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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_sort.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "buckets.h"
  13 #include "checksum.h"
  14 #include "debug.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "io.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_seq_blacklist.h"
  20 #include "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_noprof(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_noprof(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 #define btree_bounce_alloc(_c, _size, _used_mempool)            \
 127         alloc_hooks(btree_bounce_alloc_noprof(_c, _size, _used_mempool))
 128
 129 static void sort_bkey_ptrs(const struct btree *bt,
 130                            struct bkey_packed **ptrs, unsigned nr)
 131 {
 132         unsigned n = nr, a = nr / 2, b, c, d;
 133
 134         if (!a)
 135                 return;
 136
 137         /* Heap sort: see lib/sort.c: */
 138         while (1) {
 139                 if (a)
 140                         a--;
 141                 else if (--n)
 142                         swap(ptrs[0], ptrs[n]);
 143                 else
 144                         break;
 145
 146                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 147                         b = bch2_bkey_cmp_packed(bt,
 148                                             ptrs[c],
 149                                             ptrs[d]) >= 0 ? c : d;
 150                 if (d == n)
 151                         b = c;
 152
 153                 while (b != a &&
 154                        bch2_bkey_cmp_packed(bt,
 155                                        ptrs[a],
 156                                        ptrs[b]) >= 0)
 157                         b = (b - 1) / 2;
 158                 c = b;
 159                 while (b != a) {
 160                         b = (b - 1) / 2;
 161                         swap(ptrs[b], ptrs[c]);
 162                 }
 163         }
 164 }
 165
 166 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 167 {
 168         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 169         bool used_mempool = false;
 170         size_t bytes = b->whiteout_u64s * sizeof(u64);
 171
 172         if (!b->whiteout_u64s)
 173                 return;
 174
 175         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 176
 177         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 178
 179         for (k = unwritten_whiteouts_start(c, b);
 180              k != unwritten_whiteouts_end(c, b);
 181              k = bkey_p_next(k))
 182                 *--ptrs = k;
 183
 184         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 185
 186         k = new_whiteouts;
 187
 188         while (ptrs != ptrs_end) {
 189                 bkey_copy(k, *ptrs);
 190                 k = bkey_p_next(k);
 191                 ptrs++;
 192         }
 193
 194         verify_no_dups(b, new_whiteouts,
 195                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 196
 197         memcpy_u64s(unwritten_whiteouts_start(c, b),
 198                     new_whiteouts, b->whiteout_u64s);
 199
 200         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 201 }
 202
 203 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 204                                 bool compacting, enum compact_mode mode)
 205 {
 206         if (!bset_dead_u64s(b, t))
 207                 return false;
 208
 209         switch (mode) {
 210         case COMPACT_LAZY:
 211                 return should_compact_bset_lazy(b, t) ||
 212                         (compacting && !bset_written(b, bset(b, t)));
 213         case COMPACT_ALL:
 214                 return true;
 215         default:
 216                 BUG();
 217         }
 218 }
 219
 220 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 221 {
 222         struct bset_tree *t;
 223         bool ret = false;
 224
 225         for_each_bset(b, t) {
 226                 struct bset *i = bset(b, t);
 227                 struct bkey_packed *k, *n, *out, *start, *end;
 228                 struct btree_node_entry *src = NULL, *dst = NULL;
 229
 230                 if (t != b->set && !bset_written(b, i)) {
 231                         src = container_of(i, struct btree_node_entry, keys);
 232                         dst = max(write_block(b),
 233                                   (void *) btree_bkey_last(b, t - 1));
 234                 }
 235
 236                 if (src != dst)
 237                         ret = true;
 238
 239                 if (!should_compact_bset(b, t, ret, mode)) {
 240                         if (src != dst) {
 241                                 memmove(dst, src, sizeof(*src) +
 242                                         le16_to_cpu(src->keys.u64s) *
 243                                         sizeof(u64));
 244                                 i = &dst->keys;
 245                                 set_btree_bset(b, t, i);
 246                         }
 247                         continue;
 248                 }
 249
 250                 start   = btree_bkey_first(b, t);
 251                 end     = btree_bkey_last(b, t);
 252
 253                 if (src != dst) {
 254                         memmove(dst, src, sizeof(*src));
 255                         i = &dst->keys;
 256                         set_btree_bset(b, t, i);
 257                 }
 258
 259                 out = i->start;
 260
 261                 for (k = start; k != end; k = n) {
 262                         n = bkey_p_next(k);
 263
 264                         if (!bkey_deleted(k)) {
 265                                 bkey_copy(out, k);
 266                                 out = bkey_p_next(out);
 267                         } else {
 268                                 BUG_ON(k->needs_whiteout);
 269                         }
 270                 }
 271
 272                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 273                 set_btree_bset_end(b, t);
 274                 bch2_bset_set_no_aux_tree(b, t);
 275                 ret = true;
 276         }
 277
 278         bch2_verify_btree_nr_keys(b);
 279
 280         bch2_btree_build_aux_trees(b);
 281
 282         return ret;
 283 }
 284
 285 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 286                             enum compact_mode mode)
 287 {
 288         return bch2_drop_whiteouts(b, mode);
 289 }
 290
 291 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 292                             unsigned start_idx,
 293                             unsigned end_idx,
 294                             bool filter_whiteouts)
 295 {
 296         struct btree_node *out;
 297         struct sort_iter sort_iter;
 298         struct bset_tree *t;
 299         struct bset *start_bset = bset(b, &b->set[start_idx]);
 300         bool used_mempool = false;
 301         u64 start_time, seq = 0;
 302         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 303         bool sorting_entire_node = start_idx == 0 &&
 304                 end_idx == b->nsets;
 305
 306         sort_iter_init(&sort_iter, b);
 307
 308         for (t = b->set + start_idx;
 309              t < b->set + end_idx;
 310              t++) {
 311                 u64s += le16_to_cpu(bset(b, t)->u64s);
 312                 sort_iter_add(&sort_iter,
 313                               btree_bkey_first(b, t),
 314                               btree_bkey_last(b, t));
 315         }
 316
 317         bytes = sorting_entire_node
 318                 ? btree_bytes(c)
 319                 : __vstruct_bytes(struct btree_node, u64s);
 320
 321         out = btree_bounce_alloc(c, bytes, &used_mempool);
 322
 323         start_time = local_clock();
 324
 325         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 326
 327         out->keys.u64s = cpu_to_le16(u64s);
 328
 329         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 330
 331         if (sorting_entire_node)
 332                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 333                                        start_time);
 334
 335         /* Make sure we preserve bset journal_seq: */
 336         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 337                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 338         start_bset->journal_seq = cpu_to_le64(seq);
 339
 340         if (sorting_entire_node) {
 341                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 342
 343                 BUG_ON(bytes != btree_bytes(c));
 344
 345                 /*
 346                  * Our temporary buffer is the same size as the btree node's
 347                  * buffer, we can just swap buffers instead of doing a big
 348                  * memcpy()
 349                  */
 350                 *out = *b->data;
 351                 out->keys.u64s = cpu_to_le16(u64s);
 352                 swap(out, b->data);
 353                 set_btree_bset(b, b->set, &b->data->keys);
 354         } else {
 355                 start_bset->u64s = out->keys.u64s;
 356                 memcpy_u64s(start_bset->start,
 357                             out->keys.start,
 358                             le16_to_cpu(out->keys.u64s));
 359         }
 360
 361         for (i = start_idx + 1; i < end_idx; i++)
 362                 b->nr.bset_u64s[start_idx] +=
 363                         b->nr.bset_u64s[i];
 364
 365         b->nsets -= shift;
 366
 367         for (i = start_idx + 1; i < b->nsets; i++) {
 368                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 369                 b->set[i]               = b->set[i + shift];
 370         }
 371
 372         for (i = b->nsets; i < MAX_BSETS; i++)
 373                 b->nr.bset_u64s[i] = 0;
 374
 375         set_btree_bset_end(b, &b->set[start_idx]);
 376         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 377
 378         btree_bounce_free(c, bytes, used_mempool, out);
 379
 380         bch2_verify_btree_nr_keys(b);
 381 }
 382
 383 void bch2_btree_sort_into(struct bch_fs *c,
 384                          struct btree *dst,
 385                          struct btree *src)
 386 {
 387         struct btree_nr_keys nr;
 388         struct btree_node_iter src_iter;
 389         u64 start_time = local_clock();
 390
 391         BUG_ON(dst->nsets != 1);
 392
 393         bch2_bset_set_no_aux_tree(dst, dst->set);
 394
 395         bch2_btree_node_iter_init_from_start(&src_iter, src);
 396
 397         nr = bch2_sort_repack(btree_bset_first(dst),
 398                         src, &src_iter,
 399                         &dst->format,
 400                         true);
 401
 402         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 403                                start_time);
 404
 405         set_btree_bset_end(dst, dst->set);
 406
 407         dst->nr.live_u64s       += nr.live_u64s;
 408         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 409         dst->nr.packed_keys     += nr.packed_keys;
 410         dst->nr.unpacked_keys   += nr.unpacked_keys;
 411
 412         bch2_verify_btree_nr_keys(dst);
 413 }
 414
 415 #define SORT_CRIT       (4096 / sizeof(u64))
 416
 417 /*
 418  * We're about to add another bset to the btree node, so if there's currently
 419  * too many bsets - sort some of them together:
 420  */
 421 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 422 {
 423         unsigned unwritten_idx;
 424         bool ret = false;
 425
 426         for (unwritten_idx = 0;
 427              unwritten_idx < b->nsets;
 428              unwritten_idx++)
 429                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 430                         break;
 431
 432         if (b->nsets - unwritten_idx > 1) {
 433                 btree_node_sort(c, b, unwritten_idx,
 434                                 b->nsets, false);
 435                 ret = true;
 436         }
 437
 438         if (unwritten_idx > 1) {
 439                 btree_node_sort(c, b, 0, unwritten_idx, false);
 440                 ret = true;
 441         }
 442
 443         return ret;
 444 }
 445
 446 void bch2_btree_build_aux_trees(struct btree *b)
 447 {
 448         struct bset_tree *t;
 449
 450         for_each_bset(b, t)
 451                 bch2_bset_build_aux_tree(b, t,
 452                                 !bset_written(b, bset(b, t)) &&
 453                                 t == bset_tree_last(b));
 454 }
 455
 456 /*
 457  * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
 458  *
 459  * The first bset is going to be of similar order to the size of the node, the
 460  * last bset is bounded by btree_write_set_buffer(), which is set to keep the
 461  * memmove on insert from being too expensive: the middle bset should, ideally,
 462  * be the geometric mean of the first and the last.
 463  *
 464  * Returns true if the middle bset is greater than that geometric mean:
 465  */
 466 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
 467 {
 468         unsigned mid_u64s_bits =
 469                 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
 470
 471         return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
 472 }
 473
 474 /*
 475  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 476  * inserted into
 477  *
 478  * Safe to call if there already is an unwritten bset - will only add a new bset
 479  * if @b doesn't already have one.
 480  *
 481  * Returns true if we sorted (i.e. invalidated iterators
 482  */
 483 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 484 {
 485         struct bch_fs *c = trans->c;
 486         struct btree_node_entry *bne;
 487         bool reinit_iter = false;
 488
 489         EBUG_ON(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]);
 490         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 491         BUG_ON(btree_node_just_written(b));
 492
 493         if (b->nsets == MAX_BSETS &&
 494             !btree_node_write_in_flight(b) &&
 495             should_compact_all(c, b)) {
 496                 bch2_btree_node_write(c, b, SIX_LOCK_write,
 497                                       BTREE_WRITE_init_next_bset);
 498                 reinit_iter = true;
 499         }
 500
 501         if (b->nsets == MAX_BSETS &&
 502             btree_node_compact(c, b))
 503                 reinit_iter = true;
 504
 505         BUG_ON(b->nsets >= MAX_BSETS);
 506
 507         bne = want_new_bset(c, b);
 508         if (bne)
 509                 bch2_bset_init_next(c, b, bne);
 510
 511         bch2_btree_build_aux_trees(b);
 512
 513         if (reinit_iter)
 514                 bch2_trans_node_reinit_iter(trans, b);
 515 }
 516
 517 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 518                           struct btree *b)
 519 {
 520         prt_printf(out, "%s level %u/%u\n  ",
 521                bch2_btree_ids[b->c.btree_id],
 522                b->c.level,
 523                bch2_btree_id_root(c, b->c.btree_id)->level);
 524         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 525 }
 526
 527 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 528                           struct bch_dev *ca,
 529                           struct btree *b, struct bset *i,
 530                           unsigned offset, int write)
 531 {
 532         prt_printf(out, bch2_log_msg(c, "%s"),
 533                    write == READ
 534                    ? "error validating btree node "
 535                    : "corrupt btree node before write ");
 536         if (ca)
 537                 prt_printf(out, "on %s ", ca->name);
 538         prt_printf(out, "at btree ");
 539         btree_pos_to_text(out, c, b);
 540
 541         prt_printf(out, "\n  node offset %u", b->written);
 542         if (i)
 543                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 544         prt_str(out, ": ");
 545 }
 546
 547 static int __btree_err(int ret,
 548                        struct bch_fs *c,
 549                        struct bch_dev *ca,
 550                        struct btree *b,
 551                        struct bset *i,
 552                        int write,
 553                        bool have_retry,
 554                        const char *fmt, ...)
 555 {
 556         struct printbuf out = PRINTBUF;
 557         va_list args;
 558
 559         btree_err_msg(&out, c, ca, b, i, b->written, write);
 560
 561         va_start(args, fmt);
 562         prt_vprintf(&out, fmt, args);
 563         va_end(args);
 564
 565         if (write == WRITE) {
 566                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 567                 ret = c->opts.errors == BCH_ON_ERROR_continue
 568                         ? 0
 569                         : -BCH_ERR_fsck_errors_not_fixed;
 570                 goto out;
 571         }
 572
 573         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
 574                 ret = -BCH_ERR_btree_node_read_err_fixable;
 575         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
 576                 ret = -BCH_ERR_btree_node_read_err_bad_node;
 577
 578         switch (ret) {
 579         case -BCH_ERR_btree_node_read_err_fixable:
 580                 mustfix_fsck_err(c, "%s", out.buf);
 581                 ret = -BCH_ERR_fsck_fix;
 582                 break;
 583         case -BCH_ERR_btree_node_read_err_want_retry:
 584         case -BCH_ERR_btree_node_read_err_must_retry:
 585                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 586                 break;
 587         case -BCH_ERR_btree_node_read_err_bad_node:
 588                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 589                 bch2_topology_error(c);
 590                 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology) ?: -EIO;
 591                 break;
 592         case -BCH_ERR_btree_node_read_err_incompatible:
 593                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 594                 ret = -BCH_ERR_fsck_errors_not_fixed;
 595                 break;
 596         default:
 597                 BUG();
 598         }
 599 out:
 600 fsck_err:
 601         printbuf_exit(&out);
 602         return ret;
 603 }
 604
 605 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 606 ({                                                                      \
 607         int _ret = __btree_err(type, c, ca, b, i, write, have_retry, msg, ##__VA_ARGS__);\
 608                                                                         \
 609         if (_ret != -BCH_ERR_fsck_fix) {                                \
 610                 ret = _ret;                                             \
 611                 goto fsck_err;                                          \
 612         }                                                               \
 613                                                                         \
 614         *saw_error = true;                                              \
 615 })
 616
 617 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 618
 619 /*
 620  * When btree topology repair changes the start or end of a node, that might
 621  * mean we have to drop keys that are no longer inside the node:
 622  */
 623 __cold
 624 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 625 {
 626         struct bset_tree *t;
 627         struct bkey_s_c k;
 628         struct bkey unpacked;
 629         struct btree_node_iter iter;
 630
 631         for_each_bset(b, t) {
 632                 struct bset *i = bset(b, t);
 633                 struct bkey_packed *k;
 634
 635                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 636                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 637                                 break;
 638
 639                 if (k != i->start) {
 640                         unsigned shift = (u64 *) k - (u64 *) i->start;
 641
 642                         memmove_u64s_down(i->start, k,
 643                                           (u64 *) vstruct_end(i) - (u64 *) k);
 644                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 645                         set_btree_bset_end(b, t);
 646                 }
 647
 648                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 649                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 650                                 break;
 651
 652                 if (k != vstruct_last(i)) {
 653                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 654                         set_btree_bset_end(b, t);
 655                 }
 656         }
 657
 658         /*
 659          * Always rebuild search trees: eytzinger search tree nodes directly
 660          * depend on the values of min/max key:
 661          */
 662         bch2_bset_set_no_aux_tree(b, b->set);
 663         bch2_btree_build_aux_trees(b);
 664
 665         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 666                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 667                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 668         }
 669 }
 670
 671 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 672                          struct btree *b, struct bset *i,
 673                          unsigned offset, unsigned sectors,
 674                          int write, bool have_retry, bool *saw_error)
 675 {
 676         unsigned version = le16_to_cpu(i->version);
 677         struct printbuf buf1 = PRINTBUF;
 678         struct printbuf buf2 = PRINTBUF;
 679         int ret = 0;
 680
 681         btree_err_on(!bch2_version_compatible(version),
 682                      -BCH_ERR_btree_node_read_err_incompatible, c, ca, b, i,
 683                      "unsupported bset version %u.%u",
 684                      BCH_VERSION_MAJOR(version),
 685                      BCH_VERSION_MINOR(version));
 686
 687         if (btree_err_on(version < c->sb.version_min,
 688                          -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 689                          "bset version %u older than superblock version_min %u",
 690                          version, c->sb.version_min)) {
 691                 mutex_lock(&c->sb_lock);
 692                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 693                 bch2_write_super(c);
 694                 mutex_unlock(&c->sb_lock);
 695         }
 696
 697         if (btree_err_on(BCH_VERSION_MAJOR(version) >
 698                          BCH_VERSION_MAJOR(c->sb.version),
 699                          -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 700                          "bset version %u newer than superblock version %u",
 701                          version, c->sb.version)) {
 702                 mutex_lock(&c->sb_lock);
 703                 c->disk_sb.sb->version = cpu_to_le16(version);
 704                 bch2_write_super(c);
 705                 mutex_unlock(&c->sb_lock);
 706         }
 707
 708         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 709                      -BCH_ERR_btree_node_read_err_incompatible, c, ca, b, i,
 710                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 711
 712         if (btree_err_on(offset + sectors > btree_sectors(c),
 713                          -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
 714                          "bset past end of btree node")) {
 715                 i->u64s = 0;
 716                 ret = 0;
 717                 goto out;
 718         }
 719
 720         btree_err_on(offset && !i->u64s,
 721                      -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
 722                      "empty bset");
 723
 724         btree_err_on(BSET_OFFSET(i) &&
 725                      BSET_OFFSET(i) != offset,
 726                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 727                      "bset at wrong sector offset");
 728
 729         if (!offset) {
 730                 struct btree_node *bn =
 731                         container_of(i, struct btree_node, keys);
 732                 /* These indicate that we read the wrong btree node: */
 733
 734                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 735                         struct bch_btree_ptr_v2 *bp =
 736                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 737
 738                         /* XXX endianness */
 739                         btree_err_on(bp->seq != bn->keys.seq,
 740                                      -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 741                                      "incorrect sequence number (wrong btree node)");
 742                 }
 743
 744                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 745                              -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
 746                              "incorrect btree id");
 747
 748                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 749                              -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
 750                              "incorrect level");
 751
 752                 if (!write)
 753                         compat_btree_node(b->c.level, b->c.btree_id, version,
 754                                           BSET_BIG_ENDIAN(i), write, bn);
 755
 756                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 757                         struct bch_btree_ptr_v2 *bp =
 758                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 759
 760                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 761                                 b->data->min_key = bp->min_key;
 762                                 b->data->max_key = b->key.k.p;
 763                         }
 764
 765                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 766                                      -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 767                                      "incorrect min_key: got %s should be %s",
 768                                      (printbuf_reset(&buf1),
 769                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 770                                      (printbuf_reset(&buf2),
 771                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 772                 }
 773
 774                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 775                              -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
 776                              "incorrect max key %s",
 777                              (printbuf_reset(&buf1),
 778                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 779
 780                 if (write)
 781                         compat_btree_node(b->c.level, b->c.btree_id, version,
 782                                           BSET_BIG_ENDIAN(i), write, bn);
 783
 784                 btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
 785                              -BCH_ERR_btree_node_read_err_bad_node, c, ca, b, i,
 786                              "invalid bkey format: %s\n  %s", buf1.buf,
 787                              (printbuf_reset(&buf2),
 788                               bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
 789                 printbuf_reset(&buf1);
 790
 791                 compat_bformat(b->c.level, b->c.btree_id, version,
 792                                BSET_BIG_ENDIAN(i), write,
 793                                &bn->format);
 794         }
 795 out:
 796 fsck_err:
 797         printbuf_exit(&buf2);
 798         printbuf_exit(&buf1);
 799         return ret;
 800 }
 801
 802 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 803                             struct bkey_s_c k,
 804                             bool updated_range, int rw,
 805                             struct printbuf *err)
 806 {
 807         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 808                 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
 809                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 810 }
 811
 812 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 813                          struct bset *i, int write,
 814                          bool have_retry, bool *saw_error)
 815 {
 816         unsigned version = le16_to_cpu(i->version);
 817         struct bkey_packed *k, *prev = NULL;
 818         struct printbuf buf = PRINTBUF;
 819         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 820                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 821         int ret = 0;
 822
 823         for (k = i->start;
 824              k != vstruct_last(i);) {
 825                 struct bkey_s u;
 826                 struct bkey tmp;
 827
 828                 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
 829                                  -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 830                                  "key extends past end of bset")) {
 831                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 832                         break;
 833                 }
 834
 835                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 836                                  -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
 837                                  "invalid bkey format %u", k->format)) {
 838                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 839                         memmove_u64s_down(k, bkey_p_next(k),
 840                                           (u64 *) vstruct_end(i) - (u64 *) k);
 841                         continue;
 842                 }
 843
 844                 /* XXX: validate k->u64s */
 845                 if (!write)
 846                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 847                                     BSET_BIG_ENDIAN(i), write,
 848                                     &b->format, k);
 849
 850                 u = __bkey_disassemble(b, k, &tmp);
 851
 852                 printbuf_reset(&buf);
 853                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 854                         printbuf_reset(&buf);
 855                         prt_printf(&buf, "invalid bkey:  ");
 856                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 857                         prt_printf(&buf, "\n  ");
 858                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 859
 860                         btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
 861
 862                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 863                         memmove_u64s_down(k, bkey_p_next(k),
 864                                           (u64 *) vstruct_end(i) - (u64 *) k);
 865                         continue;
 866                 }
 867
 868                 if (write)
 869                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 870                                     BSET_BIG_ENDIAN(i), write,
 871                                     &b->format, k);
 872
 873                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 874                         struct bkey up = bkey_unpack_key(b, prev);
 875
 876                         printbuf_reset(&buf);
 877                         prt_printf(&buf, "keys out of order: ");
 878                         bch2_bkey_to_text(&buf, &up);
 879                         prt_printf(&buf, " > ");
 880                         bch2_bkey_to_text(&buf, u.k);
 881
 882                         bch2_dump_bset(c, b, i, 0);
 883
 884                         if (btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf)) {
 885                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 886                                 memmove_u64s_down(k, bkey_p_next(k),
 887                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 888                                 continue;
 889                         }
 890                 }
 891
 892                 prev = k;
 893                 k = bkey_p_next(k);
 894         }
 895 fsck_err:
 896         printbuf_exit(&buf);
 897         return ret;
 898 }
 899
 900 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 901                               struct btree *b, bool have_retry, bool *saw_error)
 902 {
 903         struct btree_node_entry *bne;
 904         struct sort_iter *iter;
 905         struct btree_node *sorted;
 906         struct bkey_packed *k;
 907         struct bch_extent_ptr *ptr;
 908         struct bset *i;
 909         bool used_mempool, blacklisted;
 910         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 911                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 912         unsigned u64s;
 913         unsigned blacklisted_written, nonblacklisted_written = 0;
 914         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 915         struct printbuf buf = PRINTBUF;
 916         int ret = 0, retry_read = 0, write = READ;
 917
 918         b->version_ondisk = U16_MAX;
 919         /* We might get called multiple times on read retry: */
 920         b->written = 0;
 921
 922         iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
 923         sort_iter_init(iter, b);
 924         iter->size = (btree_blocks(c) + 1) * 2;
 925
 926         if (bch2_meta_read_fault("btree"))
 927                 btree_err(-BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 928                           "dynamic fault");
 929
 930         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 931                      -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 932                      "bad magic: want %llx, got %llx",
 933                      bset_magic(c), le64_to_cpu(b->data->magic));
 934
 935         btree_err_on(!b->data->keys.seq,
 936                      -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 937                      "bad btree header: seq 0");
 938
 939         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 940                 struct bch_btree_ptr_v2 *bp =
 941                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 942
 943                 btree_err_on(b->data->keys.seq != bp->seq,
 944                              -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
 945                              "got wrong btree node (seq %llx want %llx)",
 946                              b->data->keys.seq, bp->seq);
 947         }
 948
 949         while (b->written < (ptr_written ?: btree_sectors(c))) {
 950                 unsigned sectors;
 951                 struct nonce nonce;
 952                 struct bch_csum csum;
 953                 bool first = !b->written;
 954
 955                 if (!b->written) {
 956                         i = &b->data->keys;
 957
 958                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 959                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 960                                      "unknown checksum type %llu",
 961                                      BSET_CSUM_TYPE(i));
 962
 963                         nonce = btree_nonce(i, b->written << 9);
 964                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 965
 966                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 967                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 968                                      "invalid checksum");
 969
 970                         ret = bset_encrypt(c, i, b->written << 9);
 971                         if (bch2_fs_fatal_err_on(ret, c,
 972                                         "error decrypting btree node: %i", ret))
 973                                 goto fsck_err;
 974
 975                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 976                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 977                                      -BCH_ERR_btree_node_read_err_incompatible, c, NULL, b, NULL,
 978                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 979
 980                         sectors = vstruct_sectors(b->data, c->block_bits);
 981                 } else {
 982                         bne = write_block(b);
 983                         i = &bne->keys;
 984
 985                         if (i->seq != b->data->keys.seq)
 986                                 break;
 987
 988                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 989                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 990                                      "unknown checksum type %llu",
 991                                      BSET_CSUM_TYPE(i));
 992
 993                         nonce = btree_nonce(i, b->written << 9);
 994                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 995
 996                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 997                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
 998                                      "invalid checksum");
 999
1000                         ret = bset_encrypt(c, i, b->written << 9);
1001                         if (bch2_fs_fatal_err_on(ret, c,
1002                                         "error decrypting btree node: %i\n", ret))
1003                                 goto fsck_err;
1004
1005                         sectors = vstruct_sectors(bne, c->block_bits);
1006                 }
1007
1008                 b->version_ondisk = min(b->version_ondisk,
1009                                         le16_to_cpu(i->version));
1010
1011                 ret = validate_bset(c, ca, b, i, b->written, sectors,
1012                                     READ, have_retry, saw_error);
1013                 if (ret)
1014                         goto fsck_err;
1015
1016                 if (!b->written)
1017                         btree_node_set_format(b, b->data->format);
1018
1019                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1020                 if (ret)
1021                         goto fsck_err;
1022
1023                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1024
1025                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1026                                         le64_to_cpu(i->journal_seq),
1027                                         true);
1028
1029                 btree_err_on(blacklisted && first,
1030                              -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
1031                              "first btree node bset has blacklisted journal seq (%llu)",
1032                              le64_to_cpu(i->journal_seq));
1033
1034                 btree_err_on(blacklisted && ptr_written,
1035                              -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
1036                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1037                              le64_to_cpu(i->journal_seq),
1038                              b->written, b->written + sectors, ptr_written);
1039
1040                 b->written += sectors;
1041
1042                 if (blacklisted && !first)
1043                         continue;
1044
1045                 sort_iter_add(iter,
1046                               vstruct_idx(i, 0),
1047                               vstruct_last(i));
1048
1049                 nonblacklisted_written = b->written;
1050         }
1051
1052         if (ptr_written) {
1053                 btree_err_on(b->written < ptr_written,
1054                              -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
1055                              "btree node data missing: expected %u sectors, found %u",
1056                              ptr_written, b->written);
1057         } else {
1058                 for (bne = write_block(b);
1059                      bset_byte_offset(b, bne) < btree_bytes(c);
1060                      bne = (void *) bne + block_bytes(c))
1061                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1062                                      !bch2_journal_seq_is_blacklisted(c,
1063                                                                       le64_to_cpu(bne->keys.journal_seq),
1064                                                                       true),
1065                                      -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
1066                                      "found bset signature after last bset");
1067
1068                 /*
1069                  * Blacklisted bsets are those that were written after the most recent
1070                  * (flush) journal write. Since there wasn't a flush, they may not have
1071                  * made it to all devices - which means we shouldn't write new bsets
1072                  * after them, as that could leave a gap and then reads from that device
1073                  * wouldn't find all the bsets in that btree node - which means it's
1074                  * important that we start writing new bsets after the most recent _non_
1075                  * blacklisted bset:
1076                  */
1077                 blacklisted_written = b->written;
1078                 b->written = nonblacklisted_written;
1079         }
1080
1081         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1082         sorted->keys.u64s = 0;
1083
1084         set_btree_bset(b, b->set, &b->data->keys);
1085
1086         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1087
1088         u64s = le16_to_cpu(sorted->keys.u64s);
1089         *sorted = *b->data;
1090         sorted->keys.u64s = cpu_to_le16(u64s);
1091         swap(sorted, b->data);
1092         set_btree_bset(b, b->set, &b->data->keys);
1093         b->nsets = 1;
1094
1095         BUG_ON(b->nr.live_u64s != u64s);
1096
1097         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1098
1099         if (updated_range)
1100                 bch2_btree_node_drop_keys_outside_node(b);
1101
1102         i = &b->data->keys;
1103         for (k = i->start; k != vstruct_last(i);) {
1104                 struct bkey tmp;
1105                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1106
1107                 printbuf_reset(&buf);
1108
1109                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1110                     (bch2_inject_invalid_keys &&
1111                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1112                         printbuf_reset(&buf);
1113
1114                         prt_printf(&buf, "invalid bkey: ");
1115                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1116                         prt_printf(&buf, "\n  ");
1117                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1118
1119                         btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
1120
1121                         btree_keys_account_key_drop(&b->nr, 0, k);
1122
1123                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1124                         memmove_u64s_down(k, bkey_p_next(k),
1125                                           (u64 *) vstruct_end(i) - (u64 *) k);
1126                         set_btree_bset_end(b, b->set);
1127                         continue;
1128                 }
1129
1130                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1131                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1132
1133                         bp.v->mem_ptr = 0;
1134                 }
1135
1136                 k = bkey_p_next(k);
1137         }
1138
1139         bch2_bset_build_aux_tree(b, b->set, false);
1140
1141         set_needs_whiteout(btree_bset_first(b), true);
1142
1143         btree_node_reset_sib_u64s(b);
1144
1145         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1146                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1147
1148                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1149                         set_btree_node_need_rewrite(b);
1150         }
1151
1152         if (!ptr_written)
1153                 set_btree_node_need_rewrite(b);
1154 out:
1155         mempool_free(iter, &c->fill_iter);
1156         printbuf_exit(&buf);
1157         return retry_read;
1158 fsck_err:
1159         if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1160             ret == -BCH_ERR_btree_node_read_err_must_retry)
1161                 retry_read = 1;
1162         else
1163                 set_btree_node_read_error(b);
1164         goto out;
1165 }
1166
1167 static void btree_node_read_work(struct work_struct *work)
1168 {
1169         struct btree_read_bio *rb =
1170                 container_of(work, struct btree_read_bio, work);
1171         struct bch_fs *c        = rb->c;
1172         struct btree *b         = rb->b;
1173         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1174         struct bio *bio         = &rb->bio;
1175         struct bch_io_failures failed = { .nr = 0 };
1176         struct printbuf buf = PRINTBUF;
1177         bool saw_error = false;
1178         bool retry = false;
1179         bool can_retry;
1180
1181         goto start;
1182         while (1) {
1183                 retry = true;
1184                 bch_info(c, "retrying read");
1185                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1186                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1187                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1188                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1189                 bio->bi_iter.bi_size    = btree_bytes(c);
1190
1191                 if (rb->have_ioref) {
1192                         bio_set_dev(bio, ca->disk_sb.bdev);
1193                         submit_bio_wait(bio);
1194                 } else {
1195                         bio->bi_status = BLK_STS_REMOVED;
1196                 }
1197 start:
1198                 printbuf_reset(&buf);
1199                 btree_pos_to_text(&buf, c, b);
1200                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1201                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1202                 if (rb->have_ioref)
1203                         percpu_ref_put(&ca->io_ref);
1204                 rb->have_ioref = false;
1205
1206                 bch2_mark_io_failure(&failed, &rb->pick);
1207
1208                 can_retry = bch2_bkey_pick_read_device(c,
1209                                 bkey_i_to_s_c(&b->key),
1210                                 &failed, &rb->pick) > 0;
1211
1212                 if (!bio->bi_status &&
1213                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1214                         if (retry)
1215                                 bch_info(c, "retry success");
1216                         break;
1217                 }
1218
1219                 saw_error = true;
1220
1221                 if (!can_retry) {
1222                         set_btree_node_read_error(b);
1223                         break;
1224                 }
1225         }
1226
1227         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1228                                rb->start_time);
1229         bio_put(&rb->bio);
1230         printbuf_exit(&buf);
1231
1232         if (saw_error && !btree_node_read_error(b)) {
1233                 struct printbuf buf = PRINTBUF;
1234
1235                 bch2_bpos_to_text(&buf, b->key.k.p);
1236                 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1237                          __func__, bch2_btree_ids[b->c.btree_id], b->c.level, buf.buf);
1238                 printbuf_exit(&buf);
1239
1240                 bch2_btree_node_rewrite_async(c, b);
1241         }
1242
1243         clear_btree_node_read_in_flight(b);
1244         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1245 }
1246
1247 static void btree_node_read_endio(struct bio *bio)
1248 {
1249         struct btree_read_bio *rb =
1250                 container_of(bio, struct btree_read_bio, bio);
1251         struct bch_fs *c        = rb->c;
1252
1253         if (rb->have_ioref) {
1254                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1255
1256                 bch2_latency_acct(ca, rb->start_time, READ);
1257         }
1258
1259         queue_work(c->io_complete_wq, &rb->work);
1260 }
1261
1262 struct btree_node_read_all {
1263         struct closure          cl;
1264         struct bch_fs           *c;
1265         struct btree            *b;
1266         unsigned                nr;
1267         void                    *buf[BCH_REPLICAS_MAX];
1268         struct bio              *bio[BCH_REPLICAS_MAX];
1269         blk_status_t            err[BCH_REPLICAS_MAX];
1270 };
1271
1272 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1273 {
1274         struct btree_node *bn = data;
1275         struct btree_node_entry *bne;
1276         unsigned offset = 0;
1277
1278         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1279                 return 0;
1280
1281         while (offset < btree_sectors(c)) {
1282                 if (!offset) {
1283                         offset += vstruct_sectors(bn, c->block_bits);
1284                 } else {
1285                         bne = data + (offset << 9);
1286                         if (bne->keys.seq != bn->keys.seq)
1287                                 break;
1288                         offset += vstruct_sectors(bne, c->block_bits);
1289                 }
1290         }
1291
1292         return offset;
1293 }
1294
1295 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1296 {
1297         struct btree_node *bn = data;
1298         struct btree_node_entry *bne;
1299
1300         if (!offset)
1301                 return false;
1302
1303         while (offset < btree_sectors(c)) {
1304                 bne = data + (offset << 9);
1305                 if (bne->keys.seq == bn->keys.seq)
1306                         return true;
1307                 offset++;
1308         }
1309
1310         return false;
1311         return offset;
1312 }
1313
1314 static void btree_node_read_all_replicas_done(struct closure *cl)
1315 {
1316         struct btree_node_read_all *ra =
1317                 container_of(cl, struct btree_node_read_all, cl);
1318         struct bch_fs *c = ra->c;
1319         struct btree *b = ra->b;
1320         struct printbuf buf = PRINTBUF;
1321         bool dump_bset_maps = false;
1322         bool have_retry = false;
1323         int ret = 0, best = -1, write = READ;
1324         unsigned i, written = 0, written2 = 0;
1325         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1326                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1327         bool _saw_error = false, *saw_error = &_saw_error;
1328
1329         for (i = 0; i < ra->nr; i++) {
1330                 struct btree_node *bn = ra->buf[i];
1331
1332                 if (ra->err[i])
1333                         continue;
1334
1335                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1336                     (seq && seq != bn->keys.seq))
1337                         continue;
1338
1339                 if (best < 0) {
1340                         best = i;
1341                         written = btree_node_sectors_written(c, bn);
1342                         continue;
1343                 }
1344
1345                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1346                 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1347                                  "btree node sectors written mismatch: %u != %u",
1348                                  written, written2) ||
1349                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1350                                  -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1351                                  "found bset signature after last bset") ||
1352                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1353                                  -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1354                                  "btree node replicas content mismatch"))
1355                         dump_bset_maps = true;
1356
1357                 if (written2 > written) {
1358                         written = written2;
1359                         best = i;
1360                 }
1361         }
1362 fsck_err:
1363         if (dump_bset_maps) {
1364                 for (i = 0; i < ra->nr; i++) {
1365                         struct btree_node *bn = ra->buf[i];
1366                         struct btree_node_entry *bne = NULL;
1367                         unsigned offset = 0, sectors;
1368                         bool gap = false;
1369
1370                         if (ra->err[i])
1371                                 continue;
1372
1373                         printbuf_reset(&buf);
1374
1375                         while (offset < btree_sectors(c)) {
1376                                 if (!offset) {
1377                                         sectors = vstruct_sectors(bn, c->block_bits);
1378                                 } else {
1379                                         bne = ra->buf[i] + (offset << 9);
1380                                         if (bne->keys.seq != bn->keys.seq)
1381                                                 break;
1382                                         sectors = vstruct_sectors(bne, c->block_bits);
1383                                 }
1384
1385                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1386                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1387                                                         le64_to_cpu(bne->keys.journal_seq), false))
1388                                         prt_printf(&buf, "*");
1389                                 offset += sectors;
1390                         }
1391
1392                         while (offset < btree_sectors(c)) {
1393                                 bne = ra->buf[i] + (offset << 9);
1394                                 if (bne->keys.seq == bn->keys.seq) {
1395                                         if (!gap)
1396                                                 prt_printf(&buf, " GAP");
1397                                         gap = true;
1398
1399                                         sectors = vstruct_sectors(bne, c->block_bits);
1400                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1401                                         if (bch2_journal_seq_is_blacklisted(c,
1402                                                         le64_to_cpu(bne->keys.journal_seq), false))
1403                                                 prt_printf(&buf, "*");
1404                                 }
1405                                 offset++;
1406                         }
1407
1408                         bch_err(c, "replica %u:%s", i, buf.buf);
1409                 }
1410         }
1411
1412         if (best >= 0) {
1413                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1414                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1415         } else {
1416                 ret = -1;
1417         }
1418
1419         if (ret)
1420                 set_btree_node_read_error(b);
1421         else if (*saw_error)
1422                 bch2_btree_node_rewrite_async(c, b);
1423
1424         for (i = 0; i < ra->nr; i++) {
1425                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1426                 bio_put(ra->bio[i]);
1427         }
1428
1429         closure_debug_destroy(&ra->cl);
1430         kfree(ra);
1431         printbuf_exit(&buf);
1432
1433         clear_btree_node_read_in_flight(b);
1434         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1435 }
1436
1437 static void btree_node_read_all_replicas_endio(struct bio *bio)
1438 {
1439         struct btree_read_bio *rb =
1440                 container_of(bio, struct btree_read_bio, bio);
1441         struct bch_fs *c        = rb->c;
1442         struct btree_node_read_all *ra = rb->ra;
1443
1444         if (rb->have_ioref) {
1445                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1446
1447                 bch2_latency_acct(ca, rb->start_time, READ);
1448         }
1449
1450         ra->err[rb->idx] = bio->bi_status;
1451         closure_put(&ra->cl);
1452 }
1453
1454 /*
1455  * XXX This allocates multiple times from the same mempools, and can deadlock
1456  * under sufficient memory pressure (but is only a debug path)
1457  */
1458 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1459 {
1460         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1461         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1462         const union bch_extent_entry *entry;
1463         struct extent_ptr_decoded pick;
1464         struct btree_node_read_all *ra;
1465         unsigned i;
1466
1467         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1468         if (!ra)
1469                 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1470
1471         closure_init(&ra->cl, NULL);
1472         ra->c   = c;
1473         ra->b   = b;
1474         ra->nr  = bch2_bkey_nr_ptrs(k);
1475
1476         for (i = 0; i < ra->nr; i++) {
1477                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1478                 ra->bio[i] = bio_alloc_bioset(NULL,
1479                                               buf_pages(ra->buf[i], btree_bytes(c)),
1480                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1481                                               GFP_NOFS,
1482                                               &c->btree_bio);
1483         }
1484
1485         i = 0;
1486         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1487                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1488                 struct btree_read_bio *rb =
1489                         container_of(ra->bio[i], struct btree_read_bio, bio);
1490                 rb->c                   = c;
1491                 rb->b                   = b;
1492                 rb->ra                  = ra;
1493                 rb->start_time          = local_clock();
1494                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1495                 rb->idx                 = i;
1496                 rb->pick                = pick;
1497                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1498                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1499                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1500
1501                 if (rb->have_ioref) {
1502                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1503                                      bio_sectors(&rb->bio));
1504                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1505
1506                         closure_get(&ra->cl);
1507                         submit_bio(&rb->bio);
1508                 } else {
1509                         ra->err[i] = BLK_STS_REMOVED;
1510                 }
1511
1512                 i++;
1513         }
1514
1515         if (sync) {
1516                 closure_sync(&ra->cl);
1517                 btree_node_read_all_replicas_done(&ra->cl);
1518         } else {
1519                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1520                             c->io_complete_wq);
1521         }
1522
1523         return 0;
1524 }
1525
1526 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1527                           bool sync)
1528 {
1529         struct extent_ptr_decoded pick;
1530         struct btree_read_bio *rb;
1531         struct bch_dev *ca;
1532         struct bio *bio;
1533         int ret;
1534
1535         trace_and_count(c, btree_node_read, c, b);
1536
1537         if (bch2_verify_all_btree_replicas &&
1538             !btree_node_read_all_replicas(c, b, sync))
1539                 return;
1540
1541         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1542                                          NULL, &pick);
1543
1544         if (ret <= 0) {
1545                 struct printbuf buf = PRINTBUF;
1546
1547                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1548                 btree_pos_to_text(&buf, c, b);
1549                 bch_err(c, "%s", buf.buf);
1550
1551                 if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1552                     c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1553                         bch2_fatal_error(c);
1554
1555                 set_btree_node_read_error(b);
1556                 clear_btree_node_read_in_flight(b);
1557                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1558                 printbuf_exit(&buf);
1559                 return;
1560         }
1561
1562         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1563
1564         bio = bio_alloc_bioset(NULL,
1565                                buf_pages(b->data, btree_bytes(c)),
1566                                REQ_OP_READ|REQ_SYNC|REQ_META,
1567                                GFP_NOFS,
1568                                &c->btree_bio);
1569         rb = container_of(bio, struct btree_read_bio, bio);
1570         rb->c                   = c;
1571         rb->b                   = b;
1572         rb->ra                  = NULL;
1573         rb->start_time          = local_clock();
1574         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1575         rb->pick                = pick;
1576         INIT_WORK(&rb->work, btree_node_read_work);
1577         bio->bi_iter.bi_sector  = pick.ptr.offset;
1578         bio->bi_end_io          = btree_node_read_endio;
1579         bch2_bio_map(bio, b->data, btree_bytes(c));
1580
1581         if (rb->have_ioref) {
1582                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1583                              bio_sectors(bio));
1584                 bio_set_dev(bio, ca->disk_sb.bdev);
1585
1586                 if (sync) {
1587                         submit_bio_wait(bio);
1588
1589                         btree_node_read_work(&rb->work);
1590                 } else {
1591                         submit_bio(bio);
1592                 }
1593         } else {
1594                 bio->bi_status = BLK_STS_REMOVED;
1595
1596                 if (sync)
1597                         btree_node_read_work(&rb->work);
1598                 else
1599                         queue_work(c->io_complete_wq, &rb->work);
1600         }
1601 }
1602
1603 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1604                                   const struct bkey_i *k, unsigned level)
1605 {
1606         struct bch_fs *c = trans->c;
1607         struct closure cl;
1608         struct btree *b;
1609         int ret;
1610
1611         closure_init_stack(&cl);
1612
1613         do {
1614                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1615                 closure_sync(&cl);
1616         } while (ret);
1617
1618         b = bch2_btree_node_mem_alloc(trans, level != 0);
1619         bch2_btree_cache_cannibalize_unlock(c);
1620
1621         BUG_ON(IS_ERR(b));
1622
1623         bkey_copy(&b->key, k);
1624         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1625
1626         set_btree_node_read_in_flight(b);
1627
1628         bch2_btree_node_read(c, b, true);
1629
1630         if (btree_node_read_error(b)) {
1631                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1632
1633                 mutex_lock(&c->btree_cache.lock);
1634                 list_move(&b->list, &c->btree_cache.freeable);
1635                 mutex_unlock(&c->btree_cache.lock);
1636
1637                 ret = -EIO;
1638                 goto err;
1639         }
1640
1641         bch2_btree_set_root_for_read(c, b);
1642 err:
1643         six_unlock_write(&b->c.lock);
1644         six_unlock_intent(&b->c.lock);
1645
1646         return ret;
1647 }
1648
1649 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1650                         const struct bkey_i *k, unsigned level)
1651 {
1652         return bch2_trans_run(c, __bch2_btree_root_read(&trans, id, k, level));
1653
1654 }
1655
1656 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1657                               struct btree_write *w)
1658 {
1659         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1660
1661         do {
1662                 old = new = v;
1663                 if (!(old & 1))
1664                         break;
1665
1666                 new &= ~1UL;
1667         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1668
1669         if (old & 1)
1670                 closure_put(&((struct btree_update *) new)->cl);
1671
1672         bch2_journal_pin_drop(&c->journal, &w->journal);
1673 }
1674
1675 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1676 {
1677         struct btree_write *w = btree_prev_write(b);
1678         unsigned long old, new, v;
1679         unsigned type = 0;
1680
1681         bch2_btree_complete_write(c, b, w);
1682
1683         v = READ_ONCE(b->flags);
1684         do {
1685                 old = new = v;
1686
1687                 if ((old & (1U << BTREE_NODE_dirty)) &&
1688                     (old & (1U << BTREE_NODE_need_write)) &&
1689                     !(old & (1U << BTREE_NODE_never_write)) &&
1690                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1691                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1692                         new &= ~(1U << BTREE_NODE_dirty);
1693                         new &= ~(1U << BTREE_NODE_need_write);
1694                         new |=  (1U << BTREE_NODE_write_in_flight);
1695                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1696                         new |=  (1U << BTREE_NODE_just_written);
1697                         new ^=  (1U << BTREE_NODE_write_idx);
1698
1699                         type = new & BTREE_WRITE_TYPE_MASK;
1700                         new &= ~BTREE_WRITE_TYPE_MASK;
1701                 } else {
1702                         new &= ~(1U << BTREE_NODE_write_in_flight);
1703                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1704                 }
1705         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1706
1707         if (new & (1U << BTREE_NODE_write_in_flight))
1708                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1709         else
1710                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1711 }
1712
1713 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1714 {
1715         struct btree_trans trans;
1716
1717         bch2_trans_init(&trans, c, 0, 0);
1718
1719         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
1720         __btree_node_write_done(c, b);
1721         six_unlock_read(&b->c.lock);
1722
1723         bch2_trans_exit(&trans);
1724 }
1725
1726 static void btree_node_write_work(struct work_struct *work)
1727 {
1728         struct btree_write_bio *wbio =
1729                 container_of(work, struct btree_write_bio, work);
1730         struct bch_fs *c        = wbio->wbio.c;
1731         struct btree *b         = wbio->wbio.bio.bi_private;
1732         struct bch_extent_ptr *ptr;
1733         int ret = 0;
1734
1735         btree_bounce_free(c,
1736                 wbio->data_bytes,
1737                 wbio->wbio.used_mempool,
1738                 wbio->data);
1739
1740         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1741                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1742
1743         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1744                 goto err;
1745
1746         if (wbio->wbio.first_btree_write) {
1747                 if (wbio->wbio.failed.nr) {
1748
1749                 }
1750         } else {
1751                 ret = bch2_trans_do(c, NULL, NULL, 0,
1752                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1753                                         BCH_WATERMARK_reclaim|
1754                                         BTREE_INSERT_JOURNAL_RECLAIM|
1755                                         BTREE_INSERT_NOFAIL|
1756                                         BTREE_INSERT_NOCHECK_RW,
1757                                         !wbio->wbio.failed.nr));
1758                 if (ret)
1759                         goto err;
1760         }
1761 out:
1762         bio_put(&wbio->wbio.bio);
1763         btree_node_write_done(c, b);
1764         return;
1765 err:
1766         set_btree_node_noevict(b);
1767         if (!bch2_err_matches(ret, EROFS))
1768                 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1769         goto out;
1770 }
1771
1772 static void btree_node_write_endio(struct bio *bio)
1773 {
1774         struct bch_write_bio *wbio      = to_wbio(bio);
1775         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1776         struct bch_write_bio *orig      = parent ?: wbio;
1777         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1778         struct bch_fs *c                = wbio->c;
1779         struct btree *b                 = wbio->bio.bi_private;
1780         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1781         unsigned long flags;
1782
1783         if (wbio->have_ioref)
1784                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1785
1786         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1787                                bch2_blk_status_to_str(bio->bi_status)) ||
1788             bch2_meta_write_fault("btree")) {
1789                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1790                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1791                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1792         }
1793
1794         if (wbio->have_ioref)
1795                 percpu_ref_put(&ca->io_ref);
1796
1797         if (parent) {
1798                 bio_put(bio);
1799                 bio_endio(&parent->bio);
1800                 return;
1801         }
1802
1803         clear_btree_node_write_in_flight_inner(b);
1804         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1805         INIT_WORK(&wb->work, btree_node_write_work);
1806         queue_work(c->btree_io_complete_wq, &wb->work);
1807 }
1808
1809 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1810                                    struct bset *i, unsigned sectors)
1811 {
1812         struct printbuf buf = PRINTBUF;
1813         bool saw_error;
1814         int ret;
1815
1816         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1817                                 BKEY_TYPE_btree, WRITE, &buf);
1818
1819         if (ret)
1820                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1821         printbuf_exit(&buf);
1822         if (ret)
1823                 return ret;
1824
1825         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1826                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1827         if (ret) {
1828                 bch2_inconsistent_error(c);
1829                 dump_stack();
1830         }
1831
1832         return ret;
1833 }
1834
1835 static void btree_write_submit(struct work_struct *work)
1836 {
1837         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1838         struct bch_extent_ptr *ptr;
1839         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1840
1841         bkey_copy(&tmp.k, &wbio->key);
1842
1843         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1844                 ptr->offset += wbio->sector_offset;
1845
1846         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1847                                   &tmp.k, false);
1848 }
1849
1850 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1851 {
1852         struct btree_write_bio *wbio;
1853         struct bset_tree *t;
1854         struct bset *i;
1855         struct btree_node *bn = NULL;
1856         struct btree_node_entry *bne = NULL;
1857         struct sort_iter sort_iter;
1858         struct nonce nonce;
1859         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1860         u64 seq = 0;
1861         bool used_mempool;
1862         unsigned long old, new;
1863         bool validate_before_checksum = false;
1864         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1865         void *data;
1866         int ret;
1867
1868         if (flags & BTREE_WRITE_ALREADY_STARTED)
1869                 goto do_write;
1870
1871         /*
1872          * We may only have a read lock on the btree node - the dirty bit is our
1873          * "lock" against racing with other threads that may be trying to start
1874          * a write, we do a write iff we clear the dirty bit. Since setting the
1875          * dirty bit requires a write lock, we can't race with other threads
1876          * redirtying it:
1877          */
1878         do {
1879                 old = new = READ_ONCE(b->flags);
1880
1881                 if (!(old & (1 << BTREE_NODE_dirty)))
1882                         return;
1883
1884                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1885                     !(old & (1 << BTREE_NODE_need_write)))
1886                         return;
1887
1888                 if (old &
1889                     ((1 << BTREE_NODE_never_write)|
1890                      (1 << BTREE_NODE_write_blocked)))
1891                         return;
1892
1893                 if (b->written &&
1894                     (old & (1 << BTREE_NODE_will_make_reachable)))
1895                         return;
1896
1897                 if (old & (1 << BTREE_NODE_write_in_flight))
1898                         return;
1899
1900                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1901                         type = new & BTREE_WRITE_TYPE_MASK;
1902                 new &= ~BTREE_WRITE_TYPE_MASK;
1903
1904                 new &= ~(1 << BTREE_NODE_dirty);
1905                 new &= ~(1 << BTREE_NODE_need_write);
1906                 new |=  (1 << BTREE_NODE_write_in_flight);
1907                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1908                 new |=  (1 << BTREE_NODE_just_written);
1909                 new ^=  (1 << BTREE_NODE_write_idx);
1910         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1911
1912         if (new & (1U << BTREE_NODE_need_write))
1913                 return;
1914 do_write:
1915         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1916
1917         atomic_dec(&c->btree_cache.dirty);
1918
1919         BUG_ON(btree_node_fake(b));
1920         BUG_ON((b->will_make_reachable != 0) != !b->written);
1921
1922         BUG_ON(b->written >= btree_sectors(c));
1923         BUG_ON(b->written & (block_sectors(c) - 1));
1924         BUG_ON(bset_written(b, btree_bset_last(b)));
1925         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1926         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1927
1928         bch2_sort_whiteouts(c, b);
1929
1930         sort_iter_init(&sort_iter, b);
1931
1932         bytes = !b->written
1933                 ? sizeof(struct btree_node)
1934                 : sizeof(struct btree_node_entry);
1935
1936         bytes += b->whiteout_u64s * sizeof(u64);
1937
1938         for_each_bset(b, t) {
1939                 i = bset(b, t);
1940
1941                 if (bset_written(b, i))
1942                         continue;
1943
1944                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1945                 sort_iter_add(&sort_iter,
1946                               btree_bkey_first(b, t),
1947                               btree_bkey_last(b, t));
1948                 seq = max(seq, le64_to_cpu(i->journal_seq));
1949         }
1950
1951         BUG_ON(b->written && !seq);
1952
1953         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1954         bytes += 8;
1955
1956         /* buffer must be a multiple of the block size */
1957         bytes = round_up(bytes, block_bytes(c));
1958
1959         data = btree_bounce_alloc(c, bytes, &used_mempool);
1960
1961         if (!b->written) {
1962                 bn = data;
1963                 *bn = *b->data;
1964                 i = &bn->keys;
1965         } else {
1966                 bne = data;
1967                 bne->keys = b->data->keys;
1968                 i = &bne->keys;
1969         }
1970
1971         i->journal_seq  = cpu_to_le64(seq);
1972         i->u64s         = 0;
1973
1974         sort_iter_add(&sort_iter,
1975                       unwritten_whiteouts_start(c, b),
1976                       unwritten_whiteouts_end(c, b));
1977         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1978
1979         b->whiteout_u64s = 0;
1980
1981         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1982         le16_add_cpu(&i->u64s, u64s);
1983
1984         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
1985
1986         set_needs_whiteout(i, false);
1987
1988         /* do we have data to write? */
1989         if (b->written && !i->u64s)
1990                 goto nowrite;
1991
1992         bytes_to_write = vstruct_end(i) - data;
1993         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1994
1995         if (!b->written &&
1996             b->key.k.type == KEY_TYPE_btree_ptr_v2)
1997                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
1998
1999         memset(data + bytes_to_write, 0,
2000                (sectors_to_write << 9) - bytes_to_write);
2001
2002         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2003         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2004         BUG_ON(i->seq != b->data->keys.seq);
2005
2006         i->version = cpu_to_le16(c->sb.version);
2007         SET_BSET_OFFSET(i, b->written);
2008         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2009
2010         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2011                 validate_before_checksum = true;
2012
2013         /* validate_bset will be modifying: */
2014         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2015                 validate_before_checksum = true;
2016
2017         /* if we're going to be encrypting, check metadata validity first: */
2018         if (validate_before_checksum &&
2019             validate_bset_for_write(c, b, i, sectors_to_write))
2020                 goto err;
2021
2022         ret = bset_encrypt(c, i, b->written << 9);
2023         if (bch2_fs_fatal_err_on(ret, c,
2024                         "error encrypting btree node: %i\n", ret))
2025                 goto err;
2026
2027         nonce = btree_nonce(i, b->written << 9);
2028
2029         if (bn)
2030                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2031         else
2032                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2033
2034         /* if we're not encrypting, check metadata after checksumming: */
2035         if (!validate_before_checksum &&
2036             validate_bset_for_write(c, b, i, sectors_to_write))
2037                 goto err;
2038
2039         /*
2040          * We handle btree write errors by immediately halting the journal -
2041          * after we've done that, we can't issue any subsequent btree writes
2042          * because they might have pointers to new nodes that failed to write.
2043          *
2044          * Furthermore, there's no point in doing any more btree writes because
2045          * with the journal stopped, we're never going to update the journal to
2046          * reflect that those writes were done and the data flushed from the
2047          * journal:
2048          *
2049          * Also on journal error, the pending write may have updates that were
2050          * never journalled (interior nodes, see btree_update_nodes_written()) -
2051          * it's critical that we don't do the write in that case otherwise we
2052          * will have updates visible that weren't in the journal:
2053          *
2054          * Make sure to update b->written so bch2_btree_init_next() doesn't
2055          * break:
2056          */
2057         if (bch2_journal_error(&c->journal) ||
2058             c->opts.nochanges)
2059                 goto err;
2060
2061         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2062
2063         wbio = container_of(bio_alloc_bioset(NULL,
2064                                 buf_pages(data, sectors_to_write << 9),
2065                                 REQ_OP_WRITE|REQ_META,
2066                                 GFP_NOFS,
2067                                 &c->btree_bio),
2068                             struct btree_write_bio, wbio.bio);
2069         wbio_init(&wbio->wbio.bio);
2070         wbio->data                      = data;
2071         wbio->data_bytes                = bytes;
2072         wbio->sector_offset             = b->written;
2073         wbio->wbio.c                    = c;
2074         wbio->wbio.used_mempool         = used_mempool;
2075         wbio->wbio.first_btree_write    = !b->written;
2076         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2077         wbio->wbio.bio.bi_private       = b;
2078
2079         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2080
2081         bkey_copy(&wbio->key, &b->key);
2082
2083         b->written += sectors_to_write;
2084
2085         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2086                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2087                         cpu_to_le16(b->written);
2088
2089         atomic64_inc(&c->btree_write_stats[type].nr);
2090         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2091
2092         INIT_WORK(&wbio->work, btree_write_submit);
2093         queue_work(c->io_complete_wq, &wbio->work);
2094         return;
2095 err:
2096         set_btree_node_noevict(b);
2097         b->written += sectors_to_write;
2098 nowrite:
2099         btree_bounce_free(c, bytes, used_mempool, data);
2100         __btree_node_write_done(c, b);
2101 }
2102
2103 /*
2104  * Work that must be done with write lock held:
2105  */
2106 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2107 {
2108         bool invalidated_iter = false;
2109         struct btree_node_entry *bne;
2110         struct bset_tree *t;
2111
2112         if (!btree_node_just_written(b))
2113                 return false;
2114
2115         BUG_ON(b->whiteout_u64s);
2116
2117         clear_btree_node_just_written(b);
2118
2119         /*
2120          * Note: immediately after write, bset_written() doesn't work - the
2121          * amount of data we had to write after compaction might have been
2122          * smaller than the offset of the last bset.
2123          *
2124          * However, we know that all bsets have been written here, as long as
2125          * we're still holding the write lock:
2126          */
2127
2128         /*
2129          * XXX: decide if we really want to unconditionally sort down to a
2130          * single bset:
2131          */
2132         if (b->nsets > 1) {
2133                 btree_node_sort(c, b, 0, b->nsets, true);
2134                 invalidated_iter = true;
2135         } else {
2136                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2137         }
2138
2139         for_each_bset(b, t)
2140                 set_needs_whiteout(bset(b, t), true);
2141
2142         bch2_btree_verify(c, b);
2143
2144         /*
2145          * If later we don't unconditionally sort down to a single bset, we have
2146          * to ensure this is still true:
2147          */
2148         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2149
2150         bne = want_new_bset(c, b);
2151         if (bne)
2152                 bch2_bset_init_next(c, b, bne);
2153
2154         bch2_btree_build_aux_trees(b);
2155
2156         return invalidated_iter;
2157 }
2158
2159 /*
2160  * Use this one if the node is intent locked:
2161  */
2162 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2163                            enum six_lock_type lock_type_held,
2164                            unsigned flags)
2165 {
2166         if (lock_type_held == SIX_LOCK_intent ||
2167             (lock_type_held == SIX_LOCK_read &&
2168              six_lock_tryupgrade(&b->c.lock))) {
2169                 __bch2_btree_node_write(c, b, flags);
2170
2171                 /* don't cycle lock unnecessarily: */
2172                 if (btree_node_just_written(b) &&
2173                     six_trylock_write(&b->c.lock)) {
2174                         bch2_btree_post_write_cleanup(c, b);
2175                         six_unlock_write(&b->c.lock);
2176                 }
2177
2178                 if (lock_type_held == SIX_LOCK_read)
2179                         six_lock_downgrade(&b->c.lock);
2180         } else {
2181                 __bch2_btree_node_write(c, b, flags);
2182                 if (lock_type_held == SIX_LOCK_write &&
2183                     btree_node_just_written(b))
2184                         bch2_btree_post_write_cleanup(c, b);
2185         }
2186 }
2187
2188 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2189 {
2190         struct bucket_table *tbl;
2191         struct rhash_head *pos;
2192         struct btree *b;
2193         unsigned i;
2194         bool ret = false;
2195 restart:
2196         rcu_read_lock();
2197         for_each_cached_btree(b, c, tbl, i, pos)
2198                 if (test_bit(flag, &b->flags)) {
2199                         rcu_read_unlock();
2200                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2201                         ret = true;
2202                         goto restart;
2203                 }
2204         rcu_read_unlock();
2205
2206         return ret;
2207 }
2208
2209 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2210 {
2211         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2212 }
2213
2214 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2215 {
2216         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2217 }
2218
2219 static const char * const bch2_btree_write_types[] = {
2220 #define x(t, n) [n] = #t,
2221         BCH_BTREE_WRITE_TYPES()
2222         NULL
2223 };
2224
2225 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2226 {
2227         printbuf_tabstop_push(out, 20);
2228         printbuf_tabstop_push(out, 10);
2229
2230         prt_tab(out);
2231         prt_str(out, "nr");
2232         prt_tab(out);
2233         prt_str(out, "size");
2234         prt_newline(out);
2235
2236         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2237                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2238                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2239
2240                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2241                 prt_tab(out);
2242                 prt_u64(out, nr);
2243                 prt_tab(out);
2244                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2245                 prt_newline(out);
2246         }
2247 }