git.sesse.net Git - bcachefs-tools-debian/blob - c_src/libbcachefs/btree_io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_sort.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "buckets.h"
  13 #include "checksum.h"
  14 #include "debug.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "io_write.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_seq_blacklist.h"
  20 #include "recovery.h"
  21 #include "super-io.h"
  22 #include "trace.h"
  23
  24 #include <linux/sched/mm.h>
  25
  26 void bch2_btree_node_io_unlock(struct btree *b)
  27 {
  28         EBUG_ON(!btree_node_write_in_flight(b));
  29
  30         clear_btree_node_write_in_flight_inner(b);
  31         clear_btree_node_write_in_flight(b);
  32         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
  33 }
  34
  35 void bch2_btree_node_io_lock(struct btree *b)
  36 {
  37         bch2_assert_btree_nodes_not_locked();
  38
  39         wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
  40                             TASK_UNINTERRUPTIBLE);
  41 }
  42
  43 void __bch2_btree_node_wait_on_read(struct btree *b)
  44 {
  45         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  46                        TASK_UNINTERRUPTIBLE);
  47 }
  48
  49 void __bch2_btree_node_wait_on_write(struct btree *b)
  50 {
  51         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  52                        TASK_UNINTERRUPTIBLE);
  53 }
  54
  55 void bch2_btree_node_wait_on_read(struct btree *b)
  56 {
  57         bch2_assert_btree_nodes_not_locked();
  58
  59         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  60                        TASK_UNINTERRUPTIBLE);
  61 }
  62
  63 void bch2_btree_node_wait_on_write(struct btree *b)
  64 {
  65         bch2_assert_btree_nodes_not_locked();
  66
  67         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  68                        TASK_UNINTERRUPTIBLE);
  69 }
  70
  71 static void verify_no_dups(struct btree *b,
  72                            struct bkey_packed *start,
  73                            struct bkey_packed *end)
  74 {
  75 #ifdef CONFIG_BCACHEFS_DEBUG
  76         struct bkey_packed *k, *p;
  77
  78         if (start == end)
  79                 return;
  80
  81         for (p = start, k = bkey_p_next(start);
  82              k != end;
  83              p = k, k = bkey_p_next(k)) {
  84                 struct bkey l = bkey_unpack_key(b, p);
  85                 struct bkey r = bkey_unpack_key(b, k);
  86
  87                 BUG_ON(bpos_ge(l.p, bkey_start_pos(&r)));
  88         }
  89 #endif
  90 }
  91
  92 static void set_needs_whiteout(struct bset *i, int v)
  93 {
  94         struct bkey_packed *k;
  95
  96         for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
  97                 k->needs_whiteout = v;
  98 }
  99
 100 static void btree_bounce_free(struct bch_fs *c, size_t size,
 101                               bool used_mempool, void *p)
 102 {
 103         if (used_mempool)
 104                 mempool_free(p, &c->btree_bounce_pool);
 105         else
 106                 vpfree(p, size);
 107 }
 108
 109 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
 110                                 bool *used_mempool)
 111 {
 112         unsigned flags = memalloc_nofs_save();
 113         void *p;
 114
 115         BUG_ON(size > btree_bytes(c));
 116
 117         *used_mempool = false;
 118         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
 119         if (!p) {
 120                 *used_mempool = true;
 121                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
 122         }
 123         memalloc_nofs_restore(flags);
 124         return p;
 125 }
 126
 127 static void sort_bkey_ptrs(const struct btree *bt,
 128                            struct bkey_packed **ptrs, unsigned nr)
 129 {
 130         unsigned n = nr, a = nr / 2, b, c, d;
 131
 132         if (!a)
 133                 return;
 134
 135         /* Heap sort: see lib/sort.c: */
 136         while (1) {
 137                 if (a)
 138                         a--;
 139                 else if (--n)
 140                         swap(ptrs[0], ptrs[n]);
 141                 else
 142                         break;
 143
 144                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 145                         b = bch2_bkey_cmp_packed(bt,
 146                                             ptrs[c],
 147                                             ptrs[d]) >= 0 ? c : d;
 148                 if (d == n)
 149                         b = c;
 150
 151                 while (b != a &&
 152                        bch2_bkey_cmp_packed(bt,
 153                                        ptrs[a],
 154                                        ptrs[b]) >= 0)
 155                         b = (b - 1) / 2;
 156                 c = b;
 157                 while (b != a) {
 158                         b = (b - 1) / 2;
 159                         swap(ptrs[b], ptrs[c]);
 160                 }
 161         }
 162 }
 163
 164 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 165 {
 166         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 167         bool used_mempool = false;
 168         size_t bytes = b->whiteout_u64s * sizeof(u64);
 169
 170         if (!b->whiteout_u64s)
 171                 return;
 172
 173         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 174
 175         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 176
 177         for (k = unwritten_whiteouts_start(c, b);
 178              k != unwritten_whiteouts_end(c, b);
 179              k = bkey_p_next(k))
 180                 *--ptrs = k;
 181
 182         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 183
 184         k = new_whiteouts;
 185
 186         while (ptrs != ptrs_end) {
 187                 bkey_p_copy(k, *ptrs);
 188                 k = bkey_p_next(k);
 189                 ptrs++;
 190         }
 191
 192         verify_no_dups(b, new_whiteouts,
 193                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 194
 195         memcpy_u64s(unwritten_whiteouts_start(c, b),
 196                     new_whiteouts, b->whiteout_u64s);
 197
 198         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 199 }
 200
 201 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 202                                 bool compacting, enum compact_mode mode)
 203 {
 204         if (!bset_dead_u64s(b, t))
 205                 return false;
 206
 207         switch (mode) {
 208         case COMPACT_LAZY:
 209                 return should_compact_bset_lazy(b, t) ||
 210                         (compacting && !bset_written(b, bset(b, t)));
 211         case COMPACT_ALL:
 212                 return true;
 213         default:
 214                 BUG();
 215         }
 216 }
 217
 218 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 219 {
 220         struct bset_tree *t;
 221         bool ret = false;
 222
 223         for_each_bset(b, t) {
 224                 struct bset *i = bset(b, t);
 225                 struct bkey_packed *k, *n, *out, *start, *end;
 226                 struct btree_node_entry *src = NULL, *dst = NULL;
 227
 228                 if (t != b->set && !bset_written(b, i)) {
 229                         src = container_of(i, struct btree_node_entry, keys);
 230                         dst = max(write_block(b),
 231                                   (void *) btree_bkey_last(b, t - 1));
 232                 }
 233
 234                 if (src != dst)
 235                         ret = true;
 236
 237                 if (!should_compact_bset(b, t, ret, mode)) {
 238                         if (src != dst) {
 239                                 memmove(dst, src, sizeof(*src) +
 240                                         le16_to_cpu(src->keys.u64s) *
 241                                         sizeof(u64));
 242                                 i = &dst->keys;
 243                                 set_btree_bset(b, t, i);
 244                         }
 245                         continue;
 246                 }
 247
 248                 start   = btree_bkey_first(b, t);
 249                 end     = btree_bkey_last(b, t);
 250
 251                 if (src != dst) {
 252                         memmove(dst, src, sizeof(*src));
 253                         i = &dst->keys;
 254                         set_btree_bset(b, t, i);
 255                 }
 256
 257                 out = i->start;
 258
 259                 for (k = start; k != end; k = n) {
 260                         n = bkey_p_next(k);
 261
 262                         if (!bkey_deleted(k)) {
 263                                 bkey_p_copy(out, k);
 264                                 out = bkey_p_next(out);
 265                         } else {
 266                                 BUG_ON(k->needs_whiteout);
 267                         }
 268                 }
 269
 270                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 271                 set_btree_bset_end(b, t);
 272                 bch2_bset_set_no_aux_tree(b, t);
 273                 ret = true;
 274         }
 275
 276         bch2_verify_btree_nr_keys(b);
 277
 278         bch2_btree_build_aux_trees(b);
 279
 280         return ret;
 281 }
 282
 283 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 284                             enum compact_mode mode)
 285 {
 286         return bch2_drop_whiteouts(b, mode);
 287 }
 288
 289 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 290                             unsigned start_idx,
 291                             unsigned end_idx,
 292                             bool filter_whiteouts)
 293 {
 294         struct btree_node *out;
 295         struct sort_iter_stack sort_iter;
 296         struct bset_tree *t;
 297         struct bset *start_bset = bset(b, &b->set[start_idx]);
 298         bool used_mempool = false;
 299         u64 start_time, seq = 0;
 300         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 301         bool sorting_entire_node = start_idx == 0 &&
 302                 end_idx == b->nsets;
 303
 304         sort_iter_stack_init(&sort_iter, b);
 305
 306         for (t = b->set + start_idx;
 307              t < b->set + end_idx;
 308              t++) {
 309                 u64s += le16_to_cpu(bset(b, t)->u64s);
 310                 sort_iter_add(&sort_iter.iter,
 311                               btree_bkey_first(b, t),
 312                               btree_bkey_last(b, t));
 313         }
 314
 315         bytes = sorting_entire_node
 316                 ? btree_bytes(c)
 317                 : __vstruct_bytes(struct btree_node, u64s);
 318
 319         out = btree_bounce_alloc(c, bytes, &used_mempool);
 320
 321         start_time = local_clock();
 322
 323         u64s = bch2_sort_keys(out->keys.start, &sort_iter.iter, filter_whiteouts);
 324
 325         out->keys.u64s = cpu_to_le16(u64s);
 326
 327         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 328
 329         if (sorting_entire_node)
 330                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 331                                        start_time);
 332
 333         /* Make sure we preserve bset journal_seq: */
 334         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 335                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 336         start_bset->journal_seq = cpu_to_le64(seq);
 337
 338         if (sorting_entire_node) {
 339                 u64s = le16_to_cpu(out->keys.u64s);
 340
 341                 BUG_ON(bytes != btree_bytes(c));
 342
 343                 /*
 344                  * Our temporary buffer is the same size as the btree node's
 345                  * buffer, we can just swap buffers instead of doing a big
 346                  * memcpy()
 347                  */
 348                 *out = *b->data;
 349                 out->keys.u64s = cpu_to_le16(u64s);
 350                 swap(out, b->data);
 351                 set_btree_bset(b, b->set, &b->data->keys);
 352         } else {
 353                 start_bset->u64s = out->keys.u64s;
 354                 memcpy_u64s(start_bset->start,
 355                             out->keys.start,
 356                             le16_to_cpu(out->keys.u64s));
 357         }
 358
 359         for (i = start_idx + 1; i < end_idx; i++)
 360                 b->nr.bset_u64s[start_idx] +=
 361                         b->nr.bset_u64s[i];
 362
 363         b->nsets -= shift;
 364
 365         for (i = start_idx + 1; i < b->nsets; i++) {
 366                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 367                 b->set[i]               = b->set[i + shift];
 368         }
 369
 370         for (i = b->nsets; i < MAX_BSETS; i++)
 371                 b->nr.bset_u64s[i] = 0;
 372
 373         set_btree_bset_end(b, &b->set[start_idx]);
 374         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 375
 376         btree_bounce_free(c, bytes, used_mempool, out);
 377
 378         bch2_verify_btree_nr_keys(b);
 379 }
 380
 381 void bch2_btree_sort_into(struct bch_fs *c,
 382                          struct btree *dst,
 383                          struct btree *src)
 384 {
 385         struct btree_nr_keys nr;
 386         struct btree_node_iter src_iter;
 387         u64 start_time = local_clock();
 388
 389         BUG_ON(dst->nsets != 1);
 390
 391         bch2_bset_set_no_aux_tree(dst, dst->set);
 392
 393         bch2_btree_node_iter_init_from_start(&src_iter, src);
 394
 395         nr = bch2_sort_repack(btree_bset_first(dst),
 396                         src, &src_iter,
 397                         &dst->format,
 398                         true);
 399
 400         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 401                                start_time);
 402
 403         set_btree_bset_end(dst, dst->set);
 404
 405         dst->nr.live_u64s       += nr.live_u64s;
 406         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 407         dst->nr.packed_keys     += nr.packed_keys;
 408         dst->nr.unpacked_keys   += nr.unpacked_keys;
 409
 410         bch2_verify_btree_nr_keys(dst);
 411 }
 412
 413 /*
 414  * We're about to add another bset to the btree node, so if there's currently
 415  * too many bsets - sort some of them together:
 416  */
 417 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 418 {
 419         unsigned unwritten_idx;
 420         bool ret = false;
 421
 422         for (unwritten_idx = 0;
 423              unwritten_idx < b->nsets;
 424              unwritten_idx++)
 425                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 426                         break;
 427
 428         if (b->nsets - unwritten_idx > 1) {
 429                 btree_node_sort(c, b, unwritten_idx,
 430                                 b->nsets, false);
 431                 ret = true;
 432         }
 433
 434         if (unwritten_idx > 1) {
 435                 btree_node_sort(c, b, 0, unwritten_idx, false);
 436                 ret = true;
 437         }
 438
 439         return ret;
 440 }
 441
 442 void bch2_btree_build_aux_trees(struct btree *b)
 443 {
 444         struct bset_tree *t;
 445
 446         for_each_bset(b, t)
 447                 bch2_bset_build_aux_tree(b, t,
 448                                 !bset_written(b, bset(b, t)) &&
 449                                 t == bset_tree_last(b));
 450 }
 451
 452 /*
 453  * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
 454  *
 455  * The first bset is going to be of similar order to the size of the node, the
 456  * last bset is bounded by btree_write_set_buffer(), which is set to keep the
 457  * memmove on insert from being too expensive: the middle bset should, ideally,
 458  * be the geometric mean of the first and the last.
 459  *
 460  * Returns true if the middle bset is greater than that geometric mean:
 461  */
 462 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
 463 {
 464         unsigned mid_u64s_bits =
 465                 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
 466
 467         return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
 468 }
 469
 470 /*
 471  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 472  * inserted into
 473  *
 474  * Safe to call if there already is an unwritten bset - will only add a new bset
 475  * if @b doesn't already have one.
 476  *
 477  * Returns true if we sorted (i.e. invalidated iterators
 478  */
 479 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 480 {
 481         struct bch_fs *c = trans->c;
 482         struct btree_node_entry *bne;
 483         bool reinit_iter = false;
 484
 485         EBUG_ON(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]);
 486         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 487         BUG_ON(btree_node_just_written(b));
 488
 489         if (b->nsets == MAX_BSETS &&
 490             !btree_node_write_in_flight(b) &&
 491             should_compact_all(c, b)) {
 492                 bch2_btree_node_write(c, b, SIX_LOCK_write,
 493                                       BTREE_WRITE_init_next_bset);
 494                 reinit_iter = true;
 495         }
 496
 497         if (b->nsets == MAX_BSETS &&
 498             btree_node_compact(c, b))
 499                 reinit_iter = true;
 500
 501         BUG_ON(b->nsets >= MAX_BSETS);
 502
 503         bne = want_new_bset(c, b);
 504         if (bne)
 505                 bch2_bset_init_next(c, b, bne);
 506
 507         bch2_btree_build_aux_trees(b);
 508
 509         if (reinit_iter)
 510                 bch2_trans_node_reinit_iter(trans, b);
 511 }
 512
 513 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 514                           struct bch_dev *ca,
 515                           struct btree *b, struct bset *i,
 516                           unsigned offset, int write)
 517 {
 518         prt_printf(out, bch2_log_msg(c, "%s"),
 519                    write == READ
 520                    ? "error validating btree node "
 521                    : "corrupt btree node before write ");
 522         if (ca)
 523                 prt_printf(out, "on %s ", ca->name);
 524         prt_printf(out, "at btree ");
 525         bch2_btree_pos_to_text(out, c, b);
 526
 527         prt_printf(out, "\n  node offset %u/%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                 bch2_topology_error(c);
 585                 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology) ?: -EIO;
 586                 break;
 587         case -BCH_ERR_btree_node_read_err_incompatible:
 588                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 589                 ret = -BCH_ERR_fsck_errors_not_fixed;
 590                 break;
 591         default:
 592                 BUG();
 593         }
 594 out:
 595 fsck_err:
 596         printbuf_exit(&out);
 597         return ret;
 598 }
 599
 600 #define btree_err(type, c, ca, b, i, _err_type, msg, ...)               \
 601 ({                                                                      \
 602         int _ret = __btree_err(type, c, ca, b, i, write, have_retry,    \
 603                                BCH_FSCK_ERR_##_err_type,                \
 604                                msg, ##__VA_ARGS__);                     \
 605                                                                         \
 606         if (_ret != -BCH_ERR_fsck_fix) {                                \
 607                 ret = _ret;                                             \
 608                 goto fsck_err;                                          \
 609         }                                                               \
 610                                                                         \
 611         *saw_error = true;                                              \
 612 })
 613
 614 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 615
 616 /*
 617  * When btree topology repair changes the start or end of a node, that might
 618  * mean we have to drop keys that are no longer inside the node:
 619  */
 620 __cold
 621 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 622 {
 623         struct bset_tree *t;
 624
 625         for_each_bset(b, t) {
 626                 struct bset *i = bset(b, t);
 627                 struct bkey_packed *k;
 628
 629                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 630                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 631                                 break;
 632
 633                 if (k != i->start) {
 634                         unsigned shift = (u64 *) k - (u64 *) i->start;
 635
 636                         memmove_u64s_down(i->start, k,
 637                                           (u64 *) vstruct_end(i) - (u64 *) k);
 638                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 639                         set_btree_bset_end(b, t);
 640                 }
 641
 642                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 643                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 644                                 break;
 645
 646                 if (k != vstruct_last(i)) {
 647                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 648                         set_btree_bset_end(b, t);
 649                 }
 650         }
 651
 652         /*
 653          * Always rebuild search trees: eytzinger search tree nodes directly
 654          * depend on the values of min/max key:
 655          */
 656         bch2_bset_set_no_aux_tree(b, b->set);
 657         bch2_btree_build_aux_trees(b);
 658
 659         struct bkey_s_c k;
 660         struct bkey unpacked;
 661         struct btree_node_iter iter;
 662         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 663                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 664                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 665         }
 666 }
 667
 668 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 669                          struct btree *b, struct bset *i,
 670                          unsigned offset, unsigned sectors,
 671                          int write, bool have_retry, bool *saw_error)
 672 {
 673         unsigned version = le16_to_cpu(i->version);
 674         struct printbuf buf1 = PRINTBUF;
 675         struct printbuf buf2 = PRINTBUF;
 676         int ret = 0;
 677
 678         btree_err_on(!bch2_version_compatible(version),
 679                      -BCH_ERR_btree_node_read_err_incompatible,
 680                      c, ca, b, i,
 681                      btree_node_unsupported_version,
 682                      "unsupported bset version %u.%u",
 683                      BCH_VERSION_MAJOR(version),
 684                      BCH_VERSION_MINOR(version));
 685
 686         if (btree_err_on(version < c->sb.version_min,
 687                          -BCH_ERR_btree_node_read_err_fixable,
 688                          c, NULL, b, i,
 689                          btree_node_bset_older_than_sb_min,
 690                          "bset version %u older than superblock version_min %u",
 691                          version, c->sb.version_min)) {
 692                 mutex_lock(&c->sb_lock);
 693                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 694                 bch2_write_super(c);
 695                 mutex_unlock(&c->sb_lock);
 696         }
 697
 698         if (btree_err_on(BCH_VERSION_MAJOR(version) >
 699                          BCH_VERSION_MAJOR(c->sb.version),
 700                          -BCH_ERR_btree_node_read_err_fixable,
 701                          c, NULL, b, i,
 702                          btree_node_bset_newer_than_sb,
 703                          "bset version %u newer than superblock version %u",
 704                          version, c->sb.version)) {
 705                 mutex_lock(&c->sb_lock);
 706                 c->disk_sb.sb->version = cpu_to_le16(version);
 707                 bch2_write_super(c);
 708                 mutex_unlock(&c->sb_lock);
 709         }
 710
 711         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 712                      -BCH_ERR_btree_node_read_err_incompatible,
 713                      c, ca, b, i,
 714                      btree_node_unsupported_version,
 715                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 716
 717         if (btree_err_on(offset + sectors > btree_sectors(c),
 718                          -BCH_ERR_btree_node_read_err_fixable,
 719                          c, ca, b, i,
 720                          bset_past_end_of_btree_node,
 721                          "bset past end of btree node")) {
 722                 i->u64s = 0;
 723                 ret = 0;
 724                 goto out;
 725         }
 726
 727         btree_err_on(offset && !i->u64s,
 728                      -BCH_ERR_btree_node_read_err_fixable,
 729                      c, ca, b, i,
 730                      bset_empty,
 731                      "empty bset");
 732
 733         btree_err_on(BSET_OFFSET(i) && BSET_OFFSET(i) != offset,
 734                      -BCH_ERR_btree_node_read_err_want_retry,
 735                      c, ca, b, i,
 736                      bset_wrong_sector_offset,
 737                      "bset at wrong sector offset");
 738
 739         if (!offset) {
 740                 struct btree_node *bn =
 741                         container_of(i, struct btree_node, keys);
 742                 /* These indicate that we read the wrong btree node: */
 743
 744                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 745                         struct bch_btree_ptr_v2 *bp =
 746                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 747
 748                         /* XXX endianness */
 749                         btree_err_on(bp->seq != bn->keys.seq,
 750                                      -BCH_ERR_btree_node_read_err_must_retry,
 751                                      c, ca, b, NULL,
 752                                      bset_bad_seq,
 753                                      "incorrect sequence number (wrong btree node)");
 754                 }
 755
 756                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 757                              -BCH_ERR_btree_node_read_err_must_retry,
 758                              c, ca, b, i,
 759                              btree_node_bad_btree,
 760                              "incorrect btree id");
 761
 762                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 763                              -BCH_ERR_btree_node_read_err_must_retry,
 764                              c, ca, b, i,
 765                              btree_node_bad_level,
 766                              "incorrect level");
 767
 768                 if (!write)
 769                         compat_btree_node(b->c.level, b->c.btree_id, version,
 770                                           BSET_BIG_ENDIAN(i), write, bn);
 771
 772                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 773                         struct bch_btree_ptr_v2 *bp =
 774                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 775
 776                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 777                                 b->data->min_key = bp->min_key;
 778                                 b->data->max_key = b->key.k.p;
 779                         }
 780
 781                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 782                                      -BCH_ERR_btree_node_read_err_must_retry,
 783                                      c, ca, b, NULL,
 784                                      btree_node_bad_min_key,
 785                                      "incorrect min_key: got %s should be %s",
 786                                      (printbuf_reset(&buf1),
 787                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 788                                      (printbuf_reset(&buf2),
 789                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 790                 }
 791
 792                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 793                              -BCH_ERR_btree_node_read_err_must_retry,
 794                              c, ca, b, i,
 795                              btree_node_bad_max_key,
 796                              "incorrect max key %s",
 797                              (printbuf_reset(&buf1),
 798                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 799
 800                 if (write)
 801                         compat_btree_node(b->c.level, b->c.btree_id, version,
 802                                           BSET_BIG_ENDIAN(i), write, bn);
 803
 804                 btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
 805                              -BCH_ERR_btree_node_read_err_bad_node,
 806                              c, ca, b, i,
 807                              btree_node_bad_format,
 808                              "invalid bkey format: %s\n  %s", buf1.buf,
 809                              (printbuf_reset(&buf2),
 810                               bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
 811                 printbuf_reset(&buf1);
 812
 813                 compat_bformat(b->c.level, b->c.btree_id, version,
 814                                BSET_BIG_ENDIAN(i), write,
 815                                &bn->format);
 816         }
 817 out:
 818 fsck_err:
 819         printbuf_exit(&buf2);
 820         printbuf_exit(&buf1);
 821         return ret;
 822 }
 823
 824 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 825                             struct bkey_s_c k,
 826                             bool updated_range, int rw,
 827                             struct printbuf *err)
 828 {
 829         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 830                 (!updated_range ? bch2_bkey_in_btree_node(c, b, k, err) : 0) ?:
 831                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 832 }
 833
 834 static bool __bkey_valid(struct bch_fs *c, struct btree *b,
 835                          struct bset *i, struct bkey_packed *k)
 836 {
 837         if (bkey_p_next(k) > vstruct_last(i))
 838                 return false;
 839
 840         if (k->format > KEY_FORMAT_CURRENT)
 841                 return false;
 842
 843         struct printbuf buf = PRINTBUF;
 844         struct bkey tmp;
 845         struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 846         bool ret = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b), READ, &buf);
 847         printbuf_exit(&buf);
 848         return ret;
 849 }
 850
 851 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 852                          struct bset *i, int write,
 853                          bool have_retry, bool *saw_error)
 854 {
 855         unsigned version = le16_to_cpu(i->version);
 856         struct bkey_packed *k, *prev = NULL;
 857         struct printbuf buf = PRINTBUF;
 858         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 859                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 860         int ret = 0;
 861
 862         for (k = i->start;
 863              k != vstruct_last(i);) {
 864                 struct bkey_s u;
 865                 struct bkey tmp;
 866                 unsigned next_good_key;
 867
 868                 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
 869                                  -BCH_ERR_btree_node_read_err_fixable,
 870                                  c, NULL, b, i,
 871                                  btree_node_bkey_past_bset_end,
 872                                  "key extends past end of bset")) {
 873                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 874                         break;
 875                 }
 876
 877                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 878                                  -BCH_ERR_btree_node_read_err_fixable,
 879                                  c, NULL, b, i,
 880                                  btree_node_bkey_bad_format,
 881                                  "invalid bkey format %u", k->format))
 882                         goto drop_this_key;
 883
 884                 /* XXX: validate k->u64s */
 885                 if (!write)
 886                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 887                                     BSET_BIG_ENDIAN(i), write,
 888                                     &b->format, k);
 889
 890                 u = __bkey_disassemble(b, k, &tmp);
 891
 892                 printbuf_reset(&buf);
 893                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 894                         printbuf_reset(&buf);
 895                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 896                         prt_printf(&buf, "\n  ");
 897                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 898
 899                         btree_err(-BCH_ERR_btree_node_read_err_fixable,
 900                                   c, NULL, b, i,
 901                                   btree_node_bad_bkey,
 902                                   "invalid bkey: %s", buf.buf);
 903                         goto drop_this_key;
 904                 }
 905
 906                 if (write)
 907                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 908                                     BSET_BIG_ENDIAN(i), write,
 909                                     &b->format, k);
 910
 911                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 912                         struct bkey up = bkey_unpack_key(b, prev);
 913
 914                         printbuf_reset(&buf);
 915                         prt_printf(&buf, "keys out of order: ");
 916                         bch2_bkey_to_text(&buf, &up);
 917                         prt_printf(&buf, " > ");
 918                         bch2_bkey_to_text(&buf, u.k);
 919
 920                         if (btree_err(-BCH_ERR_btree_node_read_err_fixable,
 921                                       c, NULL, b, i,
 922                                       btree_node_bkey_out_of_order,
 923                                       "%s", buf.buf))
 924                                 goto drop_this_key;
 925                 }
 926
 927                 prev = k;
 928                 k = bkey_p_next(k);
 929                 continue;
 930 drop_this_key:
 931                 next_good_key = k->u64s;
 932
 933                 if (!next_good_key ||
 934                     (BSET_BIG_ENDIAN(i) == CPU_BIG_ENDIAN &&
 935                      version >= bcachefs_metadata_version_snapshot)) {
 936                         /*
 937                          * only do scanning if bch2_bkey_compat() has nothing to
 938                          * do
 939                          */
 940
 941                         if (!__bkey_valid(c, b, i, (void *) ((u64 *) k + next_good_key))) {
 942                                 for (next_good_key = 1;
 943                                      next_good_key < (u64 *) vstruct_last(i) - (u64 *) k;
 944                                      next_good_key++)
 945                                         if (__bkey_valid(c, b, i, (void *) ((u64 *) k + next_good_key)))
 946                                                 goto got_good_key;
 947
 948                         }
 949
 950                         /*
 951                          * didn't find a good key, have to truncate the rest of
 952                          * the bset
 953                          */
 954                         next_good_key = (u64 *) vstruct_last(i) - (u64 *) k;
 955                 }
 956 got_good_key:
 957                 le16_add_cpu(&i->u64s, -next_good_key);
 958                 memmove_u64s_down(k, bkey_p_next(k), (u64 *) vstruct_end(i) - (u64 *) k);
 959         }
 960 fsck_err:
 961         printbuf_exit(&buf);
 962         return ret;
 963 }
 964
 965 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 966                               struct btree *b, bool have_retry, bool *saw_error)
 967 {
 968         struct btree_node_entry *bne;
 969         struct sort_iter *iter;
 970         struct btree_node *sorted;
 971         struct bkey_packed *k;
 972         struct bset *i;
 973         bool used_mempool, blacklisted;
 974         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 975                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 976         unsigned u64s;
 977         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 978         struct printbuf buf = PRINTBUF;
 979         int ret = 0, retry_read = 0, write = READ;
 980         u64 start_time = local_clock();
 981
 982         b->version_ondisk = U16_MAX;
 983         /* We might get called multiple times on read retry: */
 984         b->written = 0;
 985
 986         iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
 987         sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2);
 988
 989         if (bch2_meta_read_fault("btree"))
 990                 btree_err(-BCH_ERR_btree_node_read_err_must_retry,
 991                           c, ca, b, NULL,
 992                           btree_node_fault_injected,
 993                           "dynamic fault");
 994
 995         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 996                      -BCH_ERR_btree_node_read_err_must_retry,
 997                      c, ca, b, NULL,
 998                      btree_node_bad_magic,
 999                      "bad magic: want %llx, got %llx",
1000                      bset_magic(c), le64_to_cpu(b->data->magic));
1001
1002         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
1003                 struct bch_btree_ptr_v2 *bp =
1004                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
1005
1006                 bch2_bpos_to_text(&buf, b->data->min_key);
1007                 prt_str(&buf, "-");
1008                 bch2_bpos_to_text(&buf, b->data->max_key);
1009
1010                 btree_err_on(b->data->keys.seq != bp->seq,
1011                              -BCH_ERR_btree_node_read_err_must_retry,
1012                              c, ca, b, NULL,
1013                              btree_node_bad_seq,
1014                              "got wrong btree node (want %llx got %llx)\n"
1015                              "got btree %s level %llu pos %s",
1016                              bp->seq, b->data->keys.seq,
1017                              bch2_btree_id_str(BTREE_NODE_ID(b->data)),
1018                              BTREE_NODE_LEVEL(b->data),
1019                              buf.buf);
1020         } else {
1021                 btree_err_on(!b->data->keys.seq,
1022                              -BCH_ERR_btree_node_read_err_must_retry,
1023                              c, ca, b, NULL,
1024                              btree_node_bad_seq,
1025                              "bad btree header: seq 0");
1026         }
1027
1028         while (b->written < (ptr_written ?: btree_sectors(c))) {
1029                 unsigned sectors;
1030                 struct nonce nonce;
1031                 bool first = !b->written;
1032                 bool csum_bad;
1033
1034                 if (!b->written) {
1035                         i = &b->data->keys;
1036
1037                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1038                                      -BCH_ERR_btree_node_read_err_want_retry,
1039                                      c, ca, b, i,
1040                                      bset_unknown_csum,
1041                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1042
1043                         nonce = btree_nonce(i, b->written << 9);
1044
1045                         struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
1046                         csum_bad = bch2_crc_cmp(b->data->csum, csum);
1047                         if (csum_bad)
1048                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1049
1050                         btree_err_on(csum_bad,
1051                                      -BCH_ERR_btree_node_read_err_want_retry,
1052                                      c, ca, b, i,
1053                                      bset_bad_csum,
1054                                      "%s",
1055                                      (printbuf_reset(&buf),
1056                                       bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), b->data->csum, csum),
1057                                       buf.buf));
1058
1059                         ret = bset_encrypt(c, i, b->written << 9);
1060                         if (bch2_fs_fatal_err_on(ret, c,
1061                                         "error decrypting btree node: %i", ret))
1062                                 goto fsck_err;
1063
1064                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
1065                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
1066                                      -BCH_ERR_btree_node_read_err_incompatible,
1067                                      c, NULL, b, NULL,
1068                                      btree_node_unsupported_version,
1069                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
1070
1071                         sectors = vstruct_sectors(b->data, c->block_bits);
1072                 } else {
1073                         bne = write_block(b);
1074                         i = &bne->keys;
1075
1076                         if (i->seq != b->data->keys.seq)
1077                                 break;
1078
1079                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1080                                      -BCH_ERR_btree_node_read_err_want_retry,
1081                                      c, ca, b, i,
1082                                      bset_unknown_csum,
1083                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1084
1085                         nonce = btree_nonce(i, b->written << 9);
1086                         struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1087                         csum_bad = bch2_crc_cmp(bne->csum, csum);
1088                         if (csum_bad)
1089                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1090
1091                         btree_err_on(csum_bad,
1092                                      -BCH_ERR_btree_node_read_err_want_retry,
1093                                      c, ca, b, i,
1094                                      bset_bad_csum,
1095                                      "%s",
1096                                      (printbuf_reset(&buf),
1097                                       bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), bne->csum, csum),
1098                                       buf.buf));
1099
1100                         ret = bset_encrypt(c, i, b->written << 9);
1101                         if (bch2_fs_fatal_err_on(ret, c,
1102                                         "error decrypting btree node: %i\n", ret))
1103                                 goto fsck_err;
1104
1105                         sectors = vstruct_sectors(bne, c->block_bits);
1106                 }
1107
1108                 b->version_ondisk = min(b->version_ondisk,
1109                                         le16_to_cpu(i->version));
1110
1111                 ret = validate_bset(c, ca, b, i, b->written, sectors,
1112                                     READ, have_retry, saw_error);
1113                 if (ret)
1114                         goto fsck_err;
1115
1116                 if (!b->written)
1117                         btree_node_set_format(b, b->data->format);
1118
1119                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1120                 if (ret)
1121                         goto fsck_err;
1122
1123                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1124
1125                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1126                                         le64_to_cpu(i->journal_seq),
1127                                         true);
1128
1129                 btree_err_on(blacklisted && first,
1130                              -BCH_ERR_btree_node_read_err_fixable,
1131                              c, ca, b, i,
1132                              bset_blacklisted_journal_seq,
1133                              "first btree node bset has blacklisted journal seq (%llu)",
1134                              le64_to_cpu(i->journal_seq));
1135
1136                 btree_err_on(blacklisted && ptr_written,
1137                              -BCH_ERR_btree_node_read_err_fixable,
1138                              c, ca, b, i,
1139                              first_bset_blacklisted_journal_seq,
1140                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1141                              le64_to_cpu(i->journal_seq),
1142                              b->written, b->written + sectors, ptr_written);
1143
1144                 b->written += sectors;
1145
1146                 if (blacklisted && !first)
1147                         continue;
1148
1149                 sort_iter_add(iter,
1150                               vstruct_idx(i, 0),
1151                               vstruct_last(i));
1152         }
1153
1154         if (ptr_written) {
1155                 btree_err_on(b->written < ptr_written,
1156                              -BCH_ERR_btree_node_read_err_want_retry,
1157                              c, ca, b, NULL,
1158                              btree_node_data_missing,
1159                              "btree node data missing: expected %u sectors, found %u",
1160                              ptr_written, b->written);
1161         } else {
1162                 for (bne = write_block(b);
1163                      bset_byte_offset(b, bne) < btree_bytes(c);
1164                      bne = (void *) bne + block_bytes(c))
1165                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1166                                      !bch2_journal_seq_is_blacklisted(c,
1167                                                                       le64_to_cpu(bne->keys.journal_seq),
1168                                                                       true),
1169                                      -BCH_ERR_btree_node_read_err_want_retry,
1170                                      c, ca, b, NULL,
1171                                      btree_node_bset_after_end,
1172                                      "found bset signature after last bset");
1173         }
1174
1175         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1176         sorted->keys.u64s = 0;
1177
1178         set_btree_bset(b, b->set, &b->data->keys);
1179
1180         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1181
1182         u64s = le16_to_cpu(sorted->keys.u64s);
1183         *sorted = *b->data;
1184         sorted->keys.u64s = cpu_to_le16(u64s);
1185         swap(sorted, b->data);
1186         set_btree_bset(b, b->set, &b->data->keys);
1187         b->nsets = 1;
1188
1189         BUG_ON(b->nr.live_u64s != u64s);
1190
1191         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1192
1193         if (updated_range)
1194                 bch2_btree_node_drop_keys_outside_node(b);
1195
1196         i = &b->data->keys;
1197         for (k = i->start; k != vstruct_last(i);) {
1198                 struct bkey tmp;
1199                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1200
1201                 printbuf_reset(&buf);
1202
1203                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1204                     (bch2_inject_invalid_keys &&
1205                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1206                         printbuf_reset(&buf);
1207
1208                         prt_printf(&buf, "invalid bkey: ");
1209                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1210                         prt_printf(&buf, "\n  ");
1211                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1212
1213                         btree_err(-BCH_ERR_btree_node_read_err_fixable,
1214                                   c, NULL, b, i,
1215                                   btree_node_bad_bkey,
1216                                   "%s", buf.buf);
1217
1218                         btree_keys_account_key_drop(&b->nr, 0, k);
1219
1220                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1221                         memmove_u64s_down(k, bkey_p_next(k),
1222                                           (u64 *) vstruct_end(i) - (u64 *) k);
1223                         set_btree_bset_end(b, b->set);
1224                         continue;
1225                 }
1226
1227                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1228                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1229
1230                         bp.v->mem_ptr = 0;
1231                 }
1232
1233                 k = bkey_p_next(k);
1234         }
1235
1236         bch2_bset_build_aux_tree(b, b->set, false);
1237
1238         set_needs_whiteout(btree_bset_first(b), true);
1239
1240         btree_node_reset_sib_u64s(b);
1241
1242         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1243                 struct bch_dev *ca2 = bch_dev_bkey_exists(c, ptr->dev);
1244
1245                 if (ca2->mi.state != BCH_MEMBER_STATE_rw)
1246                         set_btree_node_need_rewrite(b);
1247         }
1248
1249         if (!ptr_written)
1250                 set_btree_node_need_rewrite(b);
1251 out:
1252         mempool_free(iter, &c->fill_iter);
1253         printbuf_exit(&buf);
1254         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read_done], start_time);
1255         return retry_read;
1256 fsck_err:
1257         if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1258             ret == -BCH_ERR_btree_node_read_err_must_retry)
1259                 retry_read = 1;
1260         else
1261                 set_btree_node_read_error(b);
1262         goto out;
1263 }
1264
1265 static void btree_node_read_work(struct work_struct *work)
1266 {
1267         struct btree_read_bio *rb =
1268                 container_of(work, struct btree_read_bio, work);
1269         struct bch_fs *c        = rb->c;
1270         struct btree *b         = rb->b;
1271         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1272         struct bio *bio         = &rb->bio;
1273         struct bch_io_failures failed = { .nr = 0 };
1274         struct printbuf buf = PRINTBUF;
1275         bool saw_error = false;
1276         bool retry = false;
1277         bool can_retry;
1278
1279         goto start;
1280         while (1) {
1281                 retry = true;
1282                 bch_info(c, "retrying read");
1283                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1284                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1285                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1286                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1287                 bio->bi_iter.bi_size    = btree_bytes(c);
1288
1289                 if (rb->have_ioref) {
1290                         bio_set_dev(bio, ca->disk_sb.bdev);
1291                         submit_bio_wait(bio);
1292                 } else {
1293                         bio->bi_status = BLK_STS_REMOVED;
1294                 }
1295 start:
1296                 printbuf_reset(&buf);
1297                 bch2_btree_pos_to_text(&buf, c, b);
1298                 bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_read,
1299                                    "btree read error %s for %s",
1300                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1301                 if (rb->have_ioref)
1302                         percpu_ref_put(&ca->io_ref);
1303                 rb->have_ioref = false;
1304
1305                 bch2_mark_io_failure(&failed, &rb->pick);
1306
1307                 can_retry = bch2_bkey_pick_read_device(c,
1308                                 bkey_i_to_s_c(&b->key),
1309                                 &failed, &rb->pick) > 0;
1310
1311                 if (!bio->bi_status &&
1312                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1313                         if (retry)
1314                                 bch_info(c, "retry success");
1315                         break;
1316                 }
1317
1318                 saw_error = true;
1319
1320                 if (!can_retry) {
1321                         set_btree_node_read_error(b);
1322                         break;
1323                 }
1324         }
1325
1326         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1327                                rb->start_time);
1328         bio_put(&rb->bio);
1329
1330         if (saw_error && !btree_node_read_error(b)) {
1331                 printbuf_reset(&buf);
1332                 bch2_bpos_to_text(&buf, b->key.k.p);
1333                 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1334                          __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf);
1335
1336                 bch2_btree_node_rewrite_async(c, b);
1337         }
1338
1339         printbuf_exit(&buf);
1340         clear_btree_node_read_in_flight(b);
1341         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1342 }
1343
1344 static void btree_node_read_endio(struct bio *bio)
1345 {
1346         struct btree_read_bio *rb =
1347                 container_of(bio, struct btree_read_bio, bio);
1348         struct bch_fs *c        = rb->c;
1349
1350         if (rb->have_ioref) {
1351                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1352
1353                 bch2_latency_acct(ca, rb->start_time, READ);
1354         }
1355
1356         queue_work(c->io_complete_wq, &rb->work);
1357 }
1358
1359 struct btree_node_read_all {
1360         struct closure          cl;
1361         struct bch_fs           *c;
1362         struct btree            *b;
1363         unsigned                nr;
1364         void                    *buf[BCH_REPLICAS_MAX];
1365         struct bio              *bio[BCH_REPLICAS_MAX];
1366         blk_status_t            err[BCH_REPLICAS_MAX];
1367 };
1368
1369 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1370 {
1371         struct btree_node *bn = data;
1372         struct btree_node_entry *bne;
1373         unsigned offset = 0;
1374
1375         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1376                 return 0;
1377
1378         while (offset < btree_sectors(c)) {
1379                 if (!offset) {
1380                         offset += vstruct_sectors(bn, c->block_bits);
1381                 } else {
1382                         bne = data + (offset << 9);
1383                         if (bne->keys.seq != bn->keys.seq)
1384                                 break;
1385                         offset += vstruct_sectors(bne, c->block_bits);
1386                 }
1387         }
1388
1389         return offset;
1390 }
1391
1392 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1393 {
1394         struct btree_node *bn = data;
1395         struct btree_node_entry *bne;
1396
1397         if (!offset)
1398                 return false;
1399
1400         while (offset < btree_sectors(c)) {
1401                 bne = data + (offset << 9);
1402                 if (bne->keys.seq == bn->keys.seq)
1403                         return true;
1404                 offset++;
1405         }
1406
1407         return false;
1408         return offset;
1409 }
1410
1411 static CLOSURE_CALLBACK(btree_node_read_all_replicas_done)
1412 {
1413         closure_type(ra, struct btree_node_read_all, cl);
1414         struct bch_fs *c = ra->c;
1415         struct btree *b = ra->b;
1416         struct printbuf buf = PRINTBUF;
1417         bool dump_bset_maps = false;
1418         bool have_retry = false;
1419         int ret = 0, best = -1, write = READ;
1420         unsigned i, written = 0, written2 = 0;
1421         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1422                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1423         bool _saw_error = false, *saw_error = &_saw_error;
1424
1425         for (i = 0; i < ra->nr; i++) {
1426                 struct btree_node *bn = ra->buf[i];
1427
1428                 if (ra->err[i])
1429                         continue;
1430
1431                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1432                     (seq && seq != bn->keys.seq))
1433                         continue;
1434
1435                 if (best < 0) {
1436                         best = i;
1437                         written = btree_node_sectors_written(c, bn);
1438                         continue;
1439                 }
1440
1441                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1442                 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable,
1443                                  c, NULL, b, NULL,
1444                                  btree_node_replicas_sectors_written_mismatch,
1445                                  "btree node sectors written mismatch: %u != %u",
1446                                  written, written2) ||
1447                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1448                                  -BCH_ERR_btree_node_read_err_fixable,
1449                                  c, NULL, b, NULL,
1450                                  btree_node_bset_after_end,
1451                                  "found bset signature after last bset") ||
1452                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1453                                  -BCH_ERR_btree_node_read_err_fixable,
1454                                  c, NULL, b, NULL,
1455                                  btree_node_replicas_data_mismatch,
1456                                  "btree node replicas content mismatch"))
1457                         dump_bset_maps = true;
1458
1459                 if (written2 > written) {
1460                         written = written2;
1461                         best = i;
1462                 }
1463         }
1464 fsck_err:
1465         if (dump_bset_maps) {
1466                 for (i = 0; i < ra->nr; i++) {
1467                         struct btree_node *bn = ra->buf[i];
1468                         struct btree_node_entry *bne = NULL;
1469                         unsigned offset = 0, sectors;
1470                         bool gap = false;
1471
1472                         if (ra->err[i])
1473                                 continue;
1474
1475                         printbuf_reset(&buf);
1476
1477                         while (offset < btree_sectors(c)) {
1478                                 if (!offset) {
1479                                         sectors = vstruct_sectors(bn, c->block_bits);
1480                                 } else {
1481                                         bne = ra->buf[i] + (offset << 9);
1482                                         if (bne->keys.seq != bn->keys.seq)
1483                                                 break;
1484                                         sectors = vstruct_sectors(bne, c->block_bits);
1485                                 }
1486
1487                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1488                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1489                                                         le64_to_cpu(bne->keys.journal_seq), false))
1490                                         prt_printf(&buf, "*");
1491                                 offset += sectors;
1492                         }
1493
1494                         while (offset < btree_sectors(c)) {
1495                                 bne = ra->buf[i] + (offset << 9);
1496                                 if (bne->keys.seq == bn->keys.seq) {
1497                                         if (!gap)
1498                                                 prt_printf(&buf, " GAP");
1499                                         gap = true;
1500
1501                                         sectors = vstruct_sectors(bne, c->block_bits);
1502                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1503                                         if (bch2_journal_seq_is_blacklisted(c,
1504                                                         le64_to_cpu(bne->keys.journal_seq), false))
1505                                                 prt_printf(&buf, "*");
1506                                 }
1507                                 offset++;
1508                         }
1509
1510                         bch_err(c, "replica %u:%s", i, buf.buf);
1511                 }
1512         }
1513
1514         if (best >= 0) {
1515                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1516                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1517         } else {
1518                 ret = -1;
1519         }
1520
1521         if (ret)
1522                 set_btree_node_read_error(b);
1523         else if (*saw_error)
1524                 bch2_btree_node_rewrite_async(c, b);
1525
1526         for (i = 0; i < ra->nr; i++) {
1527                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1528                 bio_put(ra->bio[i]);
1529         }
1530
1531         closure_debug_destroy(&ra->cl);
1532         kfree(ra);
1533         printbuf_exit(&buf);
1534
1535         clear_btree_node_read_in_flight(b);
1536         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1537 }
1538
1539 static void btree_node_read_all_replicas_endio(struct bio *bio)
1540 {
1541         struct btree_read_bio *rb =
1542                 container_of(bio, struct btree_read_bio, bio);
1543         struct bch_fs *c        = rb->c;
1544         struct btree_node_read_all *ra = rb->ra;
1545
1546         if (rb->have_ioref) {
1547                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1548
1549                 bch2_latency_acct(ca, rb->start_time, READ);
1550         }
1551
1552         ra->err[rb->idx] = bio->bi_status;
1553         closure_put(&ra->cl);
1554 }
1555
1556 /*
1557  * XXX This allocates multiple times from the same mempools, and can deadlock
1558  * under sufficient memory pressure (but is only a debug path)
1559  */
1560 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1561 {
1562         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1563         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1564         const union bch_extent_entry *entry;
1565         struct extent_ptr_decoded pick;
1566         struct btree_node_read_all *ra;
1567         unsigned i;
1568
1569         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1570         if (!ra)
1571                 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1572
1573         closure_init(&ra->cl, NULL);
1574         ra->c   = c;
1575         ra->b   = b;
1576         ra->nr  = bch2_bkey_nr_ptrs(k);
1577
1578         for (i = 0; i < ra->nr; i++) {
1579                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1580                 ra->bio[i] = bio_alloc_bioset(NULL,
1581                                               buf_pages(ra->buf[i], btree_bytes(c)),
1582                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1583                                               GFP_NOFS,
1584                                               &c->btree_bio);
1585         }
1586
1587         i = 0;
1588         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1589                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1590                 struct btree_read_bio *rb =
1591                         container_of(ra->bio[i], struct btree_read_bio, bio);
1592                 rb->c                   = c;
1593                 rb->b                   = b;
1594                 rb->ra                  = ra;
1595                 rb->start_time          = local_clock();
1596                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1597                 rb->idx                 = i;
1598                 rb->pick                = pick;
1599                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1600                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1601                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1602
1603                 if (rb->have_ioref) {
1604                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1605                                      bio_sectors(&rb->bio));
1606                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1607
1608                         closure_get(&ra->cl);
1609                         submit_bio(&rb->bio);
1610                 } else {
1611                         ra->err[i] = BLK_STS_REMOVED;
1612                 }
1613
1614                 i++;
1615         }
1616
1617         if (sync) {
1618                 closure_sync(&ra->cl);
1619                 btree_node_read_all_replicas_done(&ra->cl.work);
1620         } else {
1621                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1622                             c->io_complete_wq);
1623         }
1624
1625         return 0;
1626 }
1627
1628 void bch2_btree_node_read(struct btree_trans *trans, struct btree *b,
1629                           bool sync)
1630 {
1631         struct bch_fs *c = trans->c;
1632         struct extent_ptr_decoded pick;
1633         struct btree_read_bio *rb;
1634         struct bch_dev *ca;
1635         struct bio *bio;
1636         int ret;
1637
1638         trace_and_count(c, btree_node_read, trans, b);
1639
1640         if (bch2_verify_all_btree_replicas &&
1641             !btree_node_read_all_replicas(c, b, sync))
1642                 return;
1643
1644         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1645                                          NULL, &pick);
1646
1647         if (ret <= 0) {
1648                 struct printbuf buf = PRINTBUF;
1649
1650                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1651                 bch2_btree_pos_to_text(&buf, c, b);
1652                 bch_err(c, "%s", buf.buf);
1653
1654                 if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1655                     c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1656                         bch2_fatal_error(c);
1657
1658                 set_btree_node_read_error(b);
1659                 clear_btree_node_read_in_flight(b);
1660                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1661                 printbuf_exit(&buf);
1662                 return;
1663         }
1664
1665         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1666
1667         bio = bio_alloc_bioset(NULL,
1668                                buf_pages(b->data, btree_bytes(c)),
1669                                REQ_OP_READ|REQ_SYNC|REQ_META,
1670                                GFP_NOFS,
1671                                &c->btree_bio);
1672         rb = container_of(bio, struct btree_read_bio, bio);
1673         rb->c                   = c;
1674         rb->b                   = b;
1675         rb->ra                  = NULL;
1676         rb->start_time          = local_clock();
1677         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1678         rb->pick                = pick;
1679         INIT_WORK(&rb->work, btree_node_read_work);
1680         bio->bi_iter.bi_sector  = pick.ptr.offset;
1681         bio->bi_end_io          = btree_node_read_endio;
1682         bch2_bio_map(bio, b->data, btree_bytes(c));
1683
1684         if (rb->have_ioref) {
1685                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1686                              bio_sectors(bio));
1687                 bio_set_dev(bio, ca->disk_sb.bdev);
1688
1689                 if (sync) {
1690                         submit_bio_wait(bio);
1691                         bch2_latency_acct(ca, rb->start_time, READ);
1692                         btree_node_read_work(&rb->work);
1693                 } else {
1694                         submit_bio(bio);
1695                 }
1696         } else {
1697                 bio->bi_status = BLK_STS_REMOVED;
1698
1699                 if (sync)
1700                         btree_node_read_work(&rb->work);
1701                 else
1702                         queue_work(c->io_complete_wq, &rb->work);
1703         }
1704 }
1705
1706 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1707                                   const struct bkey_i *k, unsigned level)
1708 {
1709         struct bch_fs *c = trans->c;
1710         struct closure cl;
1711         struct btree *b;
1712         int ret;
1713
1714         closure_init_stack(&cl);
1715
1716         do {
1717                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
1718                 closure_sync(&cl);
1719         } while (ret);
1720
1721         b = bch2_btree_node_mem_alloc(trans, level != 0);
1722         bch2_btree_cache_cannibalize_unlock(trans);
1723
1724         BUG_ON(IS_ERR(b));
1725
1726         bkey_copy(&b->key, k);
1727         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1728
1729         set_btree_node_read_in_flight(b);
1730
1731         bch2_btree_node_read(trans, b, true);
1732
1733         if (btree_node_read_error(b)) {
1734                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1735
1736                 mutex_lock(&c->btree_cache.lock);
1737                 list_move(&b->list, &c->btree_cache.freeable);
1738                 mutex_unlock(&c->btree_cache.lock);
1739
1740                 ret = -EIO;
1741                 goto err;
1742         }
1743
1744         bch2_btree_set_root_for_read(c, b);
1745 err:
1746         six_unlock_write(&b->c.lock);
1747         six_unlock_intent(&b->c.lock);
1748
1749         return ret;
1750 }
1751
1752 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1753                         const struct bkey_i *k, unsigned level)
1754 {
1755         return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
1756 }
1757
1758 static void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1759                                       struct btree_write *w)
1760 {
1761         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1762
1763         do {
1764                 old = new = v;
1765                 if (!(old & 1))
1766                         break;
1767
1768                 new &= ~1UL;
1769         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1770
1771         if (old & 1)
1772                 closure_put(&((struct btree_update *) new)->cl);
1773
1774         bch2_journal_pin_drop(&c->journal, &w->journal);
1775 }
1776
1777 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1778 {
1779         struct btree_write *w = btree_prev_write(b);
1780         unsigned long old, new, v;
1781         unsigned type = 0;
1782
1783         bch2_btree_complete_write(c, b, w);
1784
1785         v = READ_ONCE(b->flags);
1786         do {
1787                 old = new = v;
1788
1789                 if ((old & (1U << BTREE_NODE_dirty)) &&
1790                     (old & (1U << BTREE_NODE_need_write)) &&
1791                     !(old & (1U << BTREE_NODE_never_write)) &&
1792                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1793                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1794                         new &= ~(1U << BTREE_NODE_dirty);
1795                         new &= ~(1U << BTREE_NODE_need_write);
1796                         new |=  (1U << BTREE_NODE_write_in_flight);
1797                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1798                         new |=  (1U << BTREE_NODE_just_written);
1799                         new ^=  (1U << BTREE_NODE_write_idx);
1800
1801                         type = new & BTREE_WRITE_TYPE_MASK;
1802                         new &= ~BTREE_WRITE_TYPE_MASK;
1803                 } else {
1804                         new &= ~(1U << BTREE_NODE_write_in_flight);
1805                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1806                 }
1807         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1808
1809         if (new & (1U << BTREE_NODE_write_in_flight))
1810                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1811         else
1812                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1813 }
1814
1815 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1816 {
1817         struct btree_trans *trans = bch2_trans_get(c);
1818
1819         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
1820         __btree_node_write_done(c, b);
1821         six_unlock_read(&b->c.lock);
1822
1823         bch2_trans_put(trans);
1824 }
1825
1826 static void btree_node_write_work(struct work_struct *work)
1827 {
1828         struct btree_write_bio *wbio =
1829                 container_of(work, struct btree_write_bio, work);
1830         struct bch_fs *c        = wbio->wbio.c;
1831         struct btree *b         = wbio->wbio.bio.bi_private;
1832         struct bch_extent_ptr *ptr;
1833         int ret = 0;
1834
1835         btree_bounce_free(c,
1836                 wbio->data_bytes,
1837                 wbio->wbio.used_mempool,
1838                 wbio->data);
1839
1840         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1841                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1842
1843         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key))) {
1844                 ret = -BCH_ERR_btree_write_all_failed;
1845                 goto err;
1846         }
1847
1848         if (wbio->wbio.first_btree_write) {
1849                 if (wbio->wbio.failed.nr) {
1850
1851                 }
1852         } else {
1853                 ret = bch2_trans_do(c, NULL, NULL, 0,
1854                         bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
1855                                         BCH_WATERMARK_reclaim|
1856                                         BCH_TRANS_COMMIT_journal_reclaim|
1857                                         BCH_TRANS_COMMIT_no_enospc|
1858                                         BCH_TRANS_COMMIT_no_check_rw,
1859                                         !wbio->wbio.failed.nr));
1860                 if (ret)
1861                         goto err;
1862         }
1863 out:
1864         bio_put(&wbio->wbio.bio);
1865         btree_node_write_done(c, b);
1866         return;
1867 err:
1868         set_btree_node_noevict(b);
1869         if (!bch2_err_matches(ret, EROFS))
1870                 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1871         goto out;
1872 }
1873
1874 static void btree_node_write_endio(struct bio *bio)
1875 {
1876         struct bch_write_bio *wbio      = to_wbio(bio);
1877         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1878         struct bch_write_bio *orig      = parent ?: wbio;
1879         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1880         struct bch_fs *c                = wbio->c;
1881         struct btree *b                 = wbio->bio.bi_private;
1882         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1883         unsigned long flags;
1884
1885         if (wbio->have_ioref)
1886                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1887
1888         if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1889                                "btree write error: %s",
1890                                bch2_blk_status_to_str(bio->bi_status)) ||
1891             bch2_meta_write_fault("btree")) {
1892                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1893                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1894                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1895         }
1896
1897         if (wbio->have_ioref)
1898                 percpu_ref_put(&ca->io_ref);
1899
1900         if (parent) {
1901                 bio_put(bio);
1902                 bio_endio(&parent->bio);
1903                 return;
1904         }
1905
1906         clear_btree_node_write_in_flight_inner(b);
1907         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1908         INIT_WORK(&wb->work, btree_node_write_work);
1909         queue_work(c->btree_io_complete_wq, &wb->work);
1910 }
1911
1912 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1913                                    struct bset *i, unsigned sectors)
1914 {
1915         struct printbuf buf = PRINTBUF;
1916         bool saw_error;
1917         int ret;
1918
1919         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1920                                 BKEY_TYPE_btree, WRITE, &buf);
1921
1922         if (ret)
1923                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1924         printbuf_exit(&buf);
1925         if (ret)
1926                 return ret;
1927
1928         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1929                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1930         if (ret) {
1931                 bch2_inconsistent_error(c);
1932                 dump_stack();
1933         }
1934
1935         return ret;
1936 }
1937
1938 static void btree_write_submit(struct work_struct *work)
1939 {
1940         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1941         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1942
1943         bkey_copy(&tmp.k, &wbio->key);
1944
1945         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1946                 ptr->offset += wbio->sector_offset;
1947
1948         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1949                                   &tmp.k, false);
1950 }
1951
1952 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1953 {
1954         struct btree_write_bio *wbio;
1955         struct bset_tree *t;
1956         struct bset *i;
1957         struct btree_node *bn = NULL;
1958         struct btree_node_entry *bne = NULL;
1959         struct sort_iter_stack sort_iter;
1960         struct nonce nonce;
1961         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1962         u64 seq = 0;
1963         bool used_mempool;
1964         unsigned long old, new;
1965         bool validate_before_checksum = false;
1966         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1967         void *data;
1968         int ret;
1969
1970         if (flags & BTREE_WRITE_ALREADY_STARTED)
1971                 goto do_write;
1972
1973         /*
1974          * We may only have a read lock on the btree node - the dirty bit is our
1975          * "lock" against racing with other threads that may be trying to start
1976          * a write, we do a write iff we clear the dirty bit. Since setting the
1977          * dirty bit requires a write lock, we can't race with other threads
1978          * redirtying it:
1979          */
1980         do {
1981                 old = new = READ_ONCE(b->flags);
1982
1983                 if (!(old & (1 << BTREE_NODE_dirty)))
1984                         return;
1985
1986                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1987                     !(old & (1 << BTREE_NODE_need_write)))
1988                         return;
1989
1990                 if (old &
1991                     ((1 << BTREE_NODE_never_write)|
1992                      (1 << BTREE_NODE_write_blocked)))
1993                         return;
1994
1995                 if (b->written &&
1996                     (old & (1 << BTREE_NODE_will_make_reachable)))
1997                         return;
1998
1999                 if (old & (1 << BTREE_NODE_write_in_flight))
2000                         return;
2001
2002                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
2003                         type = new & BTREE_WRITE_TYPE_MASK;
2004                 new &= ~BTREE_WRITE_TYPE_MASK;
2005
2006                 new &= ~(1 << BTREE_NODE_dirty);
2007                 new &= ~(1 << BTREE_NODE_need_write);
2008                 new |=  (1 << BTREE_NODE_write_in_flight);
2009                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
2010                 new |=  (1 << BTREE_NODE_just_written);
2011                 new ^=  (1 << BTREE_NODE_write_idx);
2012         } while (cmpxchg_acquire(&b->flags, old, new) != old);
2013
2014         if (new & (1U << BTREE_NODE_need_write))
2015                 return;
2016 do_write:
2017         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
2018
2019         atomic_dec(&c->btree_cache.dirty);
2020
2021         BUG_ON(btree_node_fake(b));
2022         BUG_ON((b->will_make_reachable != 0) != !b->written);
2023
2024         BUG_ON(b->written >= btree_sectors(c));
2025         BUG_ON(b->written & (block_sectors(c) - 1));
2026         BUG_ON(bset_written(b, btree_bset_last(b)));
2027         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
2028         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
2029
2030         bch2_sort_whiteouts(c, b);
2031
2032         sort_iter_stack_init(&sort_iter, b);
2033
2034         bytes = !b->written
2035                 ? sizeof(struct btree_node)
2036                 : sizeof(struct btree_node_entry);
2037
2038         bytes += b->whiteout_u64s * sizeof(u64);
2039
2040         for_each_bset(b, t) {
2041                 i = bset(b, t);
2042
2043                 if (bset_written(b, i))
2044                         continue;
2045
2046                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
2047                 sort_iter_add(&sort_iter.iter,
2048                               btree_bkey_first(b, t),
2049                               btree_bkey_last(b, t));
2050                 seq = max(seq, le64_to_cpu(i->journal_seq));
2051         }
2052
2053         BUG_ON(b->written && !seq);
2054
2055         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
2056         bytes += 8;
2057
2058         /* buffer must be a multiple of the block size */
2059         bytes = round_up(bytes, block_bytes(c));
2060
2061         data = btree_bounce_alloc(c, bytes, &used_mempool);
2062
2063         if (!b->written) {
2064                 bn = data;
2065                 *bn = *b->data;
2066                 i = &bn->keys;
2067         } else {
2068                 bne = data;
2069                 bne->keys = b->data->keys;
2070                 i = &bne->keys;
2071         }
2072
2073         i->journal_seq  = cpu_to_le64(seq);
2074         i->u64s         = 0;
2075
2076         sort_iter_add(&sort_iter.iter,
2077                       unwritten_whiteouts_start(c, b),
2078                       unwritten_whiteouts_end(c, b));
2079         SET_BSET_SEPARATE_WHITEOUTS(i, false);
2080
2081         b->whiteout_u64s = 0;
2082
2083         u64s = bch2_sort_keys(i->start, &sort_iter.iter, false);
2084         le16_add_cpu(&i->u64s, u64s);
2085
2086         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
2087
2088         set_needs_whiteout(i, false);
2089
2090         /* do we have data to write? */
2091         if (b->written && !i->u64s)
2092                 goto nowrite;
2093
2094         bytes_to_write = vstruct_end(i) - data;
2095         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
2096
2097         if (!b->written &&
2098             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2099                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
2100
2101         memset(data + bytes_to_write, 0,
2102                (sectors_to_write << 9) - bytes_to_write);
2103
2104         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2105         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2106         BUG_ON(i->seq != b->data->keys.seq);
2107
2108         i->version = cpu_to_le16(c->sb.version);
2109         SET_BSET_OFFSET(i, b->written);
2110         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2111
2112         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2113                 validate_before_checksum = true;
2114
2115         /* validate_bset will be modifying: */
2116         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2117                 validate_before_checksum = true;
2118
2119         /* if we're going to be encrypting, check metadata validity first: */
2120         if (validate_before_checksum &&
2121             validate_bset_for_write(c, b, i, sectors_to_write))
2122                 goto err;
2123
2124         ret = bset_encrypt(c, i, b->written << 9);
2125         if (bch2_fs_fatal_err_on(ret, c,
2126                         "error encrypting btree node: %i\n", ret))
2127                 goto err;
2128
2129         nonce = btree_nonce(i, b->written << 9);
2130
2131         if (bn)
2132                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2133         else
2134                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2135
2136         /* if we're not encrypting, check metadata after checksumming: */
2137         if (!validate_before_checksum &&
2138             validate_bset_for_write(c, b, i, sectors_to_write))
2139                 goto err;
2140
2141         /*
2142          * We handle btree write errors by immediately halting the journal -
2143          * after we've done that, we can't issue any subsequent btree writes
2144          * because they might have pointers to new nodes that failed to write.
2145          *
2146          * Furthermore, there's no point in doing any more btree writes because
2147          * with the journal stopped, we're never going to update the journal to
2148          * reflect that those writes were done and the data flushed from the
2149          * journal:
2150          *
2151          * Also on journal error, the pending write may have updates that were
2152          * never journalled (interior nodes, see btree_update_nodes_written()) -
2153          * it's critical that we don't do the write in that case otherwise we
2154          * will have updates visible that weren't in the journal:
2155          *
2156          * Make sure to update b->written so bch2_btree_init_next() doesn't
2157          * break:
2158          */
2159         if (bch2_journal_error(&c->journal) ||
2160             c->opts.nochanges)
2161                 goto err;
2162
2163         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2164
2165         wbio = container_of(bio_alloc_bioset(NULL,
2166                                 buf_pages(data, sectors_to_write << 9),
2167                                 REQ_OP_WRITE|REQ_META,
2168                                 GFP_NOFS,
2169                                 &c->btree_bio),
2170                             struct btree_write_bio, wbio.bio);
2171         wbio_init(&wbio->wbio.bio);
2172         wbio->data                      = data;
2173         wbio->data_bytes                = bytes;
2174         wbio->sector_offset             = b->written;
2175         wbio->wbio.c                    = c;
2176         wbio->wbio.used_mempool         = used_mempool;
2177         wbio->wbio.first_btree_write    = !b->written;
2178         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2179         wbio->wbio.bio.bi_private       = b;
2180
2181         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2182
2183         bkey_copy(&wbio->key, &b->key);
2184
2185         b->written += sectors_to_write;
2186
2187         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2188                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2189                         cpu_to_le16(b->written);
2190
2191         atomic64_inc(&c->btree_write_stats[type].nr);
2192         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2193
2194         INIT_WORK(&wbio->work, btree_write_submit);
2195         queue_work(c->io_complete_wq, &wbio->work);
2196         return;
2197 err:
2198         set_btree_node_noevict(b);
2199         b->written += sectors_to_write;
2200 nowrite:
2201         btree_bounce_free(c, bytes, used_mempool, data);
2202         __btree_node_write_done(c, b);
2203 }
2204
2205 /*
2206  * Work that must be done with write lock held:
2207  */
2208 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2209 {
2210         bool invalidated_iter = false;
2211         struct btree_node_entry *bne;
2212         struct bset_tree *t;
2213
2214         if (!btree_node_just_written(b))
2215                 return false;
2216
2217         BUG_ON(b->whiteout_u64s);
2218
2219         clear_btree_node_just_written(b);
2220
2221         /*
2222          * Note: immediately after write, bset_written() doesn't work - the
2223          * amount of data we had to write after compaction might have been
2224          * smaller than the offset of the last bset.
2225          *
2226          * However, we know that all bsets have been written here, as long as
2227          * we're still holding the write lock:
2228          */
2229
2230         /*
2231          * XXX: decide if we really want to unconditionally sort down to a
2232          * single bset:
2233          */
2234         if (b->nsets > 1) {
2235                 btree_node_sort(c, b, 0, b->nsets, true);
2236                 invalidated_iter = true;
2237         } else {
2238                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2239         }
2240
2241         for_each_bset(b, t)
2242                 set_needs_whiteout(bset(b, t), true);
2243
2244         bch2_btree_verify(c, b);
2245
2246         /*
2247          * If later we don't unconditionally sort down to a single bset, we have
2248          * to ensure this is still true:
2249          */
2250         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2251
2252         bne = want_new_bset(c, b);
2253         if (bne)
2254                 bch2_bset_init_next(c, b, bne);
2255
2256         bch2_btree_build_aux_trees(b);
2257
2258         return invalidated_iter;
2259 }
2260
2261 /*
2262  * Use this one if the node is intent locked:
2263  */
2264 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2265                            enum six_lock_type lock_type_held,
2266                            unsigned flags)
2267 {
2268         if (lock_type_held == SIX_LOCK_intent ||
2269             (lock_type_held == SIX_LOCK_read &&
2270              six_lock_tryupgrade(&b->c.lock))) {
2271                 __bch2_btree_node_write(c, b, flags);
2272
2273                 /* don't cycle lock unnecessarily: */
2274                 if (btree_node_just_written(b) &&
2275                     six_trylock_write(&b->c.lock)) {
2276                         bch2_btree_post_write_cleanup(c, b);
2277                         six_unlock_write(&b->c.lock);
2278                 }
2279
2280                 if (lock_type_held == SIX_LOCK_read)
2281                         six_lock_downgrade(&b->c.lock);
2282         } else {
2283                 __bch2_btree_node_write(c, b, flags);
2284                 if (lock_type_held == SIX_LOCK_write &&
2285                     btree_node_just_written(b))
2286                         bch2_btree_post_write_cleanup(c, b);
2287         }
2288 }
2289
2290 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2291 {
2292         struct bucket_table *tbl;
2293         struct rhash_head *pos;
2294         struct btree *b;
2295         unsigned i;
2296         bool ret = false;
2297 restart:
2298         rcu_read_lock();
2299         for_each_cached_btree(b, c, tbl, i, pos)
2300                 if (test_bit(flag, &b->flags)) {
2301                         rcu_read_unlock();
2302                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2303                         ret = true;
2304                         goto restart;
2305                 }
2306         rcu_read_unlock();
2307
2308         return ret;
2309 }
2310
2311 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2312 {
2313         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2314 }
2315
2316 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2317 {
2318         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2319 }
2320
2321 static const char * const bch2_btree_write_types[] = {
2322 #define x(t, n) [n] = #t,
2323         BCH_BTREE_WRITE_TYPES()
2324         NULL
2325 };
2326
2327 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2328 {
2329         printbuf_tabstop_push(out, 20);
2330         printbuf_tabstop_push(out, 10);
2331
2332         prt_tab(out);
2333         prt_str(out, "nr");
2334         prt_tab(out);
2335         prt_str(out, "size");
2336         prt_newline(out);
2337
2338         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2339                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2340                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2341
2342                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2343                 prt_tab(out);
2344                 prt_u64(out, nr);
2345                 prt_tab(out);
2346                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2347                 prt_newline(out);
2348         }
2349 }