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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_sort.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "buckets.h"
  13 #include "checksum.h"
  14 #include "debug.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "io_write.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_seq_blacklist.h"
  20 #include "recovery.h"
  21 #include "super-io.h"
  22 #include "trace.h"
  23
  24 #include <linux/sched/mm.h>
  25
  26 void bch2_btree_node_io_unlock(struct btree *b)
  27 {
  28         EBUG_ON(!btree_node_write_in_flight(b));
  29
  30         clear_btree_node_write_in_flight_inner(b);
  31         clear_btree_node_write_in_flight(b);
  32         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
  33 }
  34
  35 void bch2_btree_node_io_lock(struct btree *b)
  36 {
  37         bch2_assert_btree_nodes_not_locked();
  38
  39         wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
  40                             TASK_UNINTERRUPTIBLE);
  41 }
  42
  43 void __bch2_btree_node_wait_on_read(struct btree *b)
  44 {
  45         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  46                        TASK_UNINTERRUPTIBLE);
  47 }
  48
  49 void __bch2_btree_node_wait_on_write(struct btree *b)
  50 {
  51         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  52                        TASK_UNINTERRUPTIBLE);
  53 }
  54
  55 void bch2_btree_node_wait_on_read(struct btree *b)
  56 {
  57         bch2_assert_btree_nodes_not_locked();
  58
  59         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  60                        TASK_UNINTERRUPTIBLE);
  61 }
  62
  63 void bch2_btree_node_wait_on_write(struct btree *b)
  64 {
  65         bch2_assert_btree_nodes_not_locked();
  66
  67         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  68                        TASK_UNINTERRUPTIBLE);
  69 }
  70
  71 static void verify_no_dups(struct btree *b,
  72                            struct bkey_packed *start,
  73                            struct bkey_packed *end)
  74 {
  75 #ifdef CONFIG_BCACHEFS_DEBUG
  76         struct bkey_packed *k, *p;
  77
  78         if (start == end)
  79                 return;
  80
  81         for (p = start, k = bkey_p_next(start);
  82              k != end;
  83              p = k, k = bkey_p_next(k)) {
  84                 struct bkey l = bkey_unpack_key(b, p);
  85                 struct bkey r = bkey_unpack_key(b, k);
  86
  87                 BUG_ON(bpos_ge(l.p, bkey_start_pos(&r)));
  88         }
  89 #endif
  90 }
  91
  92 static void set_needs_whiteout(struct bset *i, int v)
  93 {
  94         struct bkey_packed *k;
  95
  96         for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
  97                 k->needs_whiteout = v;
  98 }
  99
 100 static void btree_bounce_free(struct bch_fs *c, size_t size,
 101                               bool used_mempool, void *p)
 102 {
 103         if (used_mempool)
 104                 mempool_free(p, &c->btree_bounce_pool);
 105         else
 106                 vpfree(p, size);
 107 }
 108
 109 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
 110                                 bool *used_mempool)
 111 {
 112         unsigned flags = memalloc_nofs_save();
 113         void *p;
 114
 115         BUG_ON(size > btree_bytes(c));
 116
 117         *used_mempool = false;
 118         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
 119         if (!p) {
 120                 *used_mempool = true;
 121                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
 122         }
 123         memalloc_nofs_restore(flags);
 124         return p;
 125 }
 126
 127 static void sort_bkey_ptrs(const struct btree *bt,
 128                            struct bkey_packed **ptrs, unsigned nr)
 129 {
 130         unsigned n = nr, a = nr / 2, b, c, d;
 131
 132         if (!a)
 133                 return;
 134
 135         /* Heap sort: see lib/sort.c: */
 136         while (1) {
 137                 if (a)
 138                         a--;
 139                 else if (--n)
 140                         swap(ptrs[0], ptrs[n]);
 141                 else
 142                         break;
 143
 144                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 145                         b = bch2_bkey_cmp_packed(bt,
 146                                             ptrs[c],
 147                                             ptrs[d]) >= 0 ? c : d;
 148                 if (d == n)
 149                         b = c;
 150
 151                 while (b != a &&
 152                        bch2_bkey_cmp_packed(bt,
 153                                        ptrs[a],
 154                                        ptrs[b]) >= 0)
 155                         b = (b - 1) / 2;
 156                 c = b;
 157                 while (b != a) {
 158                         b = (b - 1) / 2;
 159                         swap(ptrs[b], ptrs[c]);
 160                 }
 161         }
 162 }
 163
 164 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 165 {
 166         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 167         bool used_mempool = false;
 168         size_t bytes = b->whiteout_u64s * sizeof(u64);
 169
 170         if (!b->whiteout_u64s)
 171                 return;
 172
 173         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 174
 175         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 176
 177         for (k = unwritten_whiteouts_start(c, b);
 178              k != unwritten_whiteouts_end(c, b);
 179              k = bkey_p_next(k))
 180                 *--ptrs = k;
 181
 182         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 183
 184         k = new_whiteouts;
 185
 186         while (ptrs != ptrs_end) {
 187                 bkey_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_copy(out, k);
 264                                 out = bkey_p_next(out);
 265                         } else {
 266                                 BUG_ON(k->needs_whiteout);
 267                         }
 268                 }
 269
 270                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 271                 set_btree_bset_end(b, t);
 272                 bch2_bset_set_no_aux_tree(b, t);
 273                 ret = true;
 274         }
 275
 276         bch2_verify_btree_nr_keys(b);
 277
 278         bch2_btree_build_aux_trees(b);
 279
 280         return ret;
 281 }
 282
 283 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 284                             enum compact_mode mode)
 285 {
 286         return bch2_drop_whiteouts(b, mode);
 287 }
 288
 289 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 290                             unsigned start_idx,
 291                             unsigned end_idx,
 292                             bool filter_whiteouts)
 293 {
 294         struct btree_node *out;
 295         struct sort_iter_stack sort_iter;
 296         struct bset_tree *t;
 297         struct bset *start_bset = bset(b, &b->set[start_idx]);
 298         bool used_mempool = false;
 299         u64 start_time, seq = 0;
 300         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 301         bool sorting_entire_node = start_idx == 0 &&
 302                 end_idx == b->nsets;
 303
 304         sort_iter_stack_init(&sort_iter, b);
 305
 306         for (t = b->set + start_idx;
 307              t < b->set + end_idx;
 308              t++) {
 309                 u64s += le16_to_cpu(bset(b, t)->u64s);
 310                 sort_iter_add(&sort_iter.iter,
 311                               btree_bkey_first(b, t),
 312                               btree_bkey_last(b, t));
 313         }
 314
 315         bytes = sorting_entire_node
 316                 ? btree_bytes(c)
 317                 : __vstruct_bytes(struct btree_node, u64s);
 318
 319         out = btree_bounce_alloc(c, bytes, &used_mempool);
 320
 321         start_time = local_clock();
 322
 323         u64s = bch2_sort_keys(out->keys.start, &sort_iter.iter, filter_whiteouts);
 324
 325         out->keys.u64s = cpu_to_le16(u64s);
 326
 327         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 328
 329         if (sorting_entire_node)
 330                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 331                                        start_time);
 332
 333         /* Make sure we preserve bset journal_seq: */
 334         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 335                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 336         start_bset->journal_seq = cpu_to_le64(seq);
 337
 338         if (sorting_entire_node) {
 339                 u64s = le16_to_cpu(out->keys.u64s);
 340
 341                 BUG_ON(bytes != btree_bytes(c));
 342
 343                 /*
 344                  * Our temporary buffer is the same size as the btree node's
 345                  * buffer, we can just swap buffers instead of doing a big
 346                  * memcpy()
 347                  */
 348                 *out = *b->data;
 349                 out->keys.u64s = cpu_to_le16(u64s);
 350                 swap(out, b->data);
 351                 set_btree_bset(b, b->set, &b->data->keys);
 352         } else {
 353                 start_bset->u64s = out->keys.u64s;
 354                 memcpy_u64s(start_bset->start,
 355                             out->keys.start,
 356                             le16_to_cpu(out->keys.u64s));
 357         }
 358
 359         for (i = start_idx + 1; i < end_idx; i++)
 360                 b->nr.bset_u64s[start_idx] +=
 361                         b->nr.bset_u64s[i];
 362
 363         b->nsets -= shift;
 364
 365         for (i = start_idx + 1; i < b->nsets; i++) {
 366                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 367                 b->set[i]               = b->set[i + shift];
 368         }
 369
 370         for (i = b->nsets; i < MAX_BSETS; i++)
 371                 b->nr.bset_u64s[i] = 0;
 372
 373         set_btree_bset_end(b, &b->set[start_idx]);
 374         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 375
 376         btree_bounce_free(c, bytes, used_mempool, out);
 377
 378         bch2_verify_btree_nr_keys(b);
 379 }
 380
 381 void bch2_btree_sort_into(struct bch_fs *c,
 382                          struct btree *dst,
 383                          struct btree *src)
 384 {
 385         struct btree_nr_keys nr;
 386         struct btree_node_iter src_iter;
 387         u64 start_time = local_clock();
 388
 389         BUG_ON(dst->nsets != 1);
 390
 391         bch2_bset_set_no_aux_tree(dst, dst->set);
 392
 393         bch2_btree_node_iter_init_from_start(&src_iter, src);
 394
 395         nr = bch2_sort_repack(btree_bset_first(dst),
 396                         src, &src_iter,
 397                         &dst->format,
 398                         true);
 399
 400         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 401                                start_time);
 402
 403         set_btree_bset_end(dst, dst->set);
 404
 405         dst->nr.live_u64s       += nr.live_u64s;
 406         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 407         dst->nr.packed_keys     += nr.packed_keys;
 408         dst->nr.unpacked_keys   += nr.unpacked_keys;
 409
 410         bch2_verify_btree_nr_keys(dst);
 411 }
 412
 413 /*
 414  * We're about to add another bset to the btree node, so if there's currently
 415  * too many bsets - sort some of them together:
 416  */
 417 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 418 {
 419         unsigned unwritten_idx;
 420         bool ret = false;
 421
 422         for (unwritten_idx = 0;
 423              unwritten_idx < b->nsets;
 424              unwritten_idx++)
 425                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 426                         break;
 427
 428         if (b->nsets - unwritten_idx > 1) {
 429                 btree_node_sort(c, b, unwritten_idx,
 430                                 b->nsets, false);
 431                 ret = true;
 432         }
 433
 434         if (unwritten_idx > 1) {
 435                 btree_node_sort(c, b, 0, unwritten_idx, false);
 436                 ret = true;
 437         }
 438
 439         return ret;
 440 }
 441
 442 void bch2_btree_build_aux_trees(struct btree *b)
 443 {
 444         struct bset_tree *t;
 445
 446         for_each_bset(b, t)
 447                 bch2_bset_build_aux_tree(b, t,
 448                                 !bset_written(b, bset(b, t)) &&
 449                                 t == bset_tree_last(b));
 450 }
 451
 452 /*
 453  * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
 454  *
 455  * The first bset is going to be of similar order to the size of the node, the
 456  * last bset is bounded by btree_write_set_buffer(), which is set to keep the
 457  * memmove on insert from being too expensive: the middle bset should, ideally,
 458  * be the geometric mean of the first and the last.
 459  *
 460  * Returns true if the middle bset is greater than that geometric mean:
 461  */
 462 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
 463 {
 464         unsigned mid_u64s_bits =
 465                 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
 466
 467         return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
 468 }
 469
 470 /*
 471  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 472  * inserted into
 473  *
 474  * Safe to call if there already is an unwritten bset - will only add a new bset
 475  * if @b doesn't already have one.
 476  *
 477  * Returns true if we sorted (i.e. invalidated iterators
 478  */
 479 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 480 {
 481         struct bch_fs *c = trans->c;
 482         struct btree_node_entry *bne;
 483         bool reinit_iter = false;
 484
 485         EBUG_ON(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]);
 486         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 487         BUG_ON(btree_node_just_written(b));
 488
 489         if (b->nsets == MAX_BSETS &&
 490             !btree_node_write_in_flight(b) &&
 491             should_compact_all(c, b)) {
 492                 bch2_btree_node_write(c, b, SIX_LOCK_write,
 493                                       BTREE_WRITE_init_next_bset);
 494                 reinit_iter = true;
 495         }
 496
 497         if (b->nsets == MAX_BSETS &&
 498             btree_node_compact(c, b))
 499                 reinit_iter = true;
 500
 501         BUG_ON(b->nsets >= MAX_BSETS);
 502
 503         bne = want_new_bset(c, b);
 504         if (bne)
 505                 bch2_bset_init_next(c, b, bne);
 506
 507         bch2_btree_build_aux_trees(b);
 508
 509         if (reinit_iter)
 510                 bch2_trans_node_reinit_iter(trans, b);
 511 }
 512
 513 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 514                           struct bch_dev *ca,
 515                           struct btree *b, struct bset *i,
 516                           unsigned offset, int write)
 517 {
 518         prt_printf(out, bch2_log_msg(c, "%s"),
 519                    write == READ
 520                    ? "error validating btree node "
 521                    : "corrupt btree node before write ");
 522         if (ca)
 523                 prt_printf(out, "on %s ", ca->name);
 524         prt_printf(out, "at btree ");
 525         bch2_btree_pos_to_text(out, c, b);
 526
 527         prt_printf(out, "\n  node offset %u", b->written);
 528         if (i)
 529                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 530         prt_str(out, ": ");
 531 }
 532
 533 __printf(8, 9)
 534 static int __btree_err(int ret,
 535                        struct bch_fs *c,
 536                        struct bch_dev *ca,
 537                        struct btree *b,
 538                        struct bset *i,
 539                        int write,
 540                        bool have_retry,
 541                        const char *fmt, ...)
 542 {
 543         struct printbuf out = PRINTBUF;
 544         va_list args;
 545
 546         btree_err_msg(&out, c, ca, b, i, b->written, write);
 547
 548         va_start(args, fmt);
 549         prt_vprintf(&out, fmt, args);
 550         va_end(args);
 551
 552         if (write == WRITE) {
 553                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 554                 ret = c->opts.errors == BCH_ON_ERROR_continue
 555                         ? 0
 556                         : -BCH_ERR_fsck_errors_not_fixed;
 557                 goto out;
 558         }
 559
 560         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
 561                 ret = -BCH_ERR_btree_node_read_err_fixable;
 562         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
 563                 ret = -BCH_ERR_btree_node_read_err_bad_node;
 564
 565         switch (ret) {
 566         case -BCH_ERR_btree_node_read_err_fixable:
 567                 mustfix_fsck_err(c, "%s", out.buf);
 568                 ret = -BCH_ERR_fsck_fix;
 569                 break;
 570         case -BCH_ERR_btree_node_read_err_want_retry:
 571         case -BCH_ERR_btree_node_read_err_must_retry:
 572                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 573                 break;
 574         case -BCH_ERR_btree_node_read_err_bad_node:
 575                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 576                 bch2_topology_error(c);
 577                 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology) ?: -EIO;
 578                 break;
 579         case -BCH_ERR_btree_node_read_err_incompatible:
 580                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 581                 ret = -BCH_ERR_fsck_errors_not_fixed;
 582                 break;
 583         default:
 584                 BUG();
 585         }
 586 out:
 587 fsck_err:
 588         printbuf_exit(&out);
 589         return ret;
 590 }
 591
 592 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 593 ({                                                                      \
 594         int _ret = __btree_err(type, c, ca, b, i, write, have_retry, msg, ##__VA_ARGS__);\
 595                                                                         \
 596         if (_ret != -BCH_ERR_fsck_fix) {                                \
 597                 ret = _ret;                                             \
 598                 goto fsck_err;                                          \
 599         }                                                               \
 600                                                                         \
 601         *saw_error = true;                                              \
 602 })
 603
 604 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 605
 606 /*
 607  * When btree topology repair changes the start or end of a node, that might
 608  * mean we have to drop keys that are no longer inside the node:
 609  */
 610 __cold
 611 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 612 {
 613         struct bset_tree *t;
 614
 615         for_each_bset(b, t) {
 616                 struct bset *i = bset(b, t);
 617                 struct bkey_packed *k;
 618
 619                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 620                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 621                                 break;
 622
 623                 if (k != i->start) {
 624                         unsigned shift = (u64 *) k - (u64 *) i->start;
 625
 626                         memmove_u64s_down(i->start, k,
 627                                           (u64 *) vstruct_end(i) - (u64 *) k);
 628                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 629                         set_btree_bset_end(b, t);
 630                 }
 631
 632                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 633                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 634                                 break;
 635
 636                 if (k != vstruct_last(i)) {
 637                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 638                         set_btree_bset_end(b, t);
 639                 }
 640         }
 641
 642         /*
 643          * Always rebuild search trees: eytzinger search tree nodes directly
 644          * depend on the values of min/max key:
 645          */
 646         bch2_bset_set_no_aux_tree(b, b->set);
 647         bch2_btree_build_aux_trees(b);
 648
 649         struct bkey_s_c k;
 650         struct bkey unpacked;
 651         struct btree_node_iter iter;
 652         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 653                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 654                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 655         }
 656 }
 657
 658 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 659                          struct btree *b, struct bset *i,
 660                          unsigned offset, unsigned sectors,
 661                          int write, bool have_retry, bool *saw_error)
 662 {
 663         unsigned version = le16_to_cpu(i->version);
 664         struct printbuf buf1 = PRINTBUF;
 665         struct printbuf buf2 = PRINTBUF;
 666         int ret = 0;
 667
 668         btree_err_on(!bch2_version_compatible(version),
 669                      -BCH_ERR_btree_node_read_err_incompatible, c, ca, b, i,
 670                      "unsupported bset version %u.%u",
 671                      BCH_VERSION_MAJOR(version),
 672                      BCH_VERSION_MINOR(version));
 673
 674         if (btree_err_on(version < c->sb.version_min,
 675                          -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 676                          "bset version %u older than superblock version_min %u",
 677                          version, c->sb.version_min)) {
 678                 mutex_lock(&c->sb_lock);
 679                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 680                 bch2_write_super(c);
 681                 mutex_unlock(&c->sb_lock);
 682         }
 683
 684         if (btree_err_on(BCH_VERSION_MAJOR(version) >
 685                          BCH_VERSION_MAJOR(c->sb.version),
 686                          -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 687                          "bset version %u newer than superblock version %u",
 688                          version, c->sb.version)) {
 689                 mutex_lock(&c->sb_lock);
 690                 c->disk_sb.sb->version = cpu_to_le16(version);
 691                 bch2_write_super(c);
 692                 mutex_unlock(&c->sb_lock);
 693         }
 694
 695         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 696                      -BCH_ERR_btree_node_read_err_incompatible, c, ca, b, i,
 697                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 698
 699         if (btree_err_on(offset + sectors > btree_sectors(c),
 700                          -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
 701                          "bset past end of btree node")) {
 702                 i->u64s = 0;
 703                 ret = 0;
 704                 goto out;
 705         }
 706
 707         btree_err_on(offset && !i->u64s,
 708                      -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
 709                      "empty bset");
 710
 711         btree_err_on(BSET_OFFSET(i) &&
 712                      BSET_OFFSET(i) != offset,
 713                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 714                      "bset at wrong sector offset");
 715
 716         if (!offset) {
 717                 struct btree_node *bn =
 718                         container_of(i, struct btree_node, keys);
 719                 /* These indicate that we read the wrong btree node: */
 720
 721                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 722                         struct bch_btree_ptr_v2 *bp =
 723                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 724
 725                         /* XXX endianness */
 726                         btree_err_on(bp->seq != bn->keys.seq,
 727                                      -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 728                                      "incorrect sequence number (wrong btree node)");
 729                 }
 730
 731                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 732                              -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
 733                              "incorrect btree id");
 734
 735                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 736                              -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
 737                              "incorrect level");
 738
 739                 if (!write)
 740                         compat_btree_node(b->c.level, b->c.btree_id, version,
 741                                           BSET_BIG_ENDIAN(i), write, bn);
 742
 743                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 744                         struct bch_btree_ptr_v2 *bp =
 745                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 746
 747                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 748                                 b->data->min_key = bp->min_key;
 749                                 b->data->max_key = b->key.k.p;
 750                         }
 751
 752                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 753                                      -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 754                                      "incorrect min_key: got %s should be %s",
 755                                      (printbuf_reset(&buf1),
 756                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 757                                      (printbuf_reset(&buf2),
 758                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 759                 }
 760
 761                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 762                              -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
 763                              "incorrect max key %s",
 764                              (printbuf_reset(&buf1),
 765                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 766
 767                 if (write)
 768                         compat_btree_node(b->c.level, b->c.btree_id, version,
 769                                           BSET_BIG_ENDIAN(i), write, bn);
 770
 771                 btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
 772                              -BCH_ERR_btree_node_read_err_bad_node, c, ca, b, i,
 773                              "invalid bkey format: %s\n  %s", buf1.buf,
 774                              (printbuf_reset(&buf2),
 775                               bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
 776                 printbuf_reset(&buf1);
 777
 778                 compat_bformat(b->c.level, b->c.btree_id, version,
 779                                BSET_BIG_ENDIAN(i), write,
 780                                &bn->format);
 781         }
 782 out:
 783 fsck_err:
 784         printbuf_exit(&buf2);
 785         printbuf_exit(&buf1);
 786         return ret;
 787 }
 788
 789 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 790                             struct bkey_s_c k,
 791                             bool updated_range, int rw,
 792                             struct printbuf *err)
 793 {
 794         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 795                 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
 796                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 797 }
 798
 799 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 800                          struct bset *i, int write,
 801                          bool have_retry, bool *saw_error)
 802 {
 803         unsigned version = le16_to_cpu(i->version);
 804         struct bkey_packed *k, *prev = NULL;
 805         struct printbuf buf = PRINTBUF;
 806         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 807                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 808         int ret = 0;
 809
 810         for (k = i->start;
 811              k != vstruct_last(i);) {
 812                 struct bkey_s u;
 813                 struct bkey tmp;
 814
 815                 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
 816                                  -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 817                                  "key extends past end of bset")) {
 818                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 819                         break;
 820                 }
 821
 822                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 823                                  -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 824                                  "invalid bkey format %u", k->format)) {
 825                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 826                         memmove_u64s_down(k, bkey_p_next(k),
 827                                           (u64 *) vstruct_end(i) - (u64 *) k);
 828                         continue;
 829                 }
 830
 831                 /* XXX: validate k->u64s */
 832                 if (!write)
 833                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 834                                     BSET_BIG_ENDIAN(i), write,
 835                                     &b->format, k);
 836
 837                 u = __bkey_disassemble(b, k, &tmp);
 838
 839                 printbuf_reset(&buf);
 840                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 841                         printbuf_reset(&buf);
 842                         prt_printf(&buf, "invalid bkey:  ");
 843                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 844                         prt_printf(&buf, "\n  ");
 845                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 846
 847                         btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
 848
 849                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 850                         memmove_u64s_down(k, bkey_p_next(k),
 851                                           (u64 *) vstruct_end(i) - (u64 *) k);
 852                         continue;
 853                 }
 854
 855                 if (write)
 856                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 857                                     BSET_BIG_ENDIAN(i), write,
 858                                     &b->format, k);
 859
 860                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 861                         struct bkey up = bkey_unpack_key(b, prev);
 862
 863                         printbuf_reset(&buf);
 864                         prt_printf(&buf, "keys out of order: ");
 865                         bch2_bkey_to_text(&buf, &up);
 866                         prt_printf(&buf, " > ");
 867                         bch2_bkey_to_text(&buf, u.k);
 868
 869                         bch2_dump_bset(c, b, i, 0);
 870
 871                         if (btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf)) {
 872                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 873                                 memmove_u64s_down(k, bkey_p_next(k),
 874                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 875                                 continue;
 876                         }
 877                 }
 878
 879                 prev = k;
 880                 k = bkey_p_next(k);
 881         }
 882 fsck_err:
 883         printbuf_exit(&buf);
 884         return ret;
 885 }
 886
 887 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 888                               struct btree *b, bool have_retry, bool *saw_error)
 889 {
 890         struct btree_node_entry *bne;
 891         struct sort_iter *iter;
 892         struct btree_node *sorted;
 893         struct bkey_packed *k;
 894         struct bch_extent_ptr *ptr;
 895         struct bset *i;
 896         bool used_mempool, blacklisted;
 897         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 898                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 899         unsigned u64s;
 900         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 901         struct printbuf buf = PRINTBUF;
 902         int ret = 0, retry_read = 0, write = READ;
 903
 904         b->version_ondisk = U16_MAX;
 905         /* We might get called multiple times on read retry: */
 906         b->written = 0;
 907
 908         iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
 909         sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2);
 910
 911         if (bch2_meta_read_fault("btree"))
 912                 btree_err(-BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 913                           "dynamic fault");
 914
 915         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 916                      -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 917                      "bad magic: want %llx, got %llx",
 918                      bset_magic(c), le64_to_cpu(b->data->magic));
 919
 920         btree_err_on(!b->data->keys.seq,
 921                      -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 922                      "bad btree header: seq 0");
 923
 924         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 925                 struct bch_btree_ptr_v2 *bp =
 926                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 927
 928                 btree_err_on(b->data->keys.seq != bp->seq,
 929                              -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 930                              "got wrong btree node (seq %llx want %llx)",
 931                              b->data->keys.seq, bp->seq);
 932         }
 933
 934         while (b->written < (ptr_written ?: btree_sectors(c))) {
 935                 unsigned sectors;
 936                 struct nonce nonce;
 937                 struct bch_csum csum;
 938                 bool first = !b->written;
 939
 940                 if (!b->written) {
 941                         i = &b->data->keys;
 942
 943                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 944                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 945                                      "unknown checksum type %llu",
 946                                      BSET_CSUM_TYPE(i));
 947
 948                         nonce = btree_nonce(i, b->written << 9);
 949                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 950
 951                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 952                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 953                                      "invalid checksum");
 954
 955                         ret = bset_encrypt(c, i, b->written << 9);
 956                         if (bch2_fs_fatal_err_on(ret, c,
 957                                         "error decrypting btree node: %i", ret))
 958                                 goto fsck_err;
 959
 960                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 961                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 962                                      -BCH_ERR_btree_node_read_err_incompatible, c, NULL, b, NULL,
 963                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 964
 965                         sectors = vstruct_sectors(b->data, c->block_bits);
 966                 } else {
 967                         bne = write_block(b);
 968                         i = &bne->keys;
 969
 970                         if (i->seq != b->data->keys.seq)
 971                                 break;
 972
 973                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 974                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 975                                      "unknown checksum type %llu",
 976                                      BSET_CSUM_TYPE(i));
 977
 978                         nonce = btree_nonce(i, b->written << 9);
 979                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 980
 981                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 982                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 983                                      "invalid checksum");
 984
 985                         ret = bset_encrypt(c, i, b->written << 9);
 986                         if (bch2_fs_fatal_err_on(ret, c,
 987                                         "error decrypting btree node: %i\n", ret))
 988                                 goto fsck_err;
 989
 990                         sectors = vstruct_sectors(bne, c->block_bits);
 991                 }
 992
 993                 b->version_ondisk = min(b->version_ondisk,
 994                                         le16_to_cpu(i->version));
 995
 996                 ret = validate_bset(c, ca, b, i, b->written, sectors,
 997                                     READ, have_retry, saw_error);
 998                 if (ret)
 999                         goto fsck_err;
1000
1001                 if (!b->written)
1002                         btree_node_set_format(b, b->data->format);
1003
1004                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1005                 if (ret)
1006                         goto fsck_err;
1007
1008                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1009
1010                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1011                                         le64_to_cpu(i->journal_seq),
1012                                         true);
1013
1014                 btree_err_on(blacklisted && first,
1015                              -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
1016                              "first btree node bset has blacklisted journal seq (%llu)",
1017                              le64_to_cpu(i->journal_seq));
1018
1019                 btree_err_on(blacklisted && ptr_written,
1020                              -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
1021                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1022                              le64_to_cpu(i->journal_seq),
1023                              b->written, b->written + sectors, ptr_written);
1024
1025                 b->written += sectors;
1026
1027                 if (blacklisted && !first)
1028                         continue;
1029
1030                 sort_iter_add(iter,
1031                               vstruct_idx(i, 0),
1032                               vstruct_last(i));
1033         }
1034
1035         if (ptr_written) {
1036                 btree_err_on(b->written < ptr_written,
1037                              -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
1038                              "btree node data missing: expected %u sectors, found %u",
1039                              ptr_written, b->written);
1040         } else {
1041                 for (bne = write_block(b);
1042                      bset_byte_offset(b, bne) < btree_bytes(c);
1043                      bne = (void *) bne + block_bytes(c))
1044                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1045                                      !bch2_journal_seq_is_blacklisted(c,
1046                                                                       le64_to_cpu(bne->keys.journal_seq),
1047                                                                       true),
1048                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
1049                                      "found bset signature after last bset");
1050         }
1051
1052         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1053         sorted->keys.u64s = 0;
1054
1055         set_btree_bset(b, b->set, &b->data->keys);
1056
1057         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1058
1059         u64s = le16_to_cpu(sorted->keys.u64s);
1060         *sorted = *b->data;
1061         sorted->keys.u64s = cpu_to_le16(u64s);
1062         swap(sorted, b->data);
1063         set_btree_bset(b, b->set, &b->data->keys);
1064         b->nsets = 1;
1065
1066         BUG_ON(b->nr.live_u64s != u64s);
1067
1068         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1069
1070         if (updated_range)
1071                 bch2_btree_node_drop_keys_outside_node(b);
1072
1073         i = &b->data->keys;
1074         for (k = i->start; k != vstruct_last(i);) {
1075                 struct bkey tmp;
1076                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1077
1078                 printbuf_reset(&buf);
1079
1080                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1081                     (bch2_inject_invalid_keys &&
1082                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1083                         printbuf_reset(&buf);
1084
1085                         prt_printf(&buf, "invalid bkey: ");
1086                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1087                         prt_printf(&buf, "\n  ");
1088                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1089
1090                         btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
1091
1092                         btree_keys_account_key_drop(&b->nr, 0, k);
1093
1094                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1095                         memmove_u64s_down(k, bkey_p_next(k),
1096                                           (u64 *) vstruct_end(i) - (u64 *) k);
1097                         set_btree_bset_end(b, b->set);
1098                         continue;
1099                 }
1100
1101                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1102                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1103
1104                         bp.v->mem_ptr = 0;
1105                 }
1106
1107                 k = bkey_p_next(k);
1108         }
1109
1110         bch2_bset_build_aux_tree(b, b->set, false);
1111
1112         set_needs_whiteout(btree_bset_first(b), true);
1113
1114         btree_node_reset_sib_u64s(b);
1115
1116         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1117                 struct bch_dev *ca2 = bch_dev_bkey_exists(c, ptr->dev);
1118
1119                 if (ca2->mi.state != BCH_MEMBER_STATE_rw)
1120                         set_btree_node_need_rewrite(b);
1121         }
1122
1123         if (!ptr_written)
1124                 set_btree_node_need_rewrite(b);
1125 out:
1126         mempool_free(iter, &c->fill_iter);
1127         printbuf_exit(&buf);
1128         return retry_read;
1129 fsck_err:
1130         if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1131             ret == -BCH_ERR_btree_node_read_err_must_retry)
1132                 retry_read = 1;
1133         else
1134                 set_btree_node_read_error(b);
1135         goto out;
1136 }
1137
1138 static void btree_node_read_work(struct work_struct *work)
1139 {
1140         struct btree_read_bio *rb =
1141                 container_of(work, struct btree_read_bio, work);
1142         struct bch_fs *c        = rb->c;
1143         struct btree *b         = rb->b;
1144         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1145         struct bio *bio         = &rb->bio;
1146         struct bch_io_failures failed = { .nr = 0 };
1147         struct printbuf buf = PRINTBUF;
1148         bool saw_error = false;
1149         bool retry = false;
1150         bool can_retry;
1151
1152         goto start;
1153         while (1) {
1154                 retry = true;
1155                 bch_info(c, "retrying read");
1156                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1157                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1158                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1159                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1160                 bio->bi_iter.bi_size    = btree_bytes(c);
1161
1162                 if (rb->have_ioref) {
1163                         bio_set_dev(bio, ca->disk_sb.bdev);
1164                         submit_bio_wait(bio);
1165                 } else {
1166                         bio->bi_status = BLK_STS_REMOVED;
1167                 }
1168 start:
1169                 printbuf_reset(&buf);
1170                 bch2_btree_pos_to_text(&buf, c, b);
1171                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1172                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1173                 if (rb->have_ioref)
1174                         percpu_ref_put(&ca->io_ref);
1175                 rb->have_ioref = false;
1176
1177                 bch2_mark_io_failure(&failed, &rb->pick);
1178
1179                 can_retry = bch2_bkey_pick_read_device(c,
1180                                 bkey_i_to_s_c(&b->key),
1181                                 &failed, &rb->pick) > 0;
1182
1183                 if (!bio->bi_status &&
1184                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1185                         if (retry)
1186                                 bch_info(c, "retry success");
1187                         break;
1188                 }
1189
1190                 saw_error = true;
1191
1192                 if (!can_retry) {
1193                         set_btree_node_read_error(b);
1194                         break;
1195                 }
1196         }
1197
1198         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1199                                rb->start_time);
1200         bio_put(&rb->bio);
1201
1202         if (saw_error && !btree_node_read_error(b)) {
1203                 printbuf_reset(&buf);
1204                 bch2_bpos_to_text(&buf, b->key.k.p);
1205                 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1206                          __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf);
1207
1208                 bch2_btree_node_rewrite_async(c, b);
1209         }
1210
1211         printbuf_exit(&buf);
1212         clear_btree_node_read_in_flight(b);
1213         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1214 }
1215
1216 static void btree_node_read_endio(struct bio *bio)
1217 {
1218         struct btree_read_bio *rb =
1219                 container_of(bio, struct btree_read_bio, bio);
1220         struct bch_fs *c        = rb->c;
1221
1222         if (rb->have_ioref) {
1223                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1224
1225                 bch2_latency_acct(ca, rb->start_time, READ);
1226         }
1227
1228         queue_work(c->io_complete_wq, &rb->work);
1229 }
1230
1231 struct btree_node_read_all {
1232         struct closure          cl;
1233         struct bch_fs           *c;
1234         struct btree            *b;
1235         unsigned                nr;
1236         void                    *buf[BCH_REPLICAS_MAX];
1237         struct bio              *bio[BCH_REPLICAS_MAX];
1238         blk_status_t            err[BCH_REPLICAS_MAX];
1239 };
1240
1241 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1242 {
1243         struct btree_node *bn = data;
1244         struct btree_node_entry *bne;
1245         unsigned offset = 0;
1246
1247         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1248                 return 0;
1249
1250         while (offset < btree_sectors(c)) {
1251                 if (!offset) {
1252                         offset += vstruct_sectors(bn, c->block_bits);
1253                 } else {
1254                         bne = data + (offset << 9);
1255                         if (bne->keys.seq != bn->keys.seq)
1256                                 break;
1257                         offset += vstruct_sectors(bne, c->block_bits);
1258                 }
1259         }
1260
1261         return offset;
1262 }
1263
1264 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1265 {
1266         struct btree_node *bn = data;
1267         struct btree_node_entry *bne;
1268
1269         if (!offset)
1270                 return false;
1271
1272         while (offset < btree_sectors(c)) {
1273                 bne = data + (offset << 9);
1274                 if (bne->keys.seq == bn->keys.seq)
1275                         return true;
1276                 offset++;
1277         }
1278
1279         return false;
1280         return offset;
1281 }
1282
1283 static void btree_node_read_all_replicas_done(struct closure *cl)
1284 {
1285         struct btree_node_read_all *ra =
1286                 container_of(cl, struct btree_node_read_all, cl);
1287         struct bch_fs *c = ra->c;
1288         struct btree *b = ra->b;
1289         struct printbuf buf = PRINTBUF;
1290         bool dump_bset_maps = false;
1291         bool have_retry = false;
1292         int ret = 0, best = -1, write = READ;
1293         unsigned i, written = 0, written2 = 0;
1294         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1295                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1296         bool _saw_error = false, *saw_error = &_saw_error;
1297
1298         for (i = 0; i < ra->nr; i++) {
1299                 struct btree_node *bn = ra->buf[i];
1300
1301                 if (ra->err[i])
1302                         continue;
1303
1304                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1305                     (seq && seq != bn->keys.seq))
1306                         continue;
1307
1308                 if (best < 0) {
1309                         best = i;
1310                         written = btree_node_sectors_written(c, bn);
1311                         continue;
1312                 }
1313
1314                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1315                 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1316                                  "btree node sectors written mismatch: %u != %u",
1317                                  written, written2) ||
1318                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1319                                  -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1320                                  "found bset signature after last bset") ||
1321                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1322                                  -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1323                                  "btree node replicas content mismatch"))
1324                         dump_bset_maps = true;
1325
1326                 if (written2 > written) {
1327                         written = written2;
1328                         best = i;
1329                 }
1330         }
1331 fsck_err:
1332         if (dump_bset_maps) {
1333                 for (i = 0; i < ra->nr; i++) {
1334                         struct btree_node *bn = ra->buf[i];
1335                         struct btree_node_entry *bne = NULL;
1336                         unsigned offset = 0, sectors;
1337                         bool gap = false;
1338
1339                         if (ra->err[i])
1340                                 continue;
1341
1342                         printbuf_reset(&buf);
1343
1344                         while (offset < btree_sectors(c)) {
1345                                 if (!offset) {
1346                                         sectors = vstruct_sectors(bn, c->block_bits);
1347                                 } else {
1348                                         bne = ra->buf[i] + (offset << 9);
1349                                         if (bne->keys.seq != bn->keys.seq)
1350                                                 break;
1351                                         sectors = vstruct_sectors(bne, c->block_bits);
1352                                 }
1353
1354                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1355                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1356                                                         le64_to_cpu(bne->keys.journal_seq), false))
1357                                         prt_printf(&buf, "*");
1358                                 offset += sectors;
1359                         }
1360
1361                         while (offset < btree_sectors(c)) {
1362                                 bne = ra->buf[i] + (offset << 9);
1363                                 if (bne->keys.seq == bn->keys.seq) {
1364                                         if (!gap)
1365                                                 prt_printf(&buf, " GAP");
1366                                         gap = true;
1367
1368                                         sectors = vstruct_sectors(bne, c->block_bits);
1369                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1370                                         if (bch2_journal_seq_is_blacklisted(c,
1371                                                         le64_to_cpu(bne->keys.journal_seq), false))
1372                                                 prt_printf(&buf, "*");
1373                                 }
1374                                 offset++;
1375                         }
1376
1377                         bch_err(c, "replica %u:%s", i, buf.buf);
1378                 }
1379         }
1380
1381         if (best >= 0) {
1382                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1383                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1384         } else {
1385                 ret = -1;
1386         }
1387
1388         if (ret)
1389                 set_btree_node_read_error(b);
1390         else if (*saw_error)
1391                 bch2_btree_node_rewrite_async(c, b);
1392
1393         for (i = 0; i < ra->nr; i++) {
1394                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1395                 bio_put(ra->bio[i]);
1396         }
1397
1398         closure_debug_destroy(&ra->cl);
1399         kfree(ra);
1400         printbuf_exit(&buf);
1401
1402         clear_btree_node_read_in_flight(b);
1403         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1404 }
1405
1406 static void btree_node_read_all_replicas_endio(struct bio *bio)
1407 {
1408         struct btree_read_bio *rb =
1409                 container_of(bio, struct btree_read_bio, bio);
1410         struct bch_fs *c        = rb->c;
1411         struct btree_node_read_all *ra = rb->ra;
1412
1413         if (rb->have_ioref) {
1414                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1415
1416                 bch2_latency_acct(ca, rb->start_time, READ);
1417         }
1418
1419         ra->err[rb->idx] = bio->bi_status;
1420         closure_put(&ra->cl);
1421 }
1422
1423 /*
1424  * XXX This allocates multiple times from the same mempools, and can deadlock
1425  * under sufficient memory pressure (but is only a debug path)
1426  */
1427 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1428 {
1429         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1430         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1431         const union bch_extent_entry *entry;
1432         struct extent_ptr_decoded pick;
1433         struct btree_node_read_all *ra;
1434         unsigned i;
1435
1436         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1437         if (!ra)
1438                 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1439
1440         closure_init(&ra->cl, NULL);
1441         ra->c   = c;
1442         ra->b   = b;
1443         ra->nr  = bch2_bkey_nr_ptrs(k);
1444
1445         for (i = 0; i < ra->nr; i++) {
1446                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1447                 ra->bio[i] = bio_alloc_bioset(NULL,
1448                                               buf_pages(ra->buf[i], btree_bytes(c)),
1449                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1450                                               GFP_NOFS,
1451                                               &c->btree_bio);
1452         }
1453
1454         i = 0;
1455         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1456                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1457                 struct btree_read_bio *rb =
1458                         container_of(ra->bio[i], struct btree_read_bio, bio);
1459                 rb->c                   = c;
1460                 rb->b                   = b;
1461                 rb->ra                  = ra;
1462                 rb->start_time          = local_clock();
1463                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1464                 rb->idx                 = i;
1465                 rb->pick                = pick;
1466                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1467                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1468                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1469
1470                 if (rb->have_ioref) {
1471                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1472                                      bio_sectors(&rb->bio));
1473                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1474
1475                         closure_get(&ra->cl);
1476                         submit_bio(&rb->bio);
1477                 } else {
1478                         ra->err[i] = BLK_STS_REMOVED;
1479                 }
1480
1481                 i++;
1482         }
1483
1484         if (sync) {
1485                 closure_sync(&ra->cl);
1486                 btree_node_read_all_replicas_done(&ra->cl);
1487         } else {
1488                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1489                             c->io_complete_wq);
1490         }
1491
1492         return 0;
1493 }
1494
1495 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1496                           bool sync)
1497 {
1498         struct extent_ptr_decoded pick;
1499         struct btree_read_bio *rb;
1500         struct bch_dev *ca;
1501         struct bio *bio;
1502         int ret;
1503
1504         trace_and_count(c, btree_node_read, c, b);
1505
1506         if (bch2_verify_all_btree_replicas &&
1507             !btree_node_read_all_replicas(c, b, sync))
1508                 return;
1509
1510         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1511                                          NULL, &pick);
1512
1513         if (ret <= 0) {
1514                 struct printbuf buf = PRINTBUF;
1515
1516                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1517                 bch2_btree_pos_to_text(&buf, c, b);
1518                 bch_err(c, "%s", buf.buf);
1519
1520                 if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1521                     c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1522                         bch2_fatal_error(c);
1523
1524                 set_btree_node_read_error(b);
1525                 clear_btree_node_read_in_flight(b);
1526                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1527                 printbuf_exit(&buf);
1528                 return;
1529         }
1530
1531         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1532
1533         bio = bio_alloc_bioset(NULL,
1534                                buf_pages(b->data, btree_bytes(c)),
1535                                REQ_OP_READ|REQ_SYNC|REQ_META,
1536                                GFP_NOFS,
1537                                &c->btree_bio);
1538         rb = container_of(bio, struct btree_read_bio, bio);
1539         rb->c                   = c;
1540         rb->b                   = b;
1541         rb->ra                  = NULL;
1542         rb->start_time          = local_clock();
1543         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1544         rb->pick                = pick;
1545         INIT_WORK(&rb->work, btree_node_read_work);
1546         bio->bi_iter.bi_sector  = pick.ptr.offset;
1547         bio->bi_end_io          = btree_node_read_endio;
1548         bch2_bio_map(bio, b->data, btree_bytes(c));
1549
1550         if (rb->have_ioref) {
1551                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1552                              bio_sectors(bio));
1553                 bio_set_dev(bio, ca->disk_sb.bdev);
1554
1555                 if (sync) {
1556                         submit_bio_wait(bio);
1557
1558                         btree_node_read_work(&rb->work);
1559                 } else {
1560                         submit_bio(bio);
1561                 }
1562         } else {
1563                 bio->bi_status = BLK_STS_REMOVED;
1564
1565                 if (sync)
1566                         btree_node_read_work(&rb->work);
1567                 else
1568                         queue_work(c->io_complete_wq, &rb->work);
1569         }
1570 }
1571
1572 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1573                                   const struct bkey_i *k, unsigned level)
1574 {
1575         struct bch_fs *c = trans->c;
1576         struct closure cl;
1577         struct btree *b;
1578         int ret;
1579
1580         closure_init_stack(&cl);
1581
1582         do {
1583                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1584                 closure_sync(&cl);
1585         } while (ret);
1586
1587         b = bch2_btree_node_mem_alloc(trans, level != 0);
1588         bch2_btree_cache_cannibalize_unlock(c);
1589
1590         BUG_ON(IS_ERR(b));
1591
1592         bkey_copy(&b->key, k);
1593         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1594
1595         set_btree_node_read_in_flight(b);
1596
1597         bch2_btree_node_read(c, b, true);
1598
1599         if (btree_node_read_error(b)) {
1600                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1601
1602                 mutex_lock(&c->btree_cache.lock);
1603                 list_move(&b->list, &c->btree_cache.freeable);
1604                 mutex_unlock(&c->btree_cache.lock);
1605
1606                 ret = -EIO;
1607                 goto err;
1608         }
1609
1610         bch2_btree_set_root_for_read(c, b);
1611 err:
1612         six_unlock_write(&b->c.lock);
1613         six_unlock_intent(&b->c.lock);
1614
1615         return ret;
1616 }
1617
1618 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1619                         const struct bkey_i *k, unsigned level)
1620 {
1621         return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
1622 }
1623
1624 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1625                               struct btree_write *w)
1626 {
1627         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1628
1629         do {
1630                 old = new = v;
1631                 if (!(old & 1))
1632                         break;
1633
1634                 new &= ~1UL;
1635         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1636
1637         if (old & 1)
1638                 closure_put(&((struct btree_update *) new)->cl);
1639
1640         bch2_journal_pin_drop(&c->journal, &w->journal);
1641 }
1642
1643 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1644 {
1645         struct btree_write *w = btree_prev_write(b);
1646         unsigned long old, new, v;
1647         unsigned type = 0;
1648
1649         bch2_btree_complete_write(c, b, w);
1650
1651         v = READ_ONCE(b->flags);
1652         do {
1653                 old = new = v;
1654
1655                 if ((old & (1U << BTREE_NODE_dirty)) &&
1656                     (old & (1U << BTREE_NODE_need_write)) &&
1657                     !(old & (1U << BTREE_NODE_never_write)) &&
1658                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1659                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1660                         new &= ~(1U << BTREE_NODE_dirty);
1661                         new &= ~(1U << BTREE_NODE_need_write);
1662                         new |=  (1U << BTREE_NODE_write_in_flight);
1663                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1664                         new |=  (1U << BTREE_NODE_just_written);
1665                         new ^=  (1U << BTREE_NODE_write_idx);
1666
1667                         type = new & BTREE_WRITE_TYPE_MASK;
1668                         new &= ~BTREE_WRITE_TYPE_MASK;
1669                 } else {
1670                         new &= ~(1U << BTREE_NODE_write_in_flight);
1671                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1672                 }
1673         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1674
1675         if (new & (1U << BTREE_NODE_write_in_flight))
1676                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1677         else
1678                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1679 }
1680
1681 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1682 {
1683         struct btree_trans *trans = bch2_trans_get(c);
1684
1685         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
1686         __btree_node_write_done(c, b);
1687         six_unlock_read(&b->c.lock);
1688
1689         bch2_trans_put(trans);
1690 }
1691
1692 static void btree_node_write_work(struct work_struct *work)
1693 {
1694         struct btree_write_bio *wbio =
1695                 container_of(work, struct btree_write_bio, work);
1696         struct bch_fs *c        = wbio->wbio.c;
1697         struct btree *b         = wbio->wbio.bio.bi_private;
1698         struct bch_extent_ptr *ptr;
1699         int ret = 0;
1700
1701         btree_bounce_free(c,
1702                 wbio->data_bytes,
1703                 wbio->wbio.used_mempool,
1704                 wbio->data);
1705
1706         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1707                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1708
1709         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1710                 goto err;
1711
1712         if (wbio->wbio.first_btree_write) {
1713                 if (wbio->wbio.failed.nr) {
1714
1715                 }
1716         } else {
1717                 ret = bch2_trans_do(c, NULL, NULL, 0,
1718                         bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
1719                                         BCH_WATERMARK_reclaim|
1720                                         BTREE_INSERT_JOURNAL_RECLAIM|
1721                                         BTREE_INSERT_NOFAIL|
1722                                         BTREE_INSERT_NOCHECK_RW,
1723                                         !wbio->wbio.failed.nr));
1724                 if (ret)
1725                         goto err;
1726         }
1727 out:
1728         bio_put(&wbio->wbio.bio);
1729         btree_node_write_done(c, b);
1730         return;
1731 err:
1732         set_btree_node_noevict(b);
1733         if (!bch2_err_matches(ret, EROFS))
1734                 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1735         goto out;
1736 }
1737
1738 static void btree_node_write_endio(struct bio *bio)
1739 {
1740         struct bch_write_bio *wbio      = to_wbio(bio);
1741         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1742         struct bch_write_bio *orig      = parent ?: wbio;
1743         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1744         struct bch_fs *c                = wbio->c;
1745         struct btree *b                 = wbio->bio.bi_private;
1746         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1747         unsigned long flags;
1748
1749         if (wbio->have_ioref)
1750                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1751
1752         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1753                                bch2_blk_status_to_str(bio->bi_status)) ||
1754             bch2_meta_write_fault("btree")) {
1755                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1756                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1757                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1758         }
1759
1760         if (wbio->have_ioref)
1761                 percpu_ref_put(&ca->io_ref);
1762
1763         if (parent) {
1764                 bio_put(bio);
1765                 bio_endio(&parent->bio);
1766                 return;
1767         }
1768
1769         clear_btree_node_write_in_flight_inner(b);
1770         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1771         INIT_WORK(&wb->work, btree_node_write_work);
1772         queue_work(c->btree_io_complete_wq, &wb->work);
1773 }
1774
1775 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1776                                    struct bset *i, unsigned sectors)
1777 {
1778         struct printbuf buf = PRINTBUF;
1779         bool saw_error;
1780         int ret;
1781
1782         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1783                                 BKEY_TYPE_btree, WRITE, &buf);
1784
1785         if (ret)
1786                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1787         printbuf_exit(&buf);
1788         if (ret)
1789                 return ret;
1790
1791         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1792                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1793         if (ret) {
1794                 bch2_inconsistent_error(c);
1795                 dump_stack();
1796         }
1797
1798         return ret;
1799 }
1800
1801 static void btree_write_submit(struct work_struct *work)
1802 {
1803         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1804         struct bch_extent_ptr *ptr;
1805         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1806
1807         bkey_copy(&tmp.k, &wbio->key);
1808
1809         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1810                 ptr->offset += wbio->sector_offset;
1811
1812         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1813                                   &tmp.k, false);
1814 }
1815
1816 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1817 {
1818         struct btree_write_bio *wbio;
1819         struct bset_tree *t;
1820         struct bset *i;
1821         struct btree_node *bn = NULL;
1822         struct btree_node_entry *bne = NULL;
1823         struct sort_iter_stack sort_iter;
1824         struct nonce nonce;
1825         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1826         u64 seq = 0;
1827         bool used_mempool;
1828         unsigned long old, new;
1829         bool validate_before_checksum = false;
1830         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1831         void *data;
1832         int ret;
1833
1834         if (flags & BTREE_WRITE_ALREADY_STARTED)
1835                 goto do_write;
1836
1837         /*
1838          * We may only have a read lock on the btree node - the dirty bit is our
1839          * "lock" against racing with other threads that may be trying to start
1840          * a write, we do a write iff we clear the dirty bit. Since setting the
1841          * dirty bit requires a write lock, we can't race with other threads
1842          * redirtying it:
1843          */
1844         do {
1845                 old = new = READ_ONCE(b->flags);
1846
1847                 if (!(old & (1 << BTREE_NODE_dirty)))
1848                         return;
1849
1850                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1851                     !(old & (1 << BTREE_NODE_need_write)))
1852                         return;
1853
1854                 if (old &
1855                     ((1 << BTREE_NODE_never_write)|
1856                      (1 << BTREE_NODE_write_blocked)))
1857                         return;
1858
1859                 if (b->written &&
1860                     (old & (1 << BTREE_NODE_will_make_reachable)))
1861                         return;
1862
1863                 if (old & (1 << BTREE_NODE_write_in_flight))
1864                         return;
1865
1866                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1867                         type = new & BTREE_WRITE_TYPE_MASK;
1868                 new &= ~BTREE_WRITE_TYPE_MASK;
1869
1870                 new &= ~(1 << BTREE_NODE_dirty);
1871                 new &= ~(1 << BTREE_NODE_need_write);
1872                 new |=  (1 << BTREE_NODE_write_in_flight);
1873                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1874                 new |=  (1 << BTREE_NODE_just_written);
1875                 new ^=  (1 << BTREE_NODE_write_idx);
1876         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1877
1878         if (new & (1U << BTREE_NODE_need_write))
1879                 return;
1880 do_write:
1881         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1882
1883         atomic_dec(&c->btree_cache.dirty);
1884
1885         BUG_ON(btree_node_fake(b));
1886         BUG_ON((b->will_make_reachable != 0) != !b->written);
1887
1888         BUG_ON(b->written >= btree_sectors(c));
1889         BUG_ON(b->written & (block_sectors(c) - 1));
1890         BUG_ON(bset_written(b, btree_bset_last(b)));
1891         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1892         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1893
1894         bch2_sort_whiteouts(c, b);
1895
1896         sort_iter_stack_init(&sort_iter, b);
1897
1898         bytes = !b->written
1899                 ? sizeof(struct btree_node)
1900                 : sizeof(struct btree_node_entry);
1901
1902         bytes += b->whiteout_u64s * sizeof(u64);
1903
1904         for_each_bset(b, t) {
1905                 i = bset(b, t);
1906
1907                 if (bset_written(b, i))
1908                         continue;
1909
1910                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1911                 sort_iter_add(&sort_iter.iter,
1912                               btree_bkey_first(b, t),
1913                               btree_bkey_last(b, t));
1914                 seq = max(seq, le64_to_cpu(i->journal_seq));
1915         }
1916
1917         BUG_ON(b->written && !seq);
1918
1919         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1920         bytes += 8;
1921
1922         /* buffer must be a multiple of the block size */
1923         bytes = round_up(bytes, block_bytes(c));
1924
1925         data = btree_bounce_alloc(c, bytes, &used_mempool);
1926
1927         if (!b->written) {
1928                 bn = data;
1929                 *bn = *b->data;
1930                 i = &bn->keys;
1931         } else {
1932                 bne = data;
1933                 bne->keys = b->data->keys;
1934                 i = &bne->keys;
1935         }
1936
1937         i->journal_seq  = cpu_to_le64(seq);
1938         i->u64s         = 0;
1939
1940         sort_iter_add(&sort_iter.iter,
1941                       unwritten_whiteouts_start(c, b),
1942                       unwritten_whiteouts_end(c, b));
1943         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1944
1945         b->whiteout_u64s = 0;
1946
1947         u64s = bch2_sort_keys(i->start, &sort_iter.iter, false);
1948         le16_add_cpu(&i->u64s, u64s);
1949
1950         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
1951
1952         set_needs_whiteout(i, false);
1953
1954         /* do we have data to write? */
1955         if (b->written && !i->u64s)
1956                 goto nowrite;
1957
1958         bytes_to_write = vstruct_end(i) - data;
1959         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1960
1961         if (!b->written &&
1962             b->key.k.type == KEY_TYPE_btree_ptr_v2)
1963                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
1964
1965         memset(data + bytes_to_write, 0,
1966                (sectors_to_write << 9) - bytes_to_write);
1967
1968         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1969         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1970         BUG_ON(i->seq != b->data->keys.seq);
1971
1972         i->version = cpu_to_le16(c->sb.version);
1973         SET_BSET_OFFSET(i, b->written);
1974         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1975
1976         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1977                 validate_before_checksum = true;
1978
1979         /* validate_bset will be modifying: */
1980         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1981                 validate_before_checksum = true;
1982
1983         /* if we're going to be encrypting, check metadata validity first: */
1984         if (validate_before_checksum &&
1985             validate_bset_for_write(c, b, i, sectors_to_write))
1986                 goto err;
1987
1988         ret = bset_encrypt(c, i, b->written << 9);
1989         if (bch2_fs_fatal_err_on(ret, c,
1990                         "error encrypting btree node: %i\n", ret))
1991                 goto err;
1992
1993         nonce = btree_nonce(i, b->written << 9);
1994
1995         if (bn)
1996                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1997         else
1998                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1999
2000         /* if we're not encrypting, check metadata after checksumming: */
2001         if (!validate_before_checksum &&
2002             validate_bset_for_write(c, b, i, sectors_to_write))
2003                 goto err;
2004
2005         /*
2006          * We handle btree write errors by immediately halting the journal -
2007          * after we've done that, we can't issue any subsequent btree writes
2008          * because they might have pointers to new nodes that failed to write.
2009          *
2010          * Furthermore, there's no point in doing any more btree writes because
2011          * with the journal stopped, we're never going to update the journal to
2012          * reflect that those writes were done and the data flushed from the
2013          * journal:
2014          *
2015          * Also on journal error, the pending write may have updates that were
2016          * never journalled (interior nodes, see btree_update_nodes_written()) -
2017          * it's critical that we don't do the write in that case otherwise we
2018          * will have updates visible that weren't in the journal:
2019          *
2020          * Make sure to update b->written so bch2_btree_init_next() doesn't
2021          * break:
2022          */
2023         if (bch2_journal_error(&c->journal) ||
2024             c->opts.nochanges)
2025                 goto err;
2026
2027         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2028
2029         wbio = container_of(bio_alloc_bioset(NULL,
2030                                 buf_pages(data, sectors_to_write << 9),
2031                                 REQ_OP_WRITE|REQ_META,
2032                                 GFP_NOFS,
2033                                 &c->btree_bio),
2034                             struct btree_write_bio, wbio.bio);
2035         wbio_init(&wbio->wbio.bio);
2036         wbio->data                      = data;
2037         wbio->data_bytes                = bytes;
2038         wbio->sector_offset             = b->written;
2039         wbio->wbio.c                    = c;
2040         wbio->wbio.used_mempool         = used_mempool;
2041         wbio->wbio.first_btree_write    = !b->written;
2042         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2043         wbio->wbio.bio.bi_private       = b;
2044
2045         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2046
2047         bkey_copy(&wbio->key, &b->key);
2048
2049         b->written += sectors_to_write;
2050
2051         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2052                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2053                         cpu_to_le16(b->written);
2054
2055         atomic64_inc(&c->btree_write_stats[type].nr);
2056         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2057
2058         INIT_WORK(&wbio->work, btree_write_submit);
2059         queue_work(c->io_complete_wq, &wbio->work);
2060         return;
2061 err:
2062         set_btree_node_noevict(b);
2063         b->written += sectors_to_write;
2064 nowrite:
2065         btree_bounce_free(c, bytes, used_mempool, data);
2066         __btree_node_write_done(c, b);
2067 }
2068
2069 /*
2070  * Work that must be done with write lock held:
2071  */
2072 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2073 {
2074         bool invalidated_iter = false;
2075         struct btree_node_entry *bne;
2076         struct bset_tree *t;
2077
2078         if (!btree_node_just_written(b))
2079                 return false;
2080
2081         BUG_ON(b->whiteout_u64s);
2082
2083         clear_btree_node_just_written(b);
2084
2085         /*
2086          * Note: immediately after write, bset_written() doesn't work - the
2087          * amount of data we had to write after compaction might have been
2088          * smaller than the offset of the last bset.
2089          *
2090          * However, we know that all bsets have been written here, as long as
2091          * we're still holding the write lock:
2092          */
2093
2094         /*
2095          * XXX: decide if we really want to unconditionally sort down to a
2096          * single bset:
2097          */
2098         if (b->nsets > 1) {
2099                 btree_node_sort(c, b, 0, b->nsets, true);
2100                 invalidated_iter = true;
2101         } else {
2102                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2103         }
2104
2105         for_each_bset(b, t)
2106                 set_needs_whiteout(bset(b, t), true);
2107
2108         bch2_btree_verify(c, b);
2109
2110         /*
2111          * If later we don't unconditionally sort down to a single bset, we have
2112          * to ensure this is still true:
2113          */
2114         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2115
2116         bne = want_new_bset(c, b);
2117         if (bne)
2118                 bch2_bset_init_next(c, b, bne);
2119
2120         bch2_btree_build_aux_trees(b);
2121
2122         return invalidated_iter;
2123 }
2124
2125 /*
2126  * Use this one if the node is intent locked:
2127  */
2128 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2129                            enum six_lock_type lock_type_held,
2130                            unsigned flags)
2131 {
2132         if (lock_type_held == SIX_LOCK_intent ||
2133             (lock_type_held == SIX_LOCK_read &&
2134              six_lock_tryupgrade(&b->c.lock))) {
2135                 __bch2_btree_node_write(c, b, flags);
2136
2137                 /* don't cycle lock unnecessarily: */
2138                 if (btree_node_just_written(b) &&
2139                     six_trylock_write(&b->c.lock)) {
2140                         bch2_btree_post_write_cleanup(c, b);
2141                         six_unlock_write(&b->c.lock);
2142                 }
2143
2144                 if (lock_type_held == SIX_LOCK_read)
2145                         six_lock_downgrade(&b->c.lock);
2146         } else {
2147                 __bch2_btree_node_write(c, b, flags);
2148                 if (lock_type_held == SIX_LOCK_write &&
2149                     btree_node_just_written(b))
2150                         bch2_btree_post_write_cleanup(c, b);
2151         }
2152 }
2153
2154 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2155 {
2156         struct bucket_table *tbl;
2157         struct rhash_head *pos;
2158         struct btree *b;
2159         unsigned i;
2160         bool ret = false;
2161 restart:
2162         rcu_read_lock();
2163         for_each_cached_btree(b, c, tbl, i, pos)
2164                 if (test_bit(flag, &b->flags)) {
2165                         rcu_read_unlock();
2166                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2167                         ret = true;
2168                         goto restart;
2169                 }
2170         rcu_read_unlock();
2171
2172         return ret;
2173 }
2174
2175 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2176 {
2177         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2178 }
2179
2180 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2181 {
2182         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2183 }
2184
2185 static const char * const bch2_btree_write_types[] = {
2186 #define x(t, n) [n] = #t,
2187         BCH_BTREE_WRITE_TYPES()
2188         NULL
2189 };
2190
2191 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2192 {
2193         printbuf_tabstop_push(out, 20);
2194         printbuf_tabstop_push(out, 10);
2195
2196         prt_tab(out);
2197         prt_str(out, "nr");
2198         prt_tab(out);
2199         prt_str(out, "size");
2200         prt_newline(out);
2201
2202         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2203                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2204                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2205
2206                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2207                 prt_tab(out);
2208                 prt_u64(out, nr);
2209                 prt_tab(out);
2210                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2211                 prt_newline(out);
2212         }
2213 }