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                 kvfree(p);
 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 > c->opts.btree_node_size);
 116
 117         *used_mempool = false;
 118         p = kvmalloc(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(b);
 178              k != unwritten_whiteouts_end(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(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_buf_bytes(b)
 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                 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_buf_bytes(b));
 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         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(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/%u",
 528                    b->written, btree_ptr_sectors_written(&b->key));
 529         if (i)
 530                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 531         prt_str(out, ": ");
 532 }
 533
 534 __printf(9, 10)
 535 static int __btree_err(int ret,
 536                        struct bch_fs *c,
 537                        struct bch_dev *ca,
 538                        struct btree *b,
 539                        struct bset *i,
 540                        int write,
 541                        bool have_retry,
 542                        enum bch_sb_error_id err_type,
 543                        const char *fmt, ...)
 544 {
 545         struct printbuf out = PRINTBUF;
 546         va_list args;
 547
 548         btree_err_msg(&out, c, ca, b, i, b->written, write);
 549
 550         va_start(args, fmt);
 551         prt_vprintf(&out, fmt, args);
 552         va_end(args);
 553
 554         if (write == WRITE) {
 555                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 556                 ret = c->opts.errors == BCH_ON_ERROR_continue
 557                         ? 0
 558                         : -BCH_ERR_fsck_errors_not_fixed;
 559                 goto out;
 560         }
 561
 562         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
 563                 ret = -BCH_ERR_btree_node_read_err_fixable;
 564         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
 565                 ret = -BCH_ERR_btree_node_read_err_bad_node;
 566
 567         if (ret != -BCH_ERR_btree_node_read_err_fixable)
 568                 bch2_sb_error_count(c, err_type);
 569
 570         switch (ret) {
 571         case -BCH_ERR_btree_node_read_err_fixable:
 572                 ret = bch2_fsck_err(c, FSCK_CAN_FIX, err_type, "%s", out.buf);
 573                 if (ret != -BCH_ERR_fsck_fix &&
 574                     ret != -BCH_ERR_fsck_ignore)
 575                         goto fsck_err;
 576                 ret = -BCH_ERR_fsck_fix;
 577                 break;
 578         case -BCH_ERR_btree_node_read_err_want_retry:
 579         case -BCH_ERR_btree_node_read_err_must_retry:
 580                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 581                 break;
 582         case -BCH_ERR_btree_node_read_err_bad_node:
 583                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 584                 ret = bch2_topology_error(c);
 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 bool __bkey_valid(struct bch_fs *c, struct btree *b,
 834                          struct bset *i, struct bkey_packed *k)
 835 {
 836         if (bkey_p_next(k) > vstruct_last(i))
 837                 return false;
 838
 839         if (k->format > KEY_FORMAT_CURRENT)
 840                 return false;
 841
 842         struct printbuf buf = PRINTBUF;
 843         struct bkey tmp;
 844         struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 845         bool ret = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b), READ, &buf);
 846         printbuf_exit(&buf);
 847         return ret;
 848 }
 849
 850 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 851                          struct bset *i, int write,
 852                          bool have_retry, bool *saw_error)
 853 {
 854         unsigned version = le16_to_cpu(i->version);
 855         struct bkey_packed *k, *prev = NULL;
 856         struct printbuf buf = PRINTBUF;
 857         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 858                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 859         int ret = 0;
 860
 861         for (k = i->start;
 862              k != vstruct_last(i);) {
 863                 struct bkey_s u;
 864                 struct bkey tmp;
 865                 unsigned next_good_key;
 866
 867                 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
 868                                  -BCH_ERR_btree_node_read_err_fixable,
 869                                  c, NULL, b, i,
 870                                  btree_node_bkey_past_bset_end,
 871                                  "key extends past end of bset")) {
 872                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 873                         break;
 874                 }
 875
 876                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 877                                  -BCH_ERR_btree_node_read_err_fixable,
 878                                  c, NULL, b, i,
 879                                  btree_node_bkey_bad_format,
 880                                  "invalid bkey format %u", k->format))
 881                         goto drop_this_key;
 882
 883                 /* XXX: validate k->u64s */
 884                 if (!write)
 885                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 886                                     BSET_BIG_ENDIAN(i), write,
 887                                     &b->format, k);
 888
 889                 u = __bkey_disassemble(b, k, &tmp);
 890
 891                 printbuf_reset(&buf);
 892                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 893                         printbuf_reset(&buf);
 894                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 895                         prt_printf(&buf, "\n  ");
 896                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 897
 898                         btree_err(-BCH_ERR_btree_node_read_err_fixable,
 899                                   c, NULL, b, i,
 900                                   btree_node_bad_bkey,
 901                                   "invalid bkey: %s", buf.buf);
 902                         goto drop_this_key;
 903                 }
 904
 905                 if (write)
 906                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 907                                     BSET_BIG_ENDIAN(i), write,
 908                                     &b->format, k);
 909
 910                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 911                         struct bkey up = bkey_unpack_key(b, prev);
 912
 913                         printbuf_reset(&buf);
 914                         prt_printf(&buf, "keys out of order: ");
 915                         bch2_bkey_to_text(&buf, &up);
 916                         prt_printf(&buf, " > ");
 917                         bch2_bkey_to_text(&buf, u.k);
 918
 919                         if (btree_err(-BCH_ERR_btree_node_read_err_fixable,
 920                                       c, NULL, b, i,
 921                                       btree_node_bkey_out_of_order,
 922                                       "%s", buf.buf))
 923                                 goto drop_this_key;
 924                 }
 925
 926                 prev = k;
 927                 k = bkey_p_next(k);
 928                 continue;
 929 drop_this_key:
 930                 next_good_key = k->u64s;
 931
 932                 if (!next_good_key ||
 933                     (BSET_BIG_ENDIAN(i) == CPU_BIG_ENDIAN &&
 934                      version >= bcachefs_metadata_version_snapshot)) {
 935                         /*
 936                          * only do scanning if bch2_bkey_compat() has nothing to
 937                          * do
 938                          */
 939
 940                         if (!__bkey_valid(c, b, i, (void *) ((u64 *) k + next_good_key))) {
 941                                 for (next_good_key = 1;
 942                                      next_good_key < (u64 *) vstruct_last(i) - (u64 *) k;
 943                                      next_good_key++)
 944                                         if (__bkey_valid(c, b, i, (void *) ((u64 *) k + next_good_key)))
 945                                                 goto got_good_key;
 946
 947                         }
 948
 949                         /*
 950                          * didn't find a good key, have to truncate the rest of
 951                          * the bset
 952                          */
 953                         next_good_key = (u64 *) vstruct_last(i) - (u64 *) k;
 954                 }
 955 got_good_key:
 956                 le16_add_cpu(&i->u64s, -next_good_key);
 957                 memmove_u64s_down(k, bkey_p_next(k), (u64 *) vstruct_end(i) - (u64 *) k);
 958         }
 959 fsck_err:
 960         printbuf_exit(&buf);
 961         return ret;
 962 }
 963
 964 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 965                               struct btree *b, bool have_retry, bool *saw_error)
 966 {
 967         struct btree_node_entry *bne;
 968         struct sort_iter *iter;
 969         struct btree_node *sorted;
 970         struct bkey_packed *k;
 971         struct bset *i;
 972         bool used_mempool, blacklisted;
 973         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 974                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 975         unsigned u64s;
 976         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 977         struct printbuf buf = PRINTBUF;
 978         int ret = 0, retry_read = 0, write = READ;
 979         u64 start_time = local_clock();
 980
 981         b->version_ondisk = U16_MAX;
 982         /* We might get called multiple times on read retry: */
 983         b->written = 0;
 984
 985         iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
 986         sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2);
 987
 988         if (bch2_meta_read_fault("btree"))
 989                 btree_err(-BCH_ERR_btree_node_read_err_must_retry,
 990                           c, ca, b, NULL,
 991                           btree_node_fault_injected,
 992                           "dynamic fault");
 993
 994         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 995                      -BCH_ERR_btree_node_read_err_must_retry,
 996                      c, ca, b, NULL,
 997                      btree_node_bad_magic,
 998                      "bad magic: want %llx, got %llx",
 999                      bset_magic(c), le64_to_cpu(b->data->magic));
1000
1001         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
1002                 struct bch_btree_ptr_v2 *bp =
1003                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
1004
1005                 bch2_bpos_to_text(&buf, b->data->min_key);
1006                 prt_str(&buf, "-");
1007                 bch2_bpos_to_text(&buf, b->data->max_key);
1008
1009                 btree_err_on(b->data->keys.seq != bp->seq,
1010                              -BCH_ERR_btree_node_read_err_must_retry,
1011                              c, ca, b, NULL,
1012                              btree_node_bad_seq,
1013                              "got wrong btree node (want %llx got %llx)\n"
1014                              "got btree %s level %llu pos %s",
1015                              bp->seq, b->data->keys.seq,
1016                              bch2_btree_id_str(BTREE_NODE_ID(b->data)),
1017                              BTREE_NODE_LEVEL(b->data),
1018                              buf.buf);
1019         } else {
1020                 btree_err_on(!b->data->keys.seq,
1021                              -BCH_ERR_btree_node_read_err_must_retry,
1022                              c, ca, b, NULL,
1023                              btree_node_bad_seq,
1024                              "bad btree header: seq 0");
1025         }
1026
1027         while (b->written < (ptr_written ?: btree_sectors(c))) {
1028                 unsigned sectors;
1029                 struct nonce nonce;
1030                 bool first = !b->written;
1031                 bool csum_bad;
1032
1033                 if (!b->written) {
1034                         i = &b->data->keys;
1035
1036                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1037                                      -BCH_ERR_btree_node_read_err_want_retry,
1038                                      c, ca, b, i,
1039                                      bset_unknown_csum,
1040                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1041
1042                         nonce = btree_nonce(i, b->written << 9);
1043
1044                         struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
1045                         csum_bad = bch2_crc_cmp(b->data->csum, csum);
1046                         if (csum_bad)
1047                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1048
1049                         btree_err_on(csum_bad,
1050                                      -BCH_ERR_btree_node_read_err_want_retry,
1051                                      c, ca, b, i,
1052                                      bset_bad_csum,
1053                                      "%s",
1054                                      (printbuf_reset(&buf),
1055                                       bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), b->data->csum, csum),
1056                                       buf.buf));
1057
1058                         ret = bset_encrypt(c, i, b->written << 9);
1059                         if (bch2_fs_fatal_err_on(ret, c,
1060                                         "error decrypting btree node: %i", ret))
1061                                 goto fsck_err;
1062
1063                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
1064                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
1065                                      -BCH_ERR_btree_node_read_err_incompatible,
1066                                      c, NULL, b, NULL,
1067                                      btree_node_unsupported_version,
1068                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
1069
1070                         sectors = vstruct_sectors(b->data, c->block_bits);
1071                 } else {
1072                         bne = write_block(b);
1073                         i = &bne->keys;
1074
1075                         if (i->seq != b->data->keys.seq)
1076                                 break;
1077
1078                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1079                                      -BCH_ERR_btree_node_read_err_want_retry,
1080                                      c, ca, b, i,
1081                                      bset_unknown_csum,
1082                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1083
1084                         nonce = btree_nonce(i, b->written << 9);
1085                         struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1086                         csum_bad = bch2_crc_cmp(bne->csum, csum);
1087                         if (csum_bad)
1088                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1089
1090                         btree_err_on(csum_bad,
1091                                      -BCH_ERR_btree_node_read_err_want_retry,
1092                                      c, ca, b, i,
1093                                      bset_bad_csum,
1094                                      "%s",
1095                                      (printbuf_reset(&buf),
1096                                       bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), bne->csum, csum),
1097                                       buf.buf));
1098
1099                         ret = bset_encrypt(c, i, b->written << 9);
1100                         if (bch2_fs_fatal_err_on(ret, c,
1101                                         "error decrypting btree node: %i\n", ret))
1102                                 goto fsck_err;
1103
1104                         sectors = vstruct_sectors(bne, c->block_bits);
1105                 }
1106
1107                 b->version_ondisk = min(b->version_ondisk,
1108                                         le16_to_cpu(i->version));
1109
1110                 ret = validate_bset(c, ca, b, i, b->written, sectors,
1111                                     READ, have_retry, saw_error);
1112                 if (ret)
1113                         goto fsck_err;
1114
1115                 if (!b->written)
1116                         btree_node_set_format(b, b->data->format);
1117
1118                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1119                 if (ret)
1120                         goto fsck_err;
1121
1122                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1123
1124                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1125                                         le64_to_cpu(i->journal_seq),
1126                                         true);
1127
1128                 btree_err_on(blacklisted && first,
1129                              -BCH_ERR_btree_node_read_err_fixable,
1130                              c, ca, b, i,
1131                              bset_blacklisted_journal_seq,
1132                              "first btree node bset has blacklisted journal seq (%llu)",
1133                              le64_to_cpu(i->journal_seq));
1134
1135                 btree_err_on(blacklisted && ptr_written,
1136                              -BCH_ERR_btree_node_read_err_fixable,
1137                              c, ca, b, i,
1138                              first_bset_blacklisted_journal_seq,
1139                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1140                              le64_to_cpu(i->journal_seq),
1141                              b->written, b->written + sectors, ptr_written);
1142
1143                 b->written += sectors;
1144
1145                 if (blacklisted && !first)
1146                         continue;
1147
1148                 sort_iter_add(iter,
1149                               vstruct_idx(i, 0),
1150                               vstruct_last(i));
1151         }
1152
1153         if (ptr_written) {
1154                 btree_err_on(b->written < ptr_written,
1155                              -BCH_ERR_btree_node_read_err_want_retry,
1156                              c, ca, b, NULL,
1157                              btree_node_data_missing,
1158                              "btree node data missing: expected %u sectors, found %u",
1159                              ptr_written, b->written);
1160         } else {
1161                 for (bne = write_block(b);
1162                      bset_byte_offset(b, bne) < btree_buf_bytes(b);
1163                      bne = (void *) bne + block_bytes(c))
1164                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1165                                      !bch2_journal_seq_is_blacklisted(c,
1166                                                                       le64_to_cpu(bne->keys.journal_seq),
1167                                                                       true),
1168                                      -BCH_ERR_btree_node_read_err_want_retry,
1169                                      c, ca, b, NULL,
1170                                      btree_node_bset_after_end,
1171                                      "found bset signature after last bset");
1172         }
1173
1174         sorted = btree_bounce_alloc(c, btree_buf_bytes(b), &used_mempool);
1175         sorted->keys.u64s = 0;
1176
1177         set_btree_bset(b, b->set, &b->data->keys);
1178
1179         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1180
1181         u64s = le16_to_cpu(sorted->keys.u64s);
1182         *sorted = *b->data;
1183         sorted->keys.u64s = cpu_to_le16(u64s);
1184         swap(sorted, b->data);
1185         set_btree_bset(b, b->set, &b->data->keys);
1186         b->nsets = 1;
1187
1188         BUG_ON(b->nr.live_u64s != u64s);
1189
1190         btree_bounce_free(c, btree_buf_bytes(b), used_mempool, sorted);
1191
1192         if (updated_range)
1193                 bch2_btree_node_drop_keys_outside_node(b);
1194
1195         i = &b->data->keys;
1196         for (k = i->start; k != vstruct_last(i);) {
1197                 struct bkey tmp;
1198                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1199
1200                 printbuf_reset(&buf);
1201
1202                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1203                     (bch2_inject_invalid_keys &&
1204                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1205                         printbuf_reset(&buf);
1206
1207                         prt_printf(&buf, "invalid bkey: ");
1208                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1209                         prt_printf(&buf, "\n  ");
1210                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1211
1212                         btree_err(-BCH_ERR_btree_node_read_err_fixable,
1213                                   c, NULL, b, i,
1214                                   btree_node_bad_bkey,
1215                                   "%s", buf.buf);
1216
1217                         btree_keys_account_key_drop(&b->nr, 0, k);
1218
1219                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1220                         memmove_u64s_down(k, bkey_p_next(k),
1221                                           (u64 *) vstruct_end(i) - (u64 *) k);
1222                         set_btree_bset_end(b, b->set);
1223                         continue;
1224                 }
1225
1226                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1227                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1228
1229                         bp.v->mem_ptr = 0;
1230                 }
1231
1232                 k = bkey_p_next(k);
1233         }
1234
1235         bch2_bset_build_aux_tree(b, b->set, false);
1236
1237         set_needs_whiteout(btree_bset_first(b), true);
1238
1239         btree_node_reset_sib_u64s(b);
1240
1241         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1242                 struct bch_dev *ca2 = bch_dev_bkey_exists(c, ptr->dev);
1243
1244                 if (ca2->mi.state != BCH_MEMBER_STATE_rw)
1245                         set_btree_node_need_rewrite(b);
1246         }
1247
1248         if (!ptr_written)
1249                 set_btree_node_need_rewrite(b);
1250 out:
1251         mempool_free(iter, &c->fill_iter);
1252         printbuf_exit(&buf);
1253         time_stats_update(&c->times[BCH_TIME_btree_node_read_done], start_time);
1254         return retry_read;
1255 fsck_err:
1256         if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1257             ret == -BCH_ERR_btree_node_read_err_must_retry)
1258                 retry_read = 1;
1259         else
1260                 set_btree_node_read_error(b);
1261         goto out;
1262 }
1263
1264 static void btree_node_read_work(struct work_struct *work)
1265 {
1266         struct btree_read_bio *rb =
1267                 container_of(work, struct btree_read_bio, work);
1268         struct bch_fs *c        = rb->c;
1269         struct btree *b         = rb->b;
1270         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1271         struct bio *bio         = &rb->bio;
1272         struct bch_io_failures failed = { .nr = 0 };
1273         struct printbuf buf = PRINTBUF;
1274         bool saw_error = false;
1275         bool retry = false;
1276         bool can_retry;
1277
1278         goto start;
1279         while (1) {
1280                 retry = true;
1281                 bch_info(c, "retrying read");
1282                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1283                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1284                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1285                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1286                 bio->bi_iter.bi_size    = btree_buf_bytes(b);
1287
1288                 if (rb->have_ioref) {
1289                         bio_set_dev(bio, ca->disk_sb.bdev);
1290                         submit_bio_wait(bio);
1291                 } else {
1292                         bio->bi_status = BLK_STS_REMOVED;
1293                 }
1294 start:
1295                 printbuf_reset(&buf);
1296                 bch2_btree_pos_to_text(&buf, c, b);
1297                 bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_read,
1298                                    "btree read error %s for %s",
1299                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1300                 if (rb->have_ioref)
1301                         percpu_ref_put(&ca->io_ref);
1302                 rb->have_ioref = false;
1303
1304                 bch2_mark_io_failure(&failed, &rb->pick);
1305
1306                 can_retry = bch2_bkey_pick_read_device(c,
1307                                 bkey_i_to_s_c(&b->key),
1308                                 &failed, &rb->pick) > 0;
1309
1310                 if (!bio->bi_status &&
1311                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1312                         if (retry)
1313                                 bch_info(c, "retry success");
1314                         break;
1315                 }
1316
1317                 saw_error = true;
1318
1319                 if (!can_retry) {
1320                         set_btree_node_read_error(b);
1321                         break;
1322                 }
1323         }
1324
1325         time_stats_update(&c->times[BCH_TIME_btree_node_read],
1326                                rb->start_time);
1327         bio_put(&rb->bio);
1328
1329         if (saw_error && !btree_node_read_error(b)) {
1330                 printbuf_reset(&buf);
1331                 bch2_bpos_to_text(&buf, b->key.k.p);
1332                 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1333                          __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf);
1334
1335                 bch2_btree_node_rewrite_async(c, b);
1336         }
1337
1338         printbuf_exit(&buf);
1339         clear_btree_node_read_in_flight(b);
1340         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1341 }
1342
1343 static void btree_node_read_endio(struct bio *bio)
1344 {
1345         struct btree_read_bio *rb =
1346                 container_of(bio, struct btree_read_bio, bio);
1347         struct bch_fs *c        = rb->c;
1348
1349         if (rb->have_ioref) {
1350                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1351
1352                 bch2_latency_acct(ca, rb->start_time, READ);
1353         }
1354
1355         queue_work(c->io_complete_wq, &rb->work);
1356 }
1357
1358 struct btree_node_read_all {
1359         struct closure          cl;
1360         struct bch_fs           *c;
1361         struct btree            *b;
1362         unsigned                nr;
1363         void                    *buf[BCH_REPLICAS_MAX];
1364         struct bio              *bio[BCH_REPLICAS_MAX];
1365         blk_status_t            err[BCH_REPLICAS_MAX];
1366 };
1367
1368 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1369 {
1370         struct btree_node *bn = data;
1371         struct btree_node_entry *bne;
1372         unsigned offset = 0;
1373
1374         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1375                 return 0;
1376
1377         while (offset < btree_sectors(c)) {
1378                 if (!offset) {
1379                         offset += vstruct_sectors(bn, c->block_bits);
1380                 } else {
1381                         bne = data + (offset << 9);
1382                         if (bne->keys.seq != bn->keys.seq)
1383                                 break;
1384                         offset += vstruct_sectors(bne, c->block_bits);
1385                 }
1386         }
1387
1388         return offset;
1389 }
1390
1391 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1392 {
1393         struct btree_node *bn = data;
1394         struct btree_node_entry *bne;
1395
1396         if (!offset)
1397                 return false;
1398
1399         while (offset < btree_sectors(c)) {
1400                 bne = data + (offset << 9);
1401                 if (bne->keys.seq == bn->keys.seq)
1402                         return true;
1403                 offset++;
1404         }
1405
1406         return false;
1407         return offset;
1408 }
1409
1410 static CLOSURE_CALLBACK(btree_node_read_all_replicas_done)
1411 {
1412         closure_type(ra, struct btree_node_read_all, cl);
1413         struct bch_fs *c = ra->c;
1414         struct btree *b = ra->b;
1415         struct printbuf buf = PRINTBUF;
1416         bool dump_bset_maps = false;
1417         bool have_retry = false;
1418         int ret = 0, best = -1, write = READ;
1419         unsigned i, written = 0, written2 = 0;
1420         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1421                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1422         bool _saw_error = false, *saw_error = &_saw_error;
1423
1424         for (i = 0; i < ra->nr; i++) {
1425                 struct btree_node *bn = ra->buf[i];
1426
1427                 if (ra->err[i])
1428                         continue;
1429
1430                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1431                     (seq && seq != bn->keys.seq))
1432                         continue;
1433
1434                 if (best < 0) {
1435                         best = i;
1436                         written = btree_node_sectors_written(c, bn);
1437                         continue;
1438                 }
1439
1440                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1441                 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable,
1442                                  c, NULL, b, NULL,
1443                                  btree_node_replicas_sectors_written_mismatch,
1444                                  "btree node sectors written mismatch: %u != %u",
1445                                  written, written2) ||
1446                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1447                                  -BCH_ERR_btree_node_read_err_fixable,
1448                                  c, NULL, b, NULL,
1449                                  btree_node_bset_after_end,
1450                                  "found bset signature after last bset") ||
1451                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1452                                  -BCH_ERR_btree_node_read_err_fixable,
1453                                  c, NULL, b, NULL,
1454                                  btree_node_replicas_data_mismatch,
1455                                  "btree node replicas content mismatch"))
1456                         dump_bset_maps = true;
1457
1458                 if (written2 > written) {
1459                         written = written2;
1460                         best = i;
1461                 }
1462         }
1463 fsck_err:
1464         if (dump_bset_maps) {
1465                 for (i = 0; i < ra->nr; i++) {
1466                         struct btree_node *bn = ra->buf[i];
1467                         struct btree_node_entry *bne = NULL;
1468                         unsigned offset = 0, sectors;
1469                         bool gap = false;
1470
1471                         if (ra->err[i])
1472                                 continue;
1473
1474                         printbuf_reset(&buf);
1475
1476                         while (offset < btree_sectors(c)) {
1477                                 if (!offset) {
1478                                         sectors = vstruct_sectors(bn, c->block_bits);
1479                                 } else {
1480                                         bne = ra->buf[i] + (offset << 9);
1481                                         if (bne->keys.seq != bn->keys.seq)
1482                                                 break;
1483                                         sectors = vstruct_sectors(bne, c->block_bits);
1484                                 }
1485
1486                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1487                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1488                                                         le64_to_cpu(bne->keys.journal_seq), false))
1489                                         prt_printf(&buf, "*");
1490                                 offset += sectors;
1491                         }
1492
1493                         while (offset < btree_sectors(c)) {
1494                                 bne = ra->buf[i] + (offset << 9);
1495                                 if (bne->keys.seq == bn->keys.seq) {
1496                                         if (!gap)
1497                                                 prt_printf(&buf, " GAP");
1498                                         gap = true;
1499
1500                                         sectors = vstruct_sectors(bne, c->block_bits);
1501                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1502                                         if (bch2_journal_seq_is_blacklisted(c,
1503                                                         le64_to_cpu(bne->keys.journal_seq), false))
1504                                                 prt_printf(&buf, "*");
1505                                 }
1506                                 offset++;
1507                         }
1508
1509                         bch_err(c, "replica %u:%s", i, buf.buf);
1510                 }
1511         }
1512
1513         if (best >= 0) {
1514                 memcpy(b->data, ra->buf[best], btree_buf_bytes(b));
1515                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1516         } else {
1517                 ret = -1;
1518         }
1519
1520         if (ret)
1521                 set_btree_node_read_error(b);
1522         else if (*saw_error)
1523                 bch2_btree_node_rewrite_async(c, b);
1524
1525         for (i = 0; i < ra->nr; i++) {
1526                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1527                 bio_put(ra->bio[i]);
1528         }
1529
1530         closure_debug_destroy(&ra->cl);
1531         kfree(ra);
1532         printbuf_exit(&buf);
1533
1534         clear_btree_node_read_in_flight(b);
1535         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1536 }
1537
1538 static void btree_node_read_all_replicas_endio(struct bio *bio)
1539 {
1540         struct btree_read_bio *rb =
1541                 container_of(bio, struct btree_read_bio, bio);
1542         struct bch_fs *c        = rb->c;
1543         struct btree_node_read_all *ra = rb->ra;
1544
1545         if (rb->have_ioref) {
1546                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1547
1548                 bch2_latency_acct(ca, rb->start_time, READ);
1549         }
1550
1551         ra->err[rb->idx] = bio->bi_status;
1552         closure_put(&ra->cl);
1553 }
1554
1555 /*
1556  * XXX This allocates multiple times from the same mempools, and can deadlock
1557  * under sufficient memory pressure (but is only a debug path)
1558  */
1559 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1560 {
1561         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1562         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1563         const union bch_extent_entry *entry;
1564         struct extent_ptr_decoded pick;
1565         struct btree_node_read_all *ra;
1566         unsigned i;
1567
1568         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1569         if (!ra)
1570                 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1571
1572         closure_init(&ra->cl, NULL);
1573         ra->c   = c;
1574         ra->b   = b;
1575         ra->nr  = bch2_bkey_nr_ptrs(k);
1576
1577         for (i = 0; i < ra->nr; i++) {
1578                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1579                 ra->bio[i] = bio_alloc_bioset(NULL,
1580                                               buf_pages(ra->buf[i], btree_buf_bytes(b)),
1581                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1582                                               GFP_NOFS,
1583                                               &c->btree_bio);
1584         }
1585
1586         i = 0;
1587         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1588                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1589                 struct btree_read_bio *rb =
1590                         container_of(ra->bio[i], struct btree_read_bio, bio);
1591                 rb->c                   = c;
1592                 rb->b                   = b;
1593                 rb->ra                  = ra;
1594                 rb->start_time          = local_clock();
1595                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1596                 rb->idx                 = i;
1597                 rb->pick                = pick;
1598                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1599                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1600                 bch2_bio_map(&rb->bio, ra->buf[i], btree_buf_bytes(b));
1601
1602                 if (rb->have_ioref) {
1603                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1604                                      bio_sectors(&rb->bio));
1605                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1606
1607                         closure_get(&ra->cl);
1608                         submit_bio(&rb->bio);
1609                 } else {
1610                         ra->err[i] = BLK_STS_REMOVED;
1611                 }
1612
1613                 i++;
1614         }
1615
1616         if (sync) {
1617                 closure_sync(&ra->cl);
1618                 btree_node_read_all_replicas_done(&ra->cl.work);
1619         } else {
1620                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1621                             c->io_complete_wq);
1622         }
1623
1624         return 0;
1625 }
1626
1627 void bch2_btree_node_read(struct btree_trans *trans, struct btree *b,
1628                           bool sync)
1629 {
1630         struct bch_fs *c = trans->c;
1631         struct extent_ptr_decoded pick;
1632         struct btree_read_bio *rb;
1633         struct bch_dev *ca;
1634         struct bio *bio;
1635         int ret;
1636
1637         trace_and_count(c, btree_node_read, trans, b);
1638
1639         if (bch2_verify_all_btree_replicas &&
1640             !btree_node_read_all_replicas(c, b, sync))
1641                 return;
1642
1643         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1644                                          NULL, &pick);
1645
1646         if (ret <= 0) {
1647                 struct printbuf buf = PRINTBUF;
1648
1649                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1650                 bch2_btree_pos_to_text(&buf, c, b);
1651                 bch_err(c, "%s", buf.buf);
1652
1653                 if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1654                     c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1655                         bch2_fatal_error(c);
1656
1657                 set_btree_node_read_error(b);
1658                 clear_btree_node_read_in_flight(b);
1659                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1660                 printbuf_exit(&buf);
1661                 return;
1662         }
1663
1664         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1665
1666         bio = bio_alloc_bioset(NULL,
1667                                buf_pages(b->data, btree_buf_bytes(b)),
1668                                REQ_OP_READ|REQ_SYNC|REQ_META,
1669                                GFP_NOFS,
1670                                &c->btree_bio);
1671         rb = container_of(bio, struct btree_read_bio, bio);
1672         rb->c                   = c;
1673         rb->b                   = b;
1674         rb->ra                  = NULL;
1675         rb->start_time          = local_clock();
1676         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1677         rb->pick                = pick;
1678         INIT_WORK(&rb->work, btree_node_read_work);
1679         bio->bi_iter.bi_sector  = pick.ptr.offset;
1680         bio->bi_end_io          = btree_node_read_endio;
1681         bch2_bio_map(bio, b->data, btree_buf_bytes(b));
1682
1683         if (rb->have_ioref) {
1684                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1685                              bio_sectors(bio));
1686                 bio_set_dev(bio, ca->disk_sb.bdev);
1687
1688                 if (sync) {
1689                         submit_bio_wait(bio);
1690                         bch2_latency_acct(ca, rb->start_time, READ);
1691                         btree_node_read_work(&rb->work);
1692                 } else {
1693                         submit_bio(bio);
1694                 }
1695         } else {
1696                 bio->bi_status = BLK_STS_REMOVED;
1697
1698                 if (sync)
1699                         btree_node_read_work(&rb->work);
1700                 else
1701                         queue_work(c->io_complete_wq, &rb->work);
1702         }
1703 }
1704
1705 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1706                                   const struct bkey_i *k, unsigned level)
1707 {
1708         struct bch_fs *c = trans->c;
1709         struct closure cl;
1710         struct btree *b;
1711         int ret;
1712
1713         closure_init_stack(&cl);
1714
1715         do {
1716                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
1717                 closure_sync(&cl);
1718         } while (ret);
1719
1720         b = bch2_btree_node_mem_alloc(trans, level != 0);
1721         bch2_btree_cache_cannibalize_unlock(trans);
1722
1723         BUG_ON(IS_ERR(b));
1724
1725         bkey_copy(&b->key, k);
1726         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1727
1728         set_btree_node_read_in_flight(b);
1729
1730         bch2_btree_node_read(trans, b, true);
1731
1732         if (btree_node_read_error(b)) {
1733                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1734
1735                 mutex_lock(&c->btree_cache.lock);
1736                 list_move(&b->list, &c->btree_cache.freeable);
1737                 mutex_unlock(&c->btree_cache.lock);
1738
1739                 ret = -BCH_ERR_btree_node_read_error;
1740                 goto err;
1741         }
1742
1743         bch2_btree_set_root_for_read(c, b);
1744 err:
1745         six_unlock_write(&b->c.lock);
1746         six_unlock_intent(&b->c.lock);
1747
1748         return ret;
1749 }
1750
1751 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1752                         const struct bkey_i *k, unsigned level)
1753 {
1754         return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
1755 }
1756
1757 static void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1758                                       struct btree_write *w)
1759 {
1760         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1761
1762         do {
1763                 old = new = v;
1764                 if (!(old & 1))
1765                         break;
1766
1767                 new &= ~1UL;
1768         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1769
1770         if (old & 1)
1771                 closure_put(&((struct btree_update *) new)->cl);
1772
1773         bch2_journal_pin_drop(&c->journal, &w->journal);
1774 }
1775
1776 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1777 {
1778         struct btree_write *w = btree_prev_write(b);
1779         unsigned long old, new, v;
1780         unsigned type = 0;
1781
1782         bch2_btree_complete_write(c, b, w);
1783
1784         v = READ_ONCE(b->flags);
1785         do {
1786                 old = new = v;
1787
1788                 if ((old & (1U << BTREE_NODE_dirty)) &&
1789                     (old & (1U << BTREE_NODE_need_write)) &&
1790                     !(old & (1U << BTREE_NODE_never_write)) &&
1791                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1792                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1793                         new &= ~(1U << BTREE_NODE_dirty);
1794                         new &= ~(1U << BTREE_NODE_need_write);
1795                         new |=  (1U << BTREE_NODE_write_in_flight);
1796                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1797                         new |=  (1U << BTREE_NODE_just_written);
1798                         new ^=  (1U << BTREE_NODE_write_idx);
1799
1800                         type = new & BTREE_WRITE_TYPE_MASK;
1801                         new &= ~BTREE_WRITE_TYPE_MASK;
1802                 } else {
1803                         new &= ~(1U << BTREE_NODE_write_in_flight);
1804                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1805                 }
1806         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1807
1808         if (new & (1U << BTREE_NODE_write_in_flight))
1809                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1810         else
1811                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1812 }
1813
1814 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1815 {
1816         struct btree_trans *trans = bch2_trans_get(c);
1817
1818         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
1819         __btree_node_write_done(c, b);
1820         six_unlock_read(&b->c.lock);
1821
1822         bch2_trans_put(trans);
1823 }
1824
1825 static void btree_node_write_work(struct work_struct *work)
1826 {
1827         struct btree_write_bio *wbio =
1828                 container_of(work, struct btree_write_bio, work);
1829         struct bch_fs *c        = wbio->wbio.c;
1830         struct btree *b         = wbio->wbio.bio.bi_private;
1831         struct bch_extent_ptr *ptr;
1832         int ret = 0;
1833
1834         btree_bounce_free(c,
1835                 wbio->data_bytes,
1836                 wbio->wbio.used_mempool,
1837                 wbio->data);
1838
1839         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1840                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1841
1842         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key))) {
1843                 ret = -BCH_ERR_btree_node_write_all_failed;
1844                 goto err;
1845         }
1846
1847         if (wbio->wbio.first_btree_write) {
1848                 if (wbio->wbio.failed.nr) {
1849
1850                 }
1851         } else {
1852                 ret = bch2_trans_do(c, NULL, NULL, 0,
1853                         bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
1854                                         BCH_WATERMARK_reclaim|
1855                                         BCH_TRANS_COMMIT_journal_reclaim|
1856                                         BCH_TRANS_COMMIT_no_enospc|
1857                                         BCH_TRANS_COMMIT_no_check_rw,
1858                                         !wbio->wbio.failed.nr));
1859                 if (ret)
1860                         goto err;
1861         }
1862 out:
1863         bio_put(&wbio->wbio.bio);
1864         btree_node_write_done(c, b);
1865         return;
1866 err:
1867         set_btree_node_noevict(b);
1868         if (!bch2_err_matches(ret, EROFS))
1869                 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1870         goto out;
1871 }
1872
1873 static void btree_node_write_endio(struct bio *bio)
1874 {
1875         struct bch_write_bio *wbio      = to_wbio(bio);
1876         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1877         struct bch_write_bio *orig      = parent ?: wbio;
1878         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1879         struct bch_fs *c                = wbio->c;
1880         struct btree *b                 = wbio->bio.bi_private;
1881         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1882         unsigned long flags;
1883
1884         if (wbio->have_ioref)
1885                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1886
1887         if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1888                                "btree write error: %s",
1889                                bch2_blk_status_to_str(bio->bi_status)) ||
1890             bch2_meta_write_fault("btree")) {
1891                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1892                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1893                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1894         }
1895
1896         if (wbio->have_ioref)
1897                 percpu_ref_put(&ca->io_ref);
1898
1899         if (parent) {
1900                 bio_put(bio);
1901                 bio_endio(&parent->bio);
1902                 return;
1903         }
1904
1905         clear_btree_node_write_in_flight_inner(b);
1906         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1907         INIT_WORK(&wb->work, btree_node_write_work);
1908         queue_work(c->btree_io_complete_wq, &wb->work);
1909 }
1910
1911 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1912                                    struct bset *i, unsigned sectors)
1913 {
1914         struct printbuf buf = PRINTBUF;
1915         bool saw_error;
1916         int ret;
1917
1918         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1919                                 BKEY_TYPE_btree, WRITE, &buf);
1920
1921         if (ret)
1922                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1923         printbuf_exit(&buf);
1924         if (ret)
1925                 return ret;
1926
1927         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1928                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1929         if (ret) {
1930                 bch2_inconsistent_error(c);
1931                 dump_stack();
1932         }
1933
1934         return ret;
1935 }
1936
1937 static void btree_write_submit(struct work_struct *work)
1938 {
1939         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1940         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1941
1942         bkey_copy(&tmp.k, &wbio->key);
1943
1944         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1945                 ptr->offset += wbio->sector_offset;
1946
1947         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1948                                   &tmp.k, false);
1949 }
1950
1951 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1952 {
1953         struct btree_write_bio *wbio;
1954         struct bset_tree *t;
1955         struct bset *i;
1956         struct btree_node *bn = NULL;
1957         struct btree_node_entry *bne = NULL;
1958         struct sort_iter_stack sort_iter;
1959         struct nonce nonce;
1960         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1961         u64 seq = 0;
1962         bool used_mempool;
1963         unsigned long old, new;
1964         bool validate_before_checksum = false;
1965         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1966         void *data;
1967         int ret;
1968
1969         if (flags & BTREE_WRITE_ALREADY_STARTED)
1970                 goto do_write;
1971
1972         /*
1973          * We may only have a read lock on the btree node - the dirty bit is our
1974          * "lock" against racing with other threads that may be trying to start
1975          * a write, we do a write iff we clear the dirty bit. Since setting the
1976          * dirty bit requires a write lock, we can't race with other threads
1977          * redirtying it:
1978          */
1979         do {
1980                 old = new = READ_ONCE(b->flags);
1981
1982                 if (!(old & (1 << BTREE_NODE_dirty)))
1983                         return;
1984
1985                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1986                     !(old & (1 << BTREE_NODE_need_write)))
1987                         return;
1988
1989                 if (old &
1990                     ((1 << BTREE_NODE_never_write)|
1991                      (1 << BTREE_NODE_write_blocked)))
1992                         return;
1993
1994                 if (b->written &&
1995                     (old & (1 << BTREE_NODE_will_make_reachable)))
1996                         return;
1997
1998                 if (old & (1 << BTREE_NODE_write_in_flight))
1999                         return;
2000
2001                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
2002                         type = new & BTREE_WRITE_TYPE_MASK;
2003                 new &= ~BTREE_WRITE_TYPE_MASK;
2004
2005                 new &= ~(1 << BTREE_NODE_dirty);
2006                 new &= ~(1 << BTREE_NODE_need_write);
2007                 new |=  (1 << BTREE_NODE_write_in_flight);
2008                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
2009                 new |=  (1 << BTREE_NODE_just_written);
2010                 new ^=  (1 << BTREE_NODE_write_idx);
2011         } while (cmpxchg_acquire(&b->flags, old, new) != old);
2012
2013         if (new & (1U << BTREE_NODE_need_write))
2014                 return;
2015 do_write:
2016         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
2017
2018         atomic_dec(&c->btree_cache.dirty);
2019
2020         BUG_ON(btree_node_fake(b));
2021         BUG_ON((b->will_make_reachable != 0) != !b->written);
2022
2023         BUG_ON(b->written >= btree_sectors(c));
2024         BUG_ON(b->written & (block_sectors(c) - 1));
2025         BUG_ON(bset_written(b, btree_bset_last(b)));
2026         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
2027         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
2028
2029         bch2_sort_whiteouts(c, b);
2030
2031         sort_iter_stack_init(&sort_iter, b);
2032
2033         bytes = !b->written
2034                 ? sizeof(struct btree_node)
2035                 : sizeof(struct btree_node_entry);
2036
2037         bytes += b->whiteout_u64s * sizeof(u64);
2038
2039         for_each_bset(b, t) {
2040                 i = bset(b, t);
2041
2042                 if (bset_written(b, i))
2043                         continue;
2044
2045                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
2046                 sort_iter_add(&sort_iter.iter,
2047                               btree_bkey_first(b, t),
2048                               btree_bkey_last(b, t));
2049                 seq = max(seq, le64_to_cpu(i->journal_seq));
2050         }
2051
2052         BUG_ON(b->written && !seq);
2053
2054         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
2055         bytes += 8;
2056
2057         /* buffer must be a multiple of the block size */
2058         bytes = round_up(bytes, block_bytes(c));
2059
2060         data = btree_bounce_alloc(c, bytes, &used_mempool);
2061
2062         if (!b->written) {
2063                 bn = data;
2064                 *bn = *b->data;
2065                 i = &bn->keys;
2066         } else {
2067                 bne = data;
2068                 bne->keys = b->data->keys;
2069                 i = &bne->keys;
2070         }
2071
2072         i->journal_seq  = cpu_to_le64(seq);
2073         i->u64s         = 0;
2074
2075         sort_iter_add(&sort_iter.iter,
2076                       unwritten_whiteouts_start(b),
2077                       unwritten_whiteouts_end(b));
2078         SET_BSET_SEPARATE_WHITEOUTS(i, false);
2079
2080         b->whiteout_u64s = 0;
2081
2082         u64s = bch2_sort_keys(i->start, &sort_iter.iter, false);
2083         le16_add_cpu(&i->u64s, u64s);
2084
2085         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
2086
2087         set_needs_whiteout(i, false);
2088
2089         /* do we have data to write? */
2090         if (b->written && !i->u64s)
2091                 goto nowrite;
2092
2093         bytes_to_write = vstruct_end(i) - data;
2094         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
2095
2096         if (!b->written &&
2097             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2098                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
2099
2100         memset(data + bytes_to_write, 0,
2101                (sectors_to_write << 9) - bytes_to_write);
2102
2103         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2104         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2105         BUG_ON(i->seq != b->data->keys.seq);
2106
2107         i->version = cpu_to_le16(c->sb.version);
2108         SET_BSET_OFFSET(i, b->written);
2109         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2110
2111         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2112                 validate_before_checksum = true;
2113
2114         /* validate_bset will be modifying: */
2115         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2116                 validate_before_checksum = true;
2117
2118         /* if we're going to be encrypting, check metadata validity first: */
2119         if (validate_before_checksum &&
2120             validate_bset_for_write(c, b, i, sectors_to_write))
2121                 goto err;
2122
2123         ret = bset_encrypt(c, i, b->written << 9);
2124         if (bch2_fs_fatal_err_on(ret, c,
2125                         "error encrypting btree node: %i\n", ret))
2126                 goto err;
2127
2128         nonce = btree_nonce(i, b->written << 9);
2129
2130         if (bn)
2131                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2132         else
2133                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2134
2135         /* if we're not encrypting, check metadata after checksumming: */
2136         if (!validate_before_checksum &&
2137             validate_bset_for_write(c, b, i, sectors_to_write))
2138                 goto err;
2139
2140         /*
2141          * We handle btree write errors by immediately halting the journal -
2142          * after we've done that, we can't issue any subsequent btree writes
2143          * because they might have pointers to new nodes that failed to write.
2144          *
2145          * Furthermore, there's no point in doing any more btree writes because
2146          * with the journal stopped, we're never going to update the journal to
2147          * reflect that those writes were done and the data flushed from the
2148          * journal:
2149          *
2150          * Also on journal error, the pending write may have updates that were
2151          * never journalled (interior nodes, see btree_update_nodes_written()) -
2152          * it's critical that we don't do the write in that case otherwise we
2153          * will have updates visible that weren't in the journal:
2154          *
2155          * Make sure to update b->written so bch2_btree_init_next() doesn't
2156          * break:
2157          */
2158         if (bch2_journal_error(&c->journal) ||
2159             c->opts.nochanges)
2160                 goto err;
2161
2162         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2163
2164         wbio = container_of(bio_alloc_bioset(NULL,
2165                                 buf_pages(data, sectors_to_write << 9),
2166                                 REQ_OP_WRITE|REQ_META,
2167                                 GFP_NOFS,
2168                                 &c->btree_bio),
2169                             struct btree_write_bio, wbio.bio);
2170         wbio_init(&wbio->wbio.bio);
2171         wbio->data                      = data;
2172         wbio->data_bytes                = bytes;
2173         wbio->sector_offset             = b->written;
2174         wbio->wbio.c                    = c;
2175         wbio->wbio.used_mempool         = used_mempool;
2176         wbio->wbio.first_btree_write    = !b->written;
2177         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2178         wbio->wbio.bio.bi_private       = b;
2179
2180         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2181
2182         bkey_copy(&wbio->key, &b->key);
2183
2184         b->written += sectors_to_write;
2185
2186         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2187                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2188                         cpu_to_le16(b->written);
2189
2190         atomic64_inc(&c->btree_write_stats[type].nr);
2191         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2192
2193         INIT_WORK(&wbio->work, btree_write_submit);
2194         queue_work(c->io_complete_wq, &wbio->work);
2195         return;
2196 err:
2197         set_btree_node_noevict(b);
2198         b->written += sectors_to_write;
2199 nowrite:
2200         btree_bounce_free(c, bytes, used_mempool, data);
2201         __btree_node_write_done(c, b);
2202 }
2203
2204 /*
2205  * Work that must be done with write lock held:
2206  */
2207 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2208 {
2209         bool invalidated_iter = false;
2210         struct btree_node_entry *bne;
2211         struct bset_tree *t;
2212
2213         if (!btree_node_just_written(b))
2214                 return false;
2215
2216         BUG_ON(b->whiteout_u64s);
2217
2218         clear_btree_node_just_written(b);
2219
2220         /*
2221          * Note: immediately after write, bset_written() doesn't work - the
2222          * amount of data we had to write after compaction might have been
2223          * smaller than the offset of the last bset.
2224          *
2225          * However, we know that all bsets have been written here, as long as
2226          * we're still holding the write lock:
2227          */
2228
2229         /*
2230          * XXX: decide if we really want to unconditionally sort down to a
2231          * single bset:
2232          */
2233         if (b->nsets > 1) {
2234                 btree_node_sort(c, b, 0, b->nsets, true);
2235                 invalidated_iter = true;
2236         } else {
2237                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2238         }
2239
2240         for_each_bset(b, t)
2241                 set_needs_whiteout(bset(b, t), true);
2242
2243         bch2_btree_verify(c, b);
2244
2245         /*
2246          * If later we don't unconditionally sort down to a single bset, we have
2247          * to ensure this is still true:
2248          */
2249         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2250
2251         bne = want_new_bset(c, b);
2252         if (bne)
2253                 bch2_bset_init_next(b, bne);
2254
2255         bch2_btree_build_aux_trees(b);
2256
2257         return invalidated_iter;
2258 }
2259
2260 /*
2261  * Use this one if the node is intent locked:
2262  */
2263 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2264                            enum six_lock_type lock_type_held,
2265                            unsigned flags)
2266 {
2267         if (lock_type_held == SIX_LOCK_intent ||
2268             (lock_type_held == SIX_LOCK_read &&
2269              six_lock_tryupgrade(&b->c.lock))) {
2270                 __bch2_btree_node_write(c, b, flags);
2271
2272                 /* don't cycle lock unnecessarily: */
2273                 if (btree_node_just_written(b) &&
2274                     six_trylock_write(&b->c.lock)) {
2275                         bch2_btree_post_write_cleanup(c, b);
2276                         six_unlock_write(&b->c.lock);
2277                 }
2278
2279                 if (lock_type_held == SIX_LOCK_read)
2280                         six_lock_downgrade(&b->c.lock);
2281         } else {
2282                 __bch2_btree_node_write(c, b, flags);
2283                 if (lock_type_held == SIX_LOCK_write &&
2284                     btree_node_just_written(b))
2285                         bch2_btree_post_write_cleanup(c, b);
2286         }
2287 }
2288
2289 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2290 {
2291         struct bucket_table *tbl;
2292         struct rhash_head *pos;
2293         struct btree *b;
2294         unsigned i;
2295         bool ret = false;
2296 restart:
2297         rcu_read_lock();
2298         for_each_cached_btree(b, c, tbl, i, pos)
2299                 if (test_bit(flag, &b->flags)) {
2300                         rcu_read_unlock();
2301                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2302                         ret = true;
2303                         goto restart;
2304                 }
2305         rcu_read_unlock();
2306
2307         return ret;
2308 }
2309
2310 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2311 {
2312         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2313 }
2314
2315 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2316 {
2317         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2318 }
2319
2320 static const char * const bch2_btree_write_types[] = {
2321 #define x(t, n) [n] = #t,
2322         BCH_BTREE_WRITE_TYPES()
2323         NULL
2324 };
2325
2326 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2327 {
2328         printbuf_tabstop_push(out, 20);
2329         printbuf_tabstop_push(out, 10);
2330
2331         prt_tab(out);
2332         prt_str(out, "nr");
2333         prt_tab(out);
2334         prt_str(out, "size");
2335         prt_newline(out);
2336
2337         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2338                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2339                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2340
2341                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2342                 prt_tab(out);
2343                 prt_u64(out, nr);
2344                 prt_tab(out);
2345                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2346                 prt_newline(out);
2347         }
2348 }