = true;
 495         }
 496
 497         if (b->nsets == MAX_BSETS &&
 498             btree_node_compact(c, b))
 499                 reinit_iter = true;
 500
 501         BUG_ON(b->nsets >= MAX_BSETS);
 502
 503         bne = want_new_bset(c, b);
 504         if (bne)
 505                 bch2_bset_init_next(c, b, bne);
 506
 507         bch2_btree_build_aux_trees(b);
 508
 509         if (reinit_iter)
 510                 bch2_trans_node_reinit_iter(trans, b);
 511 }
 512
 513 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 514                           struct bch_dev *ca,
 515                           struct btree *b, struct bset *i,
 516                           unsigned offset, int write)
 517 {
 518         prt_printf(out, bch2_log_msg(c, "%s"),
 519                    write == READ
 520                    ? "error validating btree node "
 521                    : "corrupt btree node before write ");
 522         if (ca)
 523                 prt_printf(out, "on %s ", ca->name);
 524         prt_printf(out, "at btree ");
 525         bch2_btree_pos_to_text(out, c, b);
 526
 527         prt_printf(out, "\n  node offset %u", b->written);
 528         if (i)
 529                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 530         prt_str(out, ": ");
 531 }
 532
 533 __printf(9, 10)
 534 static int __btree_err(int ret,
 535                        struct bch_fs *c,
 536                        struct bch_dev *ca,
 537                        struct btree *b,
 538                        struct bset *i,
 539                        int write,
 540                        bool have_retry,
 541                        enum bch_sb_error_id err_type,
 542                        const char *fmt, ...)
 543 {
 544         struct printbuf out = PRINTBUF;
 545         va_list args;
 546
 547         btree_err_msg(&out, c, ca, b, i, b->written, write);
 548
 549         va_start(args, fmt);
 550         prt_vprintf(&out, fmt, args);
 551         va_end(args);
 552
 553         if (write == WRITE) {
 554                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 555                 ret = c->opts.errors == BCH_ON_ERROR_continue
 556                         ? 0
 557                         : -BCH_ERR_fsck_errors_not_fixed;
 558                 goto out;
 559         }
 560
 561         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
 562                 ret = -BCH_ERR_btree_node_read_err_fixable;
 563         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
 564                 ret = -BCH_ERR_btree_node_read_err_bad_node;
 565
 566         if (ret != -BCH_ERR_btree_node_read_err_fixable)
 567                 bch2_sb_error_count(c, err_type);
 568
 569         switch (ret) {
 570         case -BCH_ERR_btree_node_read_err_fixable:
 571                 ret = bch2_fsck_err(c, FSCK_CAN_FIX, err_type, "%s", out.buf);
 572                 if (ret != -BCH_ERR_fsck_fix &&
 573                     ret != -BCH_ERR_fsck_ignore)
 574                         goto fsck_err;
 575                 ret = -BCH_ERR_fsck_fix;
 576                 break;
 577         case -BCH_ERR_btree_node_read_err_want_retry:
 578         case -BCH_ERR_btree_node_read_err_must_retry:
 579                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 580                 break;
 581         case -BCH_ERR_btree_node_read_err_bad_node:
 582                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 583                 bch2_topology_error(c);
 584                 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology) ?: -EIO;
 585                 break;
 586         case -BCH_ERR_btree_node_read_err_incompatible:
 587                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 588                 ret = -BCH_ERR_fsck_errors_not_fixed;
 589                 break;
 590         default:
 591                 BUG();
 592         }
 593 out:
 594 fsck_err:
 595         printbuf_exit(&out);
 596         return ret;
 597 }
 598
 599 #define btree_err(type, c, ca, b, i, _err_type, msg, ...)               \
 600 ({                                                                      \
 601         int _ret = __btree_err(type, c, ca, b, i, write, have_retry,    \
 602                                BCH_FSCK_ERR_##_err_type,                \
 603                                msg, ##__VA_ARGS__);                     \
 604                                                                         \
 605         if (_ret != -BCH_ERR_fsck_fix) {                                \
 606                 ret = _ret;                                             \
 607                 goto fsck_err;                                          \
 608         }                                                               \
 609                                                                         \
 610         *saw_error = true;                                              \
 611 })
 612
 613 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 614
 615 /*
 616  * When btree topology repair changes the start or end of a node, that might
 617  * mean we have to drop keys that are no longer inside the node:
 618  */
 619 __cold
 620 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 621 {
 622         struct bset_tree *t;
 623
 624         for_each_bset(b, t) {
 625                 struct bset *i = bset(b, t);
 626                 struct bkey_packed *k;
 627
 628                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 629                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 630                                 break;
 631
 632                 if (k != i->start) {
 633                         unsigned shift = (u64 *) k - (u64 *) i->start;
 634
 635                         memmove_u64s_down(i->start, k,
 636                                           (u64 *) vstruct_end(i) - (u64 *) k);
 637                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 638                         set_btree_bset_end(b, t);
 639                 }
 640
 641                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 642                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 643                                 break;
 644
 645                 if (k != vstruct_last(i)) {
 646                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 647                         set_btree_bset_end(b, t);
 648                 }
 649         }
 650
 651         /*
 652          * Always rebuild search trees: eytzinger search tree nodes directly
 653          * depend on the values of min/max key:
 654          */
 655         bch2_bset_set_no_aux_tree(b, b->set);
 656         bch2_btree_build_aux_trees(b);
 657
 658         struct bkey_s_c k;
 659         struct bkey unpacked;
 660         struct btree_node_iter iter;
 661         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 662                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 663                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 664         }
 665 }
 666
 667 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 668                          struct btree *b, struct bset *i,
 669                          unsigned offset, unsigned sectors,
 670                          int write, bool have_retry, bool *saw_error)
 671 {
 672         unsigned version = le16_to_cpu(i->version);
 673         struct printbuf buf1 = PRINTBUF;
 674         struct printbuf buf2 = PRINTBUF;
 675         int ret = 0;
 676
 677         btree_err_on(!bch2_version_compatible(version),
 678                      -BCH_ERR_btree_node_read_err_incompatible,
 679                      c, ca, b, i,
 680                      btree_node_unsupported_version,
 681                      "unsupported bset version %u.%u",
 682                      BCH_VERSION_MAJOR(version),
 683                      BCH_VERSION_MINOR(version));
 684
 685         if (btree_err_on(version < c->sb.version_min,
 686                          -BCH_ERR_btree_node_read_err_fixable,
 687                          c, NULL, b, i,
 688                          btree_node_bset_older_than_sb_min,
 689                          "bset version %u older than superblock version_min %u",
 690                          version, c->sb.version_min)) {
 691                 mutex_lock(&c->sb_lock);
 692                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 693                 bch2_write_super(c);
 694                 mutex_unlock(&c->sb_lock);
 695         }
 696
 697         if (btree_err_on(BCH_VERSION_MAJOR(version) >
 698                          BCH_VERSION_MAJOR(c->sb.version),
 699                          -BCH_ERR_btree_node_read_err_fixable,
 700                          c, NULL, b, i,
 701                          btree_node_bset_newer_than_sb,
 702                          "bset version %u newer than superblock version %u",
 703                          version, c->sb.version)) {
 704                 mutex_lock(&c->sb_lock);
 705                 c->disk_sb.sb->version = cpu_to_le16(version);
 706                 bch2_write_super(c);
 707                 mutex_unlock(&c->sb_lock);
 708         }
 709
 710         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 711                      -BCH_ERR_btree_node_read_err_incompatible,
 712                      c, ca, b, i,
 713                      btree_node_unsupported_version,
 714                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 715
 716         if (btree_err_on(offset + sectors > btree_sectors(c),
 717                          -BCH_ERR_btree_node_read_err_fixable,
 718                          c, ca, b, i,
 719                          bset_past_end_of_btree_node,
 720                          "bset past end of btree node")) {
 721                 i->u64s = 0;
 722                 ret = 0;
 723                 goto out;
 724         }
 725
 726         btree_err_on(offset && !i->u64s,
 727                      -BCH_ERR_btree_node_read_err_fixable,
 728                      c, ca, b, i,
 729                      bset_empty,
 730                      "empty bset");
 731
 732         btree_err_on(BSET_OFFSET(i) && BSET_OFFSET(i) != offset,
 733                      -BCH_ERR_btree_node_read_err_want_retry,
 734                      c, ca, b, i,
 735                      bset_wrong_sector_offset,
 736                      "bset at wrong sector offset");
 737
 738         if (!offset) {
 739                 struct btree_node *bn =
 740                         container_of(i, struct btree_node, keys);
 741                 /* These indicate that we read the wrong btree node: */
 742
 743                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 744                         struct bch_btree_ptr_v2 *bp =
 745                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 746
 747                         /* XXX endianness */
 748                         btree_err_on(bp->seq != bn->keys.seq,
 749                                      -BCH_ERR_btree_node_read_err_must_retry,
 750                                      c, ca, b, NULL,
 751                                      bset_bad_seq,
 752                                      "incorrect sequence number (wrong btree node)");
 753                 }
 754
 755                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 756                              -BCH_ERR_btree_node_read_err_must_retry,
 757                              c, ca, b, i,
 758                              btree_node_bad_btree,
 759                              "incorrect btree id");
 760
 761                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 762                              -BCH_ERR_btree_node_read_err_must_retry,
 763                              c, ca, b, i,
 764                              btree_node_bad_level,
 765                              "incorrect level");
 766
 767                 if (!write)
 768                         compat_btree_node(b->c.level, b->c.btree_id, version,
 769                                           BSET_BIG_ENDIAN(i), write, bn);
 770
 771                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 772                         struct bch_btree_ptr_v2 *bp =
 773                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 774
 775                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 776                                 b->data->min_key = bp->min_key;
 777                                 b->data->max_key = b->key.k.p;
 778                         }
 779
 780                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 781                                      -BCH_ERR_btree_node_read_err_must_retry,
 782                                      c, ca, b, NULL,
 783                                      btree_node_bad_min_key,
 784                                      "incorrect min_key: got %s should be %s",
 785                                      (printbuf_reset(&buf1),
 786                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 787                                      (printbuf_reset(&buf2),
 788                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 789                 }
 790
 791                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 792                              -BCH_ERR_btree_node_read_err_must_retry,
 793                              c, ca, b, i,
 794                              btree_node_bad_max_key,
 795                              "incorrect max key %s",
 796                              (printbuf_reset(&buf1),
 797                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 798
 799                 if (write)
 800                         compat_btree_node(b->c.level, b->c.btree_id, version,
 801                                           BSET_BIG_ENDIAN(i), write, bn);
 802
 803                 btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
 804                              -BCH_ERR_btree_node_read_err_bad_node,
 805                              c, ca, b, i,
 806                              btree_node_bad_format,
 807                              "invalid bkey format: %s\n  %s", buf1.buf,
 808                              (printbuf_reset(&buf2),
 809                               bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
 810                 printbuf_reset(&buf1);
 811
 812                 compat_bformat(b->c.level, b->c.btree_id, version,
 813                                BSET_BIG_ENDIAN(i), write,
 814                                &bn->format);
 815         }
 816 out:
 817 fsck_err:
 818         printbuf_exit(&buf2);
 819         printbuf_exit(&buf1);
 820         return ret;
 821 }
 822
 823 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 824                             struct bkey_s_c k,
 825                             bool updated_range, int rw,
 826                             struct printbuf *err)
 827 {
 828         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 829                 (!updated_range ? bch2_bkey_in_btree_node(c, b, k, err) : 0) ?:
 830                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 831 }
 832
 833 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 834                          struct bset *i, int write,
 835                          bool have_retry, bool *saw_error)
 836 {
 837         unsigned version = le16_to_cpu(i->version);
 838         struct bkey_packed *k, *prev = NULL;
 839         struct printbuf buf = PRINTBUF;
 840         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 841                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 842         int ret = 0;
 843
 844         for (k = i->start;
 845              k != vstruct_last(i);) {
 846                 struct bkey_s u;
 847                 struct bkey tmp;
 848
 849                 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
 850                                  -BCH_ERR_btree_node_read_err_fixable,
 851                                  c, NULL, b, i,
 852                                  btree_node_bkey_past_bset_end,
 853                                  "key extends past end of bset")) {
 854                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 855                         break;
 856                 }
 857
 858                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 859                                  -BCH_ERR_btree_node_read_err_fixable,
 860                                  c, NULL, b, i,
 861                                  btree_node_bkey_bad_format,
 862                                  "invalid bkey format %u", k->format)) {
 863                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 864                         memmove_u64s_down(k, bkey_p_next(k),
 865                                           (u64 *) vstruct_end(i) - (u64 *) k);
 866                         continue;
 867                 }
 868
 869                 /* XXX: validate k->u64s */
 870                 if (!write)
 871                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 872                                     BSET_BIG_ENDIAN(i), write,
 873                                     &b->format, k);
 874
 875                 u = __bkey_disassemble(b, k, &tmp);
 876
 877                 printbuf_reset(&buf);
 878                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 879                         printbuf_reset(&buf);
 880                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 881                         prt_printf(&buf, "\n  ");
 882                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 883
 884                         btree_err(-BCH_ERR_btree_node_read_err_fixable,
 885                                   c, NULL, b, i,
 886                                   btree_node_bad_bkey,
 887                                   "invalid bkey: %s", buf.buf);
 888
 889                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 890                         memmove_u64s_down(k, bkey_p_next(k),
 891                                           (u64 *) vstruct_end(i) - (u64 *) k);
 892                         continue;
 893                 }
 894
 895                 if (write)
 896                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 897                                     BSET_BIG_ENDIAN(i), write,
 898                                     &b->format, k);
 899
 900                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 901                         struct bkey up = bkey_unpack_key(b, prev);
 902
 903                         printbuf_reset(&buf);
 904                         prt_printf(&buf, "keys out of order: ");
 905                         bch2_bkey_to_text(&buf, &up);
 906                         prt_printf(&buf, " > ");
 907                         bch2_bkey_to_text(&buf, u.k);
 908
 909                         bch2_dump_bset(c, b, i, 0);
 910
 911                         if (btree_err(-BCH_ERR_btree_node_read_err_fixable,
 912                                       c, NULL, b, i,
 913                                       btree_node_bkey_out_of_order,
 914                                       "%s", buf.buf)) {
 915                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 916                                 memmove_u64s_down(k, bkey_p_next(k),
 917                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 918                                 continue;
 919                         }
 920                 }
 921
 922                 prev = k;
 923                 k = bkey_p_next(k);
 924         }
 925 fsck_err:
 926         printbuf_exit(&buf);
 927         return ret;
 928 }
 929
 930 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 931                               struct btree *b, bool have_retry, bool *saw_error)
 932 {
 933         struct btree_node_entry *bne;
 934         struct sort_iter *iter;
 935         struct btree_node *sorted;
 936         struct bkey_packed *k;
 937         struct bch_extent_ptr *ptr;
 938         struct bset *i;
 939         bool used_mempool, blacklisted;
 940         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 941                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 942         unsigned u64s;
 943         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 944         struct printbuf buf = PRINTBUF;
 945         int ret = 0, retry_read = 0, write = READ;
 946
 947         b->version_ondisk = U16_MAX;
 948         /* We might get called multiple times on read retry: */
 949         b->written = 0;
 950
 951         iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
 952         sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2);
 953
 954         if (bch2_meta_read_fault("btree"))
 955                 btree_err(-BCH_ERR_btree_node_read_err_must_retry,
 956                           c, ca, b, NULL,
 957                           btree_node_fault_injected,
 958                           "dynamic fault");
 959
 960         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 961                      -BCH_ERR_btree_node_read_err_must_retry,
 962                      c, ca, b, NULL,
 963                      btree_node_bad_magic,
 964                      "bad magic: want %llx, got %llx",
 965                      bset_magic(c), le64_to_cpu(b->data->magic));
 966
 967         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 968                 struct bch_btree_ptr_v2 *bp =
 969                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 970
 971                 bch2_bpos_to_text(&buf, b->data->min_key);
 972                 prt_str(&buf, "-");
 973                 bch2_bpos_to_text(&buf, b->data->max_key);
 974
 975                 btree_err_on(b->data->keys.seq != bp->seq,
 976                              -BCH_ERR_btree_node_read_err_must_retry,
 977                              c, ca, b, NULL,
 978                              btree_node_bad_seq,
 979                              "got wrong btree node (want %llx got %llx)\n"
 980                              "got btree %s level %llu pos %s",
 981                              bp->seq, b->data->keys.seq,
 982                              bch2_btree_id_str(BTREE_NODE_ID(b->data)),
 983                              BTREE_NODE_LEVEL(b->data),
 984                              buf.buf);
 985         } else {
 986                 btree_err_on(!b->data->keys.seq,
 987                              -BCH_ERR_btree_node_read_err_must_retry,
 988                              c, ca, b, NULL,
 989                              btree_node_bad_seq,
 990                              "bad btree header: seq 0");
 991         }
 992
 993         while (b->written < (ptr_written ?: btree_sectors(c))) {
 994                 unsigned sectors;
 995                 struct nonce nonce;
 996                 bool first = !b->written;
 997                 bool csum_bad;
 998
 999                 if (!b->written) {
1000                         i = &b->data->keys;
1001
1002                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1003                                      -BCH_ERR_btree_node_read_err_want_retry,
1004                                      c, ca, b, i,
1005                                      bset_unknown_csum,
1006                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1007
1008                         nonce = btree_nonce(i, b->written << 9);
1009
1010                         csum_bad = bch2_crc_cmp(b->data->csum,
1011                                 csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data));
1012                         if (csum_bad)
1013                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1014
1015                         btree_err_on(csum_bad,
1016                                      -BCH_ERR_btree_node_read_err_want_retry,
1017                                      c, ca, b, i,
1018                                      bset_bad_csum,
1019                                      "invalid checksum");
1020
1021                         ret = bset_encrypt(c, i, b->written << 9);
1022                         if (bch2_fs_fatal_err_on(ret, c,
1023                                         "error decrypting btree node: %i", ret))
1024                                 goto fsck_err;
1025
1026                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
1027                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
1028                                      -BCH_ERR_btree_node_read_err_incompatible,
1029                                      c, NULL, b, NULL,
1030                                      btree_node_unsupported_version,
1031                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
1032
1033                         sectors = vstruct_sectors(b->data, c->block_bits);
1034                 } else {
1035                         bne = write_block(b);
1036                         i = &bne->keys;
1037
1038                         if (i->seq != b->data->keys.seq)
1039                                 break;
1040
1041                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1042                                      -BCH_ERR_btree_node_read_err_want_retry,
1043                                      c, ca, b, i,
1044                                      bset_unknown_csum,
1045                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1046
1047                         nonce = btree_nonce(i, b->written << 9);
1048                         csum_bad = bch2_crc_cmp(bne->csum,
1049                                 csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne));
1050                         if (csum_bad)
1051                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1052
1053                         btree_err_on(csum_bad,
1054                                      -BCH_ERR_btree_node_read_err_want_retry,
1055                                      c, ca, b, i,
1056                                      bset_bad_csum,
1057                                      "invalid checksum");
1058
1059                         ret = bset_encrypt(c, i, b->written << 9);
1060                         if (bch2_fs_fatal_err_on(ret, c,
1061                                         "error decrypting btree node: %i\n", ret))
1062                                 goto fsck_err;
1063
1064                         sectors = vstruct_sectors(bne, c->block_bits);
1065                 }
1066
1067                 b->version_ondisk = min(b->version_ondisk,
1068                                         le16_to_cpu(i->version));
1069
1070                 ret = validate_bset(c, ca, b, i, b->written, sectors,
1071                                     READ, have_retry, saw_error);
1072                 if (ret)
1073                         goto fsck_err;
1074
1075                 if (!b->written)
1076                         btree_node_set_format(b, b->data->format);
1077
1078                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1079                 if (ret)
1080                         goto fsck_err;
1081
1082                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1083
1084                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1085                                         le64_to_cpu(i->journal_seq),
1086                                         true);
1087
1088                 btree_err_on(blacklisted && first,
1089                              -BCH_ERR_btree_node_read_err_fixable,
1090                              c, ca, b, i,
1091                              bset_blacklisted_journal_seq,
1092                              "first btree node bset has blacklisted journal seq (%llu)",
1093                              le64_to_cpu(i->journal_seq));
1094
1095                 btree_err_on(blacklisted && ptr_written,
1096                              -BCH_ERR_btree_node_read_err_fixable,
1097                              c, ca, b, i,
1098                              first_bset_blacklisted_journal_seq,
1099                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1100                              le64_to_cpu(i->journal_seq),
1101                              b->written, b->written + sectors, ptr_written);
1102
1103                 b->written += sectors;
1104
1105                 if (blacklisted && !first)
1106                         continue;
1107
1108                 sort_iter_add(iter,
1109                               vstruct_idx(i, 0),
1110                               vstruct_last(i));
1111         }
1112
1113         if (ptr_written) {
1114                 btree_err_on(b->written < ptr_written,
1115                              -BCH_ERR_btree_node_read_err_want_retry,
1116                              c, ca, b, NULL,
1117                              btree_node_data_missing,
1118                              "btree node data missing: expected %u sectors, found %u",
1119                              ptr_written, b->written);
1120         } else {
1121                 for (bne = write_block(b);
1122                      bset_byte_offset(b, bne) < btree_bytes(c);
1123                      bne = (void *) bne + block_bytes(c))
1124                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1125                                      !bch2_journal_seq_is_blacklisted(c,
1126                                                                       le64_to_cpu(bne->keys.journal_seq),
1127                                                                       true),
1128                                      -BCH_ERR_btree_node_read_err_want_retry,
1129                                      c, ca, b, NULL,
1130                                      btree_node_bset_after_end,
1131                                      "found bset signature after last bset");
1132         }
1133
1134         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1135         sorted->keys.u64s = 0;
1136
1137         set_btree_bset(b, b->set, &b->data->keys);
1138
1139         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1140
1141         u64s = le16_to_cpu(sorted->keys.u64s);
1142         *sorted = *b->data;
1143         sorted->keys.u64s = cpu_to_le16(u64s);
1144         swap(sorted, b->data);
1145         set_btree_bset(b, b->set, &b->data->keys);
1146         b->nsets = 1;
1147
1148         BUG_ON(b->nr.live_u64s != u64s);
1149
1150         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1151
1152         if (updated_range)
1153                 bch2_btree_node_drop_keys_outside_node(b);
1154
1155         i = &b->data->keys;
1156         for (k = i->start; k != vstruct_last(i);) {
1157                 struct bkey tmp;
1158                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1159
1160                 printbuf_reset(&buf);
1161
1162                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1163                     (bch2_inject_invalid_keys &&
1164                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1165                         printbuf_reset(&buf);
1166
1167                         prt_printf(&buf, "invalid bkey: ");
1168                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1169                         prt_printf(&buf, "\n  ");
1170                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1171
1172                         btree_err(-BCH_ERR_btree_node_read_err_fixable,
1173                                   c, NULL, b, i,
1174                                   btree_node_bad_bkey,
1175                                   "%s", buf.buf);
1176
1177                         btree_keys_account_key_drop(&b->nr, 0, k);
1178
1179                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1180                         memmove_u64s_down(k, bkey_p_next(k),
1181                                           (u64 *) vstruct_end(i) - (u64 *) k);
1182                         set_btree_bset_end(b, b->set);
1183                         continue;
1184                 }
1185
1186                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1187                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1188
1189                         bp.v->mem_ptr = 0;
1190                 }
1191
1192                 k = bkey_p_next(k);
1193         }
1194
1195         bch2_bset_build_aux_tree(b, b->set, false);
1196
1197         set_needs_whiteout(btree_bset_first(b), true);
1198
1199         btree_node_reset_sib_u64s(b);
1200
1201         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1202                 struct bch_dev *ca2 = bch_dev_bkey_exists(c, ptr->dev);
1203
1204                 if (ca2->mi.state != BCH_MEMBER_STATE_rw)
1205                         set_btree_node_need_rewrite(b);
1206         }
1207
1208         if (!ptr_written)
1209                 set_btree_node_need_rewrite(b);
1210 out:
1211         mempool_free(iter, &c->fill_iter);
1212         printbuf_exit(&buf);
1213         return retry_read;
1214 fsck_err:
1215         if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1216             ret == -BCH_ERR_btree_node_read_err_must_retry)
1217                 retry_read = 1;
1218         else
1219                 set_btree_node_read_error(b);
1220         goto out;
1221 }
1222
1223 static void btree_node_read_work(struct work_struct *work)
1224 {
1225         struct btree_read_bio *rb =
1226                 container_of(work, struct btree_read_bio, work);
1227         struct bch_fs *c        = rb->c;
1228         struct btree *b         = rb->b;
1229         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1230         struct bio *bio         = &rb->bio;
1231         struct bch_io_failures failed = { .nr = 0 };
1232         struct printbuf buf = PRINTBUF;
1233         bool saw_error = false;
1234         bool retry = false;
1235         bool can_retry;
1236
1237         goto start;
1238         while (1) {
1239                 retry = true;
1240                 bch_info(c, "retrying read");
1241                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1242                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1243                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1244                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1245                 bio->bi_iter.bi_size    = btree_bytes(c);
1246
1247                 if (rb->have_ioref) {
1248                         bio_set_dev(bio, ca->disk_sb.bdev);
1249                         submit_bio_wait(bio);
1250                 } else {
1251                         bio->bi_status = BLK_STS_REMOVED;
1252                 }
1253 start:
1254                 printbuf_reset(&buf);
1255                 bch2_btree_pos_to_text(&buf, c, b);
1256                 bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_read,
1257                                    "btree read error %s for %s",
1258                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1259                 if (rb->have_ioref)
1260                         percpu_ref_put(&ca->io_ref);
1261                 rb->have_ioref = false;
1262
1263                 bch2_mark_io_failure(&failed, &rb->pick);
1264
1265                 can_retry = bch2_bkey_pick_read_device(c,
1266                                 bkey_i_to_s_c(&b->key),
1267                                 &failed, &rb->pick) > 0;
1268
1269                 if (!bio->bi_status &&
1270                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1271                         if (retry)
1272                                 bch_info(c, "retry success");
1273                         break;
1274                 }
1275
1276                 saw_error = true;
1277
1278                 if (!can_retry) {
1279                         set_btree_node_read_error(b);
1280                         break;
1281                 }
1282         }
1283
1284         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1285                                rb->start_time);
1286         bio_put(&rb->bio);
1287
1288         if (saw_error && !btree_node_read_error(b)) {
1289                 printbuf_reset(&buf);
1290                 bch2_bpos_to_text(&buf, b->key.k.p);
1291                 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1292                          __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf);
1293
1294                 bch2_btree_node_rewrite_async(c, b);
1295         }
1296
1297         printbuf_exit(&buf);
1298         clear_btree_node_read_in_flight(b);
1299         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1300 }
1301
1302 static void btree_node_read_endio(struct bio *bio)
1303 {
1304         struct btree_read_bio *rb =
1305                 container_of(bio, struct btree_read_bio, bio);
1306         struct bch_fs *c        = rb->c;
1307
1308         if (rb->have_ioref) {
1309                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1310
1311                 bch2_latency_acct(ca, rb->start_time, READ);
1312         }
1313
1314         queue_work(c->io_complete_wq, &rb->work);
1315 }
1316
1317 struct btree_node_read_all {
1318         struct closure          cl;
1319         struct bch_fs           *c;
1320         struct btree            *b;
1321         unsigned                nr;
1322         void                    *buf[BCH_REPLICAS_MAX];
1323         struct bio              *bio[BCH_REPLICAS_MAX];
1324         blk_status_t            err[BCH_REPLICAS_MAX];
1325 };
1326
1327 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1328 {
1329         struct btree_node *bn = data;
1330         struct btree_node_entry *bne;
1331         unsigned offset = 0;
1332
1333         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1334                 return 0;
1335
1336         while (offset < btree_sectors(c)) {
1337                 if (!offset) {
1338                         offset += vstruct_sectors(bn, c->block_bits);
1339                 } else {
1340                         bne = data + (offset << 9);
1341                         if (bne->keys.seq != bn->keys.seq)
1342                                 break;
1343                         offset += vstruct_sectors(bne, c->block_bits);
1344                 }
1345         }
1346
1347         return offset;
1348 }
1349
1350 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1351 {
1352         struct btree_node *bn = data;
1353         struct btree_node_entry *bne;
1354
1355         if (!offset)
1356                 return false;
1357
1358         while (offset < btree_sectors(c)) {
1359                 bne = data + (offset << 9);
1360                 if (bne->keys.seq == bn->keys.seq)
1361                         return true;
1362                 offset++;
1363         }
1364
1365         return false;
1366         return offset;
1367 }
1368
1369 static CLOSURE_CALLBACK(btree_node_read_all_replicas_done)
1370 {
1371         closure_type(ra, struct btree_node_read_all, cl);
1372         struct bch_fs *c = ra->c;
1373         struct btree *b = ra->b;
1374         struct printbuf buf = PRINTBUF;
1375         bool dump_bset_maps = false;
1376         bool have_retry = false;
1377         int ret = 0, best = -1, write = READ;
1378         unsigned i, written = 0, written2 = 0;
1379         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1380                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1381         bool _saw_error = false, *saw_error = &_saw_error;
1382
1383         for (i = 0; i < ra->nr; i++) {
1384                 struct btree_node *bn = ra->buf[i];
1385
1386                 if (ra->err[i])
1387                         continue;
1388
1389                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1390                     (seq && seq != bn->keys.seq))
1391                         continue;
1392
1393                 if (best < 0) {
1394                         best = i;
1395                         written = btree_node_sectors_written(c, bn);
1396                         continue;
1397                 }
1398
1399                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1400                 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable,
1401                                  c, NULL, b, NULL,
1402                                  btree_node_replicas_sectors_written_mismatch,
1403                                  "btree node sectors written mismatch: %u != %u",
1404                                  written, written2) ||
1405                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1406                                  -BCH_ERR_btree_node_read_err_fixable,
1407                                  c, NULL, b, NULL,
1408                                  btree_node_bset_after_end,
1409                                  "found bset signature after last bset") ||
1410                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1411                                  -BCH_ERR_btree_node_read_err_fixable,
1412                                  c, NULL, b, NULL,
1413                                  btree_node_replicas_data_mismatch,
1414                                  "btree node replicas content mismatch"))
1415                         dump_bset_maps = true;
1416
1417                 if (written2 > written) {
1418                         written = written2;
1419                         best = i;
1420                 }
1421         }
1422 fsck_err:
1423         if (dump_bset_maps) {
1424                 for (i = 0; i < ra->nr; i++) {
1425                         struct btree_node *bn = ra->buf[i];
1426                         struct btree_node_entry *bne = NULL;
1427                         unsigned offset = 0, sectors;
1428                         bool gap = false;
1429
1430                         if (ra->err[i])
1431                                 continue;
1432
1433                         printbuf_reset(&buf);
1434
1435                         while (offset < btree_sectors(c)) {
1436                                 if (!offset) {
1437                                         sectors = vstruct_sectors(bn, c->block_bits);
1438                                 } else {
1439                                         bne = ra->buf[i] + (offset << 9);
1440                                         if (bne->keys.seq != bn->keys.seq)
1441                                                 break;
1442                                         sectors = vstruct_sectors(bne, c->block_bits);
1443                                 }
1444
1445                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1446                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1447                                                         le64_to_cpu(bne->keys.journal_seq), false))
1448                                         prt_printf(&buf, "*");
1449                                 offset += sectors;
1450                         }
1451
1452                         while (offset < btree_sectors(c)) {
1453                                 bne = ra->buf[i] + (offset << 9);
1454                                 if (bne->keys.seq == bn->keys.seq) {
1455                                         if (!gap)
1456                                                 prt_printf(&buf, " GAP");
1457                                         gap = true;
1458
1459                                         sectors = vstruct_sectors(bne, c->block_bits);
1460                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1461                                         if (bch2_journal_seq_is_blacklisted(c,
1462                                                         le64_to_cpu(bne->keys.journal_seq), false))
1463                                                 prt_printf(&buf, "*");
1464                                 }
1465                                 offset++;
1466                         }
1467
1468                         bch_err(c, "replica %u:%s", i, buf.buf);
1469                 }
1470         }
1471
1472         if (best >= 0) {
1473                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1474                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1475         } else {
1476                 ret = -1;
1477         }
1478
1479         if (ret)
1480                 set_btree_node_read_error(b);
1481         else if (*saw_error)
1482                 bch2_btree_node_rewrite_async(c, b);
1483
1484         for (i = 0; i < ra->nr; i++) {
1485                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1486                 bio_put(ra->bio[i]);
1487         }
1488
1489         closure_debug_destroy(&ra->cl);
1490         kfree(ra);
1491         printbuf_exit(&buf);
1492
1493         clear_btree_node_read_in_flight(b);
1494         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1495 }
1496
1497 static void btree_node_read_all_replicas_endio(struct bio *bio)
1498 {
1499         struct btree_read_bio *rb =
1500                 container_of(bio, struct btree_read_bio, bio);
1501         struct bch_fs *c        = rb->c;
1502         struct btree_node_read_all *ra = rb->ra;
1503
1504         if (rb->have_ioref) {
1505                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1506
1507                 bch2_latency_acct(ca, rb->start_time, READ);
1508         }
1509
1510         ra->err[rb->idx] = bio->bi_status;
1511         closure_put(&ra->cl);
1512 }
1513
1514 /*
1515  * XXX This allocates multiple times from the same mempools, and can deadlock
1516  * under sufficient memory pressure (but is only a debug path)
1517  */
1518 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1519 {
1520         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1521         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1522         const union bch_extent_entry *entry;
1523         struct extent_ptr_decoded pick;
1524         struct btree_node_read_all *ra;
1525         unsigned i;
1526
1527         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1528         if (!ra)
1529                 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1530
1531         closure_init(&ra->cl, NULL);
1532         ra->c   = c;
1533         ra->b   = b;
1534         ra->nr  = bch2_bkey_nr_ptrs(k);
1535
1536         for (i = 0; i < ra->nr; i++) {
1537                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1538                 ra->bio[i] = bio_alloc_bioset(NULL,
1539                                               buf_pages(ra->buf[i], btree_bytes(c)),
1540                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1541                                               GFP_NOFS,
1542                                               &c->btree_bio);
1543         }
1544
1545         i = 0;
1546         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1547                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1548                 struct btree_read_bio *rb =
1549                         container_of(ra->bio[i], struct btree_read_bio, bio);
1550                 rb->c                   = c;
1551                 rb->b                   = b;
1552                 rb->ra                  = ra;
1553                 rb->start_time          = local_clock();
1554                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1555                 rb->idx                 = i;
1556                 rb->pick                = pick;
1557                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1558                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1559                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1560
1561                 if (rb->have_ioref) {
1562                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1563                                      bio_sectors(&rb->bio));
1564                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1565
1566                         closure_get(&ra->cl);
1567                         submit_bio(&rb->bio);
1568                 } else {
1569                         ra->err[i] = BLK_STS_REMOVED;
1570                 }
1571
1572                 i++;
1573         }
1574
1575         if (sync) {
1576                 closure_sync(&ra->cl);
1577                 btree_node_read_all_replicas_done(&ra->cl.work);
1578         } else {
1579                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1580                             c->io_complete_wq);
1581         }
1582
1583         return 0;
1584 }
1585
1586 void bch2_btree_node_read(struct btree_trans *trans, struct btree *b,
1587                           bool sync)
1588 {
1589         struct bch_fs *c = trans->c;
1590         struct extent_ptr_decoded pick;
1591         struct btree_read_bio *rb;
1592         struct bch_dev *ca;
1593         struct bio *bio;
1594         int ret;
1595
1596         trace_and_count(c, btree_node_read, trans, b);
1597
1598         if (bch2_verify_all_btree_replicas &&
1599             !btree_node_read_all_replicas(c, b, sync))
1600                 return;
1601
1602         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1603                                          NULL, &pick);
1604
1605         if (ret <= 0) {
1606                 struct printbuf buf = PRINTBUF;
1607
1608                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1609                 bch2_btree_pos_to_text(&buf, c, b);
1610                 bch_err(c, "%s", buf.buf);
1611
1612                 if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1613                     c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1614                         bch2_fatal_error(c);
1615
1616                 set_btree_node_read_error(b);
1617                 clear_btree_node_read_in_flight(b);
1618                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1619                 printbuf_exit(&buf);
1620                 return;
1621         }
1622
1623         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1624
1625         bio = bio_alloc_bioset(NULL,
1626                                buf_pages(b->data, btree_bytes(c)),
1627                                REQ_OP_READ|REQ_SYNC|REQ_META,
1628                                GFP_NOFS,
1629                                &c->btree_bio);
1630         rb = container_of(bio, struct btree_read_bio, bio);
1631         rb->c                   = c;
1632         rb->b                   = b;
1633         rb->ra                  = NULL;
1634         rb->start_time          = local_clock();
1635         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1636         rb->pick                = pick;
1637         INIT_WORK(&rb->work, btree_node_read_work);
1638         bio->bi_iter.bi_sector  = pick.ptr.offset;
1639         bio->bi_end_io          = btree_node_read_endio;
1640         bch2_bio_map(bio, b->data, btree_bytes(c));
1641
1642         if (rb->have_ioref) {
1643                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1644                              bio_sectors(bio));
1645                 bio_set_dev(bio, ca->disk_sb.bdev);
1646
1647                 if (sync) {
1648                         submit_bio_wait(bio);
1649
1650                         btree_node_read_work(&rb->work);
1651                 } else {
1652                         submit_bio(bio);
1653                 }
1654         } else {
1655                 bio->bi_status = BLK_STS_REMOVED;
1656
1657                 if (sync)
1658                         btree_node_read_work(&rb->work);
1659                 else
1660                         queue_work(c->io_complete_wq, &rb->work);
1661         }
1662 }
1663
1664 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1665                                   const struct bkey_i *k, unsigned level)
1666 {
1667         struct bch_fs *c = trans->c;
1668         struct closure cl;
1669         struct btree *b;
1670         int ret;
1671
1672         closure_init_stack(&cl);
1673
1674         do {
1675                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
1676                 closure_sync(&cl);
1677         } while (ret);
1678
1679         b = bch2_btree_node_mem_alloc(trans, level != 0);
1680         bch2_btree_cache_cannibalize_unlock(trans);
1681
1682         BUG_ON(IS_ERR(b));
1683
1684         bkey_copy(&b->key, k);
1685         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1686
1687         set_btree_node_read_in_flight(b);
1688
1689         bch2_btree_node_read(trans, b, true);
1690
1691         if (btree_node_read_error(b)) {
1692                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1693
1694                 mutex_lock(&c->btree_cache.lock);
1695                 list_move(&b->list, &c->btree_cache.freeable);
1696                 mutex_unlock(&c->btree_cache.lock);
1697
1698                 ret = -EIO;
1699                 goto err;
1700         }
1701
1702         bch2_btree_set_root_for_read(c, b);
1703 err:
1704         six_unlock_write(&b->c.lock);
1705         six_unlock_intent(&b->c.lock);
1706
1707         return ret;
1708 }
1709
1710 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1711                         const struct bkey_i *k, unsigned level)
1712 {
1713         return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
1714 }
1715
1716 static void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1717                                       struct btree_write *w)
1718 {
1719         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1720
1721         do {
1722                 old = new = v;
1723                 if (!(old & 1))
1724                         break;
1725
1726                 new &= ~1UL;
1727         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1728
1729         if (old & 1)
1730                 closure_put(&((struct btree_update *) new)->cl);
1731
1732         bch2_journal_pin_drop(&c->journal, &w->journal);
1733 }
1734
1735 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1736 {
1737         struct btree_write *w = btree_prev_write(b);
1738         unsigned long old, new, v;
1739         unsigned type = 0;
1740
1741         bch2_btree_complete_write(c, b, w);
1742
1743         v = READ_ONCE(b->flags);
1744         do {
1745                 old = new = v;
1746
1747                 if ((old & (1U << BTREE_NODE_dirty)) &&
1748                     (old & (1U << BTREE_NODE_need_write)) &&
1749                     !(old & (1U << BTREE_NODE_never_write)) &&
1750                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1751                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1752                         new &= ~(1U << BTREE_NODE_dirty);
1753                         new &= ~(1U << BTREE_NODE_need_write);
1754                         new |=  (1U << BTREE_NODE_write_in_flight);
1755                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1756                         new |=  (1U << BTREE_NODE_just_written);
1757                         new ^=  (1U << BTREE_NODE_write_idx);
1758
1759                         type = new & BTREE_WRITE_TYPE_MASK;
1760                         new &= ~BTREE_WRITE_TYPE_MASK;
1761                 } else {
1762                         new &= ~(1U << BTREE_NODE_write_in_flight);
1763                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1764                 }
1765         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1766
1767         if (new & (1U << BTREE_NODE_write_in_flight))
1768                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1769         else
1770                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1771 }
1772
1773 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1774 {
1775         struct btree_trans *trans = bch2_trans_get(c);
1776
1777         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
1778         __btree_node_write_done(c, b);
1779         six_unlock_read(&b->c.lock);
1780
1781         bch2_trans_put(trans);
1782 }
1783
1784 static void btree_node_write_work(struct work_struct *work)
1785 {
1786         struct btree_write_bio *wbio =
1787                 container_of(work, struct btree_write_bio, work);
1788         struct bch_fs *c        = wbio->wbio.c;
1789         struct btree *b         = wbio->wbio.bio.bi_private;
1790         struct bch_extent_ptr *ptr;
1791         int ret = 0;
1792
1793         btree_bounce_free(c,
1794                 wbio->data_bytes,
1795                 wbio->wbio.used_mempool,
1796                 wbio->data);
1797
1798         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1799                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1800
1801         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1802                 goto err;
1803
1804         if (wbio->wbio.first_btree_write) {
1805                 if (wbio->wbio.failed.nr) {
1806
1807                 }
1808         } else {
1809                 ret = bch2_trans_do(c, NULL, NULL, 0,
1810                         bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
1811                                         BCH_WATERMARK_reclaim|
1812                                         BCH_TRANS_COMMIT_journal_reclaim|
1813                                         BCH_TRANS_COMMIT_no_enospc|
1814                                         BCH_TRANS_COMMIT_no_check_rw,
1815                                         !wbio->wbio.failed.nr));
1816                 if (ret)
1817                         goto err;
1818         }
1819 out:
1820         bio_put(&wbio->wbio.bio);
1821         btree_node_write_done(c, b);
1822         return;
1823 err:
1824         set_btree_node_noevict(b);
1825         if (!bch2_err_matches(ret, EROFS))
1826                 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1827         goto out;
1828 }
1829
1830 static void btree_node_write_endio(struct bio *bio)
1831 {
1832         struct bch_write_bio *wbio      = to_wbio(bio);
1833         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1834         struct bch_write_bio *orig      = parent ?: wbio;
1835         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1836         struct bch_fs *c                = wbio->c;
1837         struct btree *b                 = wbio->bio.bi_private;
1838         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1839         unsigned long flags;
1840
1841         if (wbio->have_ioref)
1842                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1843
1844         if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1845                                "btree write error: %s",
1846                                bch2_blk_status_to_str(bio->bi_status)) ||
1847             bch2_meta_write_fault("btree")) {
1848                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1849                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1850                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1851         }
1852
1853         if (wbio->have_ioref)
1854                 percpu_ref_put(&ca->io_ref);
1855
1856         if (parent) {
1857                 bio_put(bio);
1858                 bio_endio(&parent->bio);
1859                 return;
1860         }
1861
1862         clear_btree_node_write_in_flight_inner(b);
1863         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1864         INIT_WORK(&wb->work, btree_node_write_work);
1865         queue_work(c->btree_io_complete_wq, &wb->work);
1866 }
1867
1868 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1869                                    struct bset *i, unsigned sectors)
1870 {
1871         struct printbuf buf = PRINTBUF;
1872         bool saw_error;
1873         int ret;
1874
1875         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1876                                 BKEY_TYPE_btree, WRITE, &buf);
1877
1878         if (ret)
1879                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1880         printbuf_exit(&buf);
1881         if (ret)
1882                 return ret;
1883
1884         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1885                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1886         if (ret) {
1887                 bch2_inconsistent_error(c);
1888                 dump_stack();
1889         }
1890
1891         return ret;
1892 }
1893
1894 static void btree_write_submit(struct work_struct *work)
1895 {
1896         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1897         struct bch_extent_ptr *ptr;
1898         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1899
1900         bkey_copy(&tmp.k, &wbio->key);
1901
1902         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1903                 ptr->offset += wbio->sector_offset;
1904
1905         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1906                                   &tmp.k, false);
1907 }
1908
1909 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1910 {
1911         struct btree_write_bio *wbio;
1912         struct bset_tree *t;
1913         struct bset *i;
1914         struct btree_node *bn = NULL;
1915         struct btree_node_entry *bne = NULL;
1916         struct sort_iter_stack sort_iter;
1917         struct nonce nonce;
1918         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1919         u64 seq = 0;
1920         bool used_mempool;
1921         unsigned long old, new;
1922         bool validate_before_checksum = false;
1923         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1924         void *data;
1925         int ret;
1926
1927         if (flags & BTREE_WRITE_ALREADY_STARTED)
1928                 goto do_write;
1929
1930         /*
1931          * We may only have a read lock on the btree node - the dirty bit is our
1932          * "lock" against racing with other threads that may be trying to start
1933          * a write, we do a write iff we clear the dirty bit. Since setting the
1934          * dirty bit requires a write lock, we can't race with other threads
1935          * redirtying it:
1936          */
1937         do {
1938                 old = new = READ_ONCE(b->flags);
1939
1940                 if (!(old & (1 << BTREE_NODE_dirty)))
1941                         return;
1942
1943                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1944                     !(old & (1 << BTREE_NODE_need_write)))
1945                         return;
1946
1947                 if (old &
1948                     ((1 << BTREE_NODE_never_write)|
1949                      (1 << BTREE_NODE_write_blocked)))
1950                         return;
1951
1952                 if (b->written &&
1953                     (old & (1 << BTREE_NODE_will_make_reachable)))
1954                         return;
1955
1956                 if (old & (1 << BTREE_NODE_write_in_flight))
1957                         return;
1958
1959                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1960                         type = new & BTREE_WRITE_TYPE_MASK;
1961                 new &= ~BTREE_WRITE_TYPE_MASK;
1962
1963                 new &= ~(1 << BTREE_NODE_dirty);
1964                 new &= ~(1 << BTREE_NODE_need_write);
1965                 new |=  (1 << BTREE_NODE_write_in_flight);
1966                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1967                 new |=  (1 << BTREE_NODE_just_written);
1968                 new ^=  (1 << BTREE_NODE_write_idx);
1969         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1970
1971         if (new & (1U << BTREE_NODE_need_write))
1972                 return;
1973 do_write:
1974         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1975
1976         atomic_dec(&c->btree_cache.dirty);
1977
1978         BUG_ON(btree_node_fake(b));
1979         BUG_ON((b->will_make_reachable != 0) != !b->written);
1980
1981         BUG_ON(b->written >= btree_sectors(c));
1982         BUG_ON(b->written & (block_sectors(c) - 1));
1983         BUG_ON(bset_written(b, btree_bset_last(b)));
1984         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1985         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1986
1987         bch2_sort_whiteouts(c, b);
1988
1989         sort_iter_stack_init(&sort_iter, b);
1990
1991         bytes = !b->written
1992                 ? sizeof(struct btree_node)
1993                 : sizeof(struct btree_node_entry);
1994
1995         bytes += b->whiteout_u64s * sizeof(u64);
1996
1997         for_each_bset(b, t) {
1998                 i = bset(b, t);
1999
2000                 if (bset_written(b, i))
2001                         continue;
2002
2003                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
2004                 sort_iter_add(&sort_iter.iter,
2005                               btree_bkey_first(b, t),
2006                               btree_bkey_last(b, t));
2007                 seq = max(seq, le64_to_cpu(i->journal_seq));
2008         }
2009
2010         BUG_ON(b->written && !seq);
2011
2012         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
2013         bytes += 8;
2014
2015         /* buffer must be a multiple of the block size */
2016         bytes = round_up(bytes, block_bytes(c));
2017
2018         if (bytes > btree_bytes(c)) {
2019                 struct printbuf buf = PRINTBUF;
2020
2021                 prt_printf(&buf, "btree node write bounce buffer overrun: %u > %zu\n",
2022                            bytes, btree_bytes(c));
2023
2024                 prt_printf(&buf, "header: %zu\n", b->written
2025                            ? sizeof(struct btree_node)
2026                            : sizeof(struct btree_node_entry));
2027                 prt_printf(&buf, "unwritten: %zu\n", b->whiteout_u64s * sizeof(u64));
2028
2029                 for_each_bset(b, t) {
2030                         i = bset(b, t);
2031
2032                         if (bset_written(b, i))
2033                                 continue;
2034                         prt_printf(&buf, "bset %zu: %zu\n", t - b->set, le16_to_cpu(i->u64s) * sizeof(u64));
2035                 }
2036
2037                 panic("%s", buf.buf);
2038                 printbuf_exit(&buf);
2039         }
2040
2041         data = btree_bounce_alloc(c, bytes, &used_mempool);
2042
2043         if (!b->written) {
2044                 bn = data;
2045                 *bn = *b->data;
2046                 i = &bn->keys;
2047         } else {
2048                 bne = data;
2049                 bne->keys = b->data->keys;
2050                 i = &bne->keys;
2051         }
2052
2053         i->journal_seq  = cpu_to_le64(seq);
2054         i->u64s         = 0;
2055
2056         sort_iter_add(&sort_iter.iter,
2057                       unwritten_whiteouts_start(c, b),
2058                       unwritten_whiteouts_end(c, b));
2059         SET_BSET_SEPARATE_WHITEOUTS(i, false);
2060
2061         b->whiteout_u64s = 0;
2062
2063         u64s = bch2_sort_keys(i->start, &sort_iter.iter, false);
2064         le16_add_cpu(&i->u64s, u64s);
2065
2066         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
2067
2068         set_needs_whiteout(i, false);
2069
2070         /* do we have data to write? */
2071         if (b->written && !i->u64s)
2072                 goto nowrite;
2073
2074         bytes_to_write = vstruct_end(i) - data;
2075         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
2076
2077         if (!b->written &&
2078             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2079                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
2080
2081         memset(data + bytes_to_write, 0,
2082                (sectors_to_write << 9) - bytes_to_write);
2083
2084         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2085         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2086         BUG_ON(i->seq != b->data->keys.seq);
2087
2088         i->version = cpu_to_le16(c->sb.version);
2089         SET_BSET_OFFSET(i, b->written);
2090         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2091
2092         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2093                 validate_before_checksum = true;
2094
2095         /* validate_bset will be modifying: */
2096         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2097                 validate_before_checksum = true;
2098
2099         /* if we're going to be encrypting, check metadata validity first: */
2100         if (validate_before_checksum &&
2101             validate_bset_for_write(c, b, i, sectors_to_write))
2102                 goto err;
2103
2104         ret = bset_encrypt(c, i, b->written << 9);
2105         if (bch2_fs_fatal_err_on(ret, c,
2106                         "error encrypting btree node: %i\n", ret))
2107                 goto err;
2108
2109         nonce = btree_nonce(i, b->written << 9);
2110
2111         if (bn)
2112                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2113         else
2114                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2115
2116         /* if we're not encrypting, check metadata after checksumming: */
2117         if (!validate_before_checksum &&
2118             validate_bset_for_write(c, b, i, sectors_to_write))
2119                 goto err;
2120
2121         /*
2122          * We handle btree write errors by immediately halting the journal -
2123          * after we've done that, we can't issue any subsequent btree writes
2124          * because they might have pointers to new nodes that failed to write.
2125          *
2126          * Furthermore, there's no point in doing any more btree writes because
2127          * with the journal stopped, we're never going to update the journal to
2128          * reflect that those writes were done and the data flushed from the
2129          * journal:
2130          *
2131          * Also on journal error, the pending write may have updates that were
2132          * never journalled (interior nodes, see btree_update_nodes_written()) -
2133          * it's critical that we don't do the write in that case otherwise we
2134          * will have updates visible that weren't in the journal:
2135          *
2136          * Make sure to update b->written so bch2_btree_init_next() doesn't
2137          * break:
2138          */
2139         if (bch2_journal_error(&c->journal) ||
2140             c->opts.nochanges)
2141                 goto err;
2142
2143         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2144
2145         wbio = container_of(bio_alloc_bioset(NULL,
2146                                 buf_pages(data, sectors_to_write << 9),
2147                                 REQ_OP_WRITE|REQ_META,
2148                                 GFP_NOFS,
2149                                 &c->btree_bio),
2150                             struct btree_write_bio, wbio.bio);
2151         wbio_init(&wbio->wbio.bio);
2152         wbio->data                      = data;
2153         wbio->data_bytes                = bytes;
2154         wbio->sector_offset             = b->written;
2155         wbio->wbio.c                    = c;
2156         wbio->wbio.used_mempool         = used_mempool;
2157         wbio->wbio.first_btree_write    = !b->written;
2158         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2159         wbio->wbio.bio.bi_private       = b;
2160
2161         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2162
2163         bkey_copy(&wbio->key, &b->key);
2164
2165         b->written += sectors_to_write;
2166
2167         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2168                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2169                         cpu_to_le16(b->written);
2170
2171         atomic64_inc(&c->btree_write_stats[type].nr);
2172         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2173
2174         INIT_WORK(&wbio->work, btree_write_submit);
2175         queue_work(c->io_complete_wq, &wbio->work);
2176         return;
2177 err:
2178         set_btree_node_noevict(b);
2179         b->written += sectors_to_write;
2180 nowrite:
2181         btree_bounce_free(c, bytes, used_mempool, data);
2182         __btree_node_write_done(c, b);
2183 }
2184
2185 /*
2186  * Work that must be done with write lock held:
2187  */
2188 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2189 {
2190         bool invalidated_iter = false;
2191         struct btree_node_entry *bne;
2192         struct bset_tree *t;
2193
2194         if (!btree_node_just_written(b))
2195                 return false;
2196
2197         BUG_ON(b->whiteout_u64s);
2198
2199         clear_btree_node_just_written(b);
2200
2201         /*
2202          * Note: immediately after write, bset_written() doesn't work - the
2203          * amount of data we had to write after compaction might have been
2204          * smaller than the offset of the last bset.
2205          *
2206          * However, we know that all bsets have been written here, as long as
2207          * we're still holding the write lock:
2208          */
2209
2210         /*
2211          * XXX: decide if we really want to unconditionally sort down to a
2212          * single bset:
2213          */
2214         if (b->nsets > 1) {
2215                 btree_node_sort(c, b, 0, b->nsets, true);
2216                 invalidated_iter = true;
2217         } else {
2218                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2219         }
2220
2221         for_each_bset(b, t)
2222                 set_needs_whiteout(bset(b, t), true);
2223
2224         bch2_btree_verify(c, b);
2225
2226         /*
2227          * If later we don't unconditionally sort down to a single bset, we have
2228          * to ensure this is still true:
2229          */
2230         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2231
2232         bne = want_new_bset(c, b);
2233         if (bne)
2234                 bch2_bset_init_next(c, b, bne);
2235
2236         bch2_btree_build_aux_trees(b);
2237
2238         return invalidated_iter;
2239 }
2240
2241 /*
2242  * Use this one if the node is intent locked:
2243  */
2244 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2245                            enum six_lock_type lock_type_held,
2246                            unsigned flags)
2247 {
2248         if (lock_type_held == SIX_LOCK_intent ||
2249             (lock_type_held == SIX_LOCK_read &&
2250              six_lock_tryupgrade(&b->c.lock))) {
2251                 __bch2_btree_node_write(c, b, flags);
2252
2253                 /* don't cycle lock unnecessarily: */
2254                 if (btree_node_just_written(b) &&
2255                     six_trylock_write(&b->c.lock)) {
2256                         bch2_btree_post_write_cleanup(c, b);
2257                         six_unlock_write(&b->c.lock);
2258                 }
2259
2260                 if (lock_type_held == SIX_LOCK_read)
2261                         six_lock_downgrade(&b->c.lock);
2262         } else {
2263                 __bch2_btree_node_write(c, b, flags);
2264                 if (lock_type_held == SIX_LOCK_write &&
2265                     btree_node_just_written(b))
2266                         bch2_btree_post_write_cleanup(c, b);
2267         }
2268 }
2269
2270 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2271 {
2272         struct bucket_table *tbl;
2273         struct rhash_head *pos;
2274         struct btree *b;
2275         unsigned i;
2276         bool ret = false;
2277 restart:
2278         rcu_read_lock();
2279         for_each_cached_btree(b, c, tbl, i, pos)
2280                 if (test_bit(flag, &b->flags)) {
2281                         rcu_read_unlock();
2282                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2283                         ret = true;
2284                         goto restart;
2285                 }
2286         rcu_read_unlock();
2287
2288         return ret;
2289 }
2290
2291 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2292 {
2293         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2294 }
2295
2296 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2297 {
2298         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2299 }
2300
2301 static const char * const bch2_btree_write_types[] = {
2302 #define x(t, n) [n] = #t,
2303         BCH_BTREE_WRITE_TYPES()
2304         NULL
2305 };
2306
2307 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2308 {
2309         printbuf_tabstop_push(out, 20);
2310         printbuf_tabstop_push(out, 10);
2311
2312         prt_tab(out);
2313         prt_str(out, "nr");
2314         prt_tab(out);
2315         prt_str(out, "size");
2316         prt_newline(out);
2317
2318         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2319                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2320                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2321
2322                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2323                 prt_tab(out);
2324                 prt_u64(out, nr);
2325                 prt_tab(out);
2326                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2327                 prt_newline(out);
2328         }
2329 }